Citation
Cross-Layer Design of Resource Aware Protocols for Heterogeneous Wireless Ad Hoc Networks

Material Information

Title:
Cross-Layer Design of Resource Aware Protocols for Heterogeneous Wireless Ad Hoc Networks
Creator:
LIU, WEI ( Author, Primary )
Copyright Date:
2008

Subjects

Subjects / Keywords:
Bandwidth ( jstor )
Delicatessens ( jstor )
Distance functions ( jstor )
Energy efficiency ( jstor )
Scheduling ( jstor )
Sensors ( jstor )
Simulations ( jstor )
Supply chain management ( jstor )
Transportation ( jstor )
Warehouses ( jstor )

Record Information

Source Institution:
University of Florida
Holding Location:
University of Florida
Rights Management:
Copyright Wei Liu. Permission granted to University of Florida to digitize and display this item for non-profit research and educational purposes. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder.
Embargo Date:
8/31/2010
Resource Identifier:
658201972 ( OCLC )

Downloads

This item has the following downloads:

liu_w ( .pdf )

liu_w_Page_024.txt

liu_w_Page_071.txt

liu_w_Page_127.txt

liu_w_Page_074.txt

liu_w_Page_015.txt

liu_w_Page_158.txt

liu_w_Page_147.txt

liu_w_Page_107.txt

liu_w_Page_018.txt

liu_w_Page_137.txt

liu_w_Page_118.txt

liu_w_Page_075.txt

liu_w_Page_123.txt

liu_w_Page_132.txt

liu_w_Page_068.txt

liu_w_Page_088.txt

liu_w_Page_144.txt

liu_w_Page_133.txt

liu_w_Page_148.txt

liu_w_Page_084.txt

liu_w_Page_031.txt

liu_w_Page_058.txt

liu_w_Page_032.txt

liu_w_Page_151.txt

liu_w_Page_164.txt

liu_w_Page_062.txt

liu_w_Page_012.txt

liu_w_Page_093.txt

liu_w_Page_042.txt

liu_w_Page_038.txt

liu_w_Page_157.txt

liu_w_Page_014.txt

liu_w_Page_152.txt

liu_w_Page_141.txt

liu_w_Page_040.txt

liu_w_Page_022.txt

liu_w_Page_150.txt

liu_w_Page_156.txt

liu_w_Page_047.txt

liu_w_Page_007.txt

liu_w_Page_030.txt

liu_w_Page_122.txt

liu_w_Page_161.txt

liu_w_Page_072.txt

liu_w_Page_162.txt

liu_w_Page_134.txt

liu_w_Page_136.txt

liu_w_Page_079.txt

liu_w_Page_002.txt

liu_w_Page_121.txt

liu_w_Page_029.txt

liu_w_Page_009.txt

liu_w_Page_070.txt

liu_w_Page_046.txt

liu_w_Page_105.txt

liu_w_Page_005.txt

liu_w_Page_041.txt

liu_w_Page_083.txt

liu_w_Page_001.txt

liu_w_Page_065.txt

liu_w_Page_064.txt

liu_w_Page_028.txt

liu_w_Page_053.txt

liu_w_Page_095.txt

liu_w_Page_131.txt

liu_w_Page_076.txt

liu_w_Page_143.txt

liu_w_Page_050.txt

liu_w_Page_098.txt

liu_w_Page_130.txt

liu_w_Page_155.txt

liu_w_Page_090.txt

liu_w_Page_108.txt

liu_w_Page_023.txt

liu_w_Page_035.txt

liu_w_Page_097.txt

liu_w_Page_013.txt

liu_w_Page_094.txt

liu_w_Page_059.txt

liu_w_Page_004.txt

liu_w_Page_096.txt

liu_w_Page_010.txt

liu_w_Page_140.txt

liu_w_Page_091.txt

liu_w_Page_017.txt

liu_w_Page_039.txt

liu_w_Page_120.txt

liu_w_Page_027.txt

liu_w_Page_066.txt

liu_w_Page_165.txt

liu_w_Page_049.txt

liu_w_Page_119.txt

liu_w_Page_011.txt

liu_w_Page_163.txt

liu_w_Page_115.txt

liu_w_Page_092.txt

liu_w_Page_055.txt

liu_w_Page_051.txt

liu_w_pdf.txt

liu_w_Page_103.txt

liu_w_Page_082.txt

liu_w_Page_087.txt

liu_w_Page_146.txt

liu_w_Page_026.txt

liu_w_Page_033.txt

liu_w_Page_089.txt

liu_w_Page_135.txt

liu_w_Page_061.txt

liu_w_Page_153.txt

liu_w_Page_048.txt

liu_w_Page_016.txt

liu_w_Page_067.txt

liu_w_Page_052.txt

liu_w_Page_154.txt

liu_w_Page_166.txt

liu_w_Page_099.txt

liu_w_Page_069.txt

liu_w_Page_025.txt

liu_w_Page_138.txt

liu_w_Page_145.txt

liu_w_Page_116.txt

liu_w_Page_060.txt

liu_w_Page_043.txt

liu_w_Page_106.txt

liu_w_Page_056.txt

liu_w_Page_077.txt

liu_w_Page_063.txt

liu_w_Page_160.txt

liu_w_Page_085.txt

liu_w_Page_037.txt

liu_w_Page_102.txt

liu_w_Page_111.txt

liu_w_Page_036.txt

liu_w_Page_142.txt

liu_w_Page_100.txt

liu_w_Page_129.txt

liu_w_Page_113.txt

liu_w_Page_021.txt

liu_w_Page_073.txt

liu_w_Page_044.txt

liu_w_Page_126.txt

liu_w_Page_139.txt

liu_w_Page_128.txt

liu_w_Page_006.txt

liu_w_Page_159.txt

liu_w_Page_125.txt

liu_w_Page_149.txt

liu_w_Page_019.txt

liu_w_Page_008.txt

liu_w_Page_081.txt

liu_w_Page_112.txt

liu_w_Page_114.txt

liu_w_Page_057.txt

liu_w_Page_110.txt

liu_w_Page_054.txt

liu_w_Page_101.txt

liu_w_Page_117.txt

liu_w_Page_080.txt

liu_w_Page_078.txt

liu_w_Page_124.txt

liu_w_Page_045.txt

liu_w_Page_086.txt

liu_w_Page_020.txt

liu_w_Page_109.txt

liu_w_Page_104.txt

liu_w_Page_003.txt

liu_w_Page_034.txt


Full Text











CROSS-LAYER DESIGN OF RESOURCE AWARE PROTOCOLS FOR HETEROGENEOUS
WIRELESS AD HOC NETWORKS
















By
Wei Liu
















A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL
OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY


UNIVERSITY OF FLORIDA


2005
































Copyright 2005

by
Wei Liu

















To my wife, Shu, and my family members in China.















ACKNOWLEDGMENTS

First and foremost, I would like to express my sincere gratitude to my advisor, Prof.

Yuguang Fang, who leaded me to this exciting research area, for his invaluable advice,

encouragement and motivation during these years of my Ph.D. studies. I also thank him for

his philosophical advice on both my academic and nonacademic life. As a mentor, he has

helped me become more mature, scholastically and personally.

I also want to thank my committee members, Prof. Shigang Chen, Prof. Sartaj K.

Sahni, Prof. John. M. Shea, and Prof. Dapeng Wu, for their attention and advices.

I would like to extend my thanks to all my colleagues in Wireless Networks Laboratory

(WINET) at University of Florida for providing me such a warm family-like environment.

The constructive discussions with them either individually or together deserve special ac-

knowledgments.

Finally, on a more personal level, I am particularly indebted to my wife Shu, for her

enduring so many late nights, and for countless sacrifices made to me. Her love, support

and encouragement have made it possible for me to carry out this work. I am also very

grateful to all my family members in China for their endless encouragement and continued

support in one way or the other. This work is dedicated to all of them.

This material is based upon work supported in part by the U.S. Office of Naval Re-

search and National Science Foundation. Any opinions, findings, and conclusions or rec-

ommendations expressed in this material are those of the authors and do not necessarily

reflect the views of the U.S. Office of Naval Research and National Science Foundation.
















TABLE OF CONTENTS


page


ACKNOWLEDGMENTS.


LIST OF TABLES ............. ............ viii


LIST OF FIGURES.


ABSTRACT. .......... .........


CHAPTERS


1 INTRODUCTION


Wireless Ad Hoc Networks.
Current Research
1.2.1 Reliable Communications
1.2.2 Energy Conservation
1.2.3 Quality of Service ...
1.2.4 Cross-Layer Design ..
1.2.5 Heterogeneity in Wireless
Our Research
Outline. ... ....


Ad Hoc Networks.


2 A DEVICE-ENERGY-LOAD AWARE RELAYING FRAMEWORK
IN HETEROGENEOUS WIRELESS AD HOC NETWORKS.


2.1 Introduction.
2.2 Related Work.
2.3 Overview of DELAR.
2.3.1 Problem Statement
2.3.2 Our Solution: DELAR. ......
2.4 Design of DELAR. .. .......
2.4.1 P-nodes' Neighbor-Selection Criteria ...
2.4.2 Routing Component of DELAR ....
2.4.3 Hybrid Transmission Scheduling ....
2.4.4 Asymmetric Media Access Control Protocol
2.5 Multiple-Packets Transmission in DELAR ...
2.5.1 Multiple-packets Transmission .. ..
2.5.2 Hierarchical Modulation .. ....
2.6 Discussion.
2.6.1 The Existence of Backward Paths ....


(A-MAC)











2.6.2 DELAR and ZRP . .... .. 39
2.6.3 The Choice of m and n .... . .. .. 39
2.6.4 Benefits of the Time Division Scheduling .. .. .. .. 40
2.7 Performance Evaluation . ... .. .. 40
2.7.1 Simulation Setup . ... .. .. 40
2.7.2 Impact of The Number of P-nodes .. .. .. .. 42
2.7.3 Impact of Node Mobility .... ... .. 44
2.7.4 Impact of Traffic Load ..... .. . .47
2.8 Summary . .... ..... .. .47

3 RESOURCE AWARE MOVEMENT IN HETEROGENEOUS
MOBILE AD HOC NETWORKS ..... .... .. 49

3.1 Introduction ......... .. ... .. 49
3.2 Related Work ........ .. .. 50
3.3 Problem Formulation . ... ... .. 51
3.3.1 General Mobility Model (GMM) ... ... .. .. 52
3.3.2 Resource Aware Movement .... ... . .. 54
3.4 Waterhunter Movement . . . .. .. 56
3.4.1 Simplified Waterhunter Movement Problem .. . 56
3.4.2 RAM-DCLC Algorithm .... . .. .. 61
3.4.3 A Routing Delay Differentiation Mechanism .. . 65
3.4.4 Incorporate RAM into DELAR ... .. .. .. .. 66
3.5 Performance Evaluation . .... .. .. 66
3.5.1 Simulation Setup . .... .. 66
3.5.2 Simulation Results ...... .. .. 68
3.6 Summary ............. ... ...... 7

4 SUPPORT DIFFERENTIATED SERVICES IN MOBILE AD HOC NETWORKS 72

4.1 Introduction ......... .. ... .. 72
4.2 Related Work ........ .... .. 74
4.3 Motivation .......... ...... ...... 76
4.4 Packet-length-based Channel Model .... ..... .. 79
4.5 Courtesy Piggybacking . ..... .. .. 82
4.5.1 System Assumptions ...... .... .. 82
4.5.2 The Courtesy Piggybacking Scheme ... .. .. .. .. 83
4.5.3 Discussion ........ .. .. 89
4.6 Performance Analysis . ...... ... .. 94
4.7 Performance Evaluation . .... .. .. 97
4.7.1 Simulation Setup . ..... .. .. 97
4.7.2 Impact of Channel Characteristics .. .. .. .. .. 99
4.7.3 Impact of Traffic Load .............10
4.7.4 Impact of Node Mobility .............11
4.7.5 Impact of Piggybacking Rules .............13
4.8 Summary .............14












5 ROBUST AND ENERGY-EFFICIENT DATA DIS SEMINATION
IN WIRELESS SENSOR NETWORKS ... ... .. 106

5.1 Introduction . .. ..... .06
5.2 Related Work .. .. .. . ... . .08
5.3 Modelling Sensor Networks as Supply Chains .. . . 111
5.3.1 Introduction to Supply Chains ... .. .. .. 111
5.3.2 How Could Supply Chains Help Us .. .. .. .. .. 113
5.4 A Hybrid Data Dissemination Framework
for Wireless Sensor Networks .... .... .. 115
5.4.1 System Model . .... .. .. .115
5.4.2 Manufacture Area . ... .. .. .117
5.4.3 Transportation Area .... .. .. .119
5.4.4 Warehouse Area and Service Area ... .. .. .. .. 123
5.4.5 Discussion . .... .. .. .. .. 126
5.5 Performance Evaluation . . ... .. .128
5.5.1 Methodology and Metrics ... ... .. 128
5.5.2 Simulation Results ...... .... .. .131
5.6 Summary ......... .. .. .. .13

6 CONCLUSIONS AND FUTURE WORK ... ... .. 138

6.1 Conclusions . .. ..... .. .38
6.2 FutureWork ............. ..........139

REFERENCES ......... . .. .. .. 141

BIOGRAPHICAL SKETCH . ..... ... . 151

















LIST OF TABLES
Table pg

3-1 RAM: resource aware movement .... .. .. 61

3-2 RAM-DCLC: a DCLC routing algorithm for the Waterhunter Movement 62

3-3 Energy consumption parameters. ..... .. . 67

4-1 Channel model statistics ..... ... .. 97

5-1 Analogue between supply chains and wireless sensor networks .. .. .. 114

5-2 Simulation configuration . ...... ... .. .. 130

















LIST OF FIGURES
Figure page

2-1 The structure of a Super Frame. . .. .. .. 18

2-2 An example of the neighbor determination process (m=4, n=2, T=3.).. .. 22

2-3 The topology in homogeneous and heterogeneous cases.. .. .. .. .. 23

2-4 A unidirectional link between A and B. .. .. .. 27

2-5 The A-MAC operation procedure. ..... .... .. 28

2-6 The multiple-packets transmission enhancement to DELAR. .. .. .. 30

2-7 General hierarchical 2/4-PSK constellation. ... .. . .. 32

2-8 Implement multiple-packets transmission with hierarchical modulation .. 33

2-9 The existence of backward path. ...... .... .. 38

2-10 The average number of nodes existing in the shaded area. .. .. .. .. 38

2-11 Simulation results with different number of P-nodes. .. .. .. .. .. .. 43

2-12 Simulation results with different maximum node speed. .. . 45

2-13 Simulation results with different traffic load. ... .. . .. 46

3-1 Multiple paths between two consecutive stops. ... .. . .. 55

3-2 An exemplary complete graph. ...... .... .. 58

3-3 An exemplary resource-aware movement. ... . . .. 59

3-4 Resource-aware movement (RAM) vs. random waypoint movement. A B-
node should consecutively visit S1-S2- 3- 4- 5 and the solid lines are
labelled by the sequences they were passed through. .. . 64

3-5 Simulation results. ......... .. .. 69

4-1 The Whittier Tunnel scenario. . .. .. .. 77

4-2 The Gilbert-E11iott channel model. ..... .. .. 80

4-3 The optimal packet length (PL) vs. SNR (y), h=128... .. .. .. .. 81

4-4 Packet-Length-Based finite-state Markov channel model. .. .. .. .. 81











4-5 A fragmentation example. . ..... .. .. 83

4-6 The total overhead with PKmax and FT. ... .... .. 84

4-7 Illustration of the courtesy piggybacking scheme. .. .. .. .. 87

4-8 Three piggybacking cases. . ..... .. .. 90

4-9 An alternative piggybacking method. .... ... .. 94

4-10 The queue model for piggybacking. .... .. .. 96

4-11 Average waiting time. ....... .. ... .. 97

4-12 Simulation results with different channel settings. .. .. . 100

4-13 Simulation results with different packet arrival rates. .. .. .. .. .. .. 101

4-14 Simulation results with different pause time. .. .. .. .. 102

5-1 An exemplary supply chain. . .... .. .. .. 111

5-2 Supply chain strategies. ....... ... .. .. 112

5-3 A system architecture for habitat monitoring .. .. .. .. .. 115

5-4 The forwarding-decision-making process of nodes in the transportation area. 120

5-5 An exemplary packet structure. ...... .... .. .. 120

5-6 The routing process in the warehouse area. .. .. .. .. .. 124

5-7 The simulated sensor field. . ..... .. .. .. 129

5-8 Event delivery ratio vs. packet error rate. ... .. .. .. .. 131

5-9 Normalized energy consumption vs. packet error rate. .. .. .. .. 131

5-10 Average event end-to-end delay vs. packet error rate. .. .. .. .. 132

5-11 Average routing overhead vs. packet error rate. .. .. .. .. .. 132

5-12 Event delivery ratio vs. packet error rate. ... .. .. .. .. 134

5-13 Energy consumption and event end-to-end delay of RRP with different d 135















Abstract of Dissertation Presented to the Graduate School
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy

CROSS-LAYER DESIGN OF RESOURCE AWARE PROTOCOLS FOR
HETEROGENEOUS WIRELESS AD HOC NETWORKS

By

Wei Liu

August 2005

Chair: Yuguang Fang
Major Department: Electrical and Computer Engineering

Their features of rapid establishment and self-organization have rendered wireless ad

hoc networks to be identified as an indispensable component to support ubiquitous com-

munications. However, the viable deployment of such networks faces many challenges

resulting from some innate unfavorable features such as the constrained resources, for ex-

ample, finite energy supply and limited bandwidth. In this dissertation we investigate the

cross-layer design of resource-aware protocols that conserve energy and support quality-

of-service (QoS) in resource-constrained wireless ad hoc networks.

We first study the issue of energy conservation in heterogenous wireless ad hoc net-

works, where most nodes are powered by batteries with small capacity, while some others

are more powerful, for example, powered by batteries with large capacity. We propose

a cross-layer designed device-energy-load aware relaying framework, called DEIAR, to

capitalize such powerful nodes to conserve energy. Different from the previous work that

addresses the energy conservation issue at one layer, DELAR is a joint design of schedul-

ing, power control and routing that can conserve energy thus prolong the network lifetime.

We further propose a resource-aware movement mechanism to utilize such powerful nodes










for data relaying. Such a resource-aware movement mechanism provides a new way to

conserve energy in wireless ad hoc networks.

We then study the issue of how to support QoS in wireless ad hoc networks. We

propose a scheme called Courtesy Piggybacking to utilize system dynamics including time-

varying channel condition and stochastic traffic characteristics, to support differentiated

services in such networks. The basic idea of Courtesy Piggybacking is to let high priority

traffic help the low priority traffic by sharing the unused residual bandwidth. As a result,

the scheme improves the overall performance for different prioritized services.

Finally, we propose an energy efficient, reliable and scalable data dissemination scheme

for wireless sensor networks. Based on supply chain management methodology, for each

sensing task, the sensor field is conceptually partitioned into several functional areas, and

then different routing schemes are applied in different areas. This scheme addresses energy-

efficiency, reliability, and scalability at the same time.















CHAPTER 1
INTRODUCTION

1.1 Wireless Ad Hoc Networks

A wireless ad hoc network is a collection of autonomous communication devices,

sometimes called nodes, which can communicate with each other by forming a multi-hop

radio network and maintaining connectivity in a decentralized manner. There are two major

types of wireless ad hoc networks that are of particular interest to academia, industry and

government: mobile ad hoc networks (MANETs) and wireless sensor networks (WSNs).

Both types of wireless ad hoc networks share several typical features in common. Wireless

ad hoc networks can be deployed quickly and on-demand. They are self-organized and

independent of infrastructure. Moreover, each node in a wireless ad hoc network functions

as both a host and a router, helping relay packets for other nodes. These favorable features

make wireless ad hoc networks very attractive in military and civil applications for which

fixed infrastructures are unavailable or unreliable, yet fast network establishment and con-

stant reconfiguration are required. As an instance of mobile ad hoc networks, after an

earthquake, a rescue team sets up a temporary wireless networks and shares information to

coordinate the rescuing work. As an example of wireless sensor networks, a large number

of small sensor nodes are deployed in a field to monitor the habitat behavior of animals.

However, there are some unfavorable features impeding wireless ad hoc networks

from widespread deployment. Due to the time-varying and error-prone wireless channel,

nodes have to contend with the adverse effects of radio communications, such as noise, fad-

ing and interference. Since nodes may depart, join, or move in the field, or even perish due

to energy depletion, the network topology is in general dynamic and may change rapidly

and unpredictable over time. In such hostile, dynamic environment, reliable communica-

tions are of great importance. Additionally, the wireless channel has limited bandwidth.










Therefore, it is a challenging task to support heterogeneous traffic with various quality-

of-service requirements. Moreover, most nodes, especially those small sensors deployed

in wireless sensor networks, are battery-powered thus with limited lifetime. These nodes

would become useless once depleting the batteries. Some adverse consequences of such

node diminution include degradation of network performance and unfavorable network

partition. Thus, energy conservation, that is, how to expend the energy resources more

frugally and more evenly so as to prolong the network lifetime, becomes a crucial issue

for MANETs. The broadcast nature of wireless communications also makes it easy for

enemies to snoop the ongoing transmissions, thus communications in wireless ad hoc net-

works have security hazard. Depending on the applications running in such networks, the

networks range from small-scale networks with a small number of nodes to large-scale net-

works with thousands even millions of nodes, for example, large-scale wireless sensor net-

works. The above unfavorable factors, especially variable wireless links, topology changes,

limited bandwidth, and constrained energy supply, make the protocol design a tremendous

challenge. Particular, when heterogenous applications are finding their niches in wireless

ad hoc networks, specially designed protocols that can overcome the above adverse factors

are desired. Such protocols, scalable themselves, should provide reliable, QoS-supporting,

energy-e~fficient, and secure communications for those eye-catching applications. In the fol-

lowing we give a nutshell of the current research for the issues of reliable communications,

energy conservation and QoS provisioning.

1.2 Current Research

1.2.1 Reliable Communications

The primary goal of protocol design for wireless ad hoc networks is to support reli-

able communications which is the basic requirement for any applications, and all the other

solutions for the issues of energy conservation and quality of service should be based on re-

liable communications. In such hostile and dynamic settings, packet losses may be caused










for various reasons including transmission errors, collisions, or route changes due to mo-

bility. Thus, there are many works addressing reliable communications at different layers.

For example, at the physical layer, modulation and error coding schemes are proposed to

combat the error-prone feature of wireless communications [1]. In order to reduce the in-

terference and transmission collisions, at the MAC layer there are many MAC protocols

that are devised to to efficiently control the channel access [2]. These protocols can be

classified into three categories: contention-based (for example, MACAW [3], FAMA [4]

and IEEE 802.11 [5]), contention-free (for example, TDMA, FDMA and CDMA) and hy-

brid. The most famous one is IEEE 802.11, and it uses CSMA/CA and four way handshake

(RTS/CTS/DATA/ACK) to coordinate the channel access and provide reliable communica-

tions [5]. At the network layer many routing protocols have been proposed to deal with

the dynamic topology and support reliable communications [6]. They can be proactive

in that they determine routes independent of traffic patterns, for example, DSDV [7] and

OLSR [8]; they can be reactive in that they maintain routes on demand, such as DSR [9]

and AODV [10]; they even can be hybrid of the above two, for example, ZRP [l l] and

LANMAR [12]. Though these schemes function at different layers, they all work for the

same goal: supporting reliable communications in wireless ad hoc networks.

1.2.2 Energy Conservation

How to lengthen the lifetime of communication devices is a crucial issue for the wire-

less ad hoc networks and many power conservation techniques are applied when hardware

and software including the protocol stack are specifically designed. Many efforts have been

made at the physical layer due to the factor that the system hardware is the primary con-

sumption object in a mobile device. Two different perspectives are commonly adopted to

approach the problem. Though no breakthrough has been experienced in the past 30 years,

the direct way is to increase the battery capacity, while keeping the weight of the battery

tolerative. Another way is, hard yet attainable, to apply low-power techniques to decrease

the energy dissipation when wireless terminals are designed. Considerable energy savings










are resulted from the hardware design; however, it is pertinent to explore other venues to

further improve the energy efficiency, that is, to design the protocol stack with energy effi-

ciency in mind [13]. For example, at the data-link layer transmission power control can be

used to reduce the interference. Power-saving mode also can be used to conserve energy

by putting the network interface into the sleep mode when no communication is needed.

At the network layer, power aware routing is used to find energy-efficient paths and route

packets over the energy-efficient paths to save energy. At the application layer, energy-

efficient OS/middleware and applications (for example, power-aware video processing) are

proposed to save energy. Because the key to energy conservation in wireless ad hoc net-

works lies within the higher levels of the wireless protocol stack [13], it would be beneficial

if these multiple layers can jointly function to conserve energy.

1.2.3 Quality of Service

It is hard to agree on a common definition of QoS, but a QoS enabled network shall

ensure that its applications and/or their users have their QoS parameters fulfilled, while

at the same time ensuring an efficient resource usage, for example the bandwidth, and

also ensure that the most important traffic still has its QoS parameters fulfilled during net-

work overload. The most important QoS parameters may include: throughput, availability,

delay, jitter and packet loss. The network can serve the applications with "hard-QoS",

"soft-QoS" or "adaptive QoS" [2, 14]. Several prominent service models or overall archi-

tectural frameworks, within which certain types of services can be provided in the network,

are suggested for ad hoc networks, such as flexible quality of service model for MANETs

(FQMM) [15] and stateless model for wireless ad-hoc networks (SWAN) [16]. Since the

resource is limited and different traffic has very different quality of service requirements,

usually Differentiated Services (DiffServ) architecture [17] is used in which traffic is clas-

sified into different classes or priority levels, and the wireless networks serve different

classes of traffic with different processing rules or policies. There are several mechanisms

proposed for supporting QoS [14], such as QoS routing, scheduling, admission control,










resource reservation, and signaling techniques. These mechanisms can coexist in a net-

work. For example, QoS routing is used to find a path satisfying certain QoS requirements,

then signalling techniques and resource reservation are used to reserve resources along the

found path. We believe that in order to efficiently utilize the resources and support QoS,

various mechanisms implemented at different layers should work together.

1.2.4 Cross-Layer Design

As a matter of fact, the issues of energy conservation and QoS provisioning are direct

results of constrained resources: energy supply and bandwidth. If nodes and networks are

endowed with plenty of bandwidth and energy supply, for example, unbounded frequency

band and heavy duty power supply with light wight, these two issues may never come to

the fore. Since the tough reality of constrained resources is there and the chance to envision

breakthrough for these two resources is rather slim, the natural response to these two issues

is to check: Do all the current protocols proposed for wireless ad hoc networks make full

use of such precious resources? The answer to the question is negative. There do have lots

of resources thar are wasted without any yield, owing to low efficiency or their overlook of

these protocols. For example, in networks with contention-based medium access control

mechanisms, lots of energy and bandwidth are wasted due to collisions.

One of the reasons for resource underutilization is that vast majority of the protocols

are based on traditional layered-design principle, which lacks the interaction and informa-

tion sharing among multiple layers. The lack of interaction causes low efficiency of those

protocols. On one hand, many resources at one layer are wasted due to undesirable de-

cisions made at another layer. For example, poor routing decisions cause congestion and

collisions at the physical layer thus waste precious resources, for example, bandwidth and

energy. On the other hand, many available resources are neglected by those protocols. For

example, time-varying channel condition leads to time-varying channel capacity, however,

due to the lack of interaction between the MAC layer and the physical layer, many MAC

protocols cannot recognize such a resource but neglect it. Hence, in recent years a new










cross-laver design philosophy emerges [18] [19]. In contrast to the layered-design phi-

losophy, cross-layer design philosophy stresses on the information sharing and interaction

among multiple layers. Cross-layer design can be performed in two ways [20]. In the

first way, parameters or information in other protocol layers are considered to improve the

performance of a protocol layer. The first way of practice does not totally abandon the

transparency between protocol layers. In contrast, the second way blurs this transparency

by merging several layers into one component to achieve much better performance.

The cross-layer design philosophy attempts to make use of the inter-layer coupling to

use the resources more efficiently. With this philosophy, the decision for resource utiliza-

tion can be made more advisably based on the information from multiple layers. Though

the cross-layer design may make the design issue more complicated than the layered design,

the appealing benefits make cross-layer design a strong candidate as the design principle in

resource-constrained wireless ad hoc networks.

1.2.5 Heterogeneity in Wireless Ad Hoc Networks

Most of the previous research assumes that networks are homogeneous or nodes are

in fully symmetrical environment. For example, all nodes have identical capabilities and

responsibilities. In reality, however, the heterogeneity of mobile devices seems to be inher-

ent and has been commonly observed in MANETs [21]. Strictly speaking, heterogeneity in

a network usually means the device heterogeneity regrading the physical hardware or soft-

ware nodes use. Nodes in the same network may differ in their communication, processing

capabilities, which can be reflected by CPU speed, memory size, battery life, transmission

range and radio, network interface, security level, operating system or protocol stack, and

so on. Nodes in the same network may differ in their responsibilities as well. They may

undertake different roles, for example, routers, servers, cluster heads, and authentication

centers, thus operate with different policies. Owing to such differences in capabilities and

responsibilities, along with the various applications running on each individual node, nodes

may also have different traffic characteristics and mobility characteristics. For example,










nodes may have different bit rate, timeliness constraints, reliability requirements, speed,

direction, and predictable/unpredictable movement patterns for each individual applica-

tion. Broadly speaking, heterogeneity in a network also means that nodes may experience

heterogeneous environment characterized by time-varying channel and network conditions.

Such environment heterogeneity may have different properties in different settings. For ex-

ample, different nodes at different locations and at different time may perceive different

channel conditions.

The heterogeneity, either device heterogeneity or environment heterogeneity, coupling

with the aforementioned unfavorable factors, makes the protocol design much more com-

plicated. However, such heterogeneity also introduces opportunities to design more ef-

ficient protocols. In fact, some types of heterogeneity are resource-related when they

themselves are resources or can be mapped into certain type(s) of resources. Moreover,

device heterogeneity and environment heterogeneity can jointly affect the usage of con-

strained resources. For example, nodes with less available battery reserve would refrain

from transmissions in poor channel condition. Thus, to recognize the inherent hetero-

geneity and further to efficiently utilize the resource-related heterogeneity to overcome the

impact stemming from those unfavorable features, for example, limited energy supply and

constrained bandwidth, is a rather challenging task. In this dissertation, we make our ef-

forts along this line to address the issues of energy conservation and QoS provisioning in

resource-constrained wireless ad hoc networks.

1.3 Our Research

As mentioned above, there is a big volume of work studying how to support reliable

communications, from the physical layer to the data link layer and to the routing layer.

Nevertheless, the research on energy conservation and QoS provisioning is still thin on

the ground. Either of the two issues is quite interesting, however, on top of the previous

research on reliability, it is pretty hard, if not impossible, to design a one-size-fit-all pro-

tocol that can simultaneously resolve these two issues. Thus, in this dissertation, based on










the previous work for reliable communications, we attempt to study these two issues and

resolve them separately, but bear in mind the possibility of future integration of these two

1SSUeS.

The cross-layer design philosophy serves as our design principle for its potential in

better utilizing the available resources and providing desired energy efficiency and QoS

provisioning to applications. More important, the cross-layer design philosophy will help

us recognize the inherent heterogeneity in wireless ad hoc networks, and further help us

utilize it to address the issues we are interested in. More specifically, we utilize the hetero-

geneity in terms of power supply or battery life to address the issues of energy conservation.

Such heterogeneity is used to design more efficient routing, power control and transmission

scheduling. We also utilize the time-varying channel condition and stochastic traffic char-

acteristics to address issues associated with QoS provisioning in wireless ad hoc networks.

For wireless sensor networks, we particularly recognize the different roles undertaken by

each individual sensor node, and utilize such heterogeneity to design more efficient com-

munication protocols.

1.4 Outline

This dissertation is organized as follows.

Chapter 2 concerns the issue of energy conservation in heterogeneous wireless ad

hoc networks in terms of power supply. In such networks, most nodes called B-nodes are

battery-powered, and some nodes called P-nodes have relatively unlimited energy supply

such as solar cell or dynamos. We propose a device-energy-load aware relaying framework,

called DELAR, to capitalize such P-nodes to conserve energy. The basic idea of DELAR

is to divide time into Super Frames, each of which is composed by several periods, and

to enable P-nodes to transmit with different power in different periods. In this framework,

device heterogeneity, nodal residual energy, and local load status are mapped into a routing

cost metric and incorporated into routing protocols to find energy-efficient paths. We also










propose a hybrid medium access control mechanism to schedule the transmission activi-

ties of P-node and B-nodes. Moreover, we introduce the concept of "mini-routing" into

the MAC layer and propose an Asymmetric MAC (A-MAC) to support reliable communi-

cations over unidirectional links resulted from the asymmetric power capabilities between

P-node and B-nodes. Further, we propose a multiple-packets transmission scheme as an

enhancement to improve the overall performance of DELAR.

Chapter 3 discusses a resource-aware movement scheme that can be incorporated

into DELAR. In this scheme, a node is enabled to determine its movement characteris-

tic (speed, direction, and so on) based on the network environment and its own residual

resources. Specifically, such kinds of movement can be mapped into a path-constrained

path-optimization problem. A heuristic algorithm is proposed to solve this NP-complete

problem. With this resource-aware movement scheme in place, B-nodes have more oppor-

tunities to utilize P-nodes and thus conserve energy and prolong the network lifetime.

Chapter 4 focuses on the issue of QoS provisioning. We consider heterogeneous mo-

bile as hoc networks on how to utilize system dynamics, a combination of environment and

device heterogeneity including time-varying channel condition and stochastic traffic char-

acteristics. A scheme called Courtesy Piggybacking is proposed to support differentiated

services in mobile ad hoc networks. The basic idea is to let the high priority traffic help

the low priority traffic by sharing the unused residual bandwidth with courtesy. Courtesy

Piggybacking is able to utilize such system dynamics to improve the overall system per-

formance for different prioritized services: improved packet delivery ratio and shortened

end-to-end delay. More important, the piggybacking scheme can be readily incorporated

into the DELAR framework.

In Charter 5, we particularly address the issues of energy efficiency and reliability in

wireless sensor networks, a special type of wireless ad hoc networks. We introduce the

concept of supply chain into wireless sensor networks and model wireless sensor networks

as supply chains. Based on supply chain management strategies, we propose an energy










efficient, reliable and scalable data dissemination scheme. Basically, for each sensing task,

the sensor field is conceptually partitioned into several functional areas, and different rout-

ing schemes are applied in different areas. Such partition strategy, hybrid strategy, and

cooperation strategy help us address the issues of energy conservation and reliable commu-

nications at the same time.

Chapter 6 concludes this dissertation and outlines the future research.















CHAPTER 2
A DEVICE-ENERGY-LOAD AWARE RELAYING FRAMEWORK
IN HETEROGENEOUS WIRELESS AD HOC NETWORKS

2.1 Introduction

In this chapter, we focus on heterogeneous wireless ad hoc networks, where most

nodes, denoted as B-nodes, are equipped with limited power sources like batteries, while

some other nodes, denoted as P-nodes, have relatively unlimited power supplies, for ex-

ample, power scavenging units such as solar cells, or dynamos when they are installed in

mobile vehicles, and so on. Our goal is to develop more energy conscious protocols by tak-

ing advantage of the heterogeneity of mobile devices, that is, being generous in using the

P-nodes while conservative in using the B-nodes. More specifically, the contributions are

mainly fourfold. First, following the cross-layer protocol design philosophy, we propose a

Device-Energy-Load Aware Relaying framework, named DELAR, to achieve energy con-

servation by utilizing the inherent heterogeneity of nodal power capabilities. Second, we

design a hybrid transmission scheduling scheme, which is a combination of reservation-

based and contention-based medium access control schemes, to coordinate the transmis-

sion activities among P-nodes and B-nodes. Such scheduling can make full use of powerful

nodes while reducing the interference and collisions. Third, we develop the "mini-routing"

and Asymmetric MAC (A-IMAC) protocols to support the MAC layer acknowledgements

over unidirectional links resulting from the asymmetric transmission power levels at P-

nodes and B-nodes. To the best of our knowledge, this is the first effort to address this

issue at the MAC layer. Last, we present a multiple-packets transmission technique to fur-

ther improve the delay performance. Detailed simulation studies are carried out to justify

the effectiveness and efficiency of the proposed framework.










The proposed DELAR can serve as a general framework where various energy con-

servation techniques such as power saving modes, transmission power control and power-

aware routing can be integrated to jointly achieve better energy conservation. In addition,

it also offers a platform to study other challenging issues, for example, quality of service

(QoS) provisioning and security support. For instance, P-nodes can act as distributed ad-

mission controllers to coordinate the access to limited network resources such as available

bandwidth. As another example, in security-sensitive MANET applications, P-nodes can

help B-nodes perform resource-hungry public-key operations. As far as we know, no simi-

lar framework has appeared elsewhere in the literature.

In the rest of this chapter, we start with the review of related work. We then introduce

the system model and the overall framework of DELAR in Section 2.3. In Section 2.4,

we elaborate the network layer components of DELAR and a hybrid transmission schedul-

ing scheme, and present a novel Asymmetric MAC protocol called A-IMAC followed by

a multiple-packet transmission scheme to further improve the performance of DELAR. In

Section 2.7, we evaluate the performance of DELAR through simulations. Finally, we

summarize this chapter in Section 2.8.

2.2 Related Work

Recent years have seen a growing body of research concerned with energy conserva-

tion in wireless ad hoc networks, among which many efforts have been made at the physical

layer to improve the hardware design of mobile devices [13]. Though important, it is still

pertinent to explore other venues to further ameliorate the energy efficiency of mobile de-

vices. For the lack of space, here we only review works performed at the MAC or network

layer of ad hoc networks, which are closely related to our study in this chapter. Based

on the mechanisms used, the solutions seen in the literature can be roughly classified into

three categories: Power-Saving Modes (PSM), Transmission Power Control (TPC) [22],

and Power-Aware Routing (PAR).










PSM is usually implemented at the MAC layer and the basic idea is to put the network

interface into the sleep mode when no communication is needed. One fundamental issue

in PSM is when to enter the sleep mode and for how long to stay in this mode. Some work

along this line includes PAMAS [23] and S-MAC [24]. In addition, Tseng et al. proposed

three asynchronous protocols, namely, Dominating-Awake-Interval, Quorum-based, and

Periodical-Fully-Awake-Interval protocols [25]. These proposals strive to efficiently and

intelligently control nodes' sleep and wake schedules and at the same time deal with such

factors as clock synchronization, neighbor discovery, and network partitions which are

inherent in multi-hop ad hoc networks [25].

TPC adapts the transmission power to the propagation and interference characteristics

experienced by the link [26, 27, 28, 29, 30, 31, 32, 33, 34, 35]. TPC sometimes is called

topology control when it attempts to control the transmission power or even turn off the

radio so that a desirable topology or connectivity can be maintained for saving energy.

Many proposals in this category are concerned with maintaining a dominant set of nodes

or forming some virtual backbone with certain clustering mechanisms. In addition, PCM

[36] was proposed to use different transmission power levels for RTS/CTS and DATA/ACK

frames on a per-packet basis.

In contrast to TPC protocols aiming at making each link as energy-efficient as possible,

a PAR protocol determines which of these links to be used for end-to-end paths so that

additional energy savings can be obtained by routing packets over energy-efficient paths.

Singh et al. proposed five power-aware routing metrics that could be incorporated into

routing protocols to reduce the per-packet energy consumption, to prolong the per-node

lifetime, or to prolong the overall network lifetime [37]. Besides, Chang and Tassiulas [38]

proposed a routing cost metric that is a combination of the transmission power level and

the residual energy. Toh [39] studied both the minimum energy and the network lifetime

issues and proposed a conditional min-max battery capacity routing protocol which tries to

strike a balance between these two competing objectives. While the link quality has been










suggested as a routing metric to reduce queuing delays and data loss rates, Banerjee and

Misra proposed a routing cost formulation that considers not only the residual power but

also the link characteristics, for example, the cost for potential retransmissions, to capture

the effects of both the link distance and the link error rate [40]. As an addition, PARO [41]

is another notable approach designed for scenarios where nodes can dynamically adjust

their transmission power. In PARO, a candidate intermediate node monitors an ongoing

direct communication between two nodes and inserts itself in the forwarding path if its

action can lead to some energy savings.

Most previous proposals do not take into account device heterogeneity inherent in

MANETs to achieve better energy conservation. How to take full advantage of P-nodes

to prolong the network lifetime as much as possible has been previously addressed in [21,

42, 43]. However, all of them focus on the network layer and many challenging issues

relating to the MAC layer are either left untouched or overlooked. For example, none of

them consider how to support the MAC-layer acknowledgements over unidirectional links

caused by different transmission power at P-nodes and B-nodes (cf. Section 2.4.4). By

contrast, our DELAR addresses the energy conservation from multiple facets, including

routing, transmission scheduling and power control.

In brief, although all the aforementioned proposals can achieve certain level of energy

conservation, how to design a comprehensive, practical framework that not only integrates

PSM, TPC, PAR, and transmission scheduling, but also makes full use of inherent device

heterogeneity remains an open, challenging problem. Our DELAR framework is proposed

to address this crucial issue.

2.3 Overview of DELAR

In this section, we first make a brief introduction to the problems we intend to address.

We then provide a high-level overview of our solution, DELAR, which stands for Device-

Energy-Load Aware Relaying framework for heterogeneous ad hoc networks.










2.3.1 Problem Statement

In this chapter, we focus on heterogeneous ad hoc networks comprised of mobile nodes

with different energy supplies, though heterogeneity may have other meanings in different

settings. We assume that besides a majority of battery-powered B-nodes, there exist some

powerful P-nodes having relatively unlimited energy supplies (cf. Section 2.1). Our objec-

tive is to develop energy conservation protocols by utilizing such heterogeneity in energy

resources. Intuitively speaking, since P-nodes are of "relatively" infinite energy reservoir

as opposed to the B-nodes with usually irreplaceable batteries, data communications should

try to utilize these P-nodes as much as possible in order to prolong the network lifetime.

Therefore, on the one hand, a packet should be forwarded to a P-node whenever an energy

saving can be expected. On the other hand, communications in the networks should avoid

using B-nodes if possible. For these purposes, it is desirable to allow P-nodes to have higher

transmission power so as to cover a larger transmission area, which can statistically reduce

the number of B-nodes involved in packet forwarding. However, this straightforward pro-

posal may pose lots of challenges to the protocol design How can a B-node be aware of

the existence of P-nodes in its vicinity? If there exist multiple paths through P-nodes to the

destination, which path should be chosen? Should the transmission range of a P-node be as

large as possible, or be kept within certain "optimal" ranges? How can the protocol support

reliable communications over the unidirectional links caused by asymmetric transmission

power at P-nodes and B-nodes along with error-prone and time-varying wireless channels?

In addition, higher transmission power often implies more reachable neighbors, decreased

spatial reuse, and increased local contention for the shared wireless medium, then how can

the protocol schedule the transmission activities so that a good balance can be struck be-

tween energy savings and other system performance factors such as packet delivery ratio

and end-to-end delay? These are all non-trivial questions and need to be answered before

we can indeed make full use of the aforementioned heterogeneity in ad hoc networks.










After a careful investigation on these interwoven issues, we believe that they are

closely related to routing, transmission scheduling, and power control. For example, with

the adjustment of the transmission power of P-nodes and B-nodes, the topology and neigh-

borhood, and thus the routing information, would change accordingly. So would the sched-

ule of transmission activities. Moreover, there exists a strong interaction between routing

and MAC layers. Apparently, our original design objective can be boiled down to designing

a joint routing, scheduling, and power control scheme, which should be addressed across

the whole protocol stack, especially at the routing and MAC layers [44][45]. To achieve

this, a cross-layer designed framework is demanded. In this framework, power control

should be implemented to optimize the transmission power of each node (both P-nodes and

B-nodes) to achieve optimal energy utilization and maintain a reasonable network topol-

agy; routing should be designed to inform all the nodes of the existence of P-nodes and find

the optimal energy-efficient routes; and transmission scheduling should be able to adjust

the transmission activities so that the energy spent on channel contentions and collisions

can be minimized. In addition, an appropriate scheduling scheme should be capable of

striking a good balance between energy efficiency and other system performance factors

such as end-to-end delay and packet delivery ratio.

2.3.2 Our Solution: DELAR

We consider a mobile ad hoc network consisting of N,, P-nodes and Nb B-nodes,

where N,, and Nb are system design parameters. We assume a single wireless broadcast

channel shared by all the nodes, though our DELAR can be easily extended to multi-

channel cases. We also adopt a simple power control scheme as follows. Each B-node

transmits omni-directionally and can maintain a circular transmission range BTR (basic

transmission range) before using up its battery, which can be properly set to achieve a good

tradeoff between energy efficiency and network connectivity [46]. In addition, we pos-

tulate that P-nodes are able to adjust their transmission power so as to cover larger areas

than B-nodes if needed. Moreover, all the P-nodes are assumed to have identical maximum










transmission range of PTRmos = M~x BTR, where M~ is a positive integer greater than

1. As revealed in [46], using common transmission power between the same type of nodes

can ensure bidirectional links and thus the correct operations of existing routing and MAC

protocols. With this simple yet efficient power control scheme, a unidirectional link only

exists between a P-node and a B-node when they use different transmission power, instead

of between any two B-nodes or P-nodesl According to [35], such simple power control

is believed to be more practical than other expensive transmission power control schemes,

either making unrealistic assumptions or having extra hardware requirements.

As mentioned before, DELAR arises from the following intuition: the P-nodes should

be utilized as much as possible. In other words, we should attempt to reduce the use of B-

nodes if we cannot avoid using them at all. Thus it is advantageous to enable a P-node to di-

rectly communicate with other P-nodes nearby or in distance by using higher transmission

power so that the number of B-nodes involved in the data forwarding can be much reduced.

However, higher transmission power or larger transmission coverage usually implies more

neighbors and increased local contention for the shared wireless channel. Therefore, in-

stead of granting these P-nodes unlimited privileges of reaching any other node at any time

at will, it makes more sense to constrain P-nodes' transmission power under certain bound

and to limit their transmission activities within some pre-planned periods in order to reduce

the collisions with other ongoing transmissions and thus maintain good channel utilization.

In order to better schedule the transmissions of P-nodes and B-nodes, we adopt a

time-division multiplexing method. We divide time into equal length time slots called

Supufmou'~lr l \. In each of the superframes, some time periods are exclusively designated

to P-nodes, while the rest are shared by all P-nodes and B-nodes in the network. More




SIn this chapter, we only consider asymmetric transmission power as the primary cause
for unidirectional links and omit others such as various collision/noise/interference levels
at different nodes.










P-to-P P-to-B P-to-B mini-slot P-to-B B-to-B

Pnd P-node 1 ... P-node k
Time Frame Duration

Figure 2-1: The structure of a Super Frame.


specifically, during one cycle of the Sep cT inlric (see Fig. 2-1), there is a P-to-P period

with length tp,, in which only P-nodes are allowed to communicate with each other by

using transmission range TR,, = m x BTR (1 < m < M~), while all B-nodes just keep

silent, as if the network were merely formed by these "mobile core" P-nodes. Addition-

ally, in one Sequ cT hana each P-node has its own exclusive period called P-to-B period

with equal length tpb, in which it can boost its transmission power to cover a range of

TRpb = a x BTR (1 < n < M~)2 The rest of one Sep cT homelc is called B-to-B period

with length tbb in which all the nodes in the network can contend for the channel and initi-

ate transmissions towards nodes in their TRbb = BTR. Obviously, all the P-nodes should

act as common B-nodes in the B-to-B period by adjusting their transmission range back to

TRbb. Notice that during one P-to-B period, since the P-node owning this period and the

B-nodes it intends to communicate with have different transmission power, unidirectional

links between them may be formed. Therefore, in contrast to the P-to-P and B-to-B periods

where some common contention-based MAC protocols such as the IEEE 802.11 can be

used, the P-to-B periods) demands an enhanced MAC protocol to support reliable com-

munications over unidirectional links. Our Asymmetric MAC protocol A-IMAC is exactly

developed for this purpose. Such rendezvous of reservation-based and contention-based

MAC schemes is able to schedule the packet transmission more efficiently, which we will

see shortly.


2 To provide reliable communications during P-to-B periods, usually n is less than m.










In DELAR, the heterogeneity of mobile nodes are also incorporated into the construc-

tion of routing tables. Routes are discovered based on routing metrics which take consid-

eration of the residual energy and the load status of mobile nodes. Once generating a data

packet, a node looks up its routing table and sends the data packet to the next hop as it does

in common ad hoc routing protocols. If residing in the forwarding path and having received

a forwarding request, a node will forward the data in an appropriate time period to the next

hop according to its own routing table. More specifically, for a B-node, when the next hop

is in its TRbb range, it can only forward the data packet during the B-to-B period. While

for a P-node, if the next hop is another P-node located in this P-node's TR,, the P-node

can forward the packet to the next hop in the P-to-P period, if the next hop is a B-node

located inside its TRyb, the P-node can forward the packet to the next hop in its exclusive

P-to-B period. In summary, with such time-division scheduling and a device-energy-load

aware routing metric, we intend to utilize P-nodes as much as possible in an efficient and

cautious way while reducing collisions and interference to an acceptable level, so that we

can achieve the expected energy conservation without harming other system performance

factors.

Several research challenges remain in supporting the seemingly simple operations of

DELAR as described above. Given a B-node (P-node) X located in a P-node P's transmis-

sion range TR,b (TR,,), for instance, what criteria should P adopt to determine if X is a

neighbor (in one hop range) or not, that is, forwarding a packet to this node X in a one-hop

manner or a multiple-hop manner? What kind of routing metric should be adopted to reflect

the heterogeneity in device types, nodal residual energy, and local load status when setting

up routing paths? How can we divide time into Supo,:,'~ll Wau. and how can one P-node

register a P-to-B period without conflicting with others' P-to-B periods? How does node

X send MAC layer acknowledgements back to P in the presence of unidirectional links

resulting from the asymmetry in transmission power? The remainder of this chapter will

address these questions one by one in more detail.










2.4 Design of DELAR

In this section, we will first discuss the neighbor-selection criteria of P-nodes. We then

illustrate the routing component of DELAR. Next, we introduce in more detail the hybrid

transmission scheduling of DELAR. Then, we present the Asymmetric Media Access Con-

trol Protocol (A-MAC) and the multiple-packet transmission scheme. Last we give some

further discussions.

2.4.1 P-nodes' Neighbor-Selection Criteria

In the literature, two nodes are usually considered as neighboring nodes of each other

when they are one hop away and they can directly communicate with each other. However,

in heterogeneous networks, we have to change the criteria to cope with the existence of

P-nodes whose transmission ranges are much larger than those of B-nodes. In this case,

any node in a P-node's TRpb COuld be a neighbor candidate of the P-node3 Nevertheless,

in order to support the MAC layer acknowledgements, not all the candidates can be finally

chosen as neighbors or be next hops in the routing table. Before presenting the rules that

guide P-nodes to make selection decisions, we first introduce the notions of Forward Path

and Backward Path. For any node pair s and t, a Forward Path indicates the path derived

from normal routing tables. For example, the Forward Path(s, t) can be represented as

s NI ~... Nk, t, where (Ni) 1 < i < k) denote the k intermediate nodes be-

tween s and t. For a given P-node P and any B-node X located in P's transmission range

TRpb, the Backward Path(P, X) is defined as the minimum-hop Forward Path(X, P) when

all the nodes have a transmission range of BTR. It is worth noting that the minimum-hop

Forward Path(X, P) is not necessarily the same as the Forward Path(X, P). Although For-

ward Pathyr\ are defined for any node pairs in the network, Backward Pathyr\ are only valid

between a P-node and the B-nodes located within the P-node's TRpb range. Furthermore,




SHowever, due to the asymmetric transmission power, the discussed P-node may not be
a neighbor of an individual node of those neighbor candidates.









for any neighbor candidate X of a given P-node P, this B-node X can be considered as

P's neighbor only when the Backward Path(P, X) satisfies the following criteria: All the

intermediate nodes along the Backward Path(P, X) should be in P's TRI>> range. In other

words, a neighbor candidate X can be considered as a P-node P's neighbor if and only if

all the intermediate nodes along the Backward Path(P, X) are P's neighbors as well.

With the above definitions, the remaining issue is how to set up these Backward rathir.

A simple way is to let a P-node broadcast a query message with a certain transmission

power, that is, covering all the B-nodes in its TRI,b = n x BTR range. Once seeing

such a query, each node broadcasts a special reply with the TTL value set to T 4 Each

node appends its own ID in the reply when relaying such a special reply. The querying

P-node will wait some time until collecting enough replies. The initiator of a reply would

be considered as a neighbor if and only if the querying P-node also receives replies initiated

from all the relaying nodes of that reply. We will see later, in order to facilitate the operation

of A-MAC, the path length of a Backward Path should be limited. We need to point out that,

even when a P-node, say P1, receives a query message initiated by another P-node, say P2,

P1 should reply like common B-nodes with a transmission range of BTR. Since our scheme

is targeted for networks with low or moderate mobility, P-nodes can execute this process

infrequently or in their respective P-to-B periods when topology changes are detected by

the MAC protocol. Therefore, the resulting overhead is expected to be affordable.

P-nodes also need to discover the neighboring relationship among themselves. To

achieve this, during the P-to-P period a P-node may send a query with appropriate trans-

mission power that is set to cover a range of TRI>; = m x BTR. P-nodes receiving this

query may directly send replies back to the requesting P-node.

Fig. 2-2 gives an example of the neighbor selection process. Suppose A is a P-node

with TRI,b = 2 x BTR and TRI~, = 4 x BTR, and the Backward Paths for neighbor


4 T can assume an integer value slightly larger than n to allow more replies.







22


/TR" O B-nod















Figure 2-2: An example of the neighbor determination process (m=4, n=2, T=3.).

candidates C, F, G and I are C B A, F E A, G F E A, and

I H C B A, respectively. Since node H does not initiate an reply to A, only

C, D, and G are considered as A's neighbors. Of course, B, D, and E are A's neighbors

as well. In this example, another P-node J is also a neighbor of node A because J is in A's

TR,, and they can directly communicate with each other.

2.4.2 Routing Component of DELAR

In homogeneous ad hoc networks, a node can only communicate with other nodes in

its BTR range, while in heterogeneous ad hoc networks, a P-node is able to reach any

other node within a larger transmission range, for example, TRpb and TR,,. Therefore, the

resulting topology and routing strategy may be quite different from that in homogeneous

networks. As an example, a network topology without P-nodes is depicted in Fig. 2-

3.a, where all the links are bi-directional and labelled with equal or unequal costs on both

directions. For instance, al/bi indicates that the link cost from A to B is al while bl from

B to A. In contrast, if one node, say A, is identified as a P-node who can reach much

further in the network, more unidirectional links may be added as shown in Fig. 2-3.b. We

label unidirectional links from P-node A to its neighbors with cost 0 to represent node A's

"unlimited" power supply.










B D




C E
a Topology in Homogeneous case



A FD




b Topology in Heterogeneous case

Figure 2-3: The topology in homogeneous and heterogeneous cases.


To cope with such heterogeneous networks, each P-node needs to maintain an internal

neighbor table recording its chosen neighbors within TRpb and the corresponding Back-

ward Pathr\ of those neighbors. In addition, each node in the network, either a P-node or

a B-node, needs to maintain a forwarding routing table similar to that in a normal table-

driven routing protocol such as DSDV [7]. For each node i, we define


P(i) = residual-en~ ,.ml(i) pL x queue_1en(i),

where residual-e 1, ml(i) indicates current remaining energy level at node i, queue _en(i)

represents the current load status at node i, and pL is a parameter representing the energy

consumption per unit data transmission Then the device-energy-load aware routing cost

metric we adopt is given in Eq. 2.1, though other cost metrics are applicable in DELAR as

well.


cost(i) = 1/P, S > (2.1)






5 In our simulation, for example, pL is equal to the average energy consumption per
packet transmission.










In the above cost metric, cost(i) is the cost of using node i as a relay, it could be

used as the cost of all directional links (arcs) starting from node i and directed to any of

its neighbors; y is a parameter used to adjust the weight of the awareness of load and

energy in the overall cost metric; constant a assumes a relatively large value to avoid us-

ing the nodes short of energy. Different types of devices may assume different values of

these parameters: pL, y and a~. For example, in this chapter, to represent a P-node's unique

power capability or device type, a P-node assumes a zero link cost6 fOr all the links to-

ward its B-node neighbors within TRpb Of P-DOde neighbors within TR,,. Ideally, in order

to find energy-efficient paths, each node should be informed about the routing costs of

other nodes as accurate and prompt as possible which may lead to excessive overhead. In

practice, however, the employed routing protocol should strike a good tradeoff between

energy efficiency and overhead. Proactive routing protocols are known for their capability

of propagating network conditions through the whole network in due course so that ap-

propriate QoS decisions, for example, admission control and route selection, can be made

intelligently. Thus, we adopt a proactive routing protocol, for example, DSDV, as the un-

derlying routing protocol to periodically exchange the routing information. We note that

other types of routing protocols can also be used in this framework. After gathering enough

routing information, a node can employ a shortest path algorithm to decide the minimum

cost paths and the related costs to all the other nodes in the network. Here the path cost is

actually the sum of the cost defined in Eq. 2.1 of all the B-nodes along a forwarding path.

Similar to those energy-aware cost metrics proposed in the literature, by choosing the

proper values of a~, pL and y, the cost function defined in Eq. 2.1 can help prolong the

network lifetime by distributing the traffic more evenly throughout the network, avoiding

the overuse of a small set of nodes, and consuming nodal energy resources in a more bal-

anced manner [37, 39]. Moreover, DELAR spontaneously incorporates P-nodes' unique


6 In practice, a very small value can be used to avoid possible routing loops.










power capabilities or device types, residual energy information, and local load status into

the routing protocol without using redirect tables in DEAR [43] any more.

2.4.3 Hybrid Transmission Scheduling

In order to reduce the interference a P-node's communication may impose on other

ongoing transmissions, it is reasonable to only allow a P-node to boost its transmission

power during some exclusively reserved periods. For this purpose, time is divided into

equal length time periods called Sequ' ,:nnlrti \ in which each P-node is assigned an exclu-

sive time interval, called a P-to-B period, to communicate with its B-node neighbors. In

addition, a P-to-P period is allocated to allow P-nodes to communicate with each other if

they are within each other's TR,, range. The rest of a Sty>
shared by all the nodes. Therefore, as shown in Fig. 2-1, a P-node has three different trans-

mission ranges in each Sty>
and TRbb in the rest of the Sep'cT; inrit In contrast, a B-node only has one transmission

range TRbb = BTR and it can only initiate transmissions during the B-to-B period.

Fig. 2-1 gives an instance of such a Srlet ,ntlric structure including multiple reserved

P-to-B periods, one for each P-node. The one-minislot-length paddings between consec-

utive P-to-B periods are used to further reduce the possible interference. The period allo-

cation of the Sty>
nodes use high transmission power to communicate and negotiate with each other, deciding

the lengths of the P-to-P period, the P-to-B period and the B-to-B period, also associating

each P-node with a P-to-B period. After finishing the negotiation, the P-nodes broadcast

such allocation information to all the B-nodes in their own TRpb. In this way, ultimately










all the nodes are informed the Supuf:T,~~ rame allocation, and can synchronize to such alloca-

tion7 In our current design, one P-to-B period is not allowed to share by multiple P-nodes

for simplicity only. However, a P-to-B period can be shared among P-nodes far away from

each other, if such sharing can ensure the transmissions conflict-free.

Since a P-node can communicate with other P-nodes in the P-to-P periods and com-

municate with the B-nodes within its BTR range in the B-to-B periods, it is natural that

in its own P-to-B period, a P-node should give priority to packets intended to its B-node

neighbors which are outside its BTR range but inside its TRI,, range. Thus, packet schedul-

ing is needed at a P-node to determine the appropriate transmission schedule for the packets

to be relayed or initiated by itself.

One may notice that the aforementioned transmission scheduling may need time syn-

chronization among the nodes, we argue that nodes equipped with GPS devices can easily

fulfill this task. This requirement is quite realistic today since such devices are inexpensive

and can provide reasonable precision. In this chapter we assume that nodes have perfect

time synchronization and leave the synchronization problem in the networks without GPS

devices as our future work.

In fact, the time-division scheduling, essentially a reservation-based access control

mechanism, and the MAC protocols employed in the three types of periods in each Super-

frame, either contention-based or reservation-based, form a hybrid transmission scheduling

for DELAR. Moreover, each type of periods may have different MAC protocol. For exam-

ple, the conventional contention-based MAC protocols, such as the IEEE 802.11 MAC

protocol, can be used during P-to-P periods and B-to-B periods. Since unidirectional links




SFor simplicity, we assume a fixed allocation scheme is used in this chapter, however,
more adaptive allocation is possible when P-nodes periodically exchange local load infor-
mation and negotiate a new allocation scheme during the P-to-P periods.







27

-L----- ~ Transmission







Figure 2-4: A unidirectional link between A and B.


are basically unavoidable in P-to-B periods, special measures are needed to deal with them.

In what follows, we will delineate the A-MAC protocol developed for P-to-B periods.

2.4.4 Asymmetric Media Access Control Protocol (A-MAC)

The presence of unidirectional links is pretty common in heterogeneous networks es-

pecially when employed power control schemes cause unequal transmission powers among

nodes. As an example, node A in Fig. 2-4 has a larger transmission range than node B.

Thus, node B can hear node A's transmission, however, node A cannot detect node B's trans-

mission. Obviously, a unidirectional link exists between node A and node B. The dilemma

is that the stop-and-wait ARQ (Automatic Repeat Request) scheme [47] employed in cur-

rent contention-based MAC protocols only works well with bidirectional links. In the face

of unidirectional links, the receiver B (Fig. 2-4) has no way to directly and successfully

send the acknowledgements back to the transmitter A, which means that the transmitter

would continuously transmit the same frame before timeout no matter whether the receiver

has received it or not. Moreover, the unidirectional links may severely affect the functional-

ities of ad hoc networks at various layers [48, 49, 50]. For example, many routing protocols

such as DSR and AODV rely on hop-wise acknowledgments for discovering route errors.

Therefore, how to support the MAC layer acknowledgements over the unidirectional links

is very important [51, 52] and has not yet been well addressed. Fortunately, we can make

use of the aforementioned Backward Pathr\ and the following "mini-routing method to

tackle this problem in an elegant way.









A (P-node)











Figure 2-5: The A-MAC operation procedure.


In current contention-based MAC protocols such as the IEEE 802.11, a receiver can

only transmit an acknowledgement frame to its one-hop-away transmitter. With the cross-

layer design methodology, we introduce a new concept of "mini-routing intoTrrack your

eBay activities the MAC layer, which requests intermediate nodes to relay the receiver's ac-

knowledgement frames, that is, CTS/ACK frames, along the established Backward Path(transmitter

receiver) in a multi-hop fashion to the transmitter (a P-node) at the MAC layer. Here the

routing information is no longer exclusively used by the network layer but shared by the

MAC and network layers.

Based on the IEEE 802.11 [5], we introduce into A-IMAC four special frames: P-

RTS, P-CTS, P-DATA, and P-ACK, all of which can only be transmitted in P-to-B pe-

riods. When a P-to-B period of a P-node comes and the P-node happens to have some

packets to transmit, it first boosts its transmission power to cover the range of TRpb =

n x BTR. With the scheduling described in Section 2.4.3, all the other nodes should re-

frain from initiating a transmission and temporarily stop transmitting usual frames, that

is, RTS/CTS/DATA/ACK. The P-node associated with this P-to-B period can send packets

to any neighboring B-node in the range of TRpb through P-RTS/P-CTS/P-DATA/P-ACK

exchanges.

Next, we illustrate the A-IMAC by using Fig. 2-5, where we assume n = 2 and P-node

A intends to send a packet to C, one of its B-node neighbors. The location relationship

among A, B, and C is also depicted in Fig. 2-5. First, A sends the P-RTS with TRpb



























































SSIFS stands for Short Inter-frame Space, and DIFS stands for DCF Inter-frame Space.


2 x BTR containing the Backward Path(A, C) (C B A). Then according to the length

of the Backward Path(A,C) in this example which is 2, A sets its waiting timer for the P-

CTS to be 2(SIFS+4T-crs+-tTprop) Where TP-CTs and Tprop are the transmission time

and air propagation time for one P-CTS respectively. Upon receiving the P-RTS destined to

it, and waiting for a SIFS period, node C will send node B a P-CTS including the addresses

of A and C. For an intermediate node residing on the backward path, it needs to set its

waiting timer according to its order in the Backward Path. For example, the ith node (the

intended receiver like C is assumed to be the Oth node on the path) on the path should set

its timer to be i x (SIFS +t TP-Crs +t Tprop). In this example, node B, when overhearing

the above P-RTS from A, it starts a timer equal to SIFS +t TP-ors + Tprop Since it is

the 1st intermediate node on the backward path. Once receiving a P-CTS from node C

before timeout, B simply appends its address and relays the modified P-CTS to the next

hop which is the P-node A in this example. Otherwise, node B sends a P-CTS containing

its own address to A after the timer expires, and the reason for doing this will be explained

later. If node A does not receive any P-CTS before timeout, it can retransmit the P-RTS

until reaching an admissible number of retries. If the same situation happens, A temporarily

saves this packet for future transmission and switches to another packet with a different next

hop. When A successfully receives a P-CTS from B containing both B 's and C's addresses,

the P-RTS/P-CTS exchange finishes. After a SIFS, A can send a P-DATA frame to node C

and set the timer to 2(SIFS+- TP-ACK -tTprop). Then the similar procedures apply. After

receiving the P-ACK from node C relayed by the intermediate node B, the P-node A can

start transmitting a new packet after a DIFS in the same manner. When its P-to-B period

expires, A lowers its transmission power and acts as a B-node in other P-nodes' P-to-B

periods and in the B-to-B period.




















Figure 2-6: The multiple-packets transmission enhancement to DELAR.


Besides the purpose of resolving the well-known hidden/exposed terminal problems,

the P-RTS/P-CTS exchange is also used to eliminate possible errors resulting from stale

routes or nodes' mobility. For example, in the above example, if node C moves out of

P-node A's 2 x BTR range while B is still in A 's BTR range, node C will not hear the P-

RTS from node A and hence A could only receive from node B a P-CTS including only B 's

address. In this case, A will think that node C is currently unreachable and may temporarily

save the packets to C for future transmissions. Another situation may happen that node C

is still in A's 2 x BTR range while node B moves out of A's BTR range, in which the

P-node A will delete node C from its neighbor table. Moreover, when node C moves into

A's BTR range, A will hear the P-CTS from node C directly. Hence A can optimize the

future transmissions to C without the help of node B any more.

2.5 Multiple-Packets Transmission in DELAR

In the basic DELAR design, time-division transmission scheduling is used to coordi-

nate the transmission activities of P-nodes and B-nodes. One undesirable consequence is

the excessive delay one packet may experience because it may be buffered at intermediate

nodes to wait for appropriate transmission periods. On the other hand, since DELAR is

energy aware and it costs P-nodes almost "nothing" to transmit a packet, many data pack-

ets may swarm to P-nodes. This may make P-nodes the "bottlenecks" of the network and

further increase the delay that packets accumulated at P-nodes would experience. In what

follows, we seek a way to alleviate this phenomenon.










2.5.1 Multiple-packets Transmission

In the basic design (see Fig. 2-5), for example, A can only transmit packets to either B

or C each time. When P-node A transmits a packet to B-node C, node B is only involved in

forwarding the control frames (for example, P-CTS/P-ACK). However, since the channel is

reserved for P-node A during its P-to-B periods and node B is also in P-node A's transmis-

sion range, node B has the capability to receive and demodulate any signal transmitted by

node A if it is allowed. Therefor, if node A can recognize B's capability and utilize it, that

is, transmitting packets to node B and C at the same time, the system performance could

be improved. This multiple-packets transmission mechanism can be illustrated in Fig. 2-6.

Suppose P-node A has some packets to B and some packets to C. Node A would put one

packet for C and another for B together, and send them in a single transmission from which

nodes B and C can acquire their own part, respectively. In this way, we expect to see the

improvement of the end-to-end delay performance of DELAR.

To do this, node A first makes sure that both B and C are within its effective range

TRpa after the P-RTS/P-CTS exchange as before. Then A can pull out from the waiting

queue one packet towards C and another packet towards B, and send them in one P-DATA

frame as depicted in Fig. 2-6. When seeing such a frame, nodes B and C can extract their

own parts and dump the rest. The same procedures as specified in A-MAC are executed

with the exception that node B also needs to indicate in the P-ACK that it has successfully

received the packet to itself. The similar procedure can be applied to the cases with more

than one interlineate nodes on the backward pathir.

With the above multiple-packets transmission mechanism, one may expect that the

total number of packets that can be packed and transmitted at one time is bounded by

the length of the Backward Path. Considering the possibility of adopting high-data rate

modulation schemes with a higher power level so as not to degrade the received signal

strength, more packets towards different receivers on the Backward Path, can be assembled

together and transmitted at the same time.
















C/ D




Figure 2-7: General hierarchical 2/4-PSK constellation.

2.5.2 Hierarchical Modulation

To support the multiple-packets transmission, hierarchical modulation (or nonuniform

modulation) schemes can be use to ensure all the receivers have enough received signal

strength to demodulate the useful information. In hierarchical modulation schemes, the

constellations consist of nonuniformly spaced symbols and allow for unequal error protec-

tion, that is, different degrees of protection for transmitted bits within a symbol are allowed

according to the importance of the information. For example, suppose that there are two

streams of data, each of which has a priority (a target BER), and QPSK hierarchical mod-

ulation (see Fig. 2-7) is used. One bit from each stream is taken to form a symbol of two

bits. The bit from bitstream with high priority (lower allowed BER) is assigned to the most

significant position, for example, the first bit in the constellation in Fig. 2-7. The bit from

the bitstream with low priority is assigned to the less significant position. By adjusting

the angle of 8, the constellation can be determined to achieve the unequal error protection

for each bitstream. Such unequal error protection has make hierarchical modulation very

attractive in multimedia services [53, 54, 55].

We notice that in the previous research, the hierarchical modulation is used in the cases

that the transmission power is fixed and the constellation is chosen in a way to allocate the

power among multiple data streams to achieve unequal error protection. In addition, in

the previous research, bitstreams usually are destined to the same receiver. However, in our









~A (P-node)







b, o




Figure 2-8: Implement multiple-packets transmission with hierarchical modulation


DELAR scheme, the transmission power of P-nodes is not fixed. Moreover, when multiple-

packets transmission is used, the bitstreams are destined to different receivers. As shown

in Fig. 2-6, multiple receivers involved in the multiple-packets transmission have different

distances to the transmitter. Therefore, in order to use hierarchical modulation to support

multiple-packets transmission at a P-node, more power ought to be allocated to far away

receivers and less power for close receivers. More specific, we need to find out the constel-

lation and the minimum overall transmission power for a P-node so that all the receivers

can successfully get their own packets. In the following we use QPSK hierarchical modula-

tion to implement the multiple-packets transmission when two receivers are involved (Fig.

2-8).

Suppose BERmay is the BER requirement for both receivers B (d2 aWay ffom trans-

mitter A) and C (dl away from transmitter A) to demodulate the received signal. Moreover,

we assume two-ray ground model with a~ = 4 is used to model the radio propagation:

PtcrG,h,2h,
P=
r 4L

where L is the system loss factor, Pt is the transmission power, Pr is the received power,

Gr(G,) is the antenna gain at the transmitter(receiver), and ht(h,) is the height of antenna

at the transmitter(receiver) For simplicity, we assume all the nodes have the same antenna

gain and height. We assume that 8 is less than 450 so that the inphase signal towards far










away receiver C is assigned with more power and quadrature signal towards close receiver

B with less power. It is easy to get the BER at C as

PtceG,h~h h
pl Q cos 8) = (1 2 cos 8) (2.2)
Nf \Nf

where N is the noise signal power, f is the transmission bit rate, W is the channel band-

width (in Hz), and Q() is Gaussian Q-function [54, 55]. Similarly, we can get the BER at

the receiver B as

PtGeG,h~h h
p2 = Q( L Sill 8) = \ 2 Sin 8). (2.3)


Thus, in order to correctly demodulate the received signal, the BER should satisfy the fol-

lowing conditions: pi < BERmax where i = 1, 2. Since we are interested in the minimum

transmission power of node A when equality is satisfied: pi = BERmax where i = 1, 2,

we define

To = Q- (BERmax)

as the SNR achieving the BERmax to facilitate us finding the boundary values of Pt and

0. From Eq. 2.2 and Eq. 2.3, it is easy to get

PtGeGThh~ h
S2 cos 8 = To (2.4)
\Nf

and
PtGeG,h h?
S2 sin 8 = To. (2.5)
\Nf
From Eq. 2.4 and Eq. 2.5, we can get


0 = arctan 2~j (2.6)


and

Pti (2.7)
mm-WGtG, ht2h









Therefore, given the distances between the transmitter and receivers, we can get the

minimum transmission power Ptain and 8 in the QPSK hierarchical modulation constel-

lation, and use this combination to implement the proposed multiple-packets transmission

and achieve the BER requirements at both receivers. In practice, P-node A may first use

a predefined transmission power covering TRpb, and transmit the RTS frame using typ-

ical BPSK. After the RTS-CTS handshake, A can determine the two parameters Ptmin

and 8 based on the feedback from the two receivers. Then A pulls out from the waiting

queue one packet towards C as the more important bitstream and one towards B as the

less important bitstream. Further it can transmit symbols using the aforementioned QPSK

hierarchical modulation. We should note that, in order to avoid the possible interference

when a P-to-B period is shared by multiple P-nodes, the involved P-nodes, before enabling

the multiple-packets transmission, should examine and make sure that Ptain is not greater

than the power to cover TRpb.

In fact, the minimum total transmission power to transmit the packets to each receiver

separately using BPSK can be written as follows:

ToN fdiL ToN fd L
ptot,, = p,1 + Pt2 = .(2.8)
WGtG,h~h~ WGtG,h~h~

Compared Eq. 2.7 to Eq. 2.8, we can see that although P,,in alone is larger than ei-

ther Ptl or Pt2, Ptmin is identical to Peroat. This verifies that multiple-packets transmission

with QPSK hierarchical modulation does not require extra transmission power 9 compared

to the case when multiple-packets transmission is not used. In fact, if multiple-packets

transmission is implemented with uniform QPSK, in order to satisfy the same BER re-

quirement at both receivers: pi < BERmax where i = 1, 2, the transmission power would

be JZPri which is much larger than Ptain. Thus, in our case, QPSK hierarchical modu-

lation is a rather reasonable option to implement multiple-packets transmission. Further,


9 Here we only consider the power used to transmit the information bits.










if we consider the energy used for control frames, the proposed multiple-packets trans-

mission requires fewer control frames thus less energy consumption, compared to the case

when packets are transmitted separately. For example, in Fig. 2-6, transmission energy for

three control frames (RTS/CTS/ACK) towards node B can be saved when multiple-packets

transmission is used.

2.6 Discussion

2.6.1 The Existence of Backward Paths

In DELAR, the P-nodes are utilized in two ways to conserve energy: enabling P-

nodes to directly communicate with other P-nodes within TR,,, and enabling P-nodes to

directly communicate with B-nodes within TRpb. Unlike the communications between P-

nodes that need no special treatment, the communications between P-nodes and B-nodes

are supported by the nodes on the backward paths as described in Section 2.4.1, thus the

backward path is very important for the proper functioning of DELAR. Then one may

question the existence of such backward paths. We believe that the existence of backward

paths is related to several factors such as basic transmission range (BTR), node density, and

the location distribution of P/B-nodes. On the other hand, the backward path should not be

too long in terms of hop count, otherwise A-MAC may not function well because a long

backward path is more subject to breakage due to node mobility or other reasons. In prac-

tice, we should include this limitation into the aforementioned neighbor-selection criteria,

for example, setting TTL value of a reply to the neighboring query to n (cf. Section 2.4.1),

such that only those B-nodes with backward paths of less than n hops can be considered as

neighbors of a given P-node.

Next we want to study the existence of such backward path in typical mobile ad hoc

networks. We assume n = 2 and T = 2 (cf. 2.4.1) so that only those nodes with backward

paths not more than 2 hops are considered as neighbors of a given P-node. Let O denote

the position of a P-node, and O' denote the position of a B-node in O's 2 x BTR = 2R

range, as shown in Fig. 2-9. We use r to indicate the distance between O and O'. It is









obvious that when r is not greater than R, O' is definitely a neighbor of O. So we shall

only consider R < r < 2R when studying the existence of backward paths for those nodes

that are r away from O. We also assume that all the nodes are uniformly and independently

distributed in a two-dimensional area with node density rl, where the probability of having

k nodes in an area of size S follows Poission distribution:


p(k, S) = e-e


The mean of the number of nodes in the area of size S is rlS. Then the existence of

backward paths between O and O' or the problem whether O' is a neighbor of O or not

can be reduced to a rather simple geometric problem: Is there any other node located in the

shadowed area S with both O and O' as neighbors?

Let -r be the average number of nodes existing in the shadowed area, and S1 and S2 be

the areas of sector A1BO and AAOB, respectively. Then we have -r = 2rl(S1 S2), where


S1 = x 2 arccos =I21/ R2 RECCOSI
2x R R

and
1 r/2
S2 -R2 Sill 2 arCCOS ~
2 R

Therefore, we have

r/2 1 r/2
-r = 2rl(R2 arCCOS 8 -2 Sill 2 arCCOS )
R2 R

Thus, the probability of existing a backward path between O and O', which is equal to the

probability that there is at least one node in the shadowed area, is 1 e-' according to the

Poission distribution. Fig. 2-10 plots this probability with different values of r, where the

total number of nodes is 50 for all network topologies. We can see that, with a reasonable

node density, the probability of existing a backward path between a given P-node and B-

node pair is pretty high, which justifies the feasibility of our A-MAC protocol and in turn

the overall DELAR scheme.





















































I


Figure 2-9: The existence of backward path.


1
I
1
BTR=~50




i


BTR=20


S1500*300 m2
h-7070 m
900*900 m2
-9-1000*1000 m2


200 250 300 350
r (m)


400 450 500


Figure 2-10: The average number of nodes existing in the shaded area.










2.6.2 DELAR and ZRP

In fact, we can borrow some ideas from ZRP [56, 11] to further improve the routing

performance of DELAR as follows. A P-node maintains the routing information within a

zone TRI~a by using the procedures described in Section 2.4.1 or any other routing proto-

col as the intra-zone routing protocol (IARP). It also maintains the information about its

neighboring P-nodes in its TRI~,. In contrast, a B-node only needs to maintain the routing

information in its TRbb = BTR, where all the nodes are its one-hop neighbors. Once a

node needs an energy-efficient route to another node, it can discover the route on demand

using the routing discovery procedure similar to that of AODV but with what is defined

in Eq. 2.1 instead of hop count as the cost metric. In this sense, DELAR can be viewed

as a special case of ZRP. The difference lies in the fact that all nodes within a P-node's

TRI,b zone, are "one-hop" away from this P-node from the routing perspective rather than

multi-hop away in the legacy ZRP. Besides, the border-casting [56, 11] technique used in

ZRP is also available in DELAR in the sense that a P-node border-casts a route request

(which we call simply a request) to all its peripheral nodes with corresponding backward

paths embedded in the route request. In this way, each peripheral node would learn the

backward path used to return a route reply if needed. Moreover, since a P-node can directly

exchange routing information with other P-nodes within its TRI>,, it also can border-cast

the route request to its P-node neighbors, which in turn can look up their own neighbor

tables to decide if the desired destination is in their own TRI>> zones. Since TRI>; is usually

larger than TRI,b, apparently, such "inter-zone routing protocol (IERP)" can further speed

up the route discovery process.

2.6.3 The Choice of m and n

Since a P-node can use higher transmission power to communicate with other P-nodes

within its m x BTR, larger m would lead to less use of B-nodes in the communications,

but also less spatial reuse. Similarly, larger n may lead to more energy savings, but also

imply possible longer backward paths and less spatial reuse, which may make A-MAC less










efficient as explained before. For the similar reasons, usually n is less than m. In order to

well balance the energy savings and other system performance factors, both m and n should

be chosen cautiously. While DELAR requires all the P-nodes have the same value of m to

avoid producing unidirectional links between P-nodes in P-to-P periods, with A-MAC in

place, they can have different values of n.

2.6.4 Benefits of the Time Division Scheduling

In fact, P-nodes can communicate with each other in one-hop or multi-hop manner

during P-to-P periods to coordinate the use of the next Sep cT f~rame. A P-node can also

notify all the other nodes within its TRpb to adjust their transmission schedules. With

these intelligent communications available, the slot allocation in each Sep cT f~~rame can be

adjusted adaptively according to traffic conditions instead of being fixed as in the given

example. It is worth pointing out that this time-division scheduling method can facilitate

the operations of PSM (cf. Section 2.2) in that it can help nodes determine their sleep and

wake schedules. For example, B-nodes can tumn off the radios during the P-to-P periods to

conserve energy.

As we discussed before, time synchronization is of importance for the correct oper-

ation of DELAR. In literature, there exist many proposals for time synchronization and

many of them can be incorporated into DELAR. In fact, P-nodes can serve as coordinators

to facilitate such synchronization by sending out some beacon information in some P-to-

P periods and P-to-B periods periodically. Subsequently, B-nodes can synchronize their

clocks to these P-nodes.

2.7 Performance Evaluation

2.7.1 Simulation Setup

In order to evaluate the performance of DELAR, we implemented our scheme includ-

ing the routing layer and the A-IMAC in the OPNET Modeler [57]. We simulated a network

with 50 nodes randomly deployed in a 1500 x 300m2 area. The BTR was 200m and the

transmission rate was 2M~bps. In our simulations, all the nodes were capable of moving










in the network according to the modified random Waypoint mobility model presented in

[58]. The pause time was set to be zero in our simulations, meaning nodes were always

moving. And each node moved with a randomly chosen speed between [ niEnc,Vm]z, where

Vni,z was fixed to be 1 m/s and V,,,,. assumed different values to reflect various network

mobility levels.

There were 20 constant bit rate (CBR) data sessions between randomly selected source

and destination pairs, and each source generated data packets of 512 bytes in length at a rate

of A packets per second. In our simulation, B-nodes had the same initial energy reservoir

6k J and their transmission and reception power were 1560m W and 930m W, respectively

[59]. The networks with 2 to 6 P-nodes were studied and these P-nodes were randomly

deployed. Besides, we chose m = 4, n = 2, tz>; = 0.02s, 1;>> = 0.02s, and tbb = 0.05S10

By varying the number of P-nodes, the maximum moving speed, and the CBR source rate,

we were able to study the performance of DELAR under different configurations. Six runs

were carried out to get an average result for each simulation configuration and each run

was executed for 900 seconds of simulation time.

Previous work [37, 39] has shown that the energy efficiency of routing protocols can

be much improved by adopting such a routing metric as what we defined in Eq. 2.1. There-

fore, we shall only compare our DELAR and DELAR with multiple-packets transmission

(denoted by DELAR+MPT) with the one referred to as EAR in this chapter, which is de-

veloped for comparison purpose and in fact a variant of DSDV with the routing cost metric

defined in Eq. 2.1. In EAR, all the P-nodes have the same transmission range as B-nodes,

but have a zero cost to their neighbors because they are assumed to have almost unlim-

ited energy reservoir and other resources. And the following metrics will be adopted in




to With advanced simulation tools such as OPNET, NS-2, and Qualnet, it is possible to
figure out a good configuration of these parameters to yield a good system performance,
before putting DELAR into practice.










comparison: average energy consumption defined as the total energy consumption for

all packet transmissions and receptions normalized by the number of delivered packets;

packet delivery ratio defined as the ratio of delivered data packets to those generated by

the sources; average packet end-to-end delay defined as the average delay from when a

packet is generated and transmitted by the source till it is received by the destination.

2.7.2 Impact of The Number of P-nodes

We first fixed the data rate to 4 packets/s and varied the number of P-nodes to study its

impact on the performance of DELAR. Here only the results for neu.=4 m/s are presented

in Fig. 2-11, though DELAR has the similar performance with other values of Exca--

Figure 2-11(a) compares average energy consumption of DELAR and EAR under

different numbers of P-nodes. Since EAR is also an energy-aware routing protocol, it is of

no surprise to see that its energy-saving performance improves with the number of P-nodes

whose routing costs to their neighboring nodes are assumed to be zero. With DELAR in

place, however, the average energy consumption can be much further reduced, for exam-

ple, with a factor of almost 50% if 6 P-nodes are available. The shown advantage comes

from the fact that DELAR makes much better use of P-nodes than EAR through intelligent

transmission scheduling and the allowance of P-nodes using different transmission power

during various periods. And the more P-nodes, the more energy savings we can expect from

DELAR. Compared to the basic DLEAR, DELAR with multiple-packets transmission can

further reduce the energy consumption. It can be contributed to the reduced transmissions

of control packets because multiple-packets transmission requires fewer control frames

than separate transmissions.

From Fig. 2-11(b), we can see that DELAR outperforms EAR in terms of packet

delivery ratio (PDR). The reason is that transmission scheduling in DELAR will lead to

less congestion, and the larger transmission ranges of P-nodes during P-to-P periods and

P-to-B periods can help reduce the number of hops a packet may travel. Again, the more

P-nodes, the more PDR is improved. Compared to the basic DLEAR, the multiple-packets









































12. I


0.00 .


SEAR
SDELAR
SDELAR+MPT


23 4
Number of P-nodes


5 6


(a) Average energy consumption


0.962-
0.960-
05-
0.958-
05-

0.952
-
~i0.950
CC: 0.948
S0.946
2 0.944
"0.942-
0.940 -
S0.938 -
,0.936 -
0.934 -
0.932 -
0.930
0.928


SEAR
SDELAR
tDELAR+MPT


234~
Number of P-nodes


5 ~6


(b) Packet delivery ratio


0.40

0.35-

0.30-

S0.25

S0.20

0i .15-




a,0.05 -


SEAR
SDELAR
SDELAR+MPT


Number of P-nodes


5 6


(c) Average packet end-to-end delay


Figure 2-11: Simulation results with different number of P-nodes.










transmission can further improve the packet delivery ratio because transmitting multiple

packets simultaneously can improve the channel utilization.

In terms of end-to-end delay, however, EAR is better than DELAR as shown in Fig. 2-

11(c). This is because, in contrast to the contention-based transmissions in EAR, DELAR

divides the time into Super Frames to schedule the transmission activities so that a packet

usually needs to be buffered at a node waiting for the coming of a proper transmission

period. In addition, we can observe that the delay of EAR decreases with the increase of

P-nodes because the existence of more P-nodes can achieve better load balance, that is,

the number of energy-efficient paths may be increased. However, this is not always the

case with DELAR. Besides the better load balance, the increase of P-nodes also means for

DELAR that the delay from the transmission scheduling becomes larger. This contributes

to the fluctuation in the delay of DELAR depicted in Fig. 2-11(c). Compared to the basic

DLEAR, DELAR with multiple-packets transmission can reduce the packet end-to-end

delay.

2.7.3 Impact of Node Mobility

In this subsection, we study the impact of the node mobility on DELAR by varying

Vm,, from 2 m/s to 16 m/s. For the reason of clarity, only the results for 4 P-nodes and a

data rate of 4 packets/s are presented.

Fig. 2-12(a) compares the average energy consumption of DELAR and EAR under

different mobility levels. As we can see, DELAR always has less energy consumption than

EAR due to the reasons stated before. Generally the higher mobility leads to less energy

consumption. After examining the average number of hops a packet may travel, we notice

that higher mobility often results in shorter route, which statistically leads to less energy

consumption because fewer transmissions and receptions are involved. As shown in Fig.

2-12(b), the packet delivery ratio decreases with the increase of the mobility, which is in

accordance with previous studies. For the similar reason we stated in previous subsection,

DELAR always has higher packet delivery ratio than EAR in all kinds of mobility. As





































































































I I I II I I I I I I I I I
0 2 4 6 8 10 12 14 16 18
Maximum Node Speed: Vmax (m/s)




(c) Average packet end-to-end delay


EAR
SDELAR
SDELAR+MTP


-


Maximum Node Speed: Vmax (m/s)




(a) Average energy consumption


SEAR
SDELAR
SDELAR+MTP


096-

0 96 -


094
; 0 92 -

S0 90 -



$ 0 66-

0 4 -

0 2 -


4


-


2 4 6 6 10 12 14 16 16
Max mum Node Speed: Vmax (m/s)


(b) Packet delivery ratio


036-
034.
032-

S028-
o 026-
,0 24 -
,0 22 -
S0 20 -
~018-
016-
S014-
S0.O12-
S010-


0 04 -


-tEAR
SDELAR
SDELAR+MTP


--~


Figure 2-12: Simulation results with different maximum node speed.














0 96-

0 95-
EAR
cc DELAR
~094-1 DEALR+MPT

0 93

092-

0 01
200 300 400 500 600 700
Traffic Load (Kbits/s)



(a) Packet delivery ratio


0 40-
0 35-
S0 30 -


S0 20-
015-
S010-
EAR
~005- *DELAR
~ ,,,~ ~DELAR+MPT

200 300 400 500 600 700
Traffic Load (Kbits/s)



(b) Average packet end-to-end delay


Figure 2-13: Simulation results with different traffic load.



the mobility increases, generally the delays of both EAR and DELAR get longer. Again,


DELAR has longer delay than EAR due to DELAR's time-division medium access con-


trol mechanism. One interesting observation is that the delays of both DELAR and EAR


fluctuate around Encer = 2m/s and Encer = 2m/s. This can be attributed to the used rout-


ing cost metric which causes lots of packets swarming to the P-nodes. This phenomenon


results in longer waiting time at P-nodes. However, nodal movement helps alleviate such


phenomenon by dispensing the traffic load. Compared to the basic DLEAR, DELAR with


multiple-packets transmission can reduce the energy consumption, improve the packet de-


livery ratio, and shorten the end-to-end delay in all kinds of mobility.










2.7.4 Impact of Traffic Load

In this subsection, we study the impact of the traffic load on DELAR by varying the

data generation rate from 3 packets/s to 8 packets/s. Since the traffic load does not much

impact average energy consumption, we only depict the simulation results for the packet

delivery ratio and the end-to-end delay in Fig. 2-12, where the number of P-nodes is four

and V,0, is equal to 4 m/s.

Figures 2-13(a) and 2-13(b) demonstrate that the packet delivery ratio decreases

and the delay increases with the increase of the traffic load for both schemes, respectively,

which are quite intuitive. Again, our DELAR is better in terms of the packet delivery ratio

but worse with regard to the end-to-end delay than EAR for the reasons stated previously.

However, Compared to the basic DLEAR, DELAR with multiple-packets transmission out-

performs DELAR due to reason described before.

In summary, DELAR is more appropriate for delay-insensitive applications, such as

file transfer and web access, which prefer higher energy-efficiency and packet delivery

ratio. With multiple-packets transmission, the performance can be further improved. Thus

it is a viable enhancement and can be effectively used in DELAR framework. With a better

tuning of system parameters DELAR can strike a good balance between energy efficiency

and other system performance factors.

2.8 Summary

In this chapter, we proposed a Device-Energy-Load Aware Relaying framework, namely

DELAR, to achieve energy conservation in mobile ad hoc networks. DELAR utilizes the

device heterogeneity inherent in ad hoc networks and features the cross layer protocol

design methodology. To take better advantage of powerful nodes (P-nodes) while reduc-

ing their interference on other ongoing communications, a hybrid transmission scheduling

mechanism is used to schedule and coordinate the transmission activities among P-nodes

and B-nodes (ordinary nodes). In addition, in order to support reliable transmissions in the










presence of unidirectional links between P-nodes and B-nodes, we introduced the "mini-

routing" technique and the novel Asymmetric MAC (A-IMAC) protocol. We demonstrated

that A-IMAC can effectively enable the MAC layer acknowledgements over unidirectional

links. To the best of our knowledge, no previous effort has been made to address this issue

at the MAC layer. Moreover, we proposed a multiple-packets transmission scheme which

can be operated with hierarchical modulation scheme to further improve the performance

of DELAR. Detailed simulations showed that DELAR can significantly reduce the energy

consumption and thus prolong the network lifetime even with a few P-nodes existing in the

network. With this framework, various energy conservation techniques such as power say-

ing modes, transmission power control and power-aware routing can be integrated to jointly

achieve better energy conservation. More important, this framework provides a platform to

study other challenging issues such as QoS provisioning and security support.















CHAPTER 3
RESOURCE AWARE MOVEMENT IN HETEROGENEOUS
MOBILE AD HOC NETWORKS

3.1 Introduction

As discussed in Chapter 2, there has been a rich literature addressing the energy con-

servation issue in MANETs, ranging from power-saving mode(PSM) [23, 60, 25, 24, 61],

to transmission power control(TPC) [26, 32, 27, 30, 33, 34], and to power-aware rout-

ing(PAM) [37, 40, 41]. Though those energy-aware MAC and routing protocols can in

general help the whole system expend the energy resources more reasonably to some ex-

tent, there are still many other aspects one can explore to further improve the system-wide

energy efficiency. As an example, similar to the traffic jam in daily life where a mass of

vehicles flocks to a single spot, unwise movement may cause local network traffic conges-

tion, thus leading to unfavorable energy waste. This observation motivates us to address

energy conservation from the perspective of node movement.

In addition to node movement, resource heterogeneity, as another inherent character-

istic of MANETs, is often either overlooked or underutilized in designing energy conserva-

tion schemes for MANETs. Though node heterogeneity can be interpreted in various ways,

we limit the scope of this chapter to heterogeneous networks in terms of energy supply. In

such a network, most nodes (called B-nodes hereafter) are furnished with lightweight bat-

teries having limited power, while a few others (called P-nodes hereafter) are powered by

almost unlimited energy supplies such as energy-scavenging devices (for example, solar

cells) and dynamos when nodes are in some mobile vehicles. In a relative sense, the energy

consumption of P-nodes can be considered as small or even negligible.

In this chapter, we propose to address energy conservation by guiding nodes's move-

ment and utilizing device heterogeneity. The basic idea is that, instead of moving in the










field blind to the network environment, for example, always following shortest-hop paths,

nodes are instructed to travel much more intelligently by keeping in mind the system-wide

objective of energy conservation and moving along "resource-aware" paths in such a way

that P-nodes can undertake as much communication tasks as possible so that less powerful

P-nodes can save energy, thus leading to the prolonging of the whole network lifetime.

The contributions in this chapter are mainly fourfold. First, we define a general mo-

bility model and formulate a general resource-aware movement problem, from which we

derive a Waterhunter Movement problem for B-nodes and a Firehunter Movement prob-

lem for P-nodes. Second, we reduce the Watehunter Movement problem to a NP-complete

distance-constrained least-cost (DCLC) routing problem and propose an efficient heuristic

solution. Third, we propose a routing delay differentiation mechanism to make full use of

the benefits provided by the resource-aware movement. Last, this resource-aware move-

ment can be incorporate into other energy conservation schemes to further improve the

energy efficiency. The effectiveness of the proposed schemes are justified and validated

through extensive simulations.

The rest of the chapter is organized as follows. We start with Section 3.2 surveying

the related work, then we formulate the resource-aware movement problem in Section 3.3.

In Section 3.4, we focus on the Waterhunter Movement problem and propose an efficient

heuristic solution. Section 3.5 evaluates the performance of the proposed schemes. Finally,

this chapter is concluded in Section 3.6.

3.2 Related Work

In this section, we brief some of related work that are closely related to this research.

Grossglauser and Tse showed that node mobility can be used to improve network

throughput [62]. After predicting the destination's location, a node forwards a data packet

to a subset of its neighboring nodes in the direction of the destination to reduce the overall

routing energy consumption [63]. Similarly, movement information was also used in [64] to

limit the flooding in a restricted area to reduce the energy consumption of route discovery.










Chakraborty et al. proposed a strategy to reduce the energy consumption by delaying

communication until a mobile node moves close to its peer target, within an application-

imposed deadline [65].

Different from the above work trying to utilize existing mobility information, there

are some work discussing how to move proactively or control node mobility to improve the

system performance. Li and Rus proposed an optimal algorithm to computer the trajectory

of nodes for minimizing message transmission delay [66]. Goldenberg at al. proposed a

distributed mobility-control scheme to guide the nodes to adjust their movements with the

purpose of attaining a potential energy-minimizing network configuration [67]. In sparse ad

hoc network, proactive node mobility can as well be used to overcome network partitions.

More specifically, nodes should buffer and carry packets during network partitions, and

forward packets to other nodes when they meet. Such a store-carry-forward paradigm was

proposed to help data delivery by making use of node mobility [68]. In addition, Zhao et al.

[69] proposed a message ferrying (MF) approach to address the similar network partition

problem. In their approach, a set of special mobile nodes, called mobile ferries, provide

data forwarding service to other nodes by either moving along the routes known a priori to

other nodes or proactively moving to meet other nodes.

Our resource-aware movement (RAM) strategy proposed in this chapter is also a

proactive movement approach. It differs from all the previous work in its unique designing

objective, that is, finding the optimal movement trails for each individual node to minimize

the total energy consumption of the whole system by taking into consideration the inherent

resource heterogeneity of MANETs.

3.3 Problem Formulation

We consider a MANET that consists of tens or even hundreds of mobile nodes, among

which there are N, regular battery-powered nodes (called B-nodes hereafter) and N, pow-

erful nodes (called P-nodes hereafter) having almost unlimited energy supplies such as

solar cells. Communication devices installed on a mobile vehicle and powered by inside










alternators are other examples of such P-nodes. Usually, N, is much smaller than Nr. We

assume that all the nodes are able to generate traffic or forward packets for others no matter

when they are at rest or in motion. Intuitively, since P-nodes have relatively infinite en-

ergy reservoir as opposed to battery-powered B-nodes, they should be utilized as much as

possible to save the scarce resources of B-nodes and thus prolong the whole network life-

time. For example, a packet should be forwarded to a P-node whenever possible if energy

savings can be expected. On the other hand, we should reduce the use of B-nodes if we

cannot completely avoid using them. How to realize this simple rationale, however, is by

no means a easy task.

In this dissertation, we intend to address the issue of energy conservation from the

viewpoint of node movement. In what follows, we first present a general mobility model

that is used to characterize nodes' movement patterns. We then introduce the resource-

aware movement problem in its general form with the consumption of energy resources as

the sole optimization objective.

3.3.1 General Mobility Model (GMM)

During an observation period T, we assume that there are some designated locations

that any node i, be it a P-node or B-node, should stop by at some designated time in-

stances. For example, a student carrying a mobile device may appear in the classroom

during schooltime while in the cafeteria during lunchtime. Let Ji denote the size of the

ordered list of locations node i should visit during T, which might be different for each

individual node. We denote by 1., -- (j) (0 < j < Ji) the jth location that node i should

stop by and by ti(j) the required time instance. Then 1., -- (0) denotes the starting point of

node i, and posi(Ji 1 ) denotes the location of its last stop during T. We will also call as an

epoch [70] the time duration from one node leaving the current stop until it reaches the next

stop henceforth. Whenever arriving at some designated location at the specified instance,

each node is assumed to pause for a while according to concrete application requirements.

Let pausei(j) indicate the time node i spends at 1., -- (j).









Based on the above definitions, the order list {ye- (j), ti(j), pausei(j),0 < j <

Ji} can well characterize the itinerary of node i during the observation period T. Let

posi(j)posi(j +t 1) denote the path travelled by node i from its jth stop to the (j 1t )th

stop. Provide that each node travels at a constant speed between two consecutive stops,

the travelling speed of node i from pos(i, j) to pos(i, j +t 1) is 1..1. tij)pouse(j ))

Notice that node i can follow potentially many different paths, for example, a straight one

or a zigzag one or even a tortuous one, as long as the time constraint is satisfied, that is,

it can reach 1., -- (j +t 1) at the time instance ti(j). However, once the path between two

consecutive stops is determined, the velocity of the node between these two stops is de-

termined and fixed. Therefore, once the paths between all pairs of consecutive stops are

determined, the movement pattern of a node during T is also determined.

The general mobility model (GMM) described above bears both similarities and dif-

ferences with the random waypoint model (RWM), which is the most commonly-used mo-

bility model in simulating MANET protocols [58, 70]. Both models are characterized by a

collection of locations of next stops, travelling speeds, and travelling time. Different from

GMM, RWM requires a node to first choose the location of its next stop and the travelling

speed, which leads to the determination of the travelling time. Our GMM actually does

the opposite by first determining the next visited location and the travelling time so as to

determine the travelling speed. The biggest difference, however, is that, in RWM nodes

always travel along the straight paths connecting two consecutive stops, while in GMM,

nodes can travel along arbitrary paths as long as they do not violate the time requirements,

that is, they should arrive at the designated stops at the specified time instances.

We believe that our GMM outperforms RWM in reflecting some practical scenarios,

which can be seen from a simple example in daily life. Suppose a student carrying a mobile

device should be in the classroom at fixed time everyday. He/she might have several options

of arriving in time at the classroom from his/her apartment: by foot through the shortest

path with the longest time, by bicycle through the second shortest path with the second










longest time, or by bus through the longest path but with the shortest time. Obviously,

RWM fails to model this case, but our GMM can.

3.3.2 Resource Aware Movement

As mentioned before, in the general mobility model there might be potentially many

different paths between any two consecutive stops. Define PATH(I., -- (0),1~ 4. (J1 1))

as node i's path set which is the concatenation of all the paths 1., -- (j J., -- (j +t 1). In this

dissertation, we are interested in finding the optimal path sets for all the nodes such that

the total energy consumption for communications by all the nodes during the observation

period T is minimized (the objective function), while all the ordered lists of visited loca-

tions and the corresponding time instances should not be violated (the constraints). To help

better understand the importance of this problem, we utilize the movement of a single node

between two consecutive stops as an example. As shown in Fig. 3-1, suppose an B-node

A should move from the current location 1 to the next location 2. It can choose the shortest

straight path (the dashed one) as it does in the random waypoint model. However, con-

sidering that node A may forward or generate packets destined for other nodes during the

movement process, the shortest straight path is not necessarily the best one for achieving

the system-wide energy efficiency. Instead, the dotted and solid paths are much better can-

didates through which node A can take advantage of more P-nodes by forwarding to the

encountered P-nodes the packets destined for other nodes and letting them finish the rest of

the task. Due to this reason, we call this problem the resource-aware movement problem in

that nodes now are moving with the system-wide resource (energy) consumption in mind

instead of moving blindly as before.

The general resource-aware movement problem itself is far too complicated to be

solvable. To render it tractable, we make some approximation and decouple it into two

relatively simpler subproblems: the Waterhunter Movement problem and the Firehunter

Movement problem. In the former, we assume that only B-nodes are capable of moving










Transmission
Range

I





eP-node () R-node

Figure 3-1: Multiple paths between two consecutive stops.


and all the P-nodes are stationary whose locations are known a priori to B-nodes. By con-

trast, in the latter, we assume that all the B-nodes are stationary and only P-nodes are able

to move.

Waterhunter Movement: In the network with N, stationary P-nodes and N, mobile

B-nodes, given all the order lists of {posi(j), ti(j), pausei(j), O < j < Ji}, the

Waterhunter Movement Probleml is to determine the optimal travelling path set for

each B-node such that the total energy consumption of the whole network during T

is minimized.

Firehunter Movement: In the network with N, mobile P-nodes and N, stationary

B-nodes, given all the order lists of {posi(j), ti(j), pausei(j), O < j < Ji}, the

Firehunter Movement Problem2 is to determine the optimal travelling path set for

each P-node such that the total energy consumption of the whole network during T

is minimized.




SIf we compare energy resources to "water", the movement of B-nodes is similar to
the behavior of water-hunters who are always looking for "fountains" (P-nodes), hence the
name.

2 If We COmpare energy resources to "water", the movement of P-nodes is similar to
the behavior of fire-hunters who are always looking for places "on fire" or "lack of water"
(B-nodes), hence the name.










The Waterhunter Movement problem is similar to the determination of people's opti-

mal travelling plans when lots of airline hubs and stops are available. On the contrary, the

Firehunter Movement problem bears similarity with the planning of hub locations when

airline companies build their global transportation networks with the purpose of providing

more convenience for passengers while reducing the overall system cost. Both problems

are pretty interesting and worthy of rigorous study. However, due to the space limitation,

we focus on finding a nearly optimal solution to the Waterhunter Movement problem in

this dissertation. Our investigation on the Firehunter Movement problem and the general

Resource-Aware Movement problem will be the future work.

3.4 Waterhunter Movement

In this section, we first present a simplified version of the Waterhunter Movement

problem, which is further reduced to a NP-complete distance-constrained least-cost (DCLC)

problem. We then present an efficient heuristic solution. Finally, we propose a routing

delay differentiation mechanism to utilize the benefits resulting from the resource-aware

movement.

3.4.1 Simplified Waterhunter Movement Problem

In the Waterhunter Movement problem, we assume that all the N, nodes are stationary

during the observation period T and are willing to forward packets for other less powerful

B-nodes. For simplicity, we do not dwell on how to place P-nodes to attain the optimal

system performance, which is believed to be a challenging problem itself and is currently

under investigation. Instead, we assume that each B-node knows the locations of all the

P-nodes and its own location at any time, and can as well adjust its moving direction at

will. For the time being, we assume here that all the P-nodes and B-nodes have the same

transmission range TR. We will discuss the case that P-nodes have the larger transmis-

sion range than B-nodes in Section 3.4.4. It is worth pointing out that the findings in this

dissertation can be easily extended to the case that each node has individual transmission

range.










The original Waterhunter Movement problem aims at minimizing the total energy con-

sumption of the whole network during the observation period T, which is a global optimiza-

tion problem and still too hard to be solvable. To make it tractable, we have to make some

approximations to get a suboptimal solution. We assume that a B-node moves at a constant

speed in one epoch, that is, between two consecutive stops, and the maximum speed it can

take is a system-wide value speedmax. Therefore, the longest path node i can travel in one

epoch is Im=, = speedmax x (ti(j +t 1) 14(j) pausei(j)). To achieve the system-wide

goal of energy conservation, instead of moving along the straight path connecting two con-

secutive stops 1.. -- (j) and posi(j +t 1), node i may travel along a resource-aware path with

the purpose of letting the encountered P-nodes forward on behalf of it as many as possible

packets destined for other nodes. For simplicity, we assume that when moving towards

a P-node, node i always goes along the straight path connecting the destined P-node and

itself. Therefore, 1., -- (j J., -- (j +t 1) is a set of zigzag straight paths if there exist multi-

ple P-nodes. Notice that the simplified target now is to find an optimal path set for each

individual node to minimize its total energy consumption during the observation period T

instead of that of the whole network. We intend to utilize the solutions to this local op-

timization problem to approximate the ones to the original global optimization problem,

which is believed to be too complicated to be tractable.

Normally, when moving between two consecutive stops posse (j) and posse (j +t 1), node

i may have several potential P-nodes to utilize. It is, however, usually unwise for node

i to pass by each of them. The reason is that, the longer path node i takes, the faster

speed it should move at, as described in the aforementioned general mobility model. It

is well-known that a faster movement speed may cause some undesirable problems such

as the instability of routing paths and the drop of packets. Therefore, some rules should

be designed to guide each B-node in deciding which P-nodes and in what order it should

pass through between two consecutive stops. A simple rule would be to only consider as

candidates the P-nodes whose distance from the direct link between any two consecutive















S~


Figure 3-2: An exemplary complete graph.


steps are no more than a threshold p. For example, p can equal 1.5 x TR, where TR is the

transmission range of each node.

With node i as an example, we can put the simplified Waterhunter Movement problem

in another way: given a source (i's current stop), a destination (i's next stop), and some

available intermediate P-nodes, and the path length constraint Imax, find a path as energy-

efficient as possible from the source to the destination, which is might be either the direct

link connecting the source and destination or a zigzag path through multiple P-nodes. Fig.

3-2 depicts such a topology, where a rectangular area, called the p-bounded rectangular

area hereafter, is formed such that only the P-nodes residing in this area are considered

as valid candidates. In addition, each link is of the forward direction from the source s

to the destination d simply because travelling backward is energy inefficient. We assign

to each link two weights, of which one represents the physical distance between two link

ends and the other indicates the virtual energy cost (defined shortly) incurred by choosing

this link. The simplified Waterhunter Movement problem can be boiled down to a distance-

constrained least-cost (DCLC) [71] routing problem which is formally defined as follows.



Consider a directed network that can be modelled as a complete graph G = (V, E),

where V is the set of vertices consisting of the source, the destination, and all the valid

candidate P-nodes, and E is the set of edges connecting each pair of nodes. V can be

further divided into two subsets, namely, ZA including the source s and destination d, and P

containing all the P-nodes. In addition, each edge e EE represents the movement from the









Transmission
Range




CI





SP-node O R-node

Figure 3-3: An exemplary resource-aware movement.


tail node to the head node. Let R*+ denote the set of non-negative real numbers. Each edge

e EE is associated with two non-negative functions: a distance function dist(e) : E R W

representing the physical distance between the end nodes of e and an energy cost function

cost(e) : E R +U {0}. More specifically, for a given edge e(vei, vy), its energy cost is
defined as

cos t(e(vi, vj))

f (dist(vi, vj)) g(dist(vi, vj)) vi = S, vj = d

f (dist(vi, vj)) g(dist(vi, vj) TR) vi = s, vj E p or vj = d, vi E p

f (dist(vi, vj)) g(dist(vi, vj) 2 x TR) vi, vj E p.

Here f is the cost, such as gas, fuel or other types of resources, required for the mechanical

movement ; g is used to reflect the cost for communications. g can be any non-decreasing

function that converts a given distance value into a non-negative cost, for example,



9(1~0 x < 0.





SIn this chapter, for simplicity, we assume that people on foot carry the communities
devices and we do not take the cost for the mechanical movement into account, that is,
f () = 0.










The motivation for the above definition of the edge energy cost is as follows. Whenever a

B-node moves into the transmission range TR of a P-node, it is capable of forwarding to

the P-node packets destined for other nodes so as to conserve energy. At one extreme, if

a B-node moves along an edge not (partially) covered by any P-node, all the packets from

this B-node would be forwarded to other energy-constrained B-nodes, which is the most

unfavorable situation. At the other extreme, if a B-node moves along an edge completely

covered by one or several P-nodes, all the packets from this node could be forwarded to the

P-node(s), which is the most desirable situation. Notice that the energy cost function given

above can well capture this effect. Though there might exist other meaningful metrics, we

believe the chosen one is very simple and tractable.

We also define the non-negative delay and cost functions for any path p as


dist(p) = dist(e)
etp

and

cost(p) = costle).
etp
Given the above definitions, the DCLC routing problem is to find a path p from s to d

such that min~cost(p), p E Pd} is achieved, where Pd is the set of all feasible paths from

s to d satisfying the distance constraint Imax, that is, dist(p) < Imax. Moreover, we define

Pra(s, d) as the path with the least distance from s to d, and Ple(s, d) as the path with the

least cost from s to d. Apparently, with the above definition of dist(e), Pla(S, d) is the

straight path directly connecting s and d.

It has been shown in [72] that the DCLC routing problem is NP-complete even for

undirected networks. In the following section we will propose an efficient heuristic algo-

rithm to provide a suboptimal solution to this DCLC problem and hence to the original

Waterhunter Movement problem.









Table 3-1: RAM: resource aware movement

1. Determine the candidate P-node set CHS
in the p-bounded rectangular area;
2. Construct a complete graph G with virtual
nodes s and d, and all the nodes in CHS;
3. Label each link e in G with cost(e) and
dist(e);
4. Find the DCLC path Pacle by calling
RAM-DCLC(G, s, t, Ims,);
5. Determine the travelling speed along Pacle
as seeddle = dist(Pacic)
as seeduic =t(i,j+1) -t(i~j) -pause(i,j) '
6. Move along Pacle at a speed of speeddcic;


3.4.2 RAM-DCLC Algorithm

As mentioned before, we assume that a B-node, say i, is aware of its own itinerary

(1** (j), ti(j), pausei(j), O < j < Ji} and the locations of all the P-nodes during the

observation period T. The procedure of node i's resource-aware movement from the cur-

rent location 1y. -- (j) to the next location 1y...- (j) is summarized in Table 3-1. Node i

first needs to determine the candidate P-nodes in the p-bounded rectangular area and

then constructs a complete graph like the one in Fig. 3-2, consisting of the vertices s

(a virtual node at 1.. -- (j)), d (a virtual node at 1., -- (j +t 1)), and all the found candi-

date P-nodes. It then proceeds to calculate the distance and the energy cost for each

link and finally generates the weighted graph G. The next step is to call the process

RAIM-DCLC given in Table 3-2 to get the DCLC path Pacle whose length is bounded by

Imaz = speedmax x (ti(j +t 1) ti(j) pausei(j)). It then moves towards 1., -- (j +t 1) at

a constant speed of speetddcic = ti"+1)(t, d- sei') along the found Pace,. Upon reach-

ing 1y. .- (j +t 1), node i pauses for a period pausei(j +t 1). Following the previous pro-

cess, it can then move towards the next stop pausei(j +t 2) until all the required stops

{posi (j), O < j < Ji} during the observation period T are visited.




















1. For each node I1 in G, find the Plc(I, ,d)
and Ple('r -, d) and
their respective next hops nid(Pic( -.d))
and nid(P~d(r I, d));
2. distSoFar = 0; Pacle = S; ThisNode = s;
3. while (ThisNode / d) do
4. if ((dist(Ple(ThisNode, d))+
distSoFar) < Imom) then
5. vu = nid(Ple(ThisNode, d));
6. distSoFar = distSoFar+
dist (This Node, u) ;
7. Pacle = Pacle + {v);
8. ThisNode = vI;
9. else
10. for each neighboring node w ( Pacle
do
11. calculate weight(ThisNode, w);
12. end for
13. v1 = extract(ThisNode);
14. distSoFar = distSoFar+
dist (This Node, u) ;
15. Pacle = Pacle + {v);
16. ThisNode = vu;
cp:ref35 17. end if
18. end while
19. Return Pacle


Table 3-2: RAM-DCLC: a DCLC routing algorithm for the Waterhunter Movement









The proposed DCLC algorithm RAIM-DCLC is summarized in Table 3-2, in which the

weight() function is defined as follows:


Cost (vi, vj) cost' (vy d) cond (1)

and

cost'(vj, d) = cotPev,))...1
cost(Pza(v ,d)) 0.w.
where

cond(1) = distSoFar dist(vi, vj) dist(P~a(vj, d)) < lmax

and

... c!ol(1) = distSoFar dist(vi, vj) dist(Ple(vj, d)) < lma,

The function extract() is used to choose the node, say w, whose weight(vei, w) is the

minimum one among all the neighboring nodes of ThisNode. If more than one node have

the same minimum value, it chooses the one with the smallest distSoFar +t dist(ve, vyj)+

dist(Pza(vj, d)).
In the RAIM-DCLC algorithm, for each node vk~ in G', the Bellman-Ford or Dijkstra

shortest-path algorithm can be used to find the PIe(r -, d) and Pra(r-, d) and their respective

next hops nid(Ple d)) and nid(Pza(r-, d)). Since the optimization objective is the path

cost, at each intermediate node vI, RAIM-DCLC always chooses the next hop w with mini-

mum cost(vl, w) + cost'(w, d) while not violating the distance constraint Imax. RAIM-DCLC

is able to find a feasible path satisfying the distance constant while keeping the path cost as

small as possible. In particular, we have the following theorems for this algorithm 4

Theorem 1: RAIM-DCLC can always find a feasible path from a source s to a destina-

tion d satisfying the given distance constraint Imax if such feasible paths exist.


4 The correctness of these theorems can be justified following the proof in [73].








64




Pnds (S 0) .(S 858.4)
SStops
250 -RAM1 3
--RWM 1
200 (Stop, Arrival time)
E (S2, 247.8)"
150- 11


( S 4 6 2 5 8 )- ,8~


.S,387. 1)
U200 400 600 800 1000 1200 1400 1600
x (m)



Figure 3-4: Resource-aware movement (RAM) vs. random waypoint movement. A B-
node should consecutively visit S1-S2-S3-S4-S5 and the solid lines are labelled by the se-
quences they were passed through.


Theorem 2: The path found by RAIM-DCLC is loop-free.

Theorem 3: RAIM-DCLC always terminates in finite time. Now we utilize the example

given in Fig. 3-3 to illustrate the resource-aware movement process using the proposed

R 4M DCLC algorithm. Suppose node A intends to move from its current location to

the location where A' resides according to its itinerary. In this example p is set to 1.5 x TR

so that there are four candidate P-nodes. Based on the output of RAM~ DCLC, node

A should move along the DCLC path denoted by the solid line instead of the straight path

denoted by the dash line. In this way, energy savings can be expected by forwarding to

the two P-nodes the packets it carries for other nodes and letting the P-nodes finish the rest

transmissions (either single-hop or multi-hop) on behalf it.

Fig. 3-4 compares a B-node's resource-aware movement trail and its random way-

point movement trail with 8 P-nodes in a 1500 x 300m2 field. The data used were gen-

erated using OPNET [57] and the pause time of the B-node was 120s. In addition, the

transmission range of each node was set to 250m. It is clear that our proposed resource-

aware movement strategy enables the B-node to have more opportunities of approaching

and utilizing the P-nodes as compared to the random waypoint movement.










3.4.3 A Routing Delay Differentiation Mechanism

The resource-aware movement strategy only enables B-nodes to move close to P-

nodes by taking the physical DCLC paths, but cannot guarantee a packet will be forwarded

from a B-node to a P-node even when they are close to each other. As a result, we need

to incorporate such nodal information as device types and residual energy into the routing

metric such that the P-nodes can be naturally selected onto the routing paths of the B-

nodes. For this purpose, we propose to map the residual energy information into the random

jitter delay for which each node has to wait before propagating a route request in a regular

MANET routing protocol such as DSR [70]. The rational here is that the less residual

energy a node has, the longer delay a routing request packet should experience at this node.

In this way, the nodes with less residual energy would be less utilized, while the nodes with

more residual energy such as P-nodes would be more utilized.

Below we illustrate the delay differentiation mechanism using DSR [70] as the under-

lying routing protocol, though our mechanism can as well be applied to other MANET rout-

ing protocols. Similar to the interframe spacing (IFS) differentiation mechanism adopted

in the IEEE 802.11e [74], we first define the Delay Diffi I~i~illaio as 6 x E,3/Et, where 6

is a delay control parameter, and Ec, and Et are the residual energy at time instance 0 and

t, respectively. When receiving a route request, each node should first compute the delay

it needs to wait for before further propagating the route request. Since the P-nodes have

relatively unlimited residual energy as compared to the B-nodes, that is, Ec, a Et, they

would forward route requests much more quickly than the B-nodes. Since the destination

usually only replies to the first arrived request, the P-nodes with powerful energy resources

would have higher chances of being chosen onto the routing paths and thus being involved

in more data forwarding activities than the B-nodes with less energy. With this delay dif-

ferentiation mechanism, the benefits resulting from the resource-aware movement can be

better utilized. The nice feature of our mechanism is that it is very easy to implement.










3.4.4 Incorporate RAM into DELAR

In our previous discussion, we assume that the P-nodes have the same transmission

range TR as the B-nodes. However, in more practical heterogeneous MANETs, P-nodes

may have greater transmission capabilities than the B-nodes. To make use of such pow-

erful P-nodes especially when P-nodes can adjust their transmission power to cover larger

range than B-nodes, in our previous work DELAR, we proposed a energy efficient relaying

framework as a joint design of scheduling, routing and power control to efficiently utilize

P-nodes to conserve energy. Since RAM and DELAR utilize powerful nodes to conserve

energy from totally orthogonal perspectives, RAM can be directly incorporated into the

DELAR scheme. With the help of DELAR, RAM can further improves the energy effi-

ciency when P-nodes and B-nodes have different transmission capabilities.

3.5 Performance Evaluation

In this section, we use simulations to evaluate the impact of the proposed schemes on

the energy conservation and other system performance factors.

3.5.1 Simulation Setup

We implemented the resource-aware mobility model, the routing delay differentiated

mechanism, and DELAR (cf. Chapter 2) in OPNET [57].

We simulated a network with N, B-nodes and Nz, P-nodes in a 1500 x 300 field,

where N, = 46 and N,, = 4. All the B-nodes were capable of moving in the field, while

all the P-nodes were fixed. Though a careful deployment of P-nodes may improve the

system performance [75], we simulated a worse scenario that the P-nodes were randomly

deployed in the field. The transmission range of the B-nodes was 250m. For the P-nodes,

we simulated two cases in which the P-nodes had the transmission ranges 250m and 500m,

respectively.

The energy consumption for the B-nodes followed the linear energy model proposed

in [76]: ener yy = m x len ythtb, where m is an incremental cost of each operation, b is the







67

Table 3-3: Energy consumption parameters.

Symbol Value Unit
mend 1.89 uW-sec/byte
bsend 246 uW-sec
meec, 0.49 uW-sec/byte
bree, 56.1 uW-sec
bsenacet 120 uW-sec
breevenl 29 uW-sec


fixed cost of each operation, and length is the size of the frame sent/received5 The (m, b)

values provided in [76] were summarized in Table 3-3 and used in all the calculations.

We intended to compare the proposed resource-aware mobility model (denoted by

RAM) with the modified random waypoint model (denoted by RWM) presented in [58],

which can guarantee the convergence of average nodal speed throughout the simulation

time. For this purpose, we first ran the simulations using RWM and recorded the stops,

the starting/arrival time instances, the moving directions, and the movement speeds of

all the movement epochs. We then used this movement profile to generate the itinerary

{posi (j), ti(j), pausei(j), O < j < Ji} for each node such that in both models each node

would stop by the same stops at the same time instances, but follow totally different move-

ment trails and take different movement speeds. Both models had the same maximum

speed 20 m/s and we adjusted nodal pause time to vary the network mobility. The traffic

used were 20 CBR connections with randomly selected source-destination pairs. All the

data packets were 64 bytes and were sent a speed of 4 packets/second. Each simulation was

executed for 15 simulated minutes and each data point represents an average of ten runs

with identical traffic models, but differently generated mobility scenarios.




5 For the A-MAC control frames P-RTS/P-CTS/P-ACK, the fixed costs bsenactl and
breevenl were used because they have the similar size.










3.5.2 Simulation Results

We compared RAM with RWM in terms of the commonly used metrics including

packet delivery ratio, average packet end-to-end delay, average packet energy consump-

tion, and average routing overhead. Motivated by the small-world phenomenon [77], we

used two additional metrics, average path length and average clustering coefficient, which

are two defining characteristics of small-world networks. The former means the average

number of hops a packet may travel through, while the latter indicates the connectivity of

an average neighborhood in the network, defined as the average node degree divided by

the network size [77]. The simulation results are presented in Fig. 3-5, where RAM-1

indicates the case that the P-nodes and B-nodes have the same transmission range 250m

and RAM-2 denotes the case that the P-nodes have a larger transmission range 500m.

Fig. 3-5(a) and 3-5(b) compare RAM and RWM with regard to average path length

and average clustering coefficient, respectively. We can see that RAM can shorten the

average path length and increase the clustering coefficient as compared to RWM. That is

because in RAM the B-nodes are always trying to move towards some P-nodes between

two consecutive stops. Such behaviors would effectively bring more B-nodes to the vicin-

ity of P-nodes, leading to shorter paths and larger clustering coefficients. In some sense,

such resource-aware movement creates a "small-world" network, which results in some

performance gains shown below. In addition, since the P-nodes in RAM-2 have a larger

transmission range and thus have more neighbors that those in RAM-1, we can observe

that RAM-2 further reduces the average path length and increases the average clustering

coefficient.

Fig. 3-5(c) compares the average packet delivery ratios of RAM and RWM, which is

defined as the ratio of delivered data packets to those generated by the sources. As we can

see, the PDR of RAM-1 or RAM-2 is always higher than that of RWM. This result is of

no surprise since the shorter average path length implies the network-wide less traffic load

and less packet drops due to the MAC-layer contention and channel errors, all of which





























































0 100 200 300 400 500 600 700 800 900
Pause t me (s)


(d) Average packet end-to-end delay.



-4- RAM-1
-8- RAM-2



-


0c-_~


5i


X-


100 200 300 400 500 600 700 800 9000
Pause tme (s)


(a) Average path length.





-a RAM-1
-8 RAM-2


(b) Average clustering coefficient.









h-RA-


~0014
0 012
0 01


S0- 006


t


0 100 200 300 400 500 600 700 800 900
Pause tme (s)


(c) Packet delivery ratio.






RAM-1
X-* I,- RAM-2


100 200 300 400 500 600 700 800 900 0 100 200 300 400 500 600 700 800 900
Pause t me (s) Pause t me (s)


(e) Average energy consumption. (f) Average routing overhead.


Figure 3-5: Simulation results.










would contribute to the increase of the PDR. Due to the same reason, RAM-2 demonstrates

a higher PDR than RAM-1.

Figure 3-5(d) depicts the comparison of average end-to-end packet delay, defined as

the time duration from when a packet is generated till it is received by the destination.

The shown advantage of RAM over RWM mainly results from the aforementioned shorter

average path length and higher clustering coefficient. Again, RAM-2 outperforms RAM-1

in reducing average packet delay because of the shorter average path length.

Figure 3-5(e) shows average energy consumption, defined as the total energy consump-

tion for transmitting and receiving all data and routing packets divided by the number of

delivered packets. Apparently, our RAM can conserve a significant amount of energy as

compared to RWM because the B-nodes are always moving along the paths through which

the P-nodes can be utilized as much as possible. Since the P-nodes in RAM-2 have a larger

transmission range, statically less B-nodes are involved in packet transmissions and thus

RAM-2 can help the system conserve more energy than RAM-1.

Figure 3-5(f) demonstrates average routing overhead, defined as the average number

of routing packets involved in delivering 100 data packets. As we can see, our RAM has

smaller routing overhead than RWM. That is because in RAM packets can be forwarded to

their destinations through shorter paths in shorter time, thus fewer routing errors occur.

To summarize, the proposed resource-aware movement strategy has many significant

and positive impacts on the system performance. It makes the network more like a "small-

world" network with shorter average path length and higher clustering coefficient. This

results in improved packet delivery ratio, shortened end-to-end delays, and most impor-

tantly, much better energy conservation. Therefore, the combination of node mobility and

heterogeneity is a valid means to address the energy conservation issue in MANETs. The

results also suggest that making the network a small-world network may have a lot of pos-

itive effects on the system performance and it deserves further investigation.










3.6 Summary

In this chapter, we studied the energy conservation problem from the new perspective

of node mobility, that is, by analyzing the impact of node movement on the system-wide

energy conservation. We proposed a novel resource-aware movement strategy to take full

advantage of some powerful nodes in heterogeneous mobile ad hoc networks. We then for-

mulated the resource-aware movement as a NP-complete distance-constrained least-cost

(DCLC) routing problem and proposed an efficient heuristic solution. Moreover, we pro-

posed a simple yet effective routing delay differentiation mechanism to virtually utilize

the benefits from the resource-aware movement. In addition, the proposed resource aware

movement strategy can be incorporated into other energy conservation schemes, for exam-

ple, DELAR, to further improve the energy efficiency. We used extensive simulations to

show the effectiveness of the proposed scheme.















CHAPTER 4
SUPPORT DIFFERENTIATED SERVICES IN MOBILE AD HOC NETWORKS

4.1 Introduction

In the literature there are proposals addressing the issue of quality of service in wired

networks, however, these proposals may not be feasible in the wireless counterpart if we

do not modify them for the wireless environments. The system dynamics [18] of multi-hop

mobile ad hoc networks, such as time-varying and error-prone wireless links, dynamic and

limited bandwidth, time-varying traffic pattern and user location, and energy constraints,

pose new challenges that do not exist in wired networks. To conquer these challenges, in re-

cent years, many researchers advocate a cross-layer design philosophy to develop protocols

and applications for MANETs. This is a departure from the traditional layered design for

the Intemet. Though the cross-layer design philosophy might not be an optimal solution,

it does provide us new network implementations that may better support the amalgamation

of user services and QoS requirements [19].

To efficiently handle heterogeneous traffic over wireless links, we need to address

two problems. The first is to handle reliable mobile communications in MANETs. This

problem has been extensively studied in recent years, and many proposed routing protocols

such as DSDV [7], DSR [9], and AODV [10], and medium access control mechanisms

such as MACAW [3], FAMA [4], and IEEE 802.11 [5], aim to achieve efficient reliable

communications. The other problem is to provide QoS provisioning for heterogeneous

traffic with different quality-of-service (QoS) requirements in terms of BER, throughput,

and delay. Since the channel bandwidth in wireless environments is limited, one strategy to

support QoS is to set up some kind of priority scheme or service differentiation mechanism

[78][79], under which delay-sensitive traffic has higher priority to access the channel over

less time-critical traffic.










In the current literature, many scheduling mechanisms for wireless networks are pro-

posed for this purpose, though most of them are not directly designed for MANETs. In

general, these scheduling mechanisms all attempt to combat the channel impairments and

to support heterogeneous traffic with the following goals: providing high wireless chan-

nel utilization, long-term fairness, bandwidth guarantees and delay bounds for flows with

error-free links or links with sporadic errors [80]. However, these algorithms may not be

practical to be implemented in MANETs. Actually, it is hard, if not impossible, to achieve

those goals simultaneously because of their conflicting nature. For example, there is a

tradeoff between the throughput and fairness or so-called inter-class effects [81] among

traffic with different priorities. Without any precautionary measures, this conflict may lead

to bandwidth starvation for low-priority traffic when the high-priority traffic load is high.

Meanwhile, most of these scheduling mechanisms are suitable for the reservation-based

MAC protocols, especially for those designed for cell-structured wireless networks. In

networks with contention-based MAC protocols such as IEEE 802.11 [5], the reservation-

based scheduling mechanisms may not be applicable because it is not easy for a node to

reserve resource mna contention manner.

In this dissertation, we attempt to avoid the conventional scheduling approach, and

propose a novel scheme called Courtesy Piggybacking (CP) to alleviate the conflict be-

tween throughput and fairness. The basic idea of CP is to let the high-priority traffic help

the low-priority traffic by sharing unused residual bandwidth with courtesy. Our scheme

closely follows the cross-layer design principle and exploits the system dynamics as much




SWe only consider the fairness problem between different classes of traffic, for example,
each class of service should be allocated some bandwidth rather than being completely
starved; while the fairness problem between different nodes, for example, each node should
have fair opportunity to access the channel in the short or long term, is out of the scope of
this dissertation.










as possible; that is, we effectively employ the dynamic channel conditions and the result-

ing dynamic bandwidth, and the dynamic characteristics of the heterogeneous traffic. Note

that not only is our scheme suitable for multi-hop mobile ad hoc networks with underlying

contention-based MAC protocols, but also it is applicable to those with reservation-based

or hybrid MAC protocols. Meanwhile, our scheme is shown to be easily implemented.

The rest of the chapter is organized as follows. We discuss some related work in

Section 4.2. In Section 4.3, we show the motivation of our proposed scheme. In Section 4.4,

we discuss the relationship between the SNR and the optimal packet length, and come up

with a Finite State Markov Chain channel model based on the packet length. Our Courtesy

Piggybacking scheme is described in Section 4.5. We present some preliminary analytical

results in Section 4.6 and evaluate our scheme with extensive simulation in Section 4.7.

Finally, we conclude the chapter in Section 4.8.

4.2 Related Work

As we mentioned in Section 4.1, scheduling is one promising way to support hetero-

gonous traffic with different QoS requirements. For scheduling mechanisms, throughput

and fairness are two main objectives to be met through bandwidth allocation with admis-

sion control and congestion control. Many scheduling algorithms such as fair queuing

scheduling [82], and virtual clock [83] are capable of providing certain QoS guarantee for

wireline networks, and many scheduling algorithms such as IWFQ [84], CIF-Q [85], CS-

DPS [86], and CSDPS + CBQ [87] are proposed for the wireless networks, especially for

wireless cellular networks. However, little progress has been made along this direction in

wireless mobile ad hoc networks with underlying contention-based MAC protocols. CS-

DPS and its improved version CSDPS+ CBQ are two of scheduling mechanisms that may

be applicable to the ad hoc networks with contention-based MAC protocols. In CSDPS,

packets to be transmitted to the same receiver are queued in the same queue and are served

in an FIFO fashion. At a node, the different queues are served according to some poli-

cies such as round robin, earliest timestamp first, or longest queue first. The basic idea










of CSDPS is as follows: when the link towards a receiver is bad, the node should defer

the transmission of packets in the queue corresponding to that receiver. With CSDPS, it is

easy to alleviate the head of line (HOL) problem when single FIFO queue is used. Since

CSDPS makes use of the channel state information, it can achieve high data throughput

and channel utilization. However, it does not address the fairness issue. To improve the

fairness in CSDPS, class-based queuing (CBQ) [88] is used together with the CSDPS. By

using CBQ, a hierarchical channel-sharing mechanism, it can achieve certain fairness, and

ensure that different traffic classes can share the overall bandwidth, while maintaining the

features of CSDPS to deal with the channel variations. Unfortunately, this scheme is also

complicated in keeping track of the amount of service each class has been served. Efficient

and less expensive mechanisms are very desirable to alleviate the conflict of throughput

and fairness in MANETs. More and comprehensive materials on scheduling can be found

in [80]. Besides, some QoS adaptive schemes such as SWAN [16] and Havana [89] are also

available in the literature. These schemes adaptively perform admission control and rate

control according to the user QoS requirements and channel states.

The main reason leading to the conflict between throughput and fairness is the lim-

ited bandwidth of the wireless link. If the system can provide plenty of bandwidth, the

conflict problem would not be so significant. Recently, many adaptive transmission tech-

niques are proposed to exploit the channel dynamics to provide more bandwidth. These

schemes can adaptively adjust the parameters such as modulation level and symbol rate to

maintain an acceptable BER without wasting much bandwidth. Koutsopoulous and Tassi-

ulas proposed to integrate adaptive transmission techniques, resource allocation and power

control for TDMA/TDD system so that higher modulation levels can be assigned to users

in good channels to enhance the throughput, while power control can be used to reduce the

interference and increase the system capacity [90]. In addition to these schemes proposed

for wireless cellular networks, some rate-adaptive schemes are also proposed to improve

the system throughput in WLANs. Holland et al. proposed a rate adaptive MAC protocol










called RBAR, which uses the RTS/CTS to exchange the channel state information and the

optimal rate on a per-packet basis [91]. Unfortunately, this scheme needs to make some

modifications to the IEEE 802.11 MAC protocols. To avoid this modification, Chevillat

et al. proposed a scheme to select the optimal rate only with the local information at the

transmitter [92]. This scheme is based on the history of attempted transmissions. It uses

one successful transmission count and one failed transmission count to indicate the chan-

nel state and to determine the optimal rate the transmitter can use. For IEEE 802.11 MAC

protocols, adaptive fragmentation schemes can also be designed with the rate adaptation to

enhance the system throughput [93][94][95].

For all the scheduling mechanisms and other channel-dependent schemes, including

our Courtesy Piggybacking scheme, designed for wireless networks, they all have to mon-

itor the channel quality based on the symbol error rate, bit error rate, and receiver signal

strength. The more accurate the channel information is, the more benefits these schemes

can bring to the system design. In general, the channel estimation can be performed by

the sender or by the receiver. Since the channel information used in all channel-dependent

schemes is the one seen by the receiver, the receiver-based channel estimation is more at-

tractive. However, the channel information needs to be sent back to the sender, which is

sometimes costly in terms of the resource used to transmit the channel information, thus

certain performance tradeoff has to be made between estimation accuracy and overhead.

More details about channel quality estimation can be found in [96].

4.3 Motivation

Consider the scenario depicted in Fig. 4-1. In a mountain area, the only way from

Anchorage to Whittier (the access to see the spectacular glacier) is to pass a tunnel near

Portage running through the Chugach Mountain Range (that is, the longest tunnel in North

America the Whittier Timnel in Alaska). The situation is the same from Seward to

Whittier. People have several choices to pass the tunnel: by train (high priority), by car,

by bicycle or on foot (low priority). Only one direction traffic is allowed during one period


















// Tunnel

( S -Road


Figure 4-1: The Whittier Tunnel scenario.


of time. To pass the tunnel, when the train approaches the tunnel, all other traffic stops

and waits until the train passes the tunnel. Often, there is a long traffic line waiting to pass

the tunnel, especially for the direction from W to P when traffic load is high, for example,

during rush hour in the afternoon. In order to quickly pass the tunnel, a better approach

for other transportation users would be to check if there is any free space left in the train.

If there is, these users could ask for permission to ride at a certain cost and according to

some rules, for example, how many free space in terms of basic units is left and what kind

of traffic (priority) the train can accommodate. After passing through the tunnel, the piggy-

backed traffic can get off the train at P and continue on its own way. Of course in the real

situations, when passengers by car, by bicycle or on foot pass through the narrow and dark

tunnel in a sequential manner, the traffic usually moves very slowly for the sake of safety.

Thus it is advisable for cars that have free space to piggyback those passengers by bicycle

or on foot according to some rules to benefit all the traffic.

We can think of these rules as being concerned with the HOW MANY-WHO problem,

that is, how much free space is available and who can enjoy such free space? If we only

consider the free space FS in the train as a function of time, then we could consider the

following scenario as an example: one person would occupy 1 basic space unit, a bike 2

units, and a car 6 units. If we have some predefined objective to meet, then we can de-

sign different piggybacking rules to solve the HOW MANY-WHO problem. For example,










suppose our objective is to maximize the revenue of the train. With different piggyback-

ing costs, for a given FS we can achieve the optimal allocation scheme for the free space

among different traffic: cars, bicycles, and pedestrians.

The above scenario is very similar to multi-hop mobile ad hoc networks supporting

differentiated services. The piggybacking strategy described above motivates us to develop

a more efficient way to alleviate the conflict between throughput and fairness for different

prioritized services. First of all, we need to identify the "free space" in a MANET. Fortu-

nately, we do have two sources that can provide us with such free space. The first one comes

from the time-varying channel conditions. In recent studies such as [90] [93], the MAC and

PHY layers adapt to the channel state by using adaptive transmission schemes to provide

higher data rates when the channel is good. With a higher data rate, the transmission time

for MAC protocol data unit (MPDU) can be shortened, leading to some potential idle time

if the transmitting node does not have further data to transmit. If the IEEE 802.11 MAC is

used, the NAV (Network Allocation Vector) setting may prevent other nodes from using the

medium, even though it is idle (the rule of virtual collision avoidance). This idle period will

be the "free space" and should be more effectively used. The second source comes from

the traffic characteristics. When we look into the traffic patterns and the stochastic traffic

behavior, sometimes the high priority traffic may not have enough data during the reserved

slots in a reservation-based system or their transmission period in a contention-based sys-

tem (for example, a network with IEEE 802.11) to fully utilize the channel capacity. For

example, consider a network with reservation-based MAC protocols. In addition to the

"free space" provided by the channel dynamics, when the packets from one high-priority

flow are not enough to fill the reserved slots, for example, during silent periods for voice

connections, some "free space" can be harvested to piggyback some bits from the queue(s)

with low priorities.

When such free space is available, the next problem would be how to make use of it

to fulfill certain objectives such as fair allocation of bandwidth. While one would think









that it should be used to better support high-priority traffic in the first place, we argue

that it may not necessarily be the case. Rather, the piggybacking rules should be properly

designed in light of specific requirements of various applications. If some delay-sensitive

applications like voice or video-telephony require that their packets get through the network

as quickly as possible, then the free space should be used to meet such needs. On the other

hand, if high-priority traffic does not need more resource than needed, a piggybacking rule

favoring low priority may be more reasonable. For streaming multimedia applications,

as an example, when the QoS requirement of one stream with high priority has already

been met, there is no need to piggyback packets belonging to this stream ahead of the

scheduled time; instead, piggybacking packets from other low-priority streams may be

more beneficial.

In the following sections, we will elaborate more on why free space exists and how

piggybacking can be used to achieve our goal alleviating the conflict between throughput

and fairness for different prioritized services.

4.4 Packet-length-based Channel Model

In the current literature, the time-varying channel is commonly modeled as the well-

known Gilbert-Elliott two-state Markov channel model (Fig. 4-2). Each state in the two-

state Markov chain model represents a binary symmetric channel (BSC). In "Good" state,

the BSC has low crossover probability, P,, and int the "Bad" state, the BSC has high

crossover probability, Pb. The transition probability matrix can be given as:



SPGGPB PGBPB


Given the transition probability, it is easy to determine that the steady state probabilities

are













PG ( Good ; ( Bad ) Pss

0 0G
0P9~ ----- lPb
G B
199 Pbl


Figure 4-2: The Gilbert-E11iott channel model.


We notice that if P, and Pb are set to 0 and 1, respectively, that is, a packet succeeds

with probability 1 in the "Good" state and is lost with probability 1 in the "Bad" state, the

two-state model is reduced to the simplified Gilbert model.

When the channel quality varies dramatically, it is not accurate enough to model the

channel as a two-state Gilbert-E11iott model. In this case, a finite-state Markov channel

(FSMC) [97] can be used. By using the received signal-to-noise-ratio (SNR) as the only

side information, the FSMC provides a mathematically tractable model for time-varying

channel. Let g* denote the received SNR that is proportional to the square of the signal

envelop. Then, for a Rayleigh fading channel, the probability density function of ]* can be

written as




where 1' is the mean of ]* (actually, it is an exponential distribution with mean 1'). In order

to build the finite state Markov chain model, we assume the received SNR remains at a

certain level for the duration of a symbol, and we partition the range of the received SNR

into a finite number of intervals. Let 0 = yo < yl
thresholds. For each interval, we associate it with a state Sp, kc = 0, 1, 2, 3, ..., K 1. The

channel is in the state SI, if g* is in the interval [ys, y*, ]. We know that there is a crossover

probability p for a given SNR y. When BPSK is used, this probability can be written as a

function of ]*:













5000-

S4000-

$3000

S2000-

1000-



y (dB)


Figure 4-3: The optimal packet length (PL) vs. SNR (y ), h=128.




0, 10 1,1 t 2 ,2 k-1,k-1


Figure 4-4: Packet-Length-Based finite-state Markov channel model.


According to [98], for a given crossover probability p, the optimal packet length,

which is a function of p, can be written as


-h ln(1 p)- -4 n1-p+h22
PL =(4.3)
2 In(1 p)

where h is the number of overhead bits per packet. Fig. 4-3 shows the relationship between

the received SNR and the optimal packet length. For a given state Sk, the average optimal

packet length PLk for this state can be derived by using (4.1), (4.2), (4.3) to be






Based on the above analysis we present our packet-length-based FSMC model in Fig.

4-4. We represent each state as the average packet length PLk, which is the packet size for

a transmission in state k. The transition probabilities between different states are denoted










as tis. Further, we can derive the state steady probability for each state as




In practice, we may use different modulation schemes (not necessarily BPSK) in dif-

ferent channel states. Moreover, by properly partitioning the range of the received SNR,

we may obtain the multiplicative relationship between the average optimal packet lengths.

4.5 Courtesy Piggybacking

In this section, we present our Courtesy Piggybacking scheme to alleviate the conflict

between throughput and fairness and to combat the starvation problem for differentiated

services.

4.5.1 System Assumptions

We consider an ad hoc network consisting of n mobile nodes uniformly distributed in

some area. Nodes can communicate with each other directly if they can hear each other

or through other relay nodes in a single broadcast channel. They employ some contention-

based MAC protocols, such as IEEE 802.11, to support their communications. Each node

can generate services with N different priorities destined to other mobile node(s). A node's

mobility follows the random waypoint model [58, 70]. At first, a node stays at a position

for duration of pause_time. After that period, the node chooses a new random position

and moves towards that position at a random speed uniformly distributed in the range from

min~speed to maxspeed. After reaching the new position, the node will stay there for an-

other pause_time. This process will continue for each node until the end of the simulation.

We assume some service differentiation mechanism is employed at the network layer.

All the heterogeneous traffic is prioritized at its originating source node. When a packet is

handed down from the network layer, it will be kept in the Tx queue corresponding to its

priority and wait for its turn to be transmitted at the MAC layer.

From the previous section, we know that the packet length is related to the received

SNR. The greater the SNR is, the greater the packet length is. In the IEEE 802.11 MAC












hMad~ r MSDU CRC hIadC SDU CRC h~ajdC MSDU CR

Figure 4-5: A fragmentation example.


protocol [5], this packet length may be called as frame length, which equals the fragmen-

tation threshold plus the length of the MAC header and the length of CRC. In the IEEE

802.11 standard, the MAC layer takes a MSDU from the Tx queue and adds MAC header

and a CRC to each MSDU to generate a MPDU. In order to reduce the probability of

transmission errors, the IEEE 802.11 limits the size of the body of a MPDU to be less

than a fixed fragmentation threshold (FT), or it will break the long MSDU into multiple

fragments, each of which will be no longer than the FT. In Fig. 4-5, we show a case

where a long MSDU is partitioned into three small MSDUs in the IEEE 802.11. Since the

length of the MAC overhead may be kept unchanged, according to the analysis in the previ-

ous section, different channel states have different frame lengths, we can say that different

channel states have different fragment thresholds (FTs). The greater the received SNR is,

the greater the fragment threshold (FT) is. Hereafter, we associate FTI, with each state SI,

of the FSMC model as depicted in Fig. 4-4. In order to improve the channel utilization, we

assume that the MAC protocol can adaptively adjust the fragmentation threshold and the

transmission rate according to the channel state. To accurately determine the channel state

when some packets need to be transmitted, we further assume that we have some channel

estimators or predictors, which can provide the accurate channel information for the proper

MAC layer fragmentation.

4.5.2 The Courtesy Piggybacking Scheme

In practice, the size of a packet generated by an application may be fixed or may vary

from a minimum allowed size to a maximum value PK,nct i We argue that the PK,,c

should be properly chosen to reduce the overall overhead. Suppose we want to transmit c

Mbits traffic. Packets are generated according to the PK,nct i We assume that each packet











4



S10 -i ~:







4000 1001200(
Fragmentation Threshold (blt) PKrnax (bit)


Figure 4-6: The total overhead with PKmax and FT.


can be correctly received without any retransmission2 Then, the total overhead should be

the sum of the overheads 04, at the IP layer (for example, 20 bytes for IPv4), Omac at the

MAC layer (for example, 34 bytes for the IEEE 802.11) and Ophy at the PHY layer (for

example, 16 bytes). Thus, the total overhead to transmit the c Mbits traffic can be written

as




where [*] is the function to round the element to the nearest integer greater than the ele-

ment. We show the relationship of overhead vs. PKmax and FT when c = 1 in Fig. 4-6.



From Fig. 4-6, we observe that PKmax should be reasonably chosen when multiple

fragmentation thresholds are used. It cannot be too small, as it may cause too much over-

head; neither can it be too large, as it may generate too many fragments when the FT is

small, which may further degrade the overall throughput. For example, a large packet may




2 For simplicity, we only consider the case without any retransmission and view the
resulting overhead as a lower bound. Apparently the overhead with retransmissions is
larger than this lower bound.










be partitioned into many small fragments. Each fragment is augmented with an individ-

ual header, sent as an independent transmission, and acknowledged individually. Hence,

a large packet will cause lots of DATA/ACK exchanges and result in suboptimal perfor-

mance. Thus, there must exist an optimal value of PKlax such that the overall overhead

associated with the successful transmission of a message is minimized. Assume we obtain

P~lax, which may not be equal to any of FTe(k = 0, 1, 2, ..., K 1); it is thus advisable

to approximate P~lax with the closest fragmentation threshold corresponding to a certain

channel state, say So, Therefore, we set PK,>e,. to FTen.

Next, we want to show where the "free space" comes from. When a packet with length

strictly less than the PK,>e,. is transmitted in the channel state with FT less than FTen, the

packet may be fragmented, and there is no "free space" available for all fragments possibly

except the last one. However, due to the time-varying nature of the channel, when the

packet is transmitted in the channel state with FT greater than FTen, one packet does not

have enough bits to utilize the full capacity the channel provides in one transmission. We

argue that we could take advantage of the "free space" to pack more bits as the channel

allows. As a matter of fact, we have shown, in our recent studies, that the fragmentation

threshold can be up to 10K bits when the SNR is close to 20 dB and 64 QAM modulation

scheme is used with targeted FER 8% (Frame Error Rate) [93]. On the other hand, we

observe that in contention-based MAC protocols, it may take a long time for a node to seize

the channel, and the node that has seized the channel should treasure every transmission

opportunity to transmit as many bits as possible, especially when the channel condition is

good. From now on, we call state SI, the free-space-effective state when k is greater than

m, otherwise the non-free-space-effective state, even though such a state may still has the

possibility to pack more bits when the traffic dynamic is taken into account.

Now we describe how the Courtesy Piggybacking scheme makes use of the free space.

When a mobile node seizes the channel, it will first check the channel state and determine

if it is in a free-space-effective state and if it is capable of piggybacking more packets in










one transmission. If it is not in a free-space-effective state, only one packet (MSDU) with

highest priority from the queues will be served, as the current MAC protocol does. If the

channel is in a free-space-effective state, the node can transmit in one transmission as many

bits as channel allows and thus can piggyback more packets (MSDUs) from the queue(s),

which may have different priorities but the same next hop in the routing table. Since the

Courtesy Piggybacking scheme follows the cross-layer design principle so that the MAC

layer has the access to the routing information, it is possible for the MAC layer to obtain

such packets from the Tx queues.

After identifying the existence of the free space, we now discuss the piggybacking

rules that guide the MAC layer to assemble enough and proper bits from the Tx queues

(the HOW MANY-WHO problem) and piggyback them to the next hop to alleviate the

conflict we intend to address. Since the channel state determines "HOW MANY" MSDUs

the node can pack and transmit in one transmission, the fundamental issue of the rules

should specify is "who" plays the role of "train" that offers the piggybacking service to

others, and "who" can enjoy such piggybacking service. Without any scheduling mecha-

nism, the role of "train" is always taken by the MSDU locating at the head of a non-empty

queue with the highest priority currently. Thus, the piggybacking rules should primarily

address "who" has the privilege to enjoy such "free" piggybacking service. As a guideline,

the basic idea for such piggybacking rules is that under different channel states, the node

assembles multiple MSDUs that may have different priorities but share the same next hop

in the routing table, to form an MPDU whose length is channel dependent. In this way we

can achieve some extent of fairness between different prioritized services. When the chan-

nel is not in a free-space-effective state, only the highest priority service in the Tx queues

is supported, and the packets are fragmented if needed and are treated as usual. When the

channel changes to a free-space-effective state, according to the rules we define, our Cour-

tesy Piggybacking scheme can pack other services, possible with lower priorities, to share

the residual bandwidth with the high-priority traffic. One such rules is to give preference
















o"queue fvo priority _+|b-MSDU |b-MSDU |. |b-MSDU|
c. 2 Asseml d MSDU
h drMSDU CRC
3 I MPDU
queue for priority level


dequeue
controller

Figure 4-7: Illustration of the courtesy piggybacking scheme.


to high-priority services. It always, if possible, packs the high-priority services destined to

the same next hop in queue(s). Only when there are no more bits from the high-priority

traffic fitting into the free-space3 will the bits from the lower-priority queue(s) be consid-

ered for piggybacking. Other rules may not prefer the high-priority service; for example,

a high-priority service may trade-off its own performance for fairer channel utilization by

its courtesy-piggybacking the low-priority service. One such rules is to always piggyback

the MSDUs from the longest Tx queue. With such piggybacking rules, the traffic dynamics

(different packet arrival time and destinations) and channel dynamics are jointly utilized to

strike a good balance between throughput and fairness. Note that the piggybacking rules

are not necessarily defined a priori; they could be designed to adapt to both channel and

traffic uncertainty in the runtime.

Intuitively, the Courtesy Piggybacking scheme can improve the performance of the

low-priority traffic, since some low-priority packets may be packed with high-priority pack-

ets and be delivered to the next hop for free; thus it can statistically reduce the time taken to

contend for accessing the channel for the low-priority services. This benefit will be more




3 When we say "no more bits in one queue fitting into the free-space", it means either
no packet is left in the queue or packets in the queue do not share the same next hop with
the MSDU who offers the piggybacking service.










pronounced in mobile ad hoc networks using service differentiation based MAC protocols

[78][99] where the MAC protocols sacrifice quality of the low-priority service to support

high-priority service through either time spacing (differentiation of Interframe Space (IFS))

or backoff parameters [5]. On the other hand, the reduction of contention from low-priority

services can in tumn benefit the high-priority services: one node's courtesy piggybacking of

low-priority services may help its neighbors to transmit high-priority traffic, because less

low-priority traffic will reduce the contention the high-priority traffic may encounter. One

may wonder why we do not simply release the channel so that other low-priority traffic can

use the channel, that is, the so-called complete sharing scheme. The problem is that the

time, for which the residual resource is available, is too short to be given to other services

due to the overhead associated with successfully seizing the channel. Besides, some MAC

protocols such as the IEEE 802.11 family forbid others to use the channel during the time

period specified by the Network Allocation Vector (NAV). Even if the NAVs are reset, the

contention process may take too long to render the harvested resource from the rate adap-

tation useless. Thus, the courtesy piggybacking by high-priority traffic flows makes more

sense. In short, our Courtesy Piggybacking is able to achieve better channel utilization

and further improves the fairness between different prioritized traffic by the following two

means: lowering the contention of the network and decreasing the overhead required for

a transmission. And as we show in simulation later, our scheme significantly improves

the performance of low-priority traffic, while improves or at least keeps unchanged the

performance of high-priority traffic with appropriate piggybacking rules.

To illustrate the Courtesy Piggybacking scheme, we demonstrate the operation of the

scheme in Fig. 4-7. First, prioritized packets called MSDUs arrive from the network

layer as b-MSDUs (basic MSDUs, the basic unit) whose lengths agree with the FTm. We

assume the maximum packet length PKmax is strictly enforced at the upper layer; if not,

an oversized MSDU will be further broken down into several b-MSDUs and the resulting

b-MSDUs will inherit the IP header of the original MSDU. The b-MSDUs are kept in the




Full Text

PAGE 1

CR OSS-LA YER DESIGN OF RESOURCE A W ARE PR O T OCOLS FOR HETER OGENEOUS WIRELESS AD HOC NETW ORKS By W ei Liu A DISSER T A TION PRESENTED T O THE GRADU A TE SCHOOL OF THE UNIVERSITY OF FLORID A IN P AR TIAL FULFILLMENT OF THE REQ UIREMENTS FOR THE DEGREE OF DOCT OR OF PHILOSOPHY UNIVERSITY OF FLORID A 2005

PAGE 2

Cop yright 2005 by W ei Liu

PAGE 3

T o my wife, Shu, and my f amily members in China.

PAGE 4

A CKNO WLEDGMENTS First and foremost, I w ould lik e to e xpress my sincere gratitude to my advisor Prof. Y uguang F ang, who leaded me to this e xciting research area, for his in v aluable advice, encouragement and moti v ation during these years of my Ph.D. studies. I also thank him for his philosophical advice on both my academic and nonacademic life. As a mentor he has helped me become more mature, scholastically and personally I also w ant to thank my committee members, Prof. Shig ang Chen, Prof. Sartaj K. Sahni, Prof. John. M. Shea, and Prof. Dapeng W u, for their attention and advices. I w ould lik e to e xtend my thanks to all my colleagues in W ireless Netw orks Laboratory (WINET) at Uni v ersity of Florida for pro viding me such a w arm f amily-lik e en vironment. The constructi v e discussions with them either indi vidually or together deserv e special ackno wledgments. Finally on a more personal le v el, I am particularly indebted to my wife Shu, for her enduring so man y late nights, and for countless sacrices made to me. Her lo v e, support and encouragement ha v e made it possible for me to carry out this w ork. I am also v ery grateful to all my f amily members in China for their endless encouragement and continued support in one w ay or the other This w ork is dedicated to all of them. This material is based upon w ork supported in part by the U.S. Of ce of Na v al Research and National Science F oundation. An y opinions, ndings, and conclusions or recommendations e xpressed in this material are those of the authors and do not necessarily reect the vie ws of the U.S. Of ce of Na v al Research and National Science F oundation. i v

PAGE 5

T ABLE OF CONTENTS pageA CKNO WLEDGMENTS. . . . . . . . . . . . . . . . i vLIST OF T ABLES. . . . . . . . . . . . . . . . . . viiiLIST OF FIGURES. . . . . . . . . . . . . . . . . . ixABSTRA CT. . . . . . . . . . . . . . . . . . . . xiCHAPTERS 1 INTR ODUCTION. . . . . . . . . . . . . . . . 11.1 W ireless Ad Hoc Netw orks. . . . . . . . . . . . 11.2 Current Research. . . . . . . . . . . . . . . 21.2.1 Reliable Communications. . . . . . . . . . . 21.2.2 Ener gy Conserv ation. . . . . . . . . . . . 31.2.3 Quality of Service. . . . . . . . . . . . . 41.2.4 Cross-Layer Design. . . . . . . . . . . . 51.2.5 Heterogeneity in W ireless Ad Hoc Netw orks. . . . . . 61.3 Our Research. . . . . . . . . . . . . . . . 71.4 Outline. . . . . . . . . . . . . . . . . . 82 A DEVICE-ENERGY -LO AD A W ARE RELA YING FRAMEW ORK IN HETER OGENEOUS WIRELESS AD HOC NETW ORKS. . . . 112.1 Introduction. . . . . . . . . . . . . . . . 112.2 Related W ork. . . . . . . . . . . . . . . . 122.3 Ov ervie w of DELAR. . . . . . . . . . . . . . 142.3.1 Problem Statement. . . . . . . . . . . . 152.3.2 Our Solution: DELAR. . . . . . . . . . . . 162.4 Design of DELAR. . . . . . . . . . . . . . . 202.4.1 P-nodes' Neighbor -Selection Criteria. . . . . . . . 202.4.2 Routing Component of DELAR. . . . . . . . . 222.4.3 Hybrid T ransmission Scheduling. . . . . . . . . 252.4.4 Asymmetric Media Access Control Protocol (A-MA C). . . 272.5 Multiple-P ack ets T ransmission in DELAR. . . . . . . . 302.5.1 Multiple-pack ets T ransmission. . . . . . . . . 312.5.2 Hierarchical Modulation. . . . . . . . . . . 322.6 Discussion. . . . . . . . . . . . . . . . . 362.6.1 The Existence of Backw ard P aths. . . . . . . . . 36 v

PAGE 6

2.6.2 DELAR and ZRP. . . . . . . . . . . . . 392.6.3 The Choice of m and n. . . . . . . . . . . 392.6.4 Benets of the T ime Di vision Scheduling. . . . . . . 402.7 Performance Ev aluation. . . . . . . . . . . . . 402.7.1 Simulation Setup. . . . . . . . . . . . . 402.7.2 Impact of The Number of P-nodes. . . . . . . . 422.7.3 Impact of Node Mobility. . . . . . . . . . . 442.7.4 Impact of T raf c Load. . . . . . . . . . . . 472.8 Summary. . . . . . . . . . . . . . . . . 473 RESOURCE A W ARE MO VEMENT IN HETER OGENEOUS MOBILE AD HOC NETW ORKS. . . . . . . . . . . . 493.1 Introduction. . . . . . . . . . . . . . . . 493.2 Related W ork. . . . . . . . . . . . . . . . 503.3 Problem F ormulation. . . . . . . . . . . . . . 513.3.1 General Mobility Model (GMM). . . . . . . . . 523.3.2 Resource A w are Mo v ement. . . . . . . . . . 543.4 W aterhunter Mo v ement. . . . . . . . . . . . . 563.4.1 Simplied W aterhunter Mo v ement Problem. . . . . . 563.4.2 RAM-DCLC Algorithm. . . . . . . . . . . 613.4.3 A Routing Delay Dif ferentiation Mechanism. . . . . . 653.4.4 Incorporate RAM into DELAR. . . . . . . . . 663.5 Performance Ev aluation. . . . . . . . . . . . . 663.5.1 Simulation Setup. . . . . . . . . . . . . 663.5.2 Simulation Results. . . . . . . . . . . . . 683.6 Summary. . . . . . . . . . . . . . . . . 714 SUPPOR T DIFFERENTIA TED SER VICES IN MOBILE AD HOC NETW ORKS724.1 Introduction. . . . . . . . . . . . . . . . 724.2 Related W ork. . . . . . . . . . . . . . . . 744.3 Moti v ation. . . . . . . . . . . . . . . . . 764.4 P ack et-length-based Channel Model. . . . . . . . . . 794.5 Courtesy Piggybacking. . . . . . . . . . . . . 824.5.1 System Assumptions. . . . . . . . . . . . 824.5.2 The Courtesy Piggybacking Scheme. . . . . . . . 834.5.3 Discussion. . . . . . . . . . . . . . . 894.6 Performance Analysis. . . . . . . . . . . . . . 944.7 Performance Ev aluation. . . . . . . . . . . . . 974.7.1 Simulation Setup. . . . . . . . . . . . . 974.7.2 Impact of Channel Characteristics. . . . . . . . 994.7.3 Impact of T raf c Load. . . . . . . . . . . 1004.7.4 Impact of Node Mobility. . . . . . . . . . . 1014.7.5 Impact of Piggybacking Rules. . . . . . . . . 1034.8 Summary. . . . . . . . . . . . . . . . . 104 vi

PAGE 7

5 R OB UST AND ENERGY -EFFICIENT D A T A DISSEMIN A TION IN WIRELESS SENSOR NETW ORKS. . . . . . . . . . 1065.1 Introduction. . . . . . . . . . . . . . . . 1065.2 Related W ork. . . . . . . . . . . . . . . . 1085.3 Modelling Sensor Netw orks as Supply Chains. . . . . . . 1115.3.1 Introduction to Supply Chains. . . . . . . . . . 1115.3.2 Ho w Could Supply Chains Help Us. . . . . . . . 1135.4 A Hybrid Data Dissemination Frame w ork for W ireless Sensor Netw orks. . . . . . . . . . . 1155.4.1 System Model. . . . . . . . . . . . . . 1155.4.2 Manuf acture Area. . . . . . . . . . . . . 1175.4.3 T ransportation Area. . . . . . . . . . . . 1195.4.4 W arehouse Area and Service Area. . . . . . . . 1235.4.5 Discussion. . . . . . . . . . . . . . . 1265.5 Performance Ev aluation. . . . . . . . . . . . . 1285.5.1 Methodology and Metrics. . . . . . . . . . . 1285.5.2 Simulation Results. . . . . . . . . . . . . 1315.6 Summary. . . . . . . . . . . . . . . . . 1376 CONCLUSIONS AND FUTURE W ORK. . . . . . . . . . 1386.1 Conclusions. . . . . . . . . . . . . . . . 1386.2 Future W ork. . . . . . . . . . . . . . . . 139REFERENCES. . . . . . . . . . . . . . . . . . . 141BIOGRAPHICAL SKETCH. . . . . . . . . . . . . . . 151 vii

PAGE 8

LIST OF T ABLES T able page3 RAM: resource a w are mo v ement. . . . . . . . . . . . 613 RAM-DCLC: a DCLC routing algorithm for the W aterhunter Mo v ement. 623 Ener gy consumption parameters.. . . . . . . . . . . . 674 Channel model statistics. . . . . . . . . . . . . . 975 Analogue between supply chains and wireless sensor netw orks. . . . 1145 Simulation conguration. . . . . . . . . . . . . . 130 viii

PAGE 9

LIST OF FIGURES Figure page2 The structure of a Super Frame.. . . . . . . . . . . . 182 An e xample of the neighbor determination process (m=4, n=2, T=3.).. . 222 The topology in homogeneous and heterogeneous cases.. . . . . . 232 A unidirectional link between A and B.. . . . . . . . . . 272 The A-MA C operation procedure.. . . . . . . . . . . 282 The multiple-pack ets transmission enhancement to DELAR.. . . . 302 General hierarchical 2/4-PSK constellation.. . . . . . . . . 322 Implement multiple-pack ets transmission with hierarchical modulation. . 332 The e xistence of backw ard path.. . . . . . . . . . . . 382 The a v erage number of nodes e xisting in the shaded area.. . . . . 382 Simulation results with dif ferent number of P-nodes.. . . . . . . 432 Simulation results with dif ferent maximum node speed.. . . . . . 452 Simulation results with dif ferent traf c load.. . . . . . . . . 463 Multiple paths between tw o consecuti v e stops.. . . . . . . . 553 An e x emplary complete graph.. . . . . . . . . . . . 583 An e x emplary resource-a w are mo v ement.. . . . . . . . . 593 Resource-a w are mo v ement (RAM) vs. random w aypoint mo v ement. A Bnode should consecuti v ely visit S 1 S 2 S 3 S 4 S 5 and the solid lines are labelled by the sequences the y were passed through.. . . . . . 643 Simulation results.. . . . . . . . . . . . . . . . 694 The Whittier T unnel scenario.. . . . . . . . . . . . . 774 The Gilbert-Elliott channel model.. . . . . . . . . . . 804 The optimal pack et length ( P L ) vs. SNR ( r ), h=128.. . . . . . 814 P ack et-Length-Based nite-state Mark o v channel model.. . . . . 81 ix

PAGE 10

4 A fragmentation e xample.. . . . . . . . . . . . . 834 The total o v erhead with P K max and F T .. . . . . . . . . . 844 Illustration of the courtesy piggybacking scheme.. . . . . . . 874 Three piggybacking cases.. . . . . . . . . . . . . 904 An alternati v e piggybacking method.. . . . . . . . . . . 944 The queue model for piggybacking.. . . . . . . . . . . 964 A v erage w aiting time.. . . . . . . . . . . . . . . 974 Simulation results with dif ferent channel settings.. . . . . . . 1004 Simulation results with dif ferent pack et arri v al rates.. . . . . . . 1014 Simulation results with dif ferent pause time.. . . . . . . . . 1025 An e x emplary supply chain.. . . . . . . . . . . . . 1115 Supply chain strate gies.. . . . . . . . . . . . . . 1125 A system architecture for habitat monitoring. . . . . . . . . 1155 The forw arding-decision-making process of nodes in the transportation area.1205 An e x emplary pack et structure.. . . . . . . . . . . . 1205 The routing process in the w arehouse area.. . . . . . . . . 1245 The simulated sensor eld.. . . . . . . . . . . . . 1295 Ev ent deli v ery ratio vs. pack et error rate.. . . . . . . . . . 1315 Normalized ener gy consumption vs. pack et error rate.. . . . . . 1315 A v erage e v ent end-to-end delay vs. pack et error rate.. . . . . . 1325 A v erage routing o v erhead vs. pack et error rate.. . . . . . . . 1325 Ev ent deli v ery ratio vs. pack et error rate.. . . . . . . . . . 1345 Ener gy consumption and e v ent end-to-end delay of RRP with dif ferent d. 135 x

PAGE 11

Abstract of Dissertation Presented to the Graduate School of the Uni v ersity of Florida in P artial Fulllment of the Requirements for the De gree of Doctor of Philosoph y CR OSS-LA YER DESIGN OF RESOURCE A W ARE PR O T OCOLS FOR HETER OGENEOUS WIRELESS AD HOC NETW ORKS By W ei Liu August 2005 Chair: Y uguang F ang Major Department: Electrical and Computer Engineering Their features of rapid establishment and self-or g anization ha v e rendered wireless ad hoc netw orks to be identied as an indispensable component to support ubiquitous communications. Ho we v er the viable deplo yment of such netw orks f aces man y challenges resulting from some innate unf a v orable features such as the constrained resources, for e xample, nite ener gy supply and limited bandwidth. In this dissertation we in v estig ate the cross-layer design of resource-a w are protocols that conserv e ener gy and support qualityof-service (QoS) in resource-constrained wireless ad hoc netw orks. W e rst study the issue of ener gy conserv ation in heterogenous wireless ad hoc netw orks, where most nodes are po wered by batteries with small capacity while some others are more po werful, for e xample, po wered by batteries with lar ge capacity W e propose a cross-layer designed de vice-ener gy-load a w are relaying frame w ork, called DELAR to capitalize such po werful nodes to conserv e ener gy Dif ferent from the pre vious w ork that addresses the ener gy conserv ation issue at one layer DELAR is a joint design of scheduling, po wer control and routing that can conserv e ener gy thus prolong the netw ork lifetime. W e further propose a resource-a w are mo v ement mechanism to utilize such po werful nodes xi

PAGE 12

for data relaying. Such a resource-a w are mo v ement mechanism pro vides a ne w w ay to conserv e ener gy in wireless ad hoc netw orks. W e then study the issue of ho w to support QoS in wireless ad hoc netw orks. W e propose a scheme called Courtesy Pig gybac king to utilize system dynamics including timev arying channel condition and stochastic traf c characteristics, to support dif ferentiated services in such netw orks. The basic idea of Courtesy Piggybacking is to let high priority traf c help the lo w priority traf c by sharing the unused residual bandwidth. As a result, the scheme impro v es the o v erall performance for dif ferent prioritized services. Finally we propose an ener gy ef cient, reliable and scalable data dissemination scheme for wireless sensor netw orks. Based on supply chain management methodology for each sensing task, the sensor eld is conceptually partitioned into se v eral functional areas, and then dif ferent routing schemes are applied in dif ferent areas. This scheme addresses ener gyef cienc y reliability and scalability at the same time. xii

PAGE 13

CHAPTER 1 INTR ODUCTION 1.1 W ireless Ad Hoc Netw orks A wireless ad hoc netw ork is a collection of autonomous communication de vices, sometimes called nodes, which can communicate with each other by forming a multi-hop radio netw ork and maintaining connecti vity in a decentralized manner There are tw o major types of wireless ad hoc netw orks that are of particular interest to academia, industry and go v ernment: mobile ad hoc netw orks (MANETs) and wireless sensor netw orks (WSNs). Both types of wireless ad hoc netw orks share se v eral typical features in common. W ireless ad hoc netw orks can be deplo yed quickly and on-demand. The y are self-or g anized and independent of infrastructure. Moreo v er each node in a wireless ad hoc netw ork functions as both a host and a router helping relay pack ets for other nodes. These f a v orable features mak e wireless ad hoc netw orks v ery attracti v e in military and ci vil applications for which x ed infrastructures are una v ailable or unreliable, yet f ast netw ork establishment and constant reconguration are required. As an instance of mobile ad hoc netw orks, after an earthquak e, a rescue team sets up a temporary wireless netw orks and shares information to coordinate the rescuing w ork. As an e xample of wireless sensor netw orks, a lar ge number of small sensor nodes are deplo yed in a eld to monitor the habitat beha vior of animals. Ho we v er there are some unf a v orable features impeding wireless ad hoc netw orks from widespread deplo yment. Due to the time-v arying and error -prone wireless channel, nodes ha v e to contend with the adv erse ef fects of radio communications, such as noise, f ading and interference. Since nodes may depart, join, or mo v e in the eld, or e v en perish due to ener gy depletion, the netw ork topology is in general dynamic and may change rapidly and unpredictable o v er time. In such hostile, dynamic en vironment, reliable communications are of great importance. Additionally the wireless channel has limited bandwidth. 1

PAGE 14

2 Therefore, it is a challenging task to support heterogeneous traf c with v arious qualityof-service requirements. Moreo v er most nodes, especially those small sensors deplo yed in wireless sensor netw orks, are battery-po wered thus with limited lifetime. These nodes w ould become useless once depleting the batteries. Some adv erse consequences of such node diminution include de gradation of netw ork performance and unf a v orable netw ork partition. Thus, ener gy conserv ation, that is, ho w to e xpend the ener gy resources more frug ally and more e v enly so as to prolong the netw ork lifetime, becomes a crucial issue for MANETs. The broadcast nature of wireless communications also mak es it easy for enemies to snoop the ongoing transmissions, thus communications in wireless ad hoc netw orks ha v e security hazard. Depending on the applications running in such netw orks, the netw orks range from small-scale netw orks with a small number of nodes to lar ge-scale netw orks with thousands e v en millions of nodes, for e xample, lar ge-scale wireless sensor netw orks. The abo v e unf a v orable f actors, especially v ariable wireless links, topology changes, limited bandwidth, and constrained ener gy supply mak e the protocol design a tremendous challenge. P articulary when heterogenous applications are nding their niches in wireless ad hoc netw orks, specially designed protocols that can o v ercome the abo v e adv erse f actors are desired. Such protocols, scalable themselv es, should pro vide r eliable QoS-supporting ener gy-ef cient and secur e communications for those e ye-catching applications. In the follo wing we gi v e a nutshell of the current research for the issues of reliable communications, ener gy conserv ation and QoS pro visioning. 1.2 Current Research 1.2.1 Reliable Communications The primary goal of protocol design for wireless ad hoc netw orks is to support reliable communications which is the basic requirement for an y applications, and all the other solutions for the issues of ener gy conserv ation and quality of service should be based on reliable communications. In such hostile and dynamic settings, pack et losses may be caused

PAGE 15

3 for v arious reasons including transmission errors, collisions, or route changes due to mobility Thus, there are man y w orks addressing reliable communications at dif ferent layers. F or e xample, at the ph ysical layer modulation and error coding schemes are proposed to combat the error -prone feature of wireless communications [1]. In order to reduce the interference and transmission collisions, at the MA C layer there are man y MA C protocols that are de vised to to ef ciently control the channel access [2]. These protocols can be classied into three cate gories: contention-based (for e xample, MA CA W [3], F AMA [4] and IEEE 802.11 [5]), contention-free (for e xample, TDMA, FDMA and CDMA) and h ybrid. The most f amous one is IEEE 802.11, and it uses CSMA/CA and four w ay handshak e (R TS/CTS/D A T A/A CK) to coordinate the channel access and pro vide reliable communications [5]. At the netw ork layer man y routing protocols ha v e been proposed to deal with the dynamic topology and support reliable communications [6]. The y can be proacti v e in that the y determine routes independent of traf c patterns, for e xample, DSD V [7] and OLSR [8]; the y can be reacti v e in that the y maintain routes on demand, such as DSR [9] and A OD V [10]; the y e v en can be h ybrid of the abo v e tw o, for e xample, ZRP [11] and LANMAR [12]. Though these schemes function at dif ferent layers, the y all w ork for the same goal: supporting reliable communications in wireless ad hoc netw orks. 1.2.2 Ener gy Conserv ation Ho w to lengthen the lifetime of communication de vices is a crucial issue for the wireless ad hoc netw orks and man y po wer conserv ation techniques are applied when hardw are and softw are including the protocol stack are specically designed. Man y ef forts ha v e been made at the ph ysical layer due to the f actor that the system hardw are is the primary consumption object in a mobile de vice. T w o dif ferent perspecti v es are commonly adopted to approach the problem. Though no breakthrough has been e xperienced in the past 30 years, the direct w ay is to increase the battery capacity while k eeping the weight of the battery tolerati v e. Another w ay is, hard yet attainable, to apply lo w-po wer techniques to decrease the ener gy dissipation when wireless terminals are designed. Considerable ener gy sa vings

PAGE 16

4 are resulted from the hardw are design; ho we v er it is pertinent to e xplore other v enues to further impro v e the ener gy ef cienc y that is, to design the protocol stack with ener gy ef cienc y in mind [13]. F or e xample, at the data-link layer transmission po wer control can be used to reduce the interference. Po wer -sa ving mode also can be used to conserv e ener gy by putting the netw ork interf ace into the sleep mode when no communication is needed. At the netw ork layer po wer a w are routing is used to nd ener gy-ef cient paths and route pack ets o v er the ener gy-ef cient paths to sa v e ener gy At the application layer ener gyef cient OS/middle w are and applications (for e xample, po wer -a w are video processing) are proposed to sa v e ener gy Because the k e y to ener gy conserv ation in wireless ad hoc netw orks lies within the higher le v els of the wireless protocol stack [13], it w ould be benecial if these multiple layers can jointly function to conserv e ener gy 1.2.3 Quality of Service It is hard to agree on a common denition of QoS, b ut a QoS enabled netw ork shall ensure that its applications and/or their users ha v e their QoS parameters fullled, while at the same time ensuring an ef cient resource usage, for e xample the bandwidth, and also ensure that the most important traf c still has its QoS parameters fullled during netw ork o v erload. The most important QoS parameters may include: throughput, a v ailability delay jitter and pack et loss. The netw ork can serv e the applications with hard-QoS, soft-QoS or adapti v e QoS [2,14]. Se v eral prominent service models or o v erall architectural frame w orks, within which certain types of services can be pro vided in the netw ork, are suggested for ad hoc netw orks, such as e xible quality of service model for MANETs (FQMM) [15] and stateless model for wireless ad-hoc netw orks (SW AN) [16]. Since the resource is limited and dif ferent traf c has v ery dif ferent quality of service requirements, usually Dif ferentiated Services (Dif fServ) architecture [17] is used in which traf c is classied into dif ferent classes or priority le v els, and the wireless netw orks serv e dif ferent classes of traf c with dif ferent processing rules or policies. There are se v eral mechanisms proposed for supporting QoS [14], such as QoS routing, scheduling, admission control,

PAGE 17

5 resource reserv ation, and signaling techniques. These mechanisms can coe xist in a netw ork. F or e xample, QoS routing is used to nd a path satisfying certain QoS requirements, then signalling techniques and resource reserv ation are used to reserv e resources along the found path. W e belie v e that in order to ef ciently utilize the resources and support QoS, v arious mechanisms implemented at dif ferent layers should w ork together 1.2.4 Cross-Layer Design As a matter of f act, the issues of ener gy conserv ation and QoS pro visioning are direct results of constrained resources: ener gy supply and bandwidth. If nodes and netw orks are endo wed with plenty of bandwidth and ener gy supply for e xample, unbounded frequenc y band and hea vy duty po wer supply with light wight, these tw o issues may ne v er come to the fore. Since the tough reality of constrained resources is there and the chance to en vision breakthrough for these tw o resources is rather slim, the natural response to these tw o issues is to check: Do all the current protocols proposed for wireless ad hoc netw orks mak e full use of such precious resources? The answer to the question is ne g ati v e. There do ha v e lots of resources thar are w asted without an y yield, o wing to lo w ef cienc y or their o v erlook of these protocols. F or e xample, in netw orks with contention-based medium access control mechanisms, lots of ener gy and bandwidth are w asted due to collisions. One of the reasons for resource underutilization is that v ast majority of the protocols are based on traditional layered-design principle, which lacks the interaction and information sharing among multiple layers. The lack of interaction causes lo w ef cienc y of those protocols. On one hand, man y resources at one layer are w asted due to undesirable decisions made at another layer F or e xample, poor routing decisions cause congestion and collisions at the ph ysical layer thus w aste precious resources, for e xample, bandwidth and ener gy On the other hand, man y a v ailable resources are ne glected by those protocols. F or e xample, time-v arying channel condition leads to time-v arying channel capacity ho we v er due to the lack of interaction between the MA C layer and the ph ysical layer man y MA C protocols cannot recognize such a resource b ut ne glect it. Hence, in recent years a ne w

PAGE 18

6 cr oss-layer design philosoph y emer ges [18] [19]. In contrast to the layered-design philosoph y cross-layer design philosoph y stresses on the information sharing and interaction among multiple layers. Cross-layer design can be performed in tw o w ays [20]. In the rst w ay parameters or information in other protocol layers are considered to impro v e the performance of a protocol layer The rst w ay of practice does not totally abandon the transparenc y between protocol layers. In contrast, the second w ay blurs this transparenc y by mer ging se v eral layers into one component to achie v e much better performance. The cross-layer design philosoph y attempts to mak e use of the inter -layer coupling to use the resources more ef ciently W ith this philosoph y the decision for resource utilization can be made more advisably based on the information from multiple layers. Though the cross-layer design may mak e the design issue more complicated than the layered design, the appealing benets mak e cross-layer design a strong candidate as the design principle in resource-constrained wireless ad hoc netw orks. 1.2.5 Heterogeneity in W ireless Ad Hoc Netw orks Most of the pre vious research assumes that netw orks are homogeneous or nodes are in fully symmetrical en vironment. F or e xample, all nodes ha v e identical capabilities and responsibilities. In reality ho we v er the heterogeneity of mobile de vices seems to be inher ent and has been commonly observ ed in MANETs [21]. Strictly speaking, heterogeneity in a netw ork usually means the de vice heter o g eneity re grading the ph ysical hardw are or softw are nodes use. Nodes in the same netw ork may dif fer in their communication, processing capabilities, which can be reected by CPU speed, memory size, battery life, transmission range and radio, netw ork interf ace, security le v el, operating system or protocol stack, and so on. Nodes in the same netw ork may dif fer in their responsibilities as well. The y may undertak e dif ferent roles, for e xample, routers, serv ers, cluster heads, and authentication centers, thus operate with dif ferent policies. Owing to such dif ferences in capabilities and responsibilities, along with the v arious applications running on each indi vidual node, nodes may also ha v e dif ferent traf c characteristics and mobility characteristics. F or e xample,

PAGE 19

7 nodes may ha v e dif ferent bit rate, timeliness constraints, reliability requirements, speed, direction, and predictable/unpredictable mo v ement patterns for each indi vidual application. Broadly speaking, heterogeneity in a netw ork also means that nodes may e xperience heterogeneous en vironment characterized by time-v arying channel and netw ork conditions. Such en vir onment heter o g eneity may ha v e dif ferent properties in dif ferent settings. F or e xample, dif ferent nodes at dif ferent locations and at dif ferent time may percei v e dif ferent channel conditions. The heterogeneity either de vice heterogeneity or en vironment heterogeneity coupling with the aforementioned unf a v orable f actors, mak es the protocol design much more complicated. Ho we v er such heterogeneity also introduces opportunities to design more efcient protocols. In f act, some types of heterogeneity are resource-related when the y themselv es are resources or can be mapped into certain type(s) of resources. Moreo v er de vice heterogeneity and en vironment heterogeneity can jointly af fect the usage of constrained resources. F or e xample, nodes with less a v ailable battery reserv e w ould refrain from transmissions in poor channel condition. Thus, to recognize the inherent heterogeneity and further to ef ciently utilize the resource-related heterogeneity to o v ercome the impact stemming from those unf a v orable features, for e xample, limited ener gy supply and constrained bandwidth, is a rather challenging task. In this dissertation, we mak e our efforts along this line to address the issues of ener gy conserv ation and QoS pro visioning in resource-constrained wireless ad hoc netw orks. 1.3 Our Research As mentioned abo v e, there is a big v olume of w ork studying ho w to support reliable communications, from the ph ysical layer to the data link layer and to the routing layer Ne v ertheless, the research on ener gy conserv ation and QoS pro visioning is still thin on the ground. Either of the tw o issues is quite interesting, ho we v er on top of the pre vious research on reliability it is pretty hard, if not impossible, to design a one-size-t-all protocol that can simultaneously resolv e these tw o issues. Thus, in this dissertation, based on

PAGE 20

8 the pre vious w ork for reliable communications, we attempt to study these tw o issues and resolv e them separately b ut bear in mind the possibility of future inte gration of these tw o issues. The cross-layer design philosoph y serv es as our design principle for its potential in better utilizing the a v ailable resources and pro viding desired ener gy ef cienc y and QoS pro visioning to applications. More important, the cross-layer design philosoph y will help us recognize the inherent heterogeneity in wireless ad hoc netw orks, and further help us utilize it to address the issues we are interested in. More specically we utilize the heterogeneity in terms of po wer supply or battery life to address the issues of ener gy conserv ation. Such heterogeneity is used to design more ef cient routing, po wer control and transmission scheduling. W e also utilize the time-v arying channel condition and stochastic traf c char acteristics to address issues associated with QoS pro visioning in wireless ad hoc netw orks. F or wireless sensor netw orks, we particularly recognize the dif ferent roles undertak en by each indi vidual sensor node, and utilize such heterogeneity to design more ef cient communication protocols. 1.4 Outline This dissertation is or g anized as follo ws. Chapter2concerns the issue of ener gy conserv ation in heterogeneous wireless ad hoc netw orks in terms of po wer supply In such netw orks, most nodes called B-nodes are battery-po wered, and some nodes called P-nodes ha v e relati v ely unlimited ener gy supply such as solar cell or dynamos. W e propose a de vice-ener gy-load a w are relaying frame w ork, called DELAR to capitalize such P-nodes to conserv e ener gy The basic idea of DELAR is to di vide time into Super F r ames each of which is composed by se v eral periods, and to enable P-nodes to transmit with dif ferent po wer in dif ferent periods. In this frame w ork, de vice heterogeneity nodal residual ener gy and local load status are mapped into a routing cost metric and incorporated into routing protocols to nd ener gy-ef cient paths. W e also

PAGE 21

9 propose a h ybrid medium access control mechanism to schedule the transmission acti vities of P-node and B-nodes. Moreo v er we introduce the concept of mini-routing into the MA C layer and propose an Asymmetric MA C (A-MA C) to support reliable communications o v er unidirectional links resulted from the asymmetric po wer capabilities between P-node and B-nodes. Further we propose a multiple-pack ets transmission scheme as an enhancement to impro v e the o v erall performance of DELAR. Chapter3discusses a resource-a w are mo v ement scheme that can be incorporated into DELAR. In this scheme, a node is enabled to determine its mo v ement characteristic (speed, direction, and so on) based on the netw ork en vironment and its o wn residual resources. Specically such kinds of mo v ement can be mapped into a path-constrained path-optimization problem. A heuristic algorithm is proposed to solv e this NP-complete problem. W ith this resource-a w are mo v ement scheme in place, B-nodes ha v e more oppor tunities to utilize P-nodes and thus conserv e ener gy and prolong the netw ork lifetime. Chapter4focuses on the issue of QoS pro visioning. W e consider heterogeneous mobile as hoc netw orks on ho w to utilize system dynamics, a combination of en vironment and de vice heterogeneity including time-v arying channel condition and stochastic traf c char acteristics. A scheme called Courtesy Pig gybac king is proposed to support dif ferentiated services in mobile ad hoc netw orks. The basic idea is to let the high priority traf c help the lo w priority traf c by sharing the unused residual bandwidth with courtesy Courtesy Piggybacking is able to utilize such system dynamics to impro v e the o v erall system per formance for dif ferent prioritized services: impro v ed pack et deli v ery ratio and shortened end-to-end delay More important, the piggybacking scheme can be readily incorporated into the DELAR frame w ork. In Charter5, we particularly address the issues of ener gy ef cienc y and reliability in wireless sensor netw orks, a special type of wireless ad hoc netw orks. W e introduce the concept of supply chain into wireless sensor netw orks and model wireless sensor netw orks as supply chains. Based on supply chain management strate gies, we propose an ener gy

PAGE 22

10 ef cient, reliable and scalable data dissemination scheme. Basically for each sensing task, the sensor eld is conceptually partitioned into se v eral functional areas, and dif ferent routing schemes are applied in dif ferent areas. Such partition strate gy h ybrid strate gy and cooperation strate gy help us address the issues of ener gy conserv ation and reliable communications at the same time. Chapter6concludes this dissertation and outlines the future research.

PAGE 23

CHAPTER 2 A DEVICE-ENERGY -LO AD A W ARE RELA YING FRAMEW ORK IN HETER OGENEOUS WIRELESS AD HOC NETW ORKS 2.1 Introduction In this chapter we focus on heterogeneous wireless ad hoc netw orks, where most nodes, denoted as B-nodes are equipped with limited po wer sources lik e batteries, while some other nodes, denoted as P-nodes ha v e relati v ely unlimited po wer supplies, for e xample, po wer sca v enging units such as solar cells, or dynamos when the y are installed in mobile v ehicles, and so on. Our goal is to de v elop more ener gy conscious protocols by taking adv antage of the heterogeneity of mobile de vices, that is, being generous in using the P-nodes while conserv ati v e in using the B-nodes. More specically the contrib utions are mainly fourfold. First, follo wing the cross-layer protocol design philosoph y we propose a De vice-Ener gy-Load A w are Relaying frame w ork, named DELAR, to achie v e ener gy conserv ation by utilizing the inherent heterogeneity of nodal po wer capabilities. Second, we design a h ybrid transmission scheduling scheme, which is a combination of reserv ationbased and contention-based medium access control schemes, to coordinate the transmission acti vities among P-nodes and B-nodes. Such scheduling can mak e full use of po werful nodes while reducing the interference and collisions. Third, we de v elop the mini-routing and Asymmetric MA C ( A-MA C ) protocols to support the MA C layer ackno wledgements o v er unidirectional links resulting from the asymmetric transmission po wer le v els at Pnodes and B-nodes. T o the best of our kno wledge, this is the rst ef fort to address this issue at the MA C layer Last, we present a multiple-pack ets transmission technique to fur ther impro v e the delay performance. Detailed simulation studies are carried out to justify the ef fecti v eness and ef cienc y of the proposed frame w ork. 11

PAGE 24

12 The proposed DELAR can serv e as a general frame w ork where v arious ener gy conserv ation techniques such as po wer sa ving modes, transmission po wer control and po wer a w are routing can be inte grated to jointly achie v e better ener gy conserv ation. In addition, it also of fers a platform to study other challenging issues, for e xample, quality of service (QoS) pro visioning and security support. F or instance, P-nodes can act as distrib uted admission controllers to coordinate the access to limited netw ork resources such as a v ailable bandwidth. As another e xample, in security-sensiti v e MANET applications, P-nodes can help B-nodes perform resource-hungry public-k e y operations. As f ar as we kno w no similar frame w ork has appeared else where in the literature. In the rest of this chapter we start with the re vie w of related w ork. W e then introduce the system model and the o v erall frame w ork of DELAR in Section2.3. In Section2.4, we elaborate the netw ork layer components of DELAR and a h ybrid transmission scheduling scheme, and present a no v el Asymmetric MA C protocol called A-MA C follo wed by a multiple-pack et transmission scheme to further impro v e the performance of DELAR. In Section2.7, we e v aluate the performance of DELAR through simulations. Finally we summarize this chapter in Section2.8. 2.2 Related W ork Recent years ha v e seen a gro wing body of research concerned with ener gy conserv ation in wireless ad hoc netw orks, among which man y ef forts ha v e been made at the ph ysical layer to impro v e the hardw are design of mobile de vices [13]. Though important, it is still pertinent to e xplore other v enues to further ameliorate the ener gy ef cienc y of mobile devices. F or the lack of space, here we only re vie w w orks performed at the MA C or netw ork layer of ad hoc netw orks, which are closely related to our study in this chapter Based on the mechanisms used, the solutions seen in the literature can be roughly classied into three cate gories: Po wer -Sa ving Modes (PSM), T ransmission Po wer Control (TPC) [22], and Po wer -A w are Routing (P AR).

PAGE 25

13 PSM is usually implemented at the MA C layer and the basic idea is to put the netw ork interf ace into the sleep mode when no communication is needed. One fundamental issue in PSM is when to enter the sleep mode and for ho w long to stay in this mode. Some w ork along this line includes P AMAS [23] and S-MA C [24]. In addition, Tseng et al. proposed three asynchronous protocols, namely Dominating-A w ak e-Interv al, Quorum-based, and Periodical-Fully-A w ak e-Interv al protocols [25]. These proposals stri v e to ef ciently and intelligently control nodes' sleep and w ak e schedules and at the same time deal with such f actors as clock synchronization, neighbor disco v ery and netw ork partitions which are inherent in multi-hop ad hoc netw orks [25]. TPC adapts the transmission po wer to the propag ation and interference characteristics e xperienced by the link [26,27,28,29,30,31,32,33,34,35]. TPC sometimes is called topology control when it attempts to control the transmission po wer or e v en turn of f the radio so that a desirable topology or connecti vity can be maintained for sa ving ener gy Man y proposals in this cate gory are concerned with maintaining a dominant set of nodes or forming some virtual backbone with certain clustering mechanisms. In addition, PCM [36] w as proposed to use dif ferent transmission po wer le v els for R TS/CTS and D A T A/A CK frames on a per -pack et basis. In contrast to TPC protocols aiming at making each link as ener gy-ef cient as possible, a P AR protocol determines which of these links to be used for end-to-end paths so that additional ener gy sa vings can be obtained by routing pack ets o v er ener gy-ef cient paths. Singh et al. proposed v e po wer -a w are routing metrics that could be incorporated into routing protocols to reduce the per -pack et ener gy consumption, to prolong the per -node lifetime, or to prolong the o v erall netw ork lifetime [37]. Besides, Chang and T assiulas [38] proposed a routing cost metric that is a combination of the transmission po wer le v el and the residual ener gy T oh [39] studied both the minimum ener gy and the netw ork lifetime issues and proposed a conditional min-max battery capacity routing protocol which tries to strik e a balance between these tw o competing objecti v es. While the link quality has been

PAGE 26

14 suggested as a routing metric to reduce queuing delays and data loss rates, Banerjee and Misra proposed a routing cost formulation that considers not only the residual po wer b ut also the link characteristics, for e xample, the cost for potential retransmissions, to capture the ef fects of both the link distance and the link error rate [40]. As an addition, P AR O [41] is another notable approach designed for scenarios where nodes can dynamically adjust their transmission po wer In P AR O, a candidate intermediate node monitors an ongoing direct communication between tw o nodes and inserts itself in the forw arding path if its action can lead to some ener gy sa vings. Most pre vious proposals do not tak e into account de vice heterogeneity inherent in MANETs to achie v e better ener gy conserv ation. Ho w to tak e full adv antage of P-nodes to prolong the netw ork lifetime as much as possible has been pre viously addressed in [21,42,43]. Ho we v er all of them focus on the netw ork layer and man y challenging issues relating to the MA C layer are either left untouched or o v erlook ed. F or e xample, none of them consider ho w to support the MA C-layer ackno wledgements o v er unidirectional links caused by dif ferent transmission po wer at P-nodes and B-nodes (cf. Section2.4.4). By contrast, our DELAR addresses the ener gy conserv ation from multiple f acets, including routing, transmission scheduling and po wer control. In brief, although all the aforementioned proposals can achie v e certain le v el of ener gy conserv ation, ho w to design a comprehensi v e, practical frame w ork that not only inte grates PSM, TPC, P AR, and transmission scheduling, b ut also mak es full use of inherent de vice heterogeneity remains an open, challenging problem. Our DELAR frame w ork is proposed to address this crucial issue. 2.3 Ov ervie w of DELAR In this section, we rst mak e a brief introduction to the problems we intend to address. W e then pro vide a high-le v el o v ervie w of our solution, DELAR which stands for De viceEner gy-Load A w are Relaying frame w ork for heterogeneous ad hoc netw orks.

PAGE 27

15 2.3.1 Problem Statement In this chapter we focus on heterogeneous ad hoc netw orks comprised of mobile nodes with dif ferent ener gy supplies, though heterogeneity may ha v e other meanings in dif ferent settings. W e assume that besides a majority of battery-po wered B-nodes, there e xist some po werful P-nodes ha ving relati v ely unlimited ener gy supplies (cf. Section2.1). Our objecti v e is to de v elop ener gy conserv ation protocols by utilizing such heterogeneity in ener gy resources. Intuiti v ely speaking, since P-nodes are of relati v ely innite ener gy reserv oir as opposed to the B-nodes with usually irreplaceable batteries, data communications should try to utilize these P-nodes as much as possible in order to prolong the netw ork lifetime. Therefore, on the one hand, a pack et should be forw arded to a P-node whene v er an ener gy sa ving can be e xpected. On the other hand, communications in the netw orks should a v oid using B-nodes if possible. F or these purposes, it is desirable to allo w P-nodes to ha v e higher transmission po wer so as to co v er a lar ger transmission area, which can statistically reduce the number of B-nodes in v olv ed in pack et forw arding. Ho we v er this straightforw ard proposal may pose lots of challenges to the protocol design Ho w can a B-node be a w are of the e xistence of P-nodes in its vicinity? If there e xist multiple paths through P-nodes to the destination, which path should be chosen? Should the transmission range of a P-node be as lar ge as possible, or be k ept within certain optimal ranges? Ho w can the protocol support reliable communications o v er the unidirectional links caused by asymmetric transmission po wer at P-nodes and B-nodes along with error -prone and time-v arying wireless channels? In addition, higher transmission po wer often implies more reachable neighbors, decreased spatial reuse, and increased local contention for the shared wireless medium, then ho w can the protocol schedule the transmission acti vities so that a good balance can be struck between ener gy sa vings and other system performance f actors such as pack et deli v ery ratio and end-to-end delay? These are all non-tri vial questions and need to be answered before we can indeed mak e full use of the aforementioned heterogeneity in ad hoc netw orks.

PAGE 28

16 After a careful in v estig ation on these interw o v en issues, we belie v e that the y are closely related to routing, transmission scheduling, and po wer control. F or e xample, with the adjustment of the transmission po wer of P-nodes and B-nodes, the topology and neighborhood, and thus the routing information, w ould change accordingly So w ould the schedule of transmission acti vities. Moreo v er there e xists a strong interaction between routing and MA C layers. Apparently our original design objecti v e can be boiled do wn to designing a joint routing, scheduling, and po wer control scheme, which should be addressed across the whole protocol stack, especially at the routing and MA C layers [44][45]. T o achie v e this, a cross-layer designed frame w ork is demanded. In this frame w ork, po wer control should be implemented to optimize the transmission po wer of each node (both P-nodes and B-nodes) to achie v e optimal ener gy utilization and maintain a reasonable netw ork topology; routing should be designed to inform all the nodes of the e xistence of P-nodes and nd the optimal ener gy-ef cient routes; and transmission scheduling should be able to adjust the transmission acti vities so that the ener gy spent on channel contentions and collisions can be minimized. In addition, an appropriate scheduling scheme should be capable of striking a good balance between ener gy ef cienc y and other system performance f actors such as end-to-end delay and pack et deli v ery ratio. 2.3.2 Our Solution: DELAR W e consider a mobile ad hoc netw ork consisting of N p P-nodes and N b B-nodes, where N p and N b are system design parameters. W e assume a single wireless broadcast channel shared by all the nodes, though our DELAR can be easily e xtended to multichannel cases. W e also adopt a simple po wer control scheme as follo ws. Each B-node transmits omni-directionally and can maintain a circular transmission range B T R (basic transmission range) before using up its battery which can be properly set to achie v e a good tradeof f between ener gy ef cienc y and netw ork connecti vity [46]. In addition, we postulate that P-nodes are able to adjust their transmission po wer so as to co v er lar ger areas than B-nodes if needed. Moreo v er all the P-nodes are assumed to ha v e identical maximum

PAGE 29

17 transmission range of P T R max = M B T R where M is a positi v e inte ger greater than 1. As re v ealed in [46], using common transmission po wer between the same type of nodes can ensure bidirectional links and thus the correct operations of e xisting routing and MA C protocols. W ith this simple yet ef cient po wer control scheme, a unidirectional link only e xists between a P-node and a B-node when the y use dif ferent transmission po wer instead of between an y tw o B-nodes or P-nodes1. According to [35], such simple po wer control is belie v ed to be more practical than other e xpensi v e transmission po wer control schemes, either making unrealistic assumptions or ha ving e xtra hardw are requirements. As mentioned before, DELAR arises from the follo wing intuition: the P-nodes should be utilized as much as possible. In other w ords, we should attempt to reduce the use of Bnodes if we cannot a v oid using them at all. Thus it is adv antageous to enable a P-node to directly communicate with other P-nodes nearby or in distance by using higher transmission po wer so that the number of B-nodes in v olv ed in the data forw arding can be much reduced. Ho we v er higher transmission po wer or lar ger transmission co v erage usually implies more neighbors and increased local contention for the shared wireless channel. Therefore, instead of granting these P-nodes unlimited pri vile ges of reaching an y other node at an y time at will, it mak es more sense to constrain P-nodes' transmission po wer under certain bound and to limit their transmission acti vities within some pre-planned periods in order to reduce the collisions with other ongoing transmissions and thus maintain good channel utilization. In order to better schedule the transmissions of P-nodes and B-nodes, we adopt a time-di vision multiple xing method. W e di vide time into equal length time slots called Superfr ames In each of the superframes, some time periods are e xclusi v ely designated to P-nodes, while the rest are shared by all P-nodes and B-nodes in the netw ork. More 1 In this chapter we only consider asymmetric transmission po wer as the primary cause for unidirectional links and omit others such as v arious collision/noise/interference le v els at dif ferent nodes.

PAGE 30

18 P n o d e 0 P n o d e 1 . P n o d e k P t o B p e r i o d 0 T i m e F r a m e D u r a t i o n P t o B p e r i o d 1 P t o B p e r i o d k P t o P p e r i o d B t o B p e r i o d m i n i s l o t Figure 2: The structure of a Super Frame. specically during one c ycle of the Superfr ame (see Fig.2), there is a P-to-P period with length t pp in which only P-nodes are allo wed to communicate with each other by using transmission range T R pp = m B T R (1 < m M ) while all B-nodes just k eep silent, as if the netw ork were merely formed by these mobile core P-nodes. Additionally in one Superfr ame each P-node has its o wn e xclusi v e period called P-to-B period with equal length t pb in which it can boost its transmission po wer to co v er a range of T R pb = n B T R (1 < n M )2. The rest of one Superfr ame is called B-to-B period with length t bb in which all the nodes in the netw ork can contend for the channel and initiate transmissions to w ards nodes in their T R bb = B T R Ob viously all the P-nodes should act as common B-nodes in the B-to-B period by adjusting their transmission range back to T R bb Notice that during one P-to-B period, since the P-node o wning this period and the B-nodes it intends to communicate with ha v e dif ferent transmission po wer unidirectional links between them may be formed. Therefore, in contrast to the P-to-P and B-to-B periods where some common contention-based MA C protocols such as the IEEE 802.11 can be used, the P-to-B period(s) demands an enhanced MA C protocol to support reliable communications o v er unidirectional links. Our Asymmetric MA C protocol A-MA C is e xactly de v eloped for this purpose. Such rendezv ous of reserv ation-based and contention-based MA C schemes is able to schedule the pack et transmission more ef ciently which we will see shortly 2 T o pro vide reliable communications during P-to-B periods, usually n is less than m .

PAGE 31

19 In DELAR, the heterogeneity of mobile nodes are also incorporated into the construction of routing tables. Routes are disco v ered based on routing metrics which tak e consideration of the residual ener gy and the load status of mobile nodes. Once generating a data pack et, a node looks up its routing table and sends the data pack et to the ne xt hop as it does in common ad hoc routing protocols. If residing in the forw arding path and ha ving recei v ed a forw arding request, a node will forw ard the data in an appropriate time period to the ne xt hop according to its o wn routing table. More specically for a B-node, when the ne xt hop is in its T R bb range, it can only forw ard the data pack et during the B-to-B period. While for a P-node, if the ne xt hop is another P-node located in this P-node' s T R pp the P-node can forw ard the pack et to the ne xt hop in the P-to-P period; if the ne xt hop is a B-node located inside its T R pb the P-node can forw ard the pack et to the ne xt hop in its e xclusi v e P-to-B period. In summary with such time-di vision scheduling and a de vice-ener gy-load a w are routing metric, we intend to utilize P-nodes as much as possible in an ef cient and cautious w ay while reducing collisions and interference to an acceptable le v el, so that we can achie v e the e xpected ener gy conserv ation without harming other system performance f actors. Se v eral research challenges remain in supporting the seemingly simple operations of DELAR as described abo v e. Gi v en a B-node (P-node) X located in a P-node P 0 s transmission range T R pb ( T R pp ), for instance, what criteria should P adopt to determine if X is a neighbor (in one hop range) or not, that is, forw arding a pack et to this node X in a one-hop manner or a multiple-hop manner? What kind of routing metric should be adopted to reect the heterogeneity in de vice types, nodal residual ener gy and local load status when setting up routing paths? Ho w can we di vide time into Superfr ames and ho w can one P-node re gister a P-to-B period without conicting with others' P-to-B periods? Ho w does node X send MA C layer ackno wledgements back to P in the presence of unidirectional links resulting from the asymmetry in transmission po wer? The remainder of this chapter will address these questions one by one in more detail.

PAGE 32

20 2.4 Design of DELAR In this section, we will rst discuss the neighbor -selection criteria of P-nodes. W e then illustrate the routing component of DELAR. Ne xt, we introduce in more detail the h ybrid transmission scheduling of DELAR. Then, we present the Asymmetric Media Access Control Protocol (A-MA C) and the multiple-pack et transmission scheme. Last we gi v e some further discussions. 2.4.1 P-nodes' Neighbor -Selection Criteria In the literature, tw o nodes are usually considered as neighboring nodes of each other when the y are one hop a w ay and the y can directly communicate with each other Ho we v er in heterogeneous netw orks, we ha v e to change the criteria to cope with the e xistence of P-nodes whose transmission ranges are much lar ger than those of B-nodes. In this case, an y node in a P-node' s T R pb could be a neighbor candidate of the P-node3. Ne v ertheless, in order to support the MA C layer ackno wledgements, not all the candidates can be nally chosen as neighbor s or be ne xt hops in the routing table. Before presenting the rules that guide P-nodes to mak e selection decisions, we rst introduce the notions of F orwar d P ath and Bac kwar d P ath F or an y node pair s and t a F orwar d P ath indicates the path deri v ed from normal routing tables. F or e xample, the F orwar d P ath ( s t ) can be represented as s N 1 ::: N k t where f N i g ( 1 i k ) denote the k intermediate nodes between s and t F or a gi v en P-node P and an y B-node X located in P 0 s transmission range T R pb the Bac kwar d P ath ( P X ) is dened as the minimum-hop F orwar d P ath ( X P ) when all the nodes ha v e a transmission range of B T R It is w orth noting that the minimum-hop F orwar d P ath ( X ; P ) is not necessarily the same as the F orwar d P ath ( X ; P ) Although F or war d P aths are dened for an y node pairs in the netw ork, Bac kwar d P aths are only v alid between a P-node and the B-nodes located within the P-node' s T R pb range. Furthermore, 3 Ho we v er due to the asymmetric transmission po wer the discussed P-node may not be a neighbor of an indi vidual node of those neighbor candidates.

PAGE 33

21 for an y neighbor candidate X of a gi v en P-node P this B-node X can be considered as P 0 s neighbor only when the Bac kwar d P ath( P X ) satises the follo wing criteria: All the intermediate nodes along the Bac kwar d P ath( P X ) should be in P 0 s T R pb range. In other w ords, a neighbor candidate X can be considered as a P-node P 0 s neighbor if and only if all the intermediate nodes along the Bac kwar d P ath( P X ) are P 0 s neighbor s as well. W ith the abo v e denitions, the remaining issue is ho w to set up these Bac kwar d P aths A simple w ay is to let a P-node broadcast a query message with a certain transmission po wer that is, co v ering all the B-nodes in its T R pb = n B T R range. Once seeing such a query each node broadcasts a special reply with the TTL v alue set to T4. Each node appends its o wn ID in the reply when relaying such a special reply The querying P-node will w ait some time until collecting enough replies. The initiator of a reply w ould be considered as a neighbor if and only if the querying P-node also recei v es replies initiated from all the relaying nodes of that reply W e will see later in order to f acilitate the operation of A-MA C, the path length of a Bac kwar d P ath should be limited. W e need to point out that, e v en when a P-node, say P 1 recei v es a query message initiated by another P-node, say P 2 P 1 should reply lik e common B-nodes with a transmission range of BTR Since our scheme is tar geted for netw orks with lo w or moderate mobility P-nodes can e x ecute this process infrequently or in their respecti v e P-to-B periods when topology changes are detected by the MA C protocol. Therefore, the resulting o v erhead is e xpected to be af fordable. P-nodes also need to disco v er the neighboring relationship among themselv es. T o achie v e this, during the P-to-P period a P-node may send a query with appropriate transmission po wer that is set to co v er a range of T R pp = m B T R P-nodes recei ving this query may directly send replies back to the requesting P-node. Fig.2gi v es an e xample of the neighbor selection process. Suppose A is a P-node with T R pb = 2 B T R and T R pp = 4 B T R and the Backw ard P aths for neighbor 4 T can assume an inte ger v alue slightly lar ger than n to allo w more replies.

PAGE 34

22 A B C D E G F I H J B T R T R p b T R p p P n o d e B n o d e Figure 2: An e xample of the neighbor determination process (m=4, n=2, T=3.). candidates C F G and I are C B A F E A G F E A and I H C B A respecti v ely Since node H does not initiate an reply to A only C D and G are considered as A 0 s neighbor s Of course, B D and E are A 0 s neighbor s as well. In this e xample, another P-node J is also a neighbor of node A because J is in A 0 s T R pp and the y can directly communicate with each other 2.4.2 Routing Component of DELAR In homogeneous ad hoc netw orks, a node can only communicate with other nodes in its B T R range, while in heterogeneous ad hoc netw orks, a P-node is able to reach an y other node within a lar ger transmission range, for e xample, T R pb and T R pp Therefore, the resulting topology and routing strate gy may be quite dif ferent from that in homogeneous netw orks. As an e xample, a netw ork topology without P-nodes is depicted in Fig.2 3.a, where all the links are bi-directional and labelled with equal or unequal costs on both directions. F or instance, a 1 =b 1 indicates that the link cost from A to B is a 1 while b 1 from B to A. In contrast, if one node, say A is identied as a P-node who can reach much further in the netw ork, more unidirectional links may be added as sho wn in Fig.2.b W e label unidirectional links from P-node A to its neighbors with cost 0 to represent node A 0 s unlimited po wer supply .

PAGE 35

23 A B C D E F 0 / b 1 0 / c 1 c 2 / e 1 b 2 / d 1 b 3 / c 3 d 3 / e 3 d 2 / f 1 e 2 / f 2 0 0 0 A B C D E F a 1 / b 1 a 2 / c 1 c 2 / e 1 b 2 / d 1 b 3 / c 3 d 3 / e 3 d 2 / f 1 e 2 / f 2 a T o p o l o g y i n H o m o g e n e o u s c a s e b T o p o l o g y i n H e t e r o g e n e o u s c a s e Figure 2: The topology in homogeneous and heterogeneous cases. T o cope with such heterogeneous netw orks, each P-node needs to maintain an internal neighbor table recording its chosen neighbors within T R pb and the corresponding Bac kwar d P aths of those neighbors. In addition, each node in the netw ork, either a P-node or a B-node, needs to maintain a forwar ding r outing table similar to that in a normal tabledri v en routing protocol such as DSD V [7]. F or each node i we dene f ( i ) = r esidual ener g y ( i ) )Tj/T1_3 11.95509 Tf11.94411 0 Td( q ueue l en ( i ) ; where r esidual ener g y ( i ) indicates current remaining ener gy le v el at node i q ueue l en ( i ) represents the current load status at node i and is a parameter representing the ener gy consumption per unit data transmission5. Then the de vice-ener gy-load a w are routing cost metric we adopt is gi v en in Eq.2.1, though other cost metrics are applicable in DELAR as well. cost ( i ) = 8 > < > : 1 =f ; f > r a; f r (2.1) 5 In our simulation, for e xample, is equal to the a v erage ener gy consumption per pack et transmission.

PAGE 36

24 In the abo v e cost metric, cost ( i ) is the cost of using node i as a relay it could be used as the cost of all directional links (arcs) starting from node i and directed to an y of its neighbors; r is a parameter used to adjust the weight of the a w areness of load and ener gy in the o v erall cost metric; constant a assumes a relati v ely lar ge v alue to a v oid using the nodes short of ener gy Dif ferent types of de vices may assume dif ferent v alues of these parameters: ; r and F or e xample, in this chapter to represent a P-node' s unique po wer capability or de vice type, a P-node assumes a zero link cost6for all the links tow ard its B-node neighbors within T R pb or P-node neighbors within T R pp Ideally in order to nd ener gy-ef cient paths, each node should be informed about the routing costs of other nodes as accurate and prompt as possible which may lead to e xcessi v e o v erhead. In practice, ho we v er the emplo yed routing protocol should strik e a good tradeof f between ener gy ef cienc y and o v erhead. Proacti v e routing protocols are kno wn for their capability of propag ating netw ork conditions through the whole netw ork in due course so that appropriate QoS decisions, for e xample, admission control and route selection, can be made intelligently Thus, we adopt a proacti v e routing protocol, for e xample, DSD V as the underlying routing protocol to periodically e xchange the routing information. W e note that other types of routing protocols can also be used in this frame w ork. After g athering enough routing information, a node can emplo y a shortest path algorithm to decide the minimum cost paths and the related costs to all the other nodes in the netw ork. Here the path cost is actually the sum of the cost dened in Eq.2.1of all the B-nodes along a forw arding path. Similar to those ener gy-a w are cost metrics proposed in the literature, by choosing the proper v alues of and r the cost function dened in Eq.2.1can help prolong the netw ork lifetime by distrib uting the traf c more e v enly throughout the netw ork, a v oiding the o v eruse of a small set of nodes, and consuming nodal ener gy resources in a more balanced manner [37,39]. Moreo v er DELAR spontaneously incorporates P-nodes' unique 6 In practice, a v ery small v alue can be used to a v oid possible routing loops.

PAGE 37

25 po wer capabilities or de vice types, residual ener gy information, and local load status into the routing protocol without using redirect tables in DEAR [43] an y more. 2.4.3 Hybrid T ransmission Scheduling In order to reduce the interference a P-node' s communication may impose on other ongoing transmissions, it is reasonable to only allo w a P-node to boost its transmission po wer during some e xclusi v ely reserv ed periods. F or this purpose, time is di vided into equal length time periods called Superfr ames in which each P-node is assigned an e xclusi v e time interv al, called a P-to-B period, to communicate with its B-node neighbors. In addition, a P-to-P period is allocated to allo w P-nodes to communicate with each other if the y are within each other' s T R pp range. The rest of a Superfr ame called a B-to-B period is shared by all the nodes. Therefore, as sho wn in Fig.2, a P-node has three dif ferent transmission ranges in each Superfr ame : T R pp in P-to-P period, T R pb in its o wn P-to-B period, and T R bb in the rest of the Superfr ame In contrast, a B-node only has one transmission range T R bb = B T R and it can only initiate transmissions during the B-to-B period. Fig.2gi v es an instance of such a Superfr ame structure including multiple reserv ed P-to-B periods, one for each P-node. The one-minislot-length paddings between consecuti v e P-to-B periods are used to further reduce the possible interference. The period allocation of the Superfr ame can be fullled as follo ws. During the netw ork startup phase, Pnodes use high transmission po wer to communicate and ne gotiate with each other deciding the lengths of the P-to-P period, the P-to-B period and the B-to-B period, also associating each P-node with a P-to-B period. After nishing the ne gotiation, the P-nodes broadcast such allocation information to all the B-nodes in their o wn T R pb In this w ay ultimately

PAGE 38

26 all the nodes are informed the Superfr ame allocation, and can synchronize to such allocation7. In our current design, one P-to-B period is not allo wed to share by multiple P-nodes for simplicity only Ho we v er a P-to-B period can be shared among P-nodes f ar a w ay from each other if such sharing can ensure the transmissions conict-free. Since a P-node can communicate with other P-nodes in the P-to-P periods and communicate with the B-nodes within its B T R range in the B-to-B periods, it is natural that in its o wn P-to-B period, a P-node should gi v e priority to pack ets intended to its B-node neighbors which are outside its BTR range b ut inside its T R pb range. Thus, pack et scheduling is needed at a P-node to determine the appropriate transmission schedule for the pack ets to be relayed or initiated by itself. One may notice that the aforementioned transmission scheduling may need time synchronization among the nodes, we ar gue that nodes equipped with GPS de vices can easily fulll this task. This requirement is quite realistic today since such de vices are ine xpensi v e and can pro vide reasonable precision. In this chapter we assume that nodes ha v e perfect time synchronization and lea v e the synchronization problem in the netw orks without GPS de vices as our future w ork. In f act, the time-di vision scheduling, essentially a reserv ation-based access control mechanism, and the MA C protocols emplo yed in the three types of periods in each Super fr ame either contention-based or reserv ation-based, form a h ybrid transmission scheduling for DELAR. Moreo v er each type of periods may ha v e dif ferent MA C protocol. F or e xample, the con v entional contention-based MA C protocols, such as the IEEE 802.11 MA C protocol, can be used during P-to-P periods and B-to-B periods. Since unidirectional links 7 F or simplicity we assume a x ed allocation scheme is used in this chapter ho we v er more adapti v e allocation is possible when P-nodes periodically e xchange local load infor mation and ne gotiate a ne w allocation scheme during the P-to-P periods.

PAGE 39

27 A B T r a n s m i s s i o n R a n g e Figure 2: A unidirectional link between A and B. are basically una v oidable in P-to-B periods, special measures are needed to deal with them. In what follo ws, we will delineate the A-MA C protocol de v eloped for P-to-B periods. 2.4.4 Asymmetric Media Access Control Protocol (A-MA C) The presence of unidirectional links is pretty common in heterogeneous netw orks especially when emplo yed po wer control schemes cause unequal transmission po wers among nodes. As an e xample, node A in Fig.2has a lar ger transmission range than node B Thus, node B can hear node A s transmission, ho we v er node A cannot detect node B s transmission. Ob viously a unidirectional link e xists between node A and node B The dilemma is that the stop-and-w ait ARQ (Automatic Repeat Request) scheme [47] emplo yed in cur rent contention-based MA C protocols only w orks well with bidirectional links. In the f ace of unidirectional links, the recei v er B (Fig.2) has no w ay to directly and successfully send the ackno wledgements back to the transmitter A which means that the transmitter w ould continuously transmit the same frame before timeout no matter whether the recei v er has recei v ed it or not. Moreo v er the unidirectional links may se v erely af fect the functionalities of ad hoc netw orks at v arious layers [48,49,50]. F or e xample, man y routing protocols such as DSR and A OD V rely on hop-wise ackno wledgments for disco v ering route errors. Therefore, ho w to support the MA C layer ackno wledgements o v er the unidirectional links is v ery important [51,52] and has not yet been well addressed. F ortunately we can mak e use of the aforementioned Bac kwar d P aths and the follo wing mini-r outing method to tackle this problem in an ele g ant w ay .

PAGE 40

28 C A ( P n o d e ) 1 2 3 4 5 6 A D I F S P R T S P D A T A S I F S P C T S C 1 4 2 B S I F S S I F S P C T S 3 S I F S P A C K 5 S I F S P A C K 6 B Figure 2: The A-MA C operation procedure. In current contention-based MA C protocols such as the IEEE 802.11, a recei v er can only transmit an ackno wledgement frame to its one-hop-a w ay transmitter W ith the crosslayer design methodology we introduce a ne w concept of mini-r outing intoT rack your eBay acti vities the MA C layer which requests intermediate nodes to relay the recei v er' s ackno wledgement frames, that is, CTS/A CK frames, along the established Bac kwar d P ath(tr ansmitter r eceiver) in a multi-hop f ashion to the transmitter (a P-node) at the MA C layer Here the routing information is no longer e xclusi v ely used by the netw ork layer b ut shared by the MA C and netw ork layers. Based on the IEEE 802.11 [5], we introduce into A-MA C four special frames: PR TS, P-CTS, P-D A T A, and P-A CK, all of which can only be transmitted in P-to-B periods. When a P-to-B period of a P-node comes and the P-node happens to ha v e some pack ets to transmit, it rst boosts its transmission po wer to co v er the range of T R pb = n B T R W ith the scheduling described in Section2.4.3, all the other nodes should refrain from initiating a transmission and temporarily stop transmitting usual frames, that is, R TS/CTS/D A T A/A CK. The P-node associated with this P-to-B period can send pack ets to an y neighboring B-node in the range of T R pb through P-R TS/P-CTS/P-D A T A/P-A CK e xchanges. Ne xt, we illustrate the A-MA C by using Fig.2, where we assume n = 2 and P-node A intends to send a pack et to C one of its B-node neighbors. The location relationship among A B and C is also depicted in Fig.2. First, A sends the P-R TS with T R pb =

PAGE 41

29 2 B T R containing the Bac kwar d P ath(A,C) ( C B A ). Then according to the length of the Bac kwar d P ath(A,C) in this e xample which is 2, A sets its w aiting timer for the PCTS to be 2( S I F S + T P )Tj/T1_5 7.97011 Tf6.5833 0 Td(C T S + T pr op )8, where T P )Tj/T1_5 7.97011 Tf6.5833 0 Td(C T S and T pr op are the transmission time and air propag ation time for one P-CTS respecti v ely Upon recei ving the P-R TS destined to it, and w aiting for a SIFS period, node C will send node B a P-CTS including the addresses of A and C F or an intermediate node residing on the backw ard path, it needs to set its w aiting timer according to its order in the Bac kwar d P ath F or e xample, the i th node (the intended recei v er lik e C is assumed to be the 0 th node on the path) on the path should set its timer to be i ( S I F S + T P )Tj/T1_5 7.97011 Tf6.5833 0 Td(C T S + T pr op ) In this e xample, node B when o v erhearing the abo v e P-R TS from A it starts a timer equal to S I F S + T P )Tj/T1_5 7.97011 Tf6.5833 0 Td(C T S + T pr op since it is the 1 st intermediate node on the backw ard path. Once recei ving a P-CTS from node C before timeout, B simply appends its address and relays the modied P-CTS to the ne xt hop which is the P-node A in this e xample. Otherwise, node B sends a P-CTS containing its o wn address to A after the timer e xpires, and the reason for doing this will be e xplained later If node A does not recei v e an y P-CTS before timeout, it can retransmit the P-R TS until reaching an admissible number of retries. If the same situation happens, A temporarily sa v es this pack et for future transmission and switches to another pack et with a dif ferent ne xt hop. When A successfully recei v es a P-CTS from B containing both B' s and C' s addresses, the P-R TS/P-CTS e xchange nishes. After a SIFS A can send a P-D A T A frame to node C and set the timer to 2( S I F S + T P )Tj/T1_5 7.97011 Tf6.5833 0 Td(AC K + T pr op ) Then the similar procedures apply After recei ving the P-A CK from node C relayed by the intermediate node B, the P-node A can start transmitting a ne w pack et after a DIFS in the same manner When its P-to-B period e xpires, A lo wers its transmission po wer and acts as a B-node in other P-nodes' P-to-B periods and in the B-to-B period. 8 SIFS stands for Short Inter -frame Space, and DIFS stands for DCF Inter -frame Space.

PAGE 42

30 C A ( P n o d e ) 1 2 3 4 5 6 A D I F S P R T S P D A T A S I F S P C T S C 1 4 2 B S I F S S I F S P C T S 3 S I F S P A C K 5 S I F S P A C K 6 B A s s e m l e d M P D U M A C h e a d e r M S D U C R C M A C h e a d e r M S D U C R C M P D U f o r B M P D U f o r C Figure 2: The multiple-pack ets transmission enhancement to DELAR. Besides the purpose of resolving the well-kno wn hidden/e xposed terminal problems, the P-R TS/P-CTS e xchange is also used to eliminate possible errors resulting from stale routes or nodes' mobility F or e xample, in the abo v e e xample, if node C mo v es out of P-node A 0 s 2 B T R range while B is still in A s B T R range, node C will not hear the PR TS from node A and hence A could only recei v e from node B a P-CTS including only B' s address. In this case, A will think that node C is currently unreachable and may temporarily sa v e the pack ets to C for future transmissions. Another situation may happen that node C is still in A 0 s 2 B T R range while node B mo v es out of A s B T R range, in which the P-node A will delete node C from its neighbor table. Moreo v er when node C mo v es into A s BTR range, A will hear the P-CTS from node C directly Hence A can optimize the future transmissions to C without the help of node B an y more. 2.5 Multiple-P ack ets T ransmission in DELAR In the basic DELAR design, time-di vision transmission scheduling is used to coordinate the transmission acti vities of P-nodes and B-nodes. One undesirable consequence is the e xcessi v e delay one pack et may e xperience because it may be b uf fered at intermediate nodes to w ait for appropriate transmission periods. On the other hand, since DELAR is ener gy a w are and it costs P-nodes almost nothing to transmit a pack et, man y data packets may sw arm to P-nodes. This may mak e P-nodes the bottlenecks of the netw ork and further increase the delay that pack ets accumulated at P-nodes w ould e xperience. In what follo ws, we seek a w ay to alle viate this phenomenon.

PAGE 43

31 2.5.1 Multiple-pack ets T ransmission In the basic design (see Fig.2), for e xample, A can only transmit pack ets to either B or C each time. When P-node A transmits a pack et to B-node C node B is only in v olv ed in forw arding the control frames (for e xample, P-CTS/P-A CK). Ho we v er since the channel is reserv ed for P-node A during its P-to-B periods and node B is also in P-node A s transmission range, node B has the capability to recei v e and demodulate an y signal transmitted by node A if it is allo wed. Therefor if node A can recognize B' s capability and utilize it, that is, transmitting pack ets to node B and C at the same time, the system performance could be impro v ed. This multiple-pack ets transmission mechanism can be illustrated in Fig.2. Suppose P-node A has some pack ets to B and some pack ets to C Node A w ould put one pack et for C and another for B together and send them in a single transmission from which nodes B and C can acquire their o wn part, respecti v ely In this w ay we e xpect to see the impro v ement of the end-to-end delay performance of DELAR. T o do this, node A rst mak es sure that both B and C are within its ef fecti v e range T R pb after the P-R TS/P-CTS e xchange as before. Then A can pull out from the w aiting queue one pack et to w ards C and another pack et to w ards B and send them in one P-D A T A frame as depicted in Fig.2. When seeing such a frame, nodes B and C can e xtract their o wn parts and dump the rest. The same procedures as specied in A-MA C are e x ecuted with the e xception that node B also needs to indicate in the P-A CK that it has successfully recei v ed the pack et to itself. The similar procedure can be applied to the cases with more than one interlineate nodes on the bac kwar d paths W ith the abo v e multiple-pack ets transmission mechanism, one may e xpect that the total number of pack ets that can be pack ed and transmitted at one time is bounded by the length of the Bac kwar d P ath Considering the possibility of adopting high-data rate modulation schemes with a higher po wer le v el so as not to de grade the recei v ed signal strength, more pack ets to w ards dif ferent recei v ers on the Bac kwar d P ath can be assembled together and transmitted at the same time.

PAGE 44

32 A 0 0 D 0 1 B 1 0 C 1 1 0 1 q Figure 2: General hierarchical 2/4-PSK constellation. 2.5.2 Hierarchical Modulation T o support the multiple-pack ets transmission, hierarchical modulation (or nonuniform modulation) schemes can be use to ensure all the recei v ers ha v e enough recei v ed signal strength to demodulate the useful information. In hierarchical modulation schemes, the constellations consist of nonuniformly spaced symbols and allo w for unequal error protection, that is, dif ferent de grees of protection for transmitted bits within a symbol are allo wed according to the importance of the information. F or e xample, suppose that there are tw o streams of data, each of which has a priority (a tar get BER), and QPSK hierarchical modulation (see Fig.2) is used. One bit from each stream is tak en to form a symbol of tw o bits. The bit from bitstream with high priority (lo wer allo wed BER) is assigned to the most signicant position, for e xample, the rst bit in the constellation in Fig.2. The bit from the bitstream with lo w priority is assigned to the less signicant position. By adjusting the angle of the constellation can be determined to achie v e the unequal error protection for each bitstream. Such unequal error protection has mak e hierarchical modulation v ery attracti v e in multimedia services [53,54,55]. W e notice that in the pre vious research, the hierarchical modulation is used in the cases that the transmission po wer is x ed and the constellation is chosen in a w ay to allocate the po wer among multiple data streams to achie v e unequal error protection. In addition, in the pre vious research, bitstreams usually are destined to the same recei v er Ho we v er in our

PAGE 45

33 0 0 0 1 1 0 1 1 q C A ( P n o d e ) d 1 d 2 d 1 2 BFigure 2: Implement multiple-pack ets transmission with hierarchical modulation DELAR scheme, the transmission po wer of P-nodes is not x ed. Moreo v er when multiplepack ets transmission is used, the bitstreams are destined to dif ferent recei v ers. As sho wn in Fig.2, multiple recei v ers in v olv ed in the multiple-pack ets transmission ha v e dif ferent distances to the transmitter Therefore, in order to use hierarchical modulation to support multiple-pack ets transmission at a P-node, more po wer ought to be allocated to f ar a w ay recei v ers and less po wer for close recei v ers. More specic, we need to nd out the constellation and the minimum o v erall transmission po wer for a P-node so that all the recei v ers can successfully get their o wn pack ets. In the follo wing we use QPSK hierarchical modulation to implement the multiple-pack ets transmission when tw o recei v ers are in v olv ed (Fig.2). Suppose B E R max is the BER requirement for both recei v ers B ( d 2 a w ay from transmitter A ) and C ( d 1 a w ay from transmitter A ) to demodulate the recei v ed signal. Moreo v er we assume tw o-ray ground model with = 4 is used to model the radio propag ation: P r = P t G t G r h 2 t h 2 r d 4 L where L is the system loss f actor P t is the transmission po wer P r is the recei v ed po wer G t ( G r ) is the antenna g ain at the transmitter(recei v er), and h t ( h r ) is the height of antenna at the transmitter(recei v er). F or simplicity we assume all the nodes ha v e the same antenna g ain and height. W e assume that is less than 45 o so that the inphase signal to w ards f ar

PAGE 46

34 a w ay recei v er C is assigned with more po wer and quadrature signal to w ards close recei v er B with less po wer It is easy to get the BER at C as p 1 = Q ( s 2 P r 1 W N f cos ) = Q ( v u u t 2 P t G t G r h 2 t h 2 r d 4 1 L W N f cos ) (2.2) where N is the noise signal po wer f is the transmission bit rate, W is the channel bandwidth (in H z ), and Q ( ) is Gaussian Q-function [54,55]. Similarly we can get the BER at the recei v er B as p 2 = Q ( s 2 P r 2 W N f sin ) = Q ( v u u t 2 P t G t G r h 2 t h 2 r d 4 2 L W N f sin ) : (2.3) Thus, in order to correctly demodulate the recei v ed signal, the BER should satisfy the follo wing conditions: p i B E R max where i = 1 ; 2 Since we are interested in the minimum transmission po wer of node A when equality is satised: p i = B E R max where i = 1 ; 2 we dene r 0 = Q )Tj/T1_2 7.97011 Tf6.5833 0 Td(1 ( B E R max ) as the SNR achie ving the B E R max to f acilitate us nding the boundary v alues of P t and From Eq.2.2and Eq.2.3, it is easy to get v u u t 2 P t G t G r h 2 t h 2 r d 4 1 L W N f cos = r 0 (2.4) and v u u t 2 P t G t G r h 2 t h 2 r d 4 2 L W N f sin = r 0 : (2.5) From Eq.2.4and Eq.2.5, we can get = arctan d 2 2 d 2 1 (2.6) and P tmin = r 0 N f ( d 4 1 + d 4 2 ) L W G t G r h 2 t h 2 r (2.7)

PAGE 47

35 Therefore, gi v en the distances between the transmitter and recei v ers, we can get the minimum transmission po wer P tmin and in the QPSK hierarchical modulation constellation, and use this combination to implement the proposed multiple-pack ets transmission and achie v e the BER requirements at both recei v ers. In practice, P-node A may rst use a predened transmission po wer co v ering T R pb and transmit the R TS frame using typical BPSK. After the R TS-CTS handshak e, A can determine the tw o parameters P tmin and based on the feedback from the tw o recei v ers. Then A pulls out from the w aiting queue one pack et to w ards C as the more important bitstream and one to w ards B as the less important bitstream. Further it can transmit symbols using the aforementioned QPSK hierarchical modulation. W e should note that, in order to a v oid the possible interference when a P-to-B period is shared by multiple P-nodes, the in v olv ed P-nodes, before enabling the multiple-pack ets transmission, should e xamine and mak e sure that P tmin is not greater than the po wer to co v er T R pb In f act, the minimum total transmission po wer to transmit the pack ets to each recei v er separately using BPSK can be written as follo ws: P ttoal = P t 1 + P t 2 = r 0 N f d 4 1 L W G t G r h 2 t h 2 r + r 0 N f d 4 2 L W G t G r h 2 t h 2 r : (2.8) Compared Eq.2.7to Eq.2.8, we can see that although P tmin alone is lar ger than either P t 1 or P t 2 P tmin is identical to P ttoal This v eries that multiple-pack ets transmission with QPSK hierarchical modulation does not require e xtra transmission po wer9compared to the case when multiple-pack ets transmission is not used. In f act, if multiple-pack ets transmission is implemented with uniform QPSK, in order to satisfy the same BER requirement at both recei v ers: p i B E R max where i = 1 ; 2 the transmission po wer w ould be p 2 P t 1 which is much lar ger than P tmin Thus, in our case, QPSK hierarchical modulation is a rather reasonable option to implement multiple-pack ets transmission. Further 9 Here we only consider the po wer used to transmit the information bits.

PAGE 48

36 if we consider the ener gy used for control frames, the proposed multiple-pack ets transmission requires fe wer control frames thus less ener gy consumption, compared to the case when pack ets are transmitted separately F or e xample, in Fig.2, transmission ener gy for three control frames (R TS/CTS/A CK) to w ards node B can be sa v ed when multiple-pack ets transmission is used. 2.6 Discussion 2.6.1 The Existence of Backw ard P aths In DELAR, the P-nodes are utilized in tw o w ays to conserv e ener gy: enabling Pnodes to directly communicate with other P-nodes within T R pp and enabling P-nodes to directly communicate with B-nodes within T R pb Unlik e the communications between Pnodes that need no special treatment, the communications between P-nodes and B-nodes are supported by the nodes on the backw ard paths as described in Section2.4.1, thus the backw ard path is v ery important for the proper functioning of DELAR. Then one may question the e xistence of such backw ard paths. W e belie v e that the e xistence of backw ard paths is related to se v eral f actors such as basic transmission range (BTR), node density and the location distrib ution of P/B-nodes. On the other hand, the backw ard path should not be too long in terms of hop count, otherwise A-MA C may not function well because a long backw ard path is more subject to breakage due to node mobility or other reasons. In practice, we should include this limitation into the aforementioned neighbor -selection criteria, for e xample, setting TTL v alue of a reply to the neighboring query to n (cf. Section2.4.1), such that only those B-nodes with backw ard paths of less than n hops can be considered as neighbors of a gi v en P-node. Ne xt we w ant to study the e xistence of such backw ard path in typical mobile ad hoc netw orks. W e assume n = 2 and T = 2 (cf.2.4.1) so that only those nodes with backw ard paths not more than 2 hops are considered as neighbors of a gi v en P-node. Let O denote the position of a P-node, and O 0 denote the position of a B-node in O s 2 B T R = 2 R range, as sho wn in Fig.2. W e use r to indicate the distance between O and O 0 It is

PAGE 49

37 ob vious that when r is not greater than R O 0 is denitely a neighbor of O So we shall only consider R < r 2 R when studying the e xistence of backw ard paths for those nodes that are r a w ay from O W e also assume that all the nodes are uniformly and independently distrib uted in a tw o-dimensional area with node density where the probability of ha ving k nodes in an area of size S follo ws Poission distrib ution: p ( k ; S ) = ( S ) k k e )Tj/T1_5 7.97011 Tf6.5833 0 Td( S : The mean of the number of nodes in the area of size S is S Then the e xistence of backw ard paths between O and O 0 or the problem whether O 0 is a neighbor of O or not can be reduced to a rather simple geometric problem: Is there an y other node located in the shado wed area S with both O and O 0 as neighbors? Let be the a v erage number of nodes e xisting in the shado wed area, and S 1 and S 2 be the areas of sector [ AB O and 4 AO B respecti v ely Then we ha v e = 2 ( S 1 )Tj/T1_1 11.95509 Tf11.4671 0 Td(S 2 ) where S 1 = R 2 2 2 arccos r = 2 R = R 2 arccos r = 2 R and S 2 = 1 2 R 2 sin 2 arccos r = 2 R : Therefore, we ha v e = 2 ( R 2 arccos r = 2 R )Tj/T1_4 11.95509 Tf13.149 8.08202 Td(1 2 R 2 sin 2 arccos r = 2 R ) : Thus, the probability of e xisting a backw ard path between O and O 0 which is equal to the probability that there is at least one node in the shado wed area, is 1 )Tj/T1_1 11.95509 Tf11.9261 0 Td(e )Tj/T1_5 7.97011 Tf6.5833 0 Td( according to the Poission distrib ution. Fig.2plots this probability with dif ferent v alues of r where the total number of nodes is 50 for all netw ork topologies. W e can see that, with a reasonable node density the probability of e xisting a backw ard path between a gi v en P-node and Bnode pair is pretty high, which justies the feasibility of our A-MA C protocol and in turn the o v erall DELAR scheme.

PAGE 50

38 R 2 R r O O A B S Figure 2: The e xistence of backw ard path. 2 0 0 2 5 0 3 0 0 3 5 0 4 0 0 4 5 0 5 0 0 0 0 1 0 2 0 3 0 4 0 5 0 6 0 7 0 8 0 9 1 r ( m ) P r o b ( E x i s t e n c e o f a b a c k w a r d p a t h ) 1 5 0 0 3 0 0 m27 0 0 7 0 0 m28 0 0 8 0 0 m29 0 0 9 0 0 m21 0 0 0 1 0 0 0 m2 B T R = 2 0 0 B T R = 2 5 0 Figure 2: The a v erage number of nodes e xisting in the shaded area.

PAGE 51

39 2.6.2 DELAR and ZRP In f act, we can borro w some ideas from ZRP [56,11] to further impro v e the routing performance of DELAR as follo ws. A P-node maintains the routing information within a zone T R pb by using the procedures described in Section2.4.1or an y other routing protocol as the intra-zone routing protocol (IARP). It also maintains the information about its neighboring P-nodes in its T R pp In contrast, a B-node only needs to maintain the routing information in its T R bb = B T R where all the nodes are its one-hop neighbors. Once a node needs an ener gy-ef cient route to another node, it can disco v er the route on demand using the routing disco v ery procedure similar to that of A OD V b ut with what is dened in Eq.2.1instead of hop count as the cost metric. In this sense, DELAR can be vie wed as a special case of ZRP The dif ference lies in the f act that all nodes within a P-node' s T R pb zone, are one-hop a w ay from this P-node from the routing perspecti v e rather than multi-hop a w ay in the le g ac y ZRP Besides, the border -casting [56,11] technique used in ZRP is also a v ailable in DELAR in the sense that a P-node border -casts a route request (which we call simply a request) to all its peripheral nodes with corresponding backw ard paths embedded in the route request. In this w ay each peripheral node w ould learn the backw ard path used to return a route reply if needed. Moreo v er since a P-node can directly e xchange routing information with other P-nodes within its T R pp it also can border -cast the route request to its P-node neighbors, which in turn can look up their o wn neighbor tables to decide if the desired destination is in their o wn T R pb zones. Since T R pp is usually lar ger than T R pb apparently such inter -zone routing protocol (IERP) can further speed up the route disco v ery process. 2.6.3 The Choice of m and n Since a P-node can use higher transmission po wer to communicate with other P-nodes within its m B T R lar ger m w ould lead to less use of B-nodes in the communications, b ut also less spatial reuse. Similarly lar ger n may lead to more ener gy sa vings, b ut also imply possible longer backw ard paths and less spatial reuse, which may mak e A-MA C less

PAGE 52

40 ef cient as e xplained before. F or the similar reasons, usually n is less than m In order to well balance the ener gy sa vings and other system performance f actors, both m and n should be chosen cautiously While DELAR requires all the P-nodes ha v e the same v alue of m to a v oid producing unidirectional links between P-nodes in P-to-P periods, with A-MA C in place, the y can ha v e dif ferent v alues of n 2.6.4 Benets of the T ime Di vision Scheduling In f act, P-nodes can communicate with each other in one-hop or multi-hop manner during P-to-P periods to coordinate the use of the ne xt Superfr ame A P-node can also notify all the other nodes within its T R pb to adjust their transmission schedules. W ith these intelligent communications a v ailable, the slot allocation in each Superfr ame can be adjusted adapti v ely according to traf c conditions instead of being x ed as in the gi v en e xample. It is w orth pointing out that this time-di vision scheduling method can f acilitate the operations of PSM (cf. Section2.2) in that it can help nodes determine their sleep and w ak e schedules. F or e xample, B-nodes can turn of f the radios during the P-to-P periods to conserv e ener gy As we discussed before, time synchronization is of importance for the correct oper ation of DELAR. In literature, there e xist man y proposals for time synchronization and man y of them can be incorporated into DELAR. In f act, P-nodes can serv e as coordinators to f acilitate such synchronization by sending out some beacon information in some P-toP periods and P-to-B periods periodically Subsequently B-nodes can synchronize their clocks to these P-nodes. 2.7 Performance Ev aluation 2.7.1 Simulation Setup In order to e v aluate the performance of DELAR, we implemented our scheme including the routing layer and the A-MA C in the OPNET Modeler [57]. W e simulated a netw ork with 50 nodes randomly deplo yed in a 1500 300 m 2 area. The B T R w as 200 m and the transmission rate w as 2 M bps In our simulations, all the nodes were capable of mo ving

PAGE 53

41 in the netw ork according to the modied random W aypoint mobility model presented in [58]. The pause time w as set to be zero in our simulations, meaning nodes were al w ays mo ving. And each node mo v ed with a randomly chosen speed between [ V min V max ], where V min w as x ed to be 1 m/s and V max assumed dif ferent v alues to reect v arious netw ork mobility le v els. There were 20 constant bit rate (CBR) data sessions between randomly selected source and destination pairs, and each source generated data pack ets of 512 bytes in length at a rate of pack ets per second. In our simulation, B-nodes had the same initial ener gy reserv oir 6 k J and their transmission and reception po wer were 1560 mW and 930 mW respecti v ely [59]. The netw orks with 2 to 6 P-nodes were studied and these P-nodes were randomly deplo yed. Besides, we chose m = 4 n = 2 t pp = 0 : 02 s t pb = 0 : 02 s and t bb = 0 : 05 s10. By v arying the number of P-nodes, the maximum mo ving speed, and the CBR source rate, we were able to study the performance of DELAR under dif ferent congurations. Six runs were carried out to get an a v erage result for each simulation conguration and each run w as e x ecuted for 900 seconds of simulation time. Pre vious w ork [37,39] has sho wn that the ener gy ef cienc y of routing protocols can be much impro v ed by adopting such a routing metric as what we dened in Eq.2.1. Therefore, we shall only compare our DELAR and DELAR with multiple-pack ets transmission (denoted by DELAR+MPT) with the one referred to as EAR in this chapter which is dev eloped for comparison purpose and in f act a v ariant of DSD V with the routing cost metric dened in Eq.2.1. In EAR, all the P-nodes ha v e the same transmission range as B-nodes, b ut ha v e a zero cost to their neighbors because the y are assumed to ha v e almost unlimited ener gy reserv oir and other resources. And the follo wing metrics will be adopted in 10 W ith adv anced simulation tools such as OPNET NS-2, and Qualnet, it is possible to gure out a good conguration of these parameters to yield a good system performance, before putting DELAR into practice.

PAGE 54

42 comparison: a v erage ener gy consumption dened as the total ener gy consumption for all pack et transmissions and receptions normalized by the number of deli v ered pack ets; pack et deli v ery ratio dened as the ratio of deli v ered data pack ets to those generated by the sources; a v erage pack et end-to-end delay dened as the a v erage delay from when a pack et is generated and transmitted by the source till it is recei v ed by the destination. 2.7.2 Impact of The Number of P-nodes W e rst x ed the data rate to 4 pack ets/s and v aried the number of P-nodes to study its impact on the performance of DELAR. Here only the results for V max =4 m=s are presented in Fig.2, though DELAR has the similar performance with other v alues of V max Figure 2(a) comparesaverageenergyconsumptionofDELARandEARunder dif ferent numbers of P-nodes. Since EAR is also an ener gy-a w are routing protocol, it is of no surprise to see that its ener gy-sa ving performance impro v es with the number of P-nodes whose routing costs to their neighboring nodes are assumed to be zero. W ith DELAR in place, ho we v er the a v erage ener gy consumption can be much further reduced, for e xample, with a f actor of almost 50% if 6 P-nodes are a v ailable. The sho wn adv antage comes from the f act that DELAR mak es much better use of P-nodes than EAR through intelligent transmission scheduling and the allo w ance of P-nodes using dif ferent transmission po wer during v arious periods. And the more P-nodes, the more ener gy sa vings we can e xpect from DELAR. Compared to the basic DLEAR, DELAR with multiple-pack ets transmission can further reduce the ener gy consumption. It can be contrib uted to the reduced transmissions of control pack ets because multiple-pack ets transmission requires fe wer control frames than separate transmissions. From Fig.2(b), we can see that DELAR outperforms EAR in terms of pack et deli v ery ratio (PDR). The reason is that transmission scheduling in DELAR will lead to less congestion, and the lar ger transmission ranges of P-nodes during P-to-P periods and P-to-B periods can help reduce the number of hops a pack et may tra v el. Ag ain, the more P-nodes, the more PDR is impro v ed. Compared to the basic DLEAR, the multiple-pack ets

PAGE 55

432 3 4 5 6 1 2 1 3 1 4 1 5 1 6 1 7 1 8 1 9 2 0 2 1 2 2 2 3 2 4 2 5 2 6 2 7 2 8 2 9 A v e r a g e P a c k e t E n e e g y C o n s u m p t i o n ( m J )N u m b e r o f P n o d e s E A R D E L A R D E L A R + M P T(a) A v erage ener gy consumption2 3 4 5 6 0 9 2 8 0 9 3 0 0 9 3 2 0 9 3 4 0 9 3 6 0 9 3 8 0 9 4 0 0 9 4 2 0 9 4 4 0 9 4 6 0 9 4 8 0 9 5 0 0 9 5 2 0 9 5 4 0 9 5 6 0 9 5 8 0 9 6 0 0 9 6 2 P a c k e t D e l i v e r y R a t i oN u m b e r o f P n o d e s E A R D E L A R D E L A R + M P T(b) P ack et deli v ery ratio2 3 4 5 6 0 0 0 0 0 5 0 1 0 0 1 5 0 2 0 0 2 5 0 3 0 0 3 5 0 4 0 A v e r a g e P a c k e t E n d t o e n d D e l a y ( s )N u m b e r o f P n o d e s E A R D E L A R D E L A R + M P T(c) A v erage pack et end-to-end delay Figure 2: Simulation results with dif ferent number of P-nodes.

PAGE 56

44 transmission can further impro v e the pack et deli v ery ratio because transmitting multiple pack ets simultaneously can impro v e the channel utilization. In terms of end-to-end delay ho we v er EAR is better than DELAR as sho wn in Fig.2 11(c). This is because, in contrast to the contention-based transmissions in EAR, DELAR di vides the time into Super F r ames to schedule the transmission acti vities so that a pack et usually needs to be b uf fered at a node w aiting for the coming of a proper transmission period. In addition, we can observ e that the delay of EAR decreases with the increase of P-nodes because the e xistence of more P-nodes can achie v e better load balance, that is, the number of ener gy-ef cient paths may be increased. Ho we v er this is not al w ays the case with DELAR. Besides the better load balance, the increase of P-nodes also means for DELAR that the delay from the transmission scheduling becomes lar ger This contrib utes to the uctuation in the delay of DELAR depicted in Fig.2(c). Compared to the basic DLEAR, DELAR with multiple-pack ets transmission can reduce the pack et end-to-end delay 2.7.3 Impact of Node Mobility In this subsection, we study the impact of the node mobility on DELAR by v arying V max from 2 m=s to 16 m=s F or the reason of clarity only the results for 4 P-nodes and a data rate of 4 pack ets=s are presented. Fig.2(a)compares the a v erage ener gy consumption of DELAR and EAR under dif ferent mobility le v els. As we can see, DELAR al w ays has less ener gy consumption than EAR due to the reasons stated before. Generally the higher mobility leads to less ener gy consumption. After e xamining the a v erage number of hops a pack et may tra v el, we notice that higher mobility often results in shorter route, which statistically leads to less ener gy consumption because fe wer transmissions and receptions are in v olv ed. As sho wn in Fig.2(b), the pack et deli v ery ratio decreases with the increase of the mobility which is in accordance with pre vious studies. F or the similar reason we stated in pre vious subsection, DELAR al w ays has higher pack et deli v ery ratio than EAR in all kinds of mobility As

PAGE 57

450 2 4 6 8 1 0 1 2 1 4 1 6 1 8 1 4 1 6 1 8 2 0 2 2 2 4 2 6 2 8 3 0 A v e r a g e P a c k e t E n e e g y C o n s u m p t i o n ( m J )M a x i m u m N o d e S p e e d : V m a x ( m / s ) E A R D E L A R D E L A R + M T P(a) A v erage ener gy consumption0 2 4 6 8 1 0 1 2 1 4 1 6 1 8 0 8 2 0 8 4 0 8 6 0 8 8 0 9 0 0 9 2 0 9 4 0 9 6 0 9 8 P a c k e t D e l i v e r y R a t i oM a x i m u m N o d e S p e e d : V m a x ( m / s ) E A R D E L A R D E L A R + M T P(b) P ack et deli v ery ratio0 2 4 6 8 1 0 1 2 1 4 1 6 1 8 0 0 4 0 0 6 0 0 8 0 1 0 0 1 2 0 1 4 0 1 6 0 1 8 0 2 0 0 2 2 0 2 4 0 2 6 0 2 8 0 3 0 0 3 2 0 3 4 0 3 6 A v e r a g e P a c k e t E n d t o e n d D e l a y ( s )M a x i m u m N o d e S p e e d : V m a x ( m / s ) E A R D E L A R D E L A R + M T P(c) A v erage pack et end-to-end delay Figure 2: Simulation results with dif ferent maximum node speed.

PAGE 58

462 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 0 9 1 0 9 2 0 9 3 0 9 4 0 9 5 0 9 6 P a c k e t D e l i v e r y R a t i oT r a f f i c L o a d ( K b i t s / s ) E A R D E L A R D E A L R + M P T(a) P ack et deli v ery ratio2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 0 0 0 0 0 5 0 1 0 0 1 5 0 2 0 0 2 5 0 3 0 0 3 5 0 4 0 A v e r a g e P a c k e t E n d t o e n d D e l a y ( s )T r a f f i c L o a d ( K b i t s / s ) E A R D E L A R D E L A R + M P T(b) A v erage pack et end-to-end delay Figure 2: Simulation results with dif ferent traf c load. the mobility increases, generally the delays of both EAR and DELAR get longer Ag ain, DELAR has longer delay than EAR due to DELAR' s time-di vision medium access control mechanism. One interesting observ ation is that the delays of both DELAR and EAR uctuate around V max = 2 m=s and V max = 2 m=s This can be attrib uted to the used routing cost metric which causes lots of pack ets sw arming to the P-nodes. This phenomenon results in longer w aiting time at P-nodes. Ho we v er nodal mo v ement helps alle viate such phenomenon by dispensing the traf c load. Compared to the basic DLEAR, DELAR with multiple-pack ets transmission can reduce the ener gy consumption, impro v e the pack et deli v ery ratio, and shorten the end-to-end delay in all kinds of mobility .

PAGE 59

47 2.7.4 Impact of T raf c Load In this subsection, we study the impact of the traf c load on DELAR by v arying the data generation rate from 3 pack ets=s to 8 pack ets=s Since the traf c load does not much impact a v erage ener gy consumption, we only depict the simulation results for the pack et deli v ery ratio and the end-to-end delay in Fig.2, where the number of P-nodes is four and V max is equal to 4 m=s Figures 2(a) and 2(b) demonstratethatthepacketdeliveryratiodecreases and the delay increases with the increase of the traf c load for both schemes, respecti v ely which are quite intuiti v e. Ag ain, our DELAR is better in terms of the pack et deli v ery ratio b ut w orse with re g ard to the end-to-end delay than EAR for the reasons stated pre viously Ho we v er Compared to the basic DLEAR, DELAR with multiple-pack ets transmission outperforms DELAR due to reason described before. In summary DELAR is more appropriate for delay-insensiti v e applications, such as le transfer and web access, which prefer higher ener gy-ef cienc y and pack et deli v ery ratio. W ith multiple-pack ets transmission, the performance can be further impro v ed. Thus it is a viable enhancement and can be ef fecti v ely used in DELAR frame w ork. W ith a better tuning of system parameters DELAR can strik e a good balance between ener gy ef cienc y and other system performance f actors. 2.8 Summary In this chapter we proposed a De vice-Ener gy-Load A w are Relaying frame w ork, namely DELAR, to achie v e ener gy conserv ation in mobile ad hoc netw orks. DELAR utilizes the de vice heterogeneity inherent in ad hoc netw orks and features the cross layer protocol design methodology T o tak e better adv antage of po werful nodes (P-nodes) while reducing their interference on other ongoing communications, a h ybrid transmission scheduling mechanism is used to schedule and coordinate the transmission acti vities among P-nodes and B-nodes (ordinary nodes). In addition, in order to support reliable transmissions in the

PAGE 60

48 presence of unidirectional links between P-nodes and B-nodes, we introduced the minirouting technique and the no v el Asymmetric MA C ( A-MA C ) protocol. W e demonstrated that A-MA C can ef fecti v ely enable the MA C layer ackno wledgements o v er unidirectional links. T o the best of our kno wledge, no pre vious ef fort has been made to address this issue at the MA C layer Moreo v er we proposed a multiple-pack ets transmission scheme which can be operated with hierarchical modulation scheme to further impro v e the performance of DELAR. Detailed simulations sho wed that DELAR can signicantly reduce the ener gy consumption and thus prolong the netw ork lifetime e v en with a fe w P-nodes e xisting in the netw ork. W ith this frame w ork, v arious ener gy conserv ation techniques such as po wer sa ving modes, transmission po wer control and po wer -a w are routing can be inte grated to jointly achie v e better ener gy conserv ation. More important, this frame w ork pro vides a platform to study other challenging issues such as QoS pro visioning and security support.

PAGE 61

CHAPTER 3 RESOURCE A W ARE MO VEMENT IN HETER OGENEOUS MOBILE AD HOC NETW ORKS 3.1 Introduction As discussed in Chapter2, there has been a rich literature addressing the ener gy conserv ation issue in MANETs, ranging from po wer -sa ving mode(PSM) [23,60,25,24,61], to transmission po wer control(TPC) [26,32,27,30,33,34], and to po wer -a w are routing(P AM) [37,40,41]. Though those ener gy-a w are MA C and routing protocols can in general help the whole system e xpend the ener gy resources more reasonably to some e xtent, there are still man y other aspects one can e xplore to further impro v e the system-wide ener gy ef cienc y As an e xample, similar to the traf c jam in daily life where a mass of v ehicles ocks to a single spot, unwise mo v ement may cause local netw ork traf c congestion, thus leading to unf a v orable ener gy w aste. This observ ation moti v ates us to address ener gy conserv ation from the perspecti v e of node mo v ement. In addition to node mo v ement, resource heterogeneity as another inherent character istic of MANETs, is often either o v erlook ed or underutilized in designing ener gy conserv ation schemes for MANETs. Though node heterogeneity can be interpreted in v arious w ays, we limit the scope of this chapter to heterogeneous netw orks in terms of ener gy supply In such a netw ork, most nodes (called B-nodes hereafter) are furnished with lightweight batteries ha ving limited po wer while a fe w others (called P-nodes hereafter) are po wered by almost unlimited ener gy supplies such as ener gy-sca v enging de vices (for e xample, solar cells) and dynamos when nodes are in some mobile v ehicles. In a relati v e sense, the ener gy consumption of P-nodes can be considered as small or e v en ne gligible. In this chapter we propose to address ener gy conserv ation by guiding nodes' s mo v ement and utilizing de vice heterogeneity The basic idea is that, instead of mo ving in the 49

PAGE 62

50 eld blind to the netw ork en vironment, for e xample, al w ays follo wing shortest-hop paths, nodes are instructed to tra v el much more intelligently by k eeping in mind the system-wide objecti v e of ener gy conserv ation and mo ving along resource-a w are paths in such a w ay that P-nodes can undertak e as much communication tasks as possible so that less po werful P-nodes can sa v e ener gy thus leading to the prolonging of the whole netw ork lifetime. The contrib utions in this chapter are mainly fourfold. First, we dene a general mobility model and formulate a general resource-a w are mo v ement problem, from which we deri v e a W aterhunter Mo v ement problem for B-nodes and a Firehunter Mo v ement problem for P-nodes. Second, we reduce the W atehunter Mo v ement problem to a NP-complete distance-constrained least-cost (DCLC) routing problem and propose an ef cient heuristic solution. Third, we propose a routing delay dif ferentiation mechanism to mak e full use of the benets pro vided by the resource-a w are mo v ement. Last, this resource-a w are mo v ement can be incorporate into other ener gy conserv ation schemes to further impro v e the ener gy ef cienc y The ef fecti v eness of the proposed schemes are justied and v alidated through e xtensi v e simulations. The rest of the chapter is or g anized as follo ws. W e start with Section3.2surv e ying the related w ork, then we formulate the resource-a w are mo v ement problem in Section3.3. In Section3.4, we focus on the W aterhunter Mo v ement problem and propose an ef cient heuristic solution. Section3.5e v aluates the performance of the proposed schemes. Finally this chapter is concluded in Section3.6. 3.2 Related W ork In this section, we brief some of related w ork that are closely related to this research. Grossglauser and Tse sho wed that node mobility can be used to impro v e netw ork throughput [62]. After predicting the destination' s location, a node forw ards a data pack et to a subset of its neighboring nodes in the direction of the destination to reduce the o v erall routing ener gy consumption [63]. Similarly mo v ement information w as also used in [64] to limit the ooding in a restricted area to reduce the ener gy consumption of route disco v ery .

PAGE 63

51 Chakraborty et al. proposed a strate gy to reduce the ener gy consumption by delaying communication until a mobile node mo v es close to its peer tar get, within an applicationimposed deadline [65]. Dif ferent from the abo v e w ork trying to utilize e xisting mobility information, there are some w ork discussing ho w to mo v e proacti v ely or control node mobility to impro v e the system performance. Li and Rus proposed an optimal algorithm to computer the trajectory of nodes for minimizing message transmission delay [66]. Goldenber g at al. proposed a distrib uted mobility-control scheme to guide the nodes to adjust their mo v ements with the purpose of attaining a potential ener gy-minimizing netw ork conguration [67]. In sparse ad hoc netw ork, proacti v e node mobility can as well be used to o v ercome netw ork partitions. More specically nodes should b uf fer and carry pack ets during netw ork partitions, and forw ard pack ets to other nodes when the y meet. Such a store-carry-forw ard paradigm w as proposed to help data deli v ery by making use of node mobility [68]. In addition, Zhao et al. [69] proposed a message ferrying (MF) approach to address the similar netw ork partition problem. In their approach, a set of special mobile nodes, called mobile ferries, pro vide data forw arding service to other nodes by either mo ving along the routes kno wn a priori to other nodes or proacti v ely mo ving to meet other nodes. Our resource-a w are mo v ement (RAM) strate gy proposed in this chapter is also a proacti v e mo v ement approach. It dif fers from all the pre vious w ork in its unique designing objecti v e, that is, nding the optimal mo v ement trails for each indi vidual node to minimize the total ener gy consumption of the whole system by taking into consideration the inherent resource heterogeneity of MANETs. 3.3 Problem F ormulation W e consider a MANET that consists of tens or e v en hundreds of mobile nodes, among which there are N r re gular battery-po wered nodes (called B-nodes hereafter) and N p po werful nodes (called P-nodes hereafter) ha ving almost unlimited ener gy supplies such as solar cells. Communication de vices installed on a mobile v ehicle and po wered by inside

PAGE 64

52 alternators are other e xamples of such P-nodes. Usually N p is much smaller than N r W e assume that all the nodes are able to generate traf c or forw ard pack ets for others no matter when the y are at rest or in motion. Intuiti v ely since P-nodes ha v e relati v ely innite ener gy reserv oir as opposed to battery-po wered B-nodes, the y should be utilized as much as possible to sa v e the scarce resources of B-nodes and thus prolong the whole netw ork lifetime. F or e xample, a pack et should be forw arded to a P-node whene v er possible if ener gy sa vings can be e xpected. On the other hand, we should reduce the use of B-nodes if we cannot completely a v oid using them. Ho w to realize this simple rationale, ho we v er is by no means a easy task. In this dissertation, we intend to address the issue of ener gy conserv ation from the vie wpoint of node mo v ement. In what follo ws, we rst present a general mobility model that is used to characterize nodes' mo v ement patterns. W e then introduce the resourcea w are mo v ement problem in its general form with the consumption of ener gy resources as the sole optimization objecti v e. 3.3.1 General Mobility Model (GMM) During an observ ation period T we assume that there are some designated locations that an y node i be it a P-node or B-node, should stop by at some designated time instances. F or e xample, a student carrying a mobile de vice may appear in the classroom during schooltime while in the cafeteria during lunchtime. Let J i denote the size of the ordered list of locations node i should visit during T which might be dif ferent for each indi vidual node. W e denote by pos i ( j ) ( 0 j < J i ) the j th location that node i should stop by and by t i ( j ) the required time instance. Then pos i (0) denotes the starting point of node i and pos i ( J i )Tj/T1_4 11.95509 Tf10.56709 0 Td(1) denotes the location of its last stop during T W e will also call as an epoc h [70] the time duration from one node lea ving the current stop until it reaches the ne xt stop henceforth. Whene v er arri ving at some designated location at the specied instance, each node is assumed to pause for a while according to concrete application requirements. Let pause i ( j ) indicate the time node i spends at pos i ( j ) .

PAGE 65

53 Based on the abo v e denitions, the order list f pos i ( j ) ; t i ( j ) ; pause i ( j ) ; 0 j < J i g can well characterize the itinerary of node i during the observ ation period T Let z }| { pos i ( j ) pos i ( j + 1) denote the path tra v elled by node i from its j th stop to the ( j + 1 )th stop. Pro vide that each node tra v els at a constant speed between tw o consecuti v e stops, the tra v elling speed of node i from pos ( i; j ) to pos ( i; j + 1) is l eng th ( z }| { pos i ( j ) pos i ( j + 1) ) t i ( j +1) )Tj/T1_3 7.97011 Tf6.5833 0 Td(t i ( j ) )Tj/T1_3 7.97011 Tf6.5833 0 Td(pause i ( j ) Notice that node i can follo w potentially man y dif ferent paths, for e xample, a straight one or a zigzag one or e v en a tortuous one, as long as the time constraint is satised, that is, it can reach pos i ( j + 1) at the time instance t i ( j ) Ho we v er once the path between tw o consecuti v e stops is determined, the v elocity of the node between these tw o stops is determined and x ed. Therefore, once the paths between all pairs of consecuti v e stops are determined, the mo v ement pattern of a node during T is also determined. The general mobility model (GMM) described abo v e bears both similarities and differences with the random w aypoint model (R WM), which is the most commonly-used mobility model in simulating MANET protocols [58,70]. Both models are characterized by a collection of locations of ne xt stops, tra v elling speeds, and tra v elling time. Dif ferent from GMM, R WM requires a node to rst choose the location of its ne xt stop and the tra v elling speed, which leads to the determination of the tra v elling time. Our GMM actually does the opposite by rst determining the ne xt visited location and the tra v elling time so as to determine the tra v elling speed. The biggest dif ference, ho we v er is that, in R WM nodes al w ays tra v el along the straight paths connecting tw o consecuti v e stops, while in GMM, nodes can tra v el along arbitrary paths as long as the y do not violate the time requirements, that is, the y should arri v e at the designated stops at the specied time instances. W e belie v e that our GMM outperforms R WM in reecting some practical scenarios, which can be seen from a simple e xample in daily life. Suppose a student carrying a mobile de vice should be in the classroom at x ed time e v eryday He/she might ha v e se v eral options of arri ving in time at the classroom from his/her apartment: by foot through the shortest path with the longest time, by bic ycle through the second shortest path with the second

PAGE 66

54 longest time, or by b us through the longest path b ut with the shortest time. Ob viously R WM f ails to model this case, b ut our GMM can. 3.3.2 Resource A w are Mo v ement As mentioned before, in the general mobility model there might be potentially man y dif ferent paths between an y tw o consecuti v e stops. Dene P AT H ( pos i (0) ; pos i ( J 1 )Tj/T1_2 11.95509 Tf12.1512 0 Td(1)) as node i s path set which is the concatenation of all the paths z }| { pos i ( j ) pos i ( j + 1) In this dissertation, we are interested in nding the optimal path sets for all the nodes such that the total ener gy consumption for communications by all the nodes during the observ ation period T is minimized (the objecti v e function), while all the ordered lists of visited locations and the corresponding time instances should not be violated (the constraints). T o help better understand the importance of this problem, we utilize the mo v ement of a single node between tw o consecuti v e stops as an e xample. As sho wn in Fig.3, suppose an B-node A should mo v e from the current location 1 to the ne xt location 2. It can choose the shortest straight path (the dashed one) as it does in the random w aypoint model. Ho we v er considering that node A may forw ard or generate pack ets destined for other nodes during the mo v ement process, the shortest straight path is not necessarily the best one for achie ving the system-wide ener gy ef cienc y Instead, the dotted and solid paths are much better candidates through which node A can tak e adv antage of more P-nodes by forw arding to the encountered P-nodes the pack ets destined for other nodes and letting them nish the rest of the task. Due to this reason, we call this problem the r esour ce-awar e mo vement problem in that nodes no w are mo ving with the system-wide resource (ener gy) consumption in mind instead of mo ving blindly as before. The general resource-a w are mo v ement problem itself is f ar too complicated to be solv able. T o render it tractable, we mak e some approximation and decouple it into tw o relati v ely simpler subproblems: the W aterhunter Mo v ement problem and the Firehunter Mo v ement problem. In the former we assume that only B-nodes are capable of mo ving

PAGE 67

55 D E F C A A B G T r a n s m i s s i o n R a n g e P n o d e R n o d e H I 2 1Figure 3: Multiple paths between tw o consecuti v e stops. and all the P-nodes are stationary whose locations are kno wn a priori to B-nodes. By contrast, in the latter we assume that all the B-nodes are stationary and only P-nodes are able to mo v e. W aterhunter Mo v ement : In the netw ork with N p stationary P-nodes and N r mobile B-nodes, gi v en all the order lists of f pos i ( j ) ; t i ( j ) ; pause i ( j ) ; 0 j < J i g the W aterhunter Mo v ement Problem1is to determine the optimal tra v elling path set for each B-node such that the total ener gy consumption of the whole netw ork during T is minimized. Fir ehunter Mo v ement : In the netw ork with N p mobile P-nodes and N r stationary B-nodes, gi v en all the order lists of f pos i ( j ) ; t i ( j ) ; pause i ( j ) ; 0 j < J i g the Firehunter Mo v ement Problem2is to determine the optimal tra v elling path set for each P-node such that the total ener gy consumption of the whole netw ork during T is minimized. 1 If we compare ener gy resources to w ater, the mo v ement of B-nodes is similar to the beha vior of w ater -hunters who are al w ays looking for fountains (P-nodes), hence the name. 2 If we compare ener gy resources to w ater, the mo v ement of P-nodes is similar to the beha vior of re-hunters who are al w ays looking for places on re or lack of w ater (B-nodes), hence the name.

PAGE 68

56 The W aterhunter Mo v ement problem is similar to the determination of people' s optimal tra v elling plans when lots of airline hubs and stops are a v ailable. On the contrary the Firehunter Mo v ement problem bears similarity with the planning of hub locations when airline companies b uild their global transportation netw orks with the purpose of pro viding more con v enience for passengers while reducing the o v erall system cost. Both problems are pretty interesting and w orth y of rigorous study Ho we v er due to the space limitation, we focus on nding a nearly optimal solution to the W aterhunter Mo v ement problem in this dissertation. Our in v estig ation on the Firehunter Mo v ement problem and the general Resource-A w are Mo v ement problem will be the future w ork. 3.4 W aterhunter Mo v ement In this section, we rst present a simplied v ersion of the W aterhunter Mo v ement problem, which is further reduced to a NP-complete distance-constrained least-cost (DCLC) problem. W e then present an ef cient heuristic solution. Finally we propose a routing delay dif ferentiation mechanism to utilize the benets resulting from the resource-a w are mo v ement. 3.4.1 Simplied W aterhunter Mo v ement Problem In the W aterhunter Mo v ement problem, we assume that all the N p nodes are stationary during the observ ation period T and are willing to forw ard pack ets for other less po werful B-nodes. F or simplicity we do not dwell on ho w to place P-nodes to attain the optimal system performance, which is belie v ed to be a challenging problem itself and is currently under in v estig ation. Instead, we assume that each B-node kno ws the locations of all the P-nodes and its o wn location at an y time, and can as well adjust its mo ving direction at will. F or the time being, we assume here that all the P-nodes and B-nodes ha v e the same transmission range T R W e will discuss the case that P-nodes ha v e the lar ger transmission range than B-nodes in Section3.4.4. It is w orth pointing out that the ndings in this dissertation can be easily e xtended to the case that each node has indi vidual transmission range.

PAGE 69

57 The original W aterhunter Mo v ement problem aims at minimizing the total ener gy consumption of the whole netw ork during the observ ation period T which is a global optimization problem and still too hard to be solv able. T o mak e it tractable, we ha v e to mak e some approximations to get a suboptimal solution. W e assume that a B-node mo v es at a constant speed in one epoch, that is, between tw o consecuti v e stops, and the maximum speed it can tak e is a system-wide v alue speed max Therefore, the longest path node i can tra v el in one epoch is l max = speed max ( t i ( j + 1) )Tj/T1_1 11.95509 Tf12.0791 0 Td(t i ( j ) )Tj/T1_1 11.95509 Tf12.0791 0 Td(pause i ( j )) T o achie v e the system-wide goal of ener gy conserv ation, instead of mo ving along the straight path connecting tw o consecuti v e stops pos i ( j ) and pos i ( j + 1) node i may tra v el along a resource-a w are path with the purpose of letting the encountered P-nodes forw ard on behalf of it as man y as possible pack ets destined for other nodes. F or simplicity we assume that when mo ving to w ards a P-node, node i al w ays goes along the straight path connecting the destined P-node and itself. Therefore, z }| { pos i ( j ) pos i ( j + 1) is a set of zigzag straight paths if there e xist multiple P -nodes. Notice that the simplied tar get no w is to nd an optimal path set for each indi vidual node to minimize its total ener gy consumption during the observ ation period T instead of that of the whole netw ork. W e intend to utilize the solutions to this local optimization problem to approximate the ones to the original global optimization problem, which is belie v ed to be too complicated to be tractable. Normally when mo ving between tw o consecuti v e stops pos i ( j ) and pos i ( j + 1) node i may ha v e se v eral potential P-nodes to utilize. It is, ho we v er usually unwise for node i to pass by each of them. The reason is that, the longer path node i tak es, the f aster speed it should mo v e at, as described in the aforementioned general mobility model. It is well-kno wn that a f aster mo v ement speed may cause some undesirable problems such as the instability of routing paths and the drop of pack ets. Therefore, some rules should be designed to guide each B-node in deciding which P-nodes and in what order it should pass through between tw o consecuti v e stops. A simple rule w ould be to only consider as candidates the P-nodes whose distance from the direct link between an y tw o consecuti v e

PAGE 70

58 B E d H s A r Figure 3: An e x emplary complete graph. steps are no more than a threshold F or e xample, can equal 1 : 5 T R where T R is the transmission range of each node. W ith node i as an e xample, we can put the simplied W aterhunter Mo v ement problem in another w ay: gi v en a source ( i s current stop), a destination ( i s ne xt stop), and some a v ailable intermediate P-nodes, and the path length constraint l max nd a path as ener gyef cient as possible from the source to the destination, which is might be either the direct link connecting the source and destination or a zigzag path through multiple P-nodes. Fig.3depicts such a topology where a rectangular area, called the -bounded r ectangular ar ea hereafter is formed such that only the P-nodes residing in this area are considered as v alid candidates. In addition, each link is of the forw ard direction from the source s to the destination d simply because tra v elling backw ard is ener gy inef cient. W e assign to each link tw o weights, of which one represents the ph ysical distance between tw o link ends and the other indicates the virtual ener gy cost (dened shortly) incurred by choosing this link. The simplied W aterhunter Mo v ement problem can be boiled do wn to a distanceconstrained least-cost (DCLC) [71] routing problem which is formally dened as follo ws. Consider a directed netw ork that can be modelled as a complete graph G = ( V ; E ) where V is the set of v ertices consisting of the source, the destination, and all the v alid candidate P-nodes, and E is the set of edges connecting each pair of nodes. V can be further di vided into tw o subsets, namely U including the source s and destination d and P containing all the P-nodes. In addition, each edge e 2 E represents the mo v ement from the

PAGE 71

59 D E F C A A B G T r a n s m i s s i o n R a n g e P n o d e R n o d e H IFigure 3: An e x emplary resource-a w are mo v ement. tail node to the head node. Let R + denote the set of non-ne g ati v e real numbers. Each edge e 2 E is associated with tw o non-ne g ati v e functions: a distance function dist ( e ) : E R + representing the ph ysical distance between the end nodes of e and an ener gy cost function cost ( e ) : E R + S f 0 g More specically for a gi v en edge e ( v i ; v j ) its ener gy cost is dened as cos t ( e ( v i ; v j )) = 8 > > > > < > > > > : f ( dist ( v i ; v j )) + g ( dist ( v i ; v j )) v i = s; v j = d f ( dist ( v i ; v j )) + g ( dist ( v i ; v j ) )Tj/T1_10 10.9091 Tf10.89729 0 Td(T R ) v i = s; v j 2 P or v j = d; v i 2 P f ( dist ( v i ; v j )) + g ( dist ( v i ; v j ) )Tj/T1_9 10.9091 Tf10.89729 0 Td(2 T R ) v i ; v j 2 P : Here f is the cost, such as g as, fuel or other types of resources, required for the mechanical mo v ement3; g is used to reect the cost for communications. g can be an y non-decreasing function that con v erts a gi v en distance v alue into a non-ne g ati v e cost, for e xample, g ( x ) = 8 > < > : x x > 0 0 x 0 : 3 In this chapter for simplicity we assume that people on foot carry the communities de vices and we do not tak e the cost for the mechanical mo v ement into account, that is, f ( ) = 0 .

PAGE 72

60 The moti v ation for the abo v e denition of the edge ener gy cost is as follo ws. Whene v er a B-node mo v es into the transmission range TR of a P-node, it is capable of forw arding to the P-node pack ets destined for other nodes so as to conserv e ener gy At one e xtreme, if a B-node mo v es along an edge not (partially) co v ered by an y P-node, all the pack ets from this B-node w ould be forw arded to other ener gy-constrained B-nodes, which is the most unf a v orable situation. At the other e xtreme, if a B-node mo v es along an edge completely co v ered by one or se v eral P-nodes, all the pack ets from this node could be forw arded to the P-node(s), which is the most desirable situation. Notice that the ener gy cost function gi v en abo v e can well capture this ef fect. Though there might e xist other meaningful metrics, we belie v e the chosen one is v ery simple and tractable. W e also dene the non-ne g ati v e delay and cost functions for an y path p as dist ( p ) = X e 2 p dist ( e ) and cost ( p ) = X e 2 p cost ( e ) : Gi v en the abo v e denitions, the DCLC routing problem is to nd a path p from s to d such that min f cost ( p ) ; p 2 P d g is achie v ed, where P d is the set of all feasible paths from s to d satisfying the distance constraint l max that is, dist ( p ) l max Moreo v er we dene P l d ( s; d ) as the path with the least distance from s to d and P l c ( s; d ) as the path with the least cost from s to d Apparently with the abo v e denition of dist ( e ) P l d ( s; d ) is the straight path directly connecting s and d It has been sho wn in [72] that the DCLC routing problem is NP-complete e v en for undirected netw orks. In the follo wing section we will propose an ef cient heuristic algorithm to pro vide a suboptimal solution to this DCLC problem and hence to the original W aterhunter Mo v ement problem.

PAGE 73

61 T able 3: RAM: resource a w are mo v ement 1. Determine the candidate P-node set C H S in the bounded r ectangular ar ea ; 2. Construct a complete graph G with virtual nodes s and d and all the nodes in C H S ; 3. Label each link e in G with cost ( e ) and dist ( e ) ; 4. Find the DCLC path P dcl c by calling RAM-DCLC( G s t l max ); 5. Determine the tra v elling speed along P dcl c as speed dcl c = dist ( P dclc ) t ( i;j +1) )Tj/T1_4 7.97011 Tf6.5833 0 Td(t ( i;j ) )Tj/T1_4 7.97011 Tf6.5833 0 Td(pause ( i;j ) ; 6. Mo v e along P dcl c at a speed of speed dcl c ; 3.4.2 RAM-DCLC Algorithm As mentioned before, we assume that a B-node, say i is a w are of its o wn itinerary f pos i ( j ) ; t i ( j ) ; pause i ( j ) ; 0 j < J i g and the locations of all the P-nodes during the observ ation period T The procedure of node i s resource-a w are mo v ement from the cur rent location pos i ( j ) to the ne xt location pos i ( j ) is summarized in T able3. Node i rst needs to determine the candidate P-nodes in the -bounded r ectangular ar ea and then constructs a complete graph lik e the one in Fig.3, consisting of the v ertices s (a virtual node at pos i ( j ) ), d (a virtual node at pos i ( j + 1) ), and all the found candidate P-nodes. It then proceeds to calculate the distance and the ener gy cost for each link and nally generates the weighted graph G The ne xt step is to call the process RAM-DCLC gi v en in T able3to get the DCLC path P dcl c whose length is bounded by l max = speed max ( t i ( j + 1) )Tj/T1_1 11.95509 Tf12.0611 0 Td(t i ( j ) )Tj/T1_1 11.95509 Tf12.05209 0 Td(pause i ( j )) It then mo v es to w ards pos i ( j + 1) at a constant speed of speed dcl c = dist ( P dclc ) t ( i;j +1) )Tj/T1_4 7.97011 Tf6.5833 0 Td(t ( i;j ) )Tj/T1_4 7.97011 Tf6.5833 0 Td(pause ( i;j ) along the found P dcl c Upon reaching pos i ( j + 1) node i pauses for a period pause i ( j + 1) F ollo wing the pre vious process, it can then mo v e to w ards the ne xt stop pause i ( j + 2) until all the required stops f pos i ( j ) ; 0 j < J i g during the observ ation period T are visited.

PAGE 74

62 T able 3: RAM-DCLC: a DCLC routing algorithm for the W aterhunter Mo v ement 1. F or each node v k in G nd the P l c ( v k ; d ) and P l d ( v k ; d ) and their respecti v e ne xt hops nid ( P l c ( v k ; d )) and nid ( P l d ( v k ; d )) ; 2. distS oF ar = 0 ; P dcl c = s ; T hisN ode = s ; 3. while ( T hisN ode 6= d ) do 4. if ( ( dist ( P l c ( T hisN ode; d )) + distS oF ar ) l max ) then 5. v = nid ( P l c ( T hisN ode; d )) ; 6. distS oF ar = distS oF ar + dist ( T hisN ode; v ) ; 7. P dcl c = P dcl c + f v g ; 8. T hisN ode = v ; 9. else 10. f or each neighboring node w = 2 P dcl c do 11. calculate w eig ht ( T hisN ode; w ) ; 12. end f or 13. v = extr act ( T hisN ode ) ; 14. distS oF ar = distS oF ar + dist ( T hisN ode; v ) ; 15. P dcl c = P dcl c + f v g ; 16. T hisN ode = v ; cp:ref35 17. end if 18. end while 19. Retur n P dcl c

PAGE 75

63 The proposed DCLC algorithm RAM-DCLC is summarized in T able3, in which the w eig ht () function is dened as follo ws: w eig ht ( v i ; v j ) = 8 > < > : cost ( v i ; v j ) + cost 0 ( v j ; d ) cond (1) + 1 o:w : ; and cost 0 ( v j ; d ) = 8 > < > : cost ( P l c ( v j ; d )) cond (2) cost ( P l d ( v j ; d )) o:w : ; where cond (1) = distS oF ar + dist ( v i ; v j ) + dist ( P l d ( v j ; d )) l max and cond (2) = distS oF ar + dist ( v i ; v j ) + dist ( P l c ( v j ; d )) l max : The function extr act () is used to choose the node, say w whose w eig ht ( v i ; w ) is the minimum one among all the neighboring nodes of T hisN ode If more than one node ha v e the same minimum v alue, it chooses the one with the smallest distS oF ar + dist ( v i ; v j ) + dist ( P l d ( v j ; d )) In the RAM-DCLC algorithm, for each node v k in G the Bellman-F ord or Dijkstra shortest-path algorithm can be used to nd the P l c ( v k ; d ) and P l d ( v k ; d ) and their respecti v e ne xt hops nid ( P l c ( v k ; d )) and nid ( P l d ( v k ; d )) Since the optimization objecti v e is the path cost, at each intermediate node v RAM-DCLC al w ays chooses the ne xt hop w with minimum cost ( v ; w ) + cost 0 ( w ; d ) while not violating the distance constraint l max RAM-DCLC is able to nd a feasible path satisfying the distance constant while k eeping the path cost as small as possible. In particular we ha v e the follo wing theorems for this algorithm4. Theor em 1 : RAM-DCLC can al w ays nd a feasible path from a source s to a destination d satisfying the gi v en distance constraint l max if such feasible paths e xist. 4 The correctness of these theorems can be justied follo wing the proof in [73].

PAGE 76

640 2 0 0 4 0 0 6 0 0 8 0 0 1 0 0 0 1 2 0 0 1 4 0 0 1 6 0 0 0 5 0 1 0 0 1 5 0 2 0 0 2 5 0 3 0 0 1 3 1 2 1 1 1 0 9 8 7 6 5 4 3 2 1 ( S t o p A r r i v a l t i m e ) ( S5, 8 5 8 4 ) ( S4, 6 2 5 8 ) ( S3, 3 8 7 4 ) ( S2, 2 4 7 8 ) ( S1, 0 ) R A M R W M P n o d e s S t o p s Y ( m )X ( m )Figure 3: Resource-a w are mo v ement (RAM) vs. random w aypoint mo v ement. A Bnode should consecuti v ely visit S 1 S 2 S 3 S 4 S 5 and the solid lines are labelled by the sequences the y were passed through. Theor em 2 : The path found by RAM-DCLC is loop-free. Theor em 3 : RAM-DCLC al w ays terminates in nite time. No w we utilize the e xample gi v en in Fig.3to illustrate the resource-a w are mo v ement process using the proposed R AM )Tj/T1_2 11.95509 Tf12.2141 0 Td(D C LC algorithm. Suppose node A intends to mo v e from its current location to the location where A resides according to its itinerary In this e xample is set to 1 : 5 T R so that there are four candidate P-nodes. Based on the output of R AM )Tj/T1_2 11.95509 Tf12.5741 0 Td(D C LC node A should mo v e along the DCLC path denoted by the solid line instead of the straight path denoted by the dash line. In this w ay ener gy sa vings can be e xpected by forw arding to the tw o P-nodes the pack ets it carries for other nodes and letting the P-nodes nish the rest transmissions (either single-hop or multi-hop) on behalf it. Fig.3compares a B-node' s resource-a w are mo v ement trail and its random w aypoint mo v ement trail with 8 P-nodes in a 1500 300 m 2 eld. The data used were generated using OPNET [57] and the pause time of the B-node w as 120s. In addition, the transmission range of each node w as set to 250m. It is clear that our proposed resourcea w are mo v ement strate gy enables the B-node to ha v e more opportunities of approaching and utilizing the P-nodes as compared to the random w aypoint mo v ement.

PAGE 77

65 3.4.3 A Routing Delay Dif ferentiation Mechanism The resource-a w are mo v ement strate gy only enables B-nodes to mo v e close to Pnodes by taking the ph ysical DCLC paths, b ut cannot guarantee a pack et will be forw arded from a B-node to a P-node e v en when the y are close to each other As a result, we need to incorporate such nodal information as de vice types and residual ener gy into the routing metric such that the P-nodes can be naturally selected onto the routing paths of the Bnodes. F or this purpose, we propose to map the residual ener gy information into the random jitter delay for which each node has to w ait before propag ating a route request in a re gular MANET routing protocol such as DSR [70]. The rational here is that the less residual ener gy a node has, the longer delay a routing request pack et should e xperience at this node. In this w ay the nodes with less residual ener gy w ould be less utilized, while the nodes with more residual ener gy such as P-nodes w ould be more utilized. Belo w we illustrate the delay dif ferentiation mechanism using DSR [70] as the under lying routing protocol, though our mechanism can as well be applied to other MANET routing protocols. Similar to the interframe spacing (IFS) dif ferentiation mechanism adopted in the IEEE 802.11e [74], we rst dene the Delay Dif fer entiation as E 0 =E t where is a delay control parameter and E 0 and E t are the residual ener gy at time instance 0 and t respecti v ely When recei ving a route request, each node should rst compute the delay it needs to w ait for before further propag ating the route request. Since the P-nodes ha v e relati v ely unlimited residual ener gy as compared to the B-nodes, that is, E 0 E t the y w ould forw ard route requests much more quickly than the B-nodes. Since the destination usually only replies to the rst arri v ed request, the P-nodes with po werful ener gy resources w ould ha v e higher chances of being chosen onto the routing paths and thus being in v olv ed in more data forw arding acti vities than the B-nodes with less ener gy W ith this delay differentiation mechanism, the benets resulting from the resource-a w are mo v ement can be better utilized. The nice feature of our mechanism is that it is v ery easy to implement.

PAGE 78

66 3.4.4 Incorporate RAM into DELAR In our pre vious discussion, we assume that the P-nodes ha v e the same transmission range T R as the B-nodes. Ho we v er in more practical heterogeneous MANETs, P-nodes may ha v e greater transmission capabilities than the B-nodes. T o mak e use of such po werful P-nodes especially when P-nodes can adjust their transmission po wer to co v er lar ger range than B-nodes, in our pre vious w ork DELAR, we proposed a ener gy ef cient relaying frame w ork as a joint design of scheduling, routing and po wer control to ef ciently utilize P-nodes to conserv e ener gy Since RAM and DELAR utilize po werful nodes to conserv e ener gy from totally orthogonal perspecti v es, RAM can be directly incorporated into the DELAR scheme. W ith the help of DELAR, RAM can further impro v es the ener gy ef cienc y when P-nodes and B-nodes ha v e dif ferent transmission capabilities. 3.5 Performance Ev aluation In this section, we use simulations to e v aluate the impact of the proposed schemes on the ener gy conserv ation and other system performance f actors. 3.5.1 Simulation Setup W e implemented the resource-a w are mobility model, the routing delay dif ferentiated mechanism, and DELAR (cf. Chapter2) in OPNET [57]. W e simulated a netw ork with N r B-nodes and N p P-nodes in a 1500 300 eld, where N r = 46 and N p = 4 All the B-nodes were capable of mo ving in the eld, while all the P-nodes were x ed. Though a careful deplo yment of P-nodes may impro v e the system performance [75], we simulated a w orse scenario that the P-nodes were randomly deplo yed in the eld. The transmission range of the B-nodes w as 250 m. F or the P-nodes, we simulated tw o cases in which the P-nodes had the transmission ranges 250 m and 500 m, respecti v ely The ener gy consumption for the B-nodes follo wed the linear ener gy model proposed in [76]: ener g y = m l eng th + b where m is an incremental cost of each operation, b is the

PAGE 79

67 T able 3: Ener gy consumption parameters. Symbol V alue Unit m send 1.89 uW -sec/byte b send 246 uW -sec m r ecv 0.49 uW -sec/byte b r ecv 56.1 uW -sec b sendctl 120 uW -sec b r ecv ctl 29 uW -sec x ed cost of each operation, and l eng th is the size of the frame sent/recei v ed5. The ( m; b ) v alues pro vided in [76] were summarized in T able3and used in all the calculations. W e intended to compare the proposed resource-a w are mobility model (denoted by RAM) with the modied random w aypoint model (denoted by R WM) presented in [58], which can guarantee the con v er gence of a v erage nodal speed throughout the simulation time. F or this purpose, we rst ran the simulations using R WM and recorded the stops, the starting/arri v al time instances, the mo ving directions, and the mo v ement speeds of all the mo v ement epochs. W e then used this mo v ement prole to generate the itinerary f pos i ( j ) ; t i ( j ) ; pause i ( j ) ; 0 j < J i g for each node such that in both models each node w ould stop by the same stops at the same time instances, b ut follo w totally dif ferent mo v ement trails and tak e dif ferent mo v ement speeds. Both models had the same maximum speed 20 m/s and we adjusted nodal pause time to v ary the netw ork mobility The traf c used were 20 CBR connections with randomly selected source-destination pairs. All the data pack ets were 64 bytes and were sent a speed of 4 pack ets/second. Each simulation w as e x ecuted for 15 simulated minutes and each data point represents an a v erage of ten runs with identical traf c models, b ut dif ferently generated mobility scenarios. 5 F or the A-MA C control frames P-R TS/P-CTS/P-A CK, the x ed costs b sendctl and b r ecv ctl were used because the y ha v e the similar size.

PAGE 80

68 3.5.2 Simulation Results W e compared RAM with R WM in terms of the commonly used metrics including pack et deli v ery ratio, a v erage pack et end-to-end delay a v erage pack et ener gy consumption, and a v erage routing o v erhead. Moti v ated by the small-w orld phenomenon [77], we used tw o additional metrics, a v erage path length and a v erage clustering coef cient, which are tw o dening characteristics of small-w orld netw orks. The former means the a v erage number of hops a pack et may tra v el through, while the latter indicates the connecti vity of an a v erage neighborhood in the netw ork, dened as the a v erage node de gree di vided by the netw ork size [77]. The simulation results are presented in Fig.3, where RAM-1 indicates the case that the P-nodes and B-nodes ha v e the same transmission range 250 m and RAM-2 denotes the case that the P-nodes ha v e a lar ger transmission range 500 m. Fig.3(a)and3(b)compare RAM and R WM with re g ard to a v erage path length and a v erage clustering coef cient, respecti v ely W e can see that RAM can shorten the a v erage path length and increase the clustering coef cient as compared to R WM. That is because in RAM the B-nodes are al w ays trying to mo v e to w ards some P-nodes between tw o consecuti v e stops. Such beha viors w ould ef fecti v ely bring more B-nodes to the vicinity of P-nodes, leading to shorter paths and lar ger clustering coef cients. In some sense, such resource-a w are mo v ement creates a small-w orld netw ork, which results in some performance g ains sho wn belo w In addition, since the P-nodes in RAM-2 ha v e a lar ger transmission range and thus ha v e more neighbors that those in RAM-1, we can observ e that RAM-2 further reduces the a v erage path length and increases the a v erage clustering coef cient. Fig.3(c)compares the a v erage pack et deli v ery ratios of RAM and R WM, which is dened as the ratio of deli v ered data pack ets to those generated by the sources. As we can see, the PDR of RAM-1 or RAM-2 is al w ays higher than that of R WM. This result is of no surprise since the shorter a v erage path length implies the netw ork-wide less traf c load and less pack et drops due to the MA C-layer contention and channel errors, all of which

PAGE 81

69 0 1 0 0 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 8 0 0 9 0 0 1 8 2 2 2 2 4 2 6 2 8 3 3 2 3 4 3 6 P a u s e t i m e ( s )A v e r a g e p a t h L e n g t h R W M R A M 1 R A M 2 (a) A v erage path length. 0 1 0 0 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 8 0 0 9 0 0 0 2 6 0 2 8 0 3 0 3 2 0 3 4 0 3 6 0 3 8 0 4 0 4 2 0 4 4 P a u s e t i m e ( s )C l u s t e r i n g c o e f f i c i e n t ( m / s ) R W M R A M 1 R A M 2 (b) A v erage clustering coef cient. 0 1 0 0 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 8 0 0 9 0 0 0 9 5 0 9 5 5 0 9 6 0 9 6 5 0 9 7 0 9 7 5 0 9 8 0 9 8 5 0 9 9 0 9 9 5 1 P a u s e t i m e ( s )P a c k e t d e l i v e r y r a t i o R W M R A M 1 R A M 2 (c) P ack et deli v ery ratio. 0 1 0 0 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 8 0 0 9 0 0 0 0 0 2 0 0 0 4 0 0 0 6 0 0 0 8 0 0 1 0 0 1 2 0 0 1 4 0 0 1 6 0 0 1 8 P a u s e t i m e ( s )P a c k e t e n d t o e n d d e l a y ( s ) R W M R A M 1 R A M 2 (d) A v erage pack et end-to-end delay 0 1 0 0 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 8 0 0 9 0 0 1 1 5 2 2 5 3 3 5 P a u s e t i m e ( s )A v e r a g e e n e r g y c o n s u m p t i o n ( m W ) R W M R A M 1 R A M 2 (e) A v erage ener gy consumption. 0 1 0 0 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 8 0 0 9 0 0 0 5 1 0 1 5 P a u s e t i m e ( s )A v e r a g e r o u t i n g o v e r h e a d R W M R A M 1 R A M 2 (f) A v erage routing o v erhead. Figure 3: Simulation results.

PAGE 82

70 w ould contrib ute to the increase of the PDR. Due to the same reason, RAM-2 demonstrates a higher PDR than RAM-1. Figure 3-5(d) depictsthecomparisonofaverageend-to-endpacketdelay,denedas the time duration from when a pack et is generated till it is recei v ed by the destination. The sho wn adv antage of RAM o v er R WM mainly results from the aforementioned shorter a v erage path length and higher clustering coef cient. Ag ain, RAM-2 outperforms RAM-1 in reducing a v erage pack et delay because of the shorter a v erage path length. Figure 3-5(e) showsaverageenergyconsumption,denedasthetotalenergyconsumption for transmitting and recei ving all data and routing pack ets di vided by the number of deli v ered pack ets. Apparently our RAM can conserv e a signicant amount of ener gy as compared to R WM because the B-nodes are al w ays mo ving along the paths through which the P-nodes can be utilized as much as possible. Since the P-nodes in RAM-2 ha v e a lar ger transmission range, statically less B-nodes are in v olv ed in pack et transmissions and thus RAM-2 can help the system conserv e more ener gy than RAM-1. Figure 3-5(f) demonstratesaverageroutingoverhead,denedastheaveragenumber of routing pack ets in v olv ed in deli v ering 100 data pack ets. As we can see, our RAM has smaller routing o v erhead than R WM. That is because in RAM pack ets can be forw arded to their destinations through shorter paths in shorter time, thus fe wer routing errors occur T o summarize, the proposed resource-a w are mo v ement strate gy has man y signicant and positi v e impacts on the system performance. It mak es the netw ork more lik e a smallw orld netw ork with shorter a v erage path length and higher clustering coef cient. This results in impro v ed pack et deli v ery ratio, shortened end-to-end delays, and most impor tantly much better ener gy conserv ation. Therefore, the combination of node mobility and heterogeneity is a v alid means to address the ener gy conserv ation issue in MANETs. The results also suggest that making the netw ork a small-w orld netw ork may ha v e a lot of positi v e ef fects on the system performance and it deserv es further in v estig ation.

PAGE 83

71 3.6 Summary In this chapter we studied the ener gy conserv ation problem from the ne w perspecti v e of node mobility that is, by analyzing the impact of node mo v ement on the system-wide ener gy conserv ation. W e proposed a no v el resource-a w are mo v ement strate gy to tak e full adv antage of some po werful nodes in heterogeneous mobile ad hoc netw orks. W e then for mulated the resource-a w are mo v ement as a NP-complete distance-constrained least-cost (DCLC) routing problem and proposed an ef cient heuristic solution. Moreo v er we proposed a simple yet ef fecti v e routing delay dif ferentiation mechanism to virtually utilize the benets from the resource-a w are mo v ement. In addition, the proposed resource a w are mo v ement strate gy can be incorporated into other ener gy conserv ation schemes, for e xample, DELAR, to further impro v e the ener gy ef cienc y W e used e xtensi v e simulations to sho w the ef fecti v eness of the proposed scheme.

PAGE 84

CHAPTER 4 SUPPOR T DIFFERENTIA TED SER VICES IN MOBILE AD HOC NETW ORKS 4.1 Introduction In the literature there are proposals addressing the issue of quality of service in wired netw orks, ho we v er these proposals may not be feasible in the wireless counterpart if we do not modify them for the wireless en vironments. The system dynamics [18] of multi-hop mobile ad hoc netw orks, such as time-v arying and error -prone wireless links, dynamic and limited bandwidth, time-v arying traf c pattern and user location, and ener gy constraints, pose ne w challenges that do not e xist in wired netw orks. T o conquer these challenges, in recent years, man y researchers adv ocate a cross-layer design philosoph y to de v elop protocols and applications for MANETs. This is a departure from the traditional layered design for the Internet. Though the cross-layer design philosoph y might not be an optimal solution, it does pro vide us ne w netw ork implementations that may better support the amalg amation of user services and QoS requirements [19]. T o ef ciently handle heterogeneous traf c o v er wireless links, we need to address tw o problems. The rst is to handle reliable mobile communications in MANETs. This problem has been e xtensi v ely studied in recent years, and man y proposed routing protocols such as DSD V [7], DSR [9], and A OD V [10], and medium access control mechanisms such as MA CA W [3], F AMA [4], and IEEE 802.11 [5], aim to achie v e ef cient reliable communications. The other problem is to pro vide QoS pro visioning for heterogeneous traf c with dif ferent quality-of-service (QoS) requirements in terms of BER, throughput, and delay Since the channel bandwidth in wireless en vironments is limited, one strate gy to support QoS is to set up some kind of priority scheme or service dif ferentiation mechanism [78][79], under which delay-sensiti v e traf c has higher priority to access the channel o v er less time-critical traf c. 72

PAGE 85

73 In the current literature, man y scheduling mechanisms for wireless netw orks are proposed for this purpose, though most of them are not directly designed for MANETs. In general, these scheduling mechanisms all attempt to combat the channel impairments and to support heterogeneous traf c with the follo wing goals: pro viding high wireless channel utilization, long-term f airness, bandwidth guarantees and delay bounds for o ws with error -free links or links with sporadic errors [80]. Ho we v er these algorithms may not be practical to be implemented in MANETs. Actually it is hard, if not impossible, to achie v e those goals simultaneously because of their conicting nature. F or e xample, there is a tradeof f between the throughput and f airness1or so-called inter -class ef fects [81] among traf c with dif ferent priorities. W ithout an y precautionary measures, this conict may lead to bandwidth starv ation for lo w-priority traf c when the high-priority traf c load is high. Meanwhile, most of these scheduling mechanisms are suitable for the reserv ation-based MA C protocols, especially for those designed for cell-structured wireless netw orks. In netw orks with contention-based MA C protocols such as IEEE 802.11 [5], the reserv ationbased scheduling mechanisms may not be applicable because it is not easy for a node to reserv e resource in a contention manner In this dissertation, we attempt to a v oid the con v entional scheduling approach, and propose a no v el scheme called Courtesy Piggybacking (CP) to alle viate the conict between throughput and f airness. The basic idea of CP is to let the high-priority traf c help the lo w-priority traf c by sharing unused residual bandwidth with courtesy Our scheme closely follo ws the cross-layer design principle and e xploits the system dynamics as much 1 W e only consider the f airness problem between dif ferent classes of traf c, for e xample, each class of service should be allocated some bandwidth rather than being completely starv ed; while the f airness problem between dif ferent nodes, for e xample, each node should ha v e f air opportunity to access the channel in the short or long term, is out of the scope of this dissertation.

PAGE 86

74 as possible; that is, we ef fecti v ely emplo y the dynamic channel conditions and the resulting dynamic bandwidth, and the dynamic characteristics of the heterogeneous traf c. Note that not only is our scheme suitable for multi-hop mobile ad hoc netw orks with underlying contention-based MA C protocols, b ut also it is applicable to those with reserv ation-based or h ybrid MA C protocols. Meanwhile, our scheme is sho wn to be easily implemented. The rest of the chapter is or g anized as follo ws. W e discuss some related w ork in Section4.2. In Section4.3, we sho w the moti v ation of our proposed scheme. In Section4.4, we discuss the relationship between the SNR and the optimal pack et length, and come up with a Finite State Mark o v Chain channel model based on the pack et length. Our Courtesy Piggybacking scheme is described in Section4.5. W e present some preliminary analytical results in Section4.6and e v aluate our scheme with e xtensi v e simulation in Section4.7. Finally we conclude the chapter in Section4.8. 4.2 Related W ork As we mentioned in Section4.1, scheduling is one promising w ay to support heterogonous traf c with dif ferent QoS requirements. F or scheduling mechanisms, throughput and f airness are tw o main objecti v es to be met through bandwidth allocation with admission control and congestion control. Man y scheduling algorithms such as f air queuing scheduling [82], and virtual clock [83] are capable of pro viding certain QoS guarantee for wireline netw orks, and man y scheduling algorithms such as IWFQ [84], CIF-Q [85], CSDPS [86], and CSDPS + CBQ [87] are proposed for the wireless netw orks, especially for wireless cellular netw orks. Ho we v er little progress has been made along this direction in wireless mobile ad hoc netw orks with underlying contention-based MA C protocols. CSDPS and its impro v ed v ersion CSDPS+ CBQ are tw o of scheduling mechanisms that may be applicable to the ad hoc netw orks with contention-based MA C protocols. In CSDPS, pack ets to be transmitted to the same recei v er are queued in the same queue and are serv ed in an FIFO f ashion. At a node, the dif ferent queues are serv ed according to some policies such as round robin, earliest timestamp rst, or longest queue rst. The basic idea

PAGE 87

75 of CSDPS is as follo ws: when the link to w ards a recei v er is bad, the node should defer the transmission of pack ets in the queue corresponding to that recei v er W ith CSDPS, it is easy to alle viate the head of line (HOL) problem when single FIFO queue is used. Since CSDPS mak es use of the channel state information, it can achie v e high data throughput and channel utilization. Ho we v er it does not address the f airness issue. T o impro v e the f airness in CSDPS, class-based queuing (CBQ) [88] is used together with the CSDPS. By using CBQ, a hierarchical channel-sharing mechanism, it can achie v e certain f airness, and ensure that dif ferent traf c classes can share the o v erall bandwidth, while maintaining the features of CSDPS to deal with the channel v ariations. Unfortunately this scheme is also complicated in k eeping track of the amount of service each class has been serv ed. Ef cient and less e xpensi v e mechanisms are v ery desirable to alle viate the conict of throughput and f airness in MANETs. More and comprehensi v e materials on scheduling can be found in [80]. Besides, some QoS adapti v e schemes such as SW AN [16] and Ha v ana [89] are also a v ailable in the literature. These schemes adapti v ely perform admission control and rate control according to the user QoS requirements and channel states. The main reason leading to the conict between throughput and f airness is the limited bandwidth of the wireless link. If the system can pro vide plenty of bandwidth, the conict problem w ould not be so signicant. Recently man y adapti v e transmission techniques are proposed to e xploit the channel dynamics to pro vide more bandwidth. These schemes can adapti v ely adjust the parameters such as modulation le v el and symbol rate to maintain an acceptable BER without w asting much bandwidth. K outsopoulous and T assiulas proposed to inte grate adapti v e transmission techniques, resource allocation and po wer control for TDMA/TDD system so that higher modulation le v els can be assigned to users in good channels to enhance the throughput, while po wer control can be used to reduce the interference and increase the system capacity [90]. In addition to these schemes proposed for wireless cellular netw orks, some rate-adapti v e schemes are also proposed to impro v e the system throughput in WLANs. Holland et al. proposed a rate adapti v e MA C protocol

PAGE 88

76 called RB AR, which uses the R TS/CTS to e xchange the channel state information and the optimal rate on a per -pack et basis [91]. Unfortunately this scheme needs to mak e some modications to the IEEE 802.11 MA C protocols. T o a v oid this modication, Che villat et al. proposed a scheme to select the optimal rate only with the local information at the transmitter [92]. This scheme is based on the history of attempted transmissions. It uses one successful transmission count and one f ailed transmission count to indicate the channel state and to determine the optimal rate the transmitter can use. F or IEEE 802.11 MA C protocols, adapti v e fragmentation schemes can also be designed with the rate adaptation to enhance the system throughput [93][94][95]. F or all the scheduling mechanisms and other channel-dependent schemes, including our Courtesy Piggybacking scheme, designed for wireless netw orks, the y all ha v e to monitor the channel quality based on the symbol error rate, bit error rate, and recei v er signal strength. The more accurate the channel information is, the more benets these schemes can bring to the system design. In general, the channel estimation can be performed by the sender or by the recei v er Since the channel information used in all channel-dependent schemes is the one seen by the recei v er the recei v er -based channel estimation is more attracti v e. Ho we v er the channel information needs to be sent back to the sender which is sometimes costly in terms of the resource used to transmit the channel information, thus certain performance tradeof f has to be made between estimation accurac y and o v erhead. More details about channel quality estimation can be found in [96]. 4.3 Moti v ation Consider the scenario depicted in Fig.4. In a mountain area, the only w ay from A nchorage to W hittier (the access to see the spectacular glacier) is to pass a tunnel near P ortage running through the C hug ach Mountain Range (that is, the longest tunnel in North America the Whittier T unnel in Alaska). The situation is the same from S e w ard to W hittier People ha v e se v eral choices to pass the tunnel: by train (high priority), by car by bic ycle or on foot (lo w priority). Only one direction traf c is allo wed during one period

PAGE 89

77 A P W S T u n n e l R a i l R o a d R o a d Figure 4: The Whittier T unnel scenario. of time. T o pass the tunnel, when the train approaches the tunnel, all other traf c stops and w aits until the train passes the tunnel. Often, there is a long traf c line w aiting to pass the tunnel, especially for the direction from W to P when traf c load is high, for e xample, during rush hour in the afternoon. In order to quickly pass the tunnel, a better approach for other transportation users w ould be to check if there is an y free space left in the train. If there is, these users could ask for permission to ride at a certain cost and according to some rules, for e xample, ho w man y free space in terms of basic units is left and what kind of traf c (priority) the train can accommodate. After passing through the tunnel, the piggyback ed traf c can get of f the train at P and continue on its o wn w ay Of course in the real situations, when passengers by car by bic ycle or on foot pass through the narro w and dark tunnel in a sequential manner the traf c usually mo v es v ery slo wly for the sak e of safety Thus it is advisable for cars that ha v e free space to piggyback those passengers by bic ycle or on foot according to some rules to benet all the traf c. W e can think of these rules as being concerned with the HO W MANY -WHO problem, that is, ho w much free space is a v ailable and who can enjo y such free space? If we only consider the free space F S in the train as a function of time, then we could consider the follo wing scenario as an e xample: one person w ould occup y 1 basic space unit, a bik e 2 units, and a car 6 units. If we ha v e some predened objecti v e to meet, then we can design dif ferent piggybacking rules to solv e the HO W MANY -WHO problem. F or e xample,

PAGE 90

78 suppose our objecti v e is to maximize the re v enue of the train. W ith dif ferent piggybacking costs, for a gi v en F S we can achie v e the optimal allocation scheme for the free space among dif ferent traf c: cars, bic ycles, and pedestrians. The abo v e scenario is v ery similar to multi-hop mobile ad hoc netw orks supporting dif ferentiated services. The piggybacking strate gy described abo v e moti v ates us to de v elop a more ef cient w ay to alle viate the conict between throughput and f airness for dif ferent prioritized services. First of all, we need to identify the free space in a MANET F ortunately we do ha v e tw o sources that can pro vide us with such free space. The rst one comes from the time-v arying channel conditions. In recent studies such as [90] [93], the MA C and PHY layers adapt to the channel state by using adapti v e transmission schemes to pro vide higher data rates when the channel is good. W ith a higher data rate, the transmission time for MA C protocol data unit (MPDU) can be shortened, leading to some potential idle time if the transmitting node does not ha v e further data to transmit. If the IEEE 802.11 MA C is used, the N A V (Netw ork Allocation V ector) setting may pre v ent other nodes from using the medium, e v en though it is idle (the rule of virtual collision a v oidance). This idle period will be the free space and should be more ef fecti v ely used. The second source comes from the traf c characteristics. When we look into the traf c patterns and the stochastic traf c beha vior sometimes the high priority traf c may not ha v e enough data during the reserv ed slots in a reserv ation-based system or their transmission period in a contention-based system (for e xample, a netw ork with IEEE 802.11) to fully utilize the channel capacity F or e xample, consider a netw ork with reserv ation-based MA C protocols. In addition to the free space pro vided by the channel dynamics, when the pack ets from one high-priority o w are not enough to ll the reserv ed slots, for e xample, during silent periods for v oice connections, some free space can be harv ested to piggyback some bits from the queue(s) with lo w priorities. When such free space is a v ailable, the ne xt problem w ould be ho w to mak e use of it to fulll certain objecti v es such as f air allocation of bandwidth. While one w ould think

PAGE 91

79 that it should be used to better support high-priority traf c in the rst place, we ar gue that it may not necessarily be the case. Rather the piggybacking rules should be properly designed in light of specic requirements of v arious applications. If some delay-sensiti v e applications lik e v oice or video-telephon y require that their pack ets get through the netw ork as quickly as possible, then the free space should be used to meet such needs. On the other hand, if high-priority traf c does not need more resource than needed, a piggybacking rule f a v oring lo w priority may be more reasonable. F or streaming multimedia applications, as an e xample, when the QoS requirement of one stream with high priority has already been met, there is no need to piggyback pack ets belonging to this stream ahead of the scheduled time; instead, piggybacking pack ets from other lo w-priority streams may be more benecial. In the follo wing sections, we will elaborate more on wh y free space e xists and ho w piggybacking can be used to achie v e our goal alle viating the conict between throughput and f airness for dif ferent prioritized services. 4.4 P ack et-length-based Channel Model In the current literature, the time-v arying channel is commonly modeled as the wellkno wn Gilbert-Elliott tw o-state Mark o v channel model (Fig.4). Each state in the tw ostate Mark o v chain model represents a binary symmetric channel (BSC). In Good state, the BSC has lo w crosso v er probability P g and int the Bad state, the BSC has high crosso v er probability P b The transition probability matrix can be gi v en as: < = 2 6 4 P GG P GB P B G P B B 3 7 5 : Gi v en the transition probability it is easy to determine that the steady state probabilities are = P B G P B G + P GB P GB P B G + P GB :

PAGE 92

80 G o o d B a d P G B P B G P B B P G G 1 P g 1 P g P g P g 0 1 0 1 1 P b 1 P b P b P b 0 1 0 1 G B Figure 4: The Gilbert-Elliott channel model. W e notice that if P g and P b are set to 0 and 1, respecti v ely that is, a pack et succeeds with probability 1 in the Good state and is lost with probability 1 in the Bad state, the tw o-state model is reduced to the simplied Gilbert model. When the channel quality v aries dramatically it is not accurate enough to model the channel as a tw o-state Gilbert-Elliott model. In this case, a nite-state Mark o v channel (FSMC) [97] can be used. By using the recei v ed signal-to-noise-ratio (SNR) as the only side information, the FSMC pro vides a mathematically tractable model for time-v arying channel. Let r denote the recei v ed SNR that is proportional to the square of the signal en v elop. Then, for a Rayleigh f ading channel, the probability density function of r can be written as f r = 1 )Tj/T1_2 11.95509 Tf-0.39601 -7.54201 Td(r e )Tj/T1_6 5.97758 Tf8.51399 3.69 Td(r )Tj/T1_6 5.97758 Tf0.53999 -4.06799 Td(r ; (4.1) where )Tj/T1_2 11.95509 Tf-0.39601 -7.54201 Td(r is the mean of r (actually it is an e xponential distrib ution with mean )Tj/T1_2 11.95509 Tf-0.39601 -7.54201 Td(r ). In order to b uild the nite state Mark o v chain model, we assume the recei v ed SNR remains at a certain le v el for the duration of a symbol, and we partition the range of the recei v ed SNR into a nite number of interv als. Let 0 = r 0 < r 1 < r 2 < :::::: < r K )Tj/T1_8 7.97011 Tf6.5833 0 Td(1 < r K = 1 be the thresholds. F or each interv al, we associate it with a state S k ; k = 0 ; 1 ; 2 ; 3 ; :::; K )Tj/T1_4 11.95509 Tf12.0791 0 Td(1 The channel is in the state S k if r is in the interv al [ r k ; r k +1 ] W e kno w that there is a crosso v er probability p for a gi v en SNR r When BPSK is used, this probability can be written as a function of r : p ( r ) = 1 )Tj/T1_4 11.95509 Tf11.94411 0 Td(b( p 2 r ) ; b( r ) = R r 1 p 2 e )Tj/T1_6 5.97758 Tf7.785 3.26701 Td(t 2 2 dt: (4.2)

PAGE 93

81 3 4 5 6 7 8 9 1 0 0 1 0 0 0 2 0 0 0 3 0 0 0 4 0 0 0 5 0 0 0 6 0 0 0 g ( d B )p a c k e t l e n g t h ( b i t )Figure 4: The optimal pack et length ( P L ) vs. SNR ( r ), h=128. P L 0 P L 1 P L 2 P L k 1 t 0 1 t 1 0 t 1 2 t 2 1 t 2 3 t 3 2 t k 2 k 1 t k 1 k 2 t 0 0 t 1 1 t 2 2 t k 1 k 1 Figure 4: P ack et-Length-Based nite-state Mark o v channel model. According to [98], for a gi v en crosso v er probability p the optimal pack et length, which is a function of p can be written as P L = )Tj/T1_3 11.95509 Tf9.28912 0 Td(h ln(1 )Tj/T1_3 11.95509 Tf11.95309 0 Td(p ) )Tj/T1_10 11.95509 Tf12.0327 14.472 Td(q )Tj/T1_8 11.95509 Tf9.28912 0 Td(4 h ln (1 )Tj/T1_3 11.95509 Tf11.94411 0 Td(p ) + h 2 ln 2 (1 )Tj/T1_3 11.95509 Tf11.94411 0 Td(p ) 2 ln (1 )Tj/T1_3 11.95509 Tf11.95309 0 Td(p ) ; (4.3) where h is the number of o v erhead bits per pack et. Fig.4sho ws the relationship between the recei v ed SNR and the optimal pack et length. F or a gi v en state S k the a v erage optimal pack et length P L k for this state can be deri v ed by using (4.1), (4.2), (4.3) to be R r k +1 r k 1 r e r r )Tj/T1_12 7.97011 Tf6.5833 0 Td(h ln(b( p 2 r )) )Tj/T1_18 10.9091 Tf6.60638 7.92899 Td(p )Tj/T1_11 7.97011 Tf6.5833 0 Td(4 h ln(b( p 2 r ))+ h 2 ln 2 (1 )Tj/T1_11 7.97011 Tf6.5833 0 Td((1 )Tj/T1_11 7.97011 Tf6.5833 0 Td(b( p 2 r ))) 2 ln(b( p 2 r )) dr R r k +1 r k 1 r e )Tj/T1_14 5.97758 Tf7.785 3.69 Td(r r dr : (4.4) Based on the abo v e analysis we present our pack et-length-based FSMC model in Fig.4. W e represent each state as the a v erage pack et length P L k which is the pack et size for a transmission in state k The transition probabilities between dif ferent states are denoted

PAGE 94

82 as t ij Further we can deri v e the state steady probability for each state as k = Z r k +1 r k 1 r e )Tj/T1_6 5.97758 Tf7.785 3.69902 Td(r r dr = e )Tj/T1_6 5.97758 Tf7.785 4.62604 Td(r k r )Tj/T1_3 10.9091 Tf10.89729 0 Td(e )Tj/T1_6 5.97758 Tf7.785 5.45403 Td(r k +1 r ; k = 0 ; 2 ; :::; K )Tj/T1_4 10.9091 Tf10.89729 0 Td(1 : (4.5) In practice, we may use dif ferent modulation schemes (not necessarily BPSK) in different channel states. Moreo v er by properly partitioning the range of the recei v ed SNR, we may obtain the multiplicati v e relationship between the a v erage optimal pack et lengths. 4.5 Courtesy Piggybacking In this section, we present our Courtesy Piggybacking scheme to alle viate the conict between throughput and f airness and to combat the starv ation problem for dif ferentiated services. 4.5.1 System Assumptions W e consider an ad hoc netw ork consisting of n mobile nodes uniformly distrib uted in some area. Nodes can communicate with each other directly if the y can hear each other or through other relay nodes in a single broadcast channel. The y emplo y some contentionbased MA C protocols, such as IEEE 802.11, to support their communications. Each node can generate services with N dif ferent priorities destined to other mobile node(s). A node' s mobility follo ws the random w aypoint model [58,70]. At rst, a node stays at a position for duration of pause time After that period, the node chooses a ne w random position and mo v es to w ards that position at a random speed uniformly distrib uted in the range from min speed to max speed After reaching the ne w position, the node will stay there for another pause time This process will continue for each node until the end of the simulation. W e assume some service dif ferentiation mechanism is emplo yed at the netw ork layer All the heterogeneous traf c is prioritized at its originating source node. When a pack et is handed do wn from the netw ork layer it will be k ept in the Tx queue corresponding to its priority and w ait for its turn to be transmitted at the MA C layer From the pre vious section, we kno w that the pack et length is related to the recei v ed SNR. The greater the SNR is, the greater the pack et length is. In the IEEE 802.11 MA C

PAGE 95

83 M S D U M A C h e a d e r M S D U C R C M A C h e a d e r M S D U C R C M A C h e a d e r M S D U C R C Figure 4: A fragmentation e xample. protocol [5], this pack et length may be called as frame length, which equals the fragmentation threshold plus the length of the MA C header and the length of CRC. In the IEEE 802.11 standard, the MA C layer tak es a MSDU from the Tx queue and adds MA C header and a CRC to each MSDU to generate a MPDU. In order to reduce the probability of transmission errors, the IEEE 802.11 limits the size of the body of a MPDU to be less than a x ed fragmentation threshold ( F T ), or it will break the long MSDU into multiple fragments, each of which will be no longer than the F T In Fig.4, we sho w a case where a long MSDU is partitioned into three small MSDUs in the IEEE 802.11. Since the length of the MA C o v erhead may be k ept unchanged, according to the analysis in the pre vious section, dif ferent channel states ha v e dif ferent frame lengths, we can say that dif ferent channel states ha v e dif ferent fragment thresholds ( F T s ). The greater the recei v ed SNR is, the greater the fragment threshold ( F T ) is. Hereafter we associate F T k with each state S k of the FSMC model as depicted in Fig.4. In order to impro v e the channel utilization, we assume that the MA C protocol can adapti v ely adjust the fragmentation threshold and the transmission rate according to the channel state. T o accurately determine the channel state when some pack ets need to be transmitted, we further assume that we ha v e some channel estimators or predictors, which can pro vide the accurate channel information for the proper MA C layer fragmentation. 4.5.2 The Courtesy Piggybacking Scheme In practice, the size of a pack et generated by an application may be x ed or may v ary from a minimum allo wed size to a maximum v alue P K max W e ar gue that the P K max should be properly chosen to reduce the o v erall o v erhead. Suppose we w ant to transmit c Mbits traf c. P ack ets are generated according to the P K max W e assume that each pack et

PAGE 96

84 Figure 4: The total o v erhead with P K max and F T can be correctly recei v ed without an y retransmission2. Then, the total o v erhead should be the sum of the o v erheads O ip at the IP layer (for e xample, 20 bytes for IPv4), O mac at the MA C layer (for e xample, 34 bytes for the IEEE 802.11) and O phy at the PHY layer (for e xample, 16 bytes). Thus, the total o v erhead to transmit the c Mbits traf c can be written as c P K max O ip + P K max F T ( O mac + O phy ) ; where d e is the function to round the element to the nearest inte ger greater than the element. W e sho w the relationship of o v erhead vs. P K max and F T when c = 1 in Fig.4. From Fig.4, we observ e that P K max should be reasonably chosen when multiple fragmentation thresholds are used. It cannot be too small, as it may cause too much o v er head; neither can it be too lar ge, as it may generate too man y fragments when the F T is small, which may further de grade the o v erall throughput. F or e xample, a lar ge pack et may 2 F or simplicity we only consider the case without an y retransmission and vie w the resulting o v erhead as a lo wer bound. Apparently the o v erhead with retransmissions is lar ger than this lo wer bound.

PAGE 97

85 be partitioned into man y small fragments. Each fragment is augmented with an indi vidual header sent as an independent transmission, and ackno wledged indi vidually Hence, a lar ge pack et will cause lots of D A T A/A CK e xchanges and result in suboptimal perfor mance. Thus, there must e xist an optimal v alue of P K max such that the o v erall o v erhead associated with the successful transmission of a message is minimized. Assume we obtain P K max which may not be equal to an y of F T k ( k = 0 ; 1 ; 2 ; :::; K )Tj/T1_5 11.95509 Tf11.98911 0 Td(1) ; it is thus advisable to approximate P K max with the closest fragmentation threshold corresponding to a certain channel state, say S m Therefore, we set P K max to F T m Ne xt, we w ant to sho w where the free space comes from. When a pack et with length strictly less than the P K max is transmitted in the channel state with F T less than F T m the pack et may be fragmented, and there is no free space a v ailable for all fragments possibly e xcept the last one. Ho we v er due to the time-v arying nature of the channel, when the pack et is transmitted in the channel state with F T greater than F T m one pack et does not ha v e enough bits to utilize the full capacity the channel pro vides in one transmission. W e ar gue that we could tak e adv antage of the free space to pack more bits as the channel allo ws. As a matter of f act, we ha v e sho wn, in our recent studies, that the fragmentation threshold can be up to 10K bits when the SNR is close to 20 dB and 64 QAM modulation scheme is used with tar geted FER 8% (Frame Error Rate) [93]. On the other hand, we observ e that in contention-based MA C protocols, it may tak e a long time for a node to seize the channel, and the node that has seized the channel should treasure e v ery transmission opportunity to transmit as man y bits as possible, especially when the channel condition is good. From no w on, we call state S k the free-space-ef fecti v e state when k is greater than m otherwise the non-free-space-ef fecti v e state, e v en though such a state may still has the possibility to pack more bits when the traf c dynamic is tak en into account. No w we describe ho w the Courtesy Piggybacking scheme mak es use of the free space. When a mobile node seizes the channel, it will rst check the channel state and determine if it is in a free-space-ef fecti v e state and if it is capable of piggybacking more pack ets in

PAGE 98

86 one transmission. If it is not in a free-space-ef fecti v e state, only one pack et (MSDU) with highest priority from the queues will be serv ed, as the current MA C protocol does. If the channel is in a free-space-ef fecti v e state, the node can transmit in one transmission as man y bits as channel allo ws and thus can piggyback more pack ets (MSDUs) from the queue(s), which may ha v e dif ferent priorities b ut the same ne xt hop in the routing table. Since the Courtesy Piggybacking scheme follo ws the cross-layer design principle so that the MA C layer has the access to the routing information, it is possible for the MA C layer to obtain such pack ets from the Tx queues. After identifying the e xistence of the free space, we no w discuss the piggybacking rules that guide the MA C layer to assemble enough and proper bits from the Tx queues (the HO W MANY -WHO problem) and piggyback them to the ne xt hop to alle viate the conict we intend to address. Since the channel state determines HO W MANY MSDUs the node can pack and transmit in one transmission, the fundamental issue of the rules should specify is who plays the role of train that of fers the piggybacking service to others, and who can enjo y such piggybacking service. W ithout an y scheduling mechanism, the role of train is al w ays tak en by the MSDU locating at the head of a non-empty queue with the highest priority currently Thus, the piggybacking rules should primarily address who has the pri vile ge to enjo y such free piggybacking service. As a guideline, the basic idea for such piggybacking rules is that under dif ferent channel states, the node assembles multiple MSDUs that may ha v e dif ferent priorities b ut share the same ne xt hop in the routing table, to form an MPDU whose length is channel dependent. In this w ay we can achie v e some e xtent of f airness between dif ferent prioritized services. When the channel is not in a free-space-ef fecti v e state, only the highest priority service in the Tx queues is supported, and the pack ets are fragmented if needed and are treated as usual. When the channel changes to a free-space-ef fecti v e state, according to the rules we dene, our Cour tesy Piggybacking scheme can pack other services, possible with lo wer priorities, to share the residual bandwidth with the high-priority traf c. One such rules is to gi v e preference

PAGE 99

87 q u e u e f o r p r i o r i t y l e v e l 0 q u e u e f o r p r i o r i t y l e v e l 1 q u e u e f o r p r i o r i t y l e v e l 2 q u e u e f o r p r i o r i t y l e v e l N 1 I n c o m i n g d a t a p a c k e t s f r o m r o u t i n g l a y e r w i t h a s s i g n e d p r i o i t y l e v e l d e q u e u e c o n t r o l l e r b M S D U b M S D U b M S D U A s s e m l e d M S D U M A C h e a d e r M S D U C R C M P D U B r e a k i n t o b M S D U i f n e c e s s a r yFigure 4: Illustration of the courtesy piggybacking scheme. to high-priority services. It al w ays, if possible, packs the high-priority services destined to the same ne xt hop in queue(s). Only when there are no more bits from the high-priority traf c tting into the free-space3will the bits from the lo wer -priority queue(s) be considered for piggybacking. Other rules may not prefer the high-priority service; for e xample, a high-priority service may trade-of f its o wn performance for f airer channel utilization by its courtesy piggybacking the lo w-priority service. One such rules is to al w ays piggyback the MSDUs from the longest Tx queue. W ith such piggybacking rules, the traf c dynamics (dif ferent pack et arri v al time and destinations) and channel dynamics are jointly utilized to strik e a good balance between throughput and f airness. Note that the piggybacking rules are not necessarily dened a priori ; the y could be designed to adapt to both channel and traf c uncertainty in the runtime. Intuiti v ely the Courtesy Piggybacking scheme can impro v e the performance of the lo w-priority traf c, since some lo w-priority pack ets may be pack ed with high-priority packets and be deli v ered to the ne xt hop for free; thus it can statistically reduce the time tak en to contend for accessing the channel for the lo w-priority services. This benet will be more 3 When we say no more bits in one queue tting into the free-space, it means either no pack et is left in the queue or pack ets in the queue do not share the same ne xt hop with the MSDU who of fers the piggybacking service.

PAGE 100

88 pronounced in mobile ad hoc netw orks using service dif ferentiation based MA C protocols [78][99] where the MA C protocols sacrice quality of the lo w-priority service to support high-priority service through either time spacing (dif ferentiation of Interframe Space (IFS)) or back of f parameters [5]. On the other hand, the reduction of contention from lo w-priority services can in turn benet the high-priority services: one node' s courtesy piggybacking of lo w-priority services may help its neighbors to transmit high-priority traf c, because less lo w-priority traf c will reduce the contention the high-priority traf c may encounter One may w onder wh y we do not simply release the channel so that other lo w-priority traf c can use the channel, that is, the so-called complete sharing scheme. The problem is that the time, for which the residual resource is a v ailable, is too short to be gi v en to other services due to the o v erhead associated with successfully seizing the channel. Besides, some MA C protocols such as the IEEE 802.11 f amily forbid others to use the channel during the time period specied by the Netw ork Allocation V ector (N A V). Ev en if the N A Vs are reset, the contention process may tak e too long to render the harv ested resource from the rate adaptation useless. Thus, the courtesy piggybacking by high-priority traf c o ws mak es more sense. In short, our Courtesy Piggybacking is able to achie v e better channel utilization and further impro v es the f airness between dif ferent prioritized traf c by the follo wing tw o means: lo wering the contention of the netw ork and decreasing the o v erhead required for a transmission. And as we sho w in simulation later our scheme signicantly impro v es the performance of lo w-priority traf c, while impro v es or at least k eeps unchanged the performance of high-priority traf c with appropriate piggybacking rules. T o illustrate the Courtesy Piggybacking scheme, we demonstrate the operation of the scheme in Fig.4. First, prioritized pack ets called MSDUs arri v e from the netw ork layer as b-MSDUs (basic MSDUs, the basic unit) whose lengths agree with the F T m W e assume the maximum pack et length P K max is strictly enforced at the upper layer; if not, an o v ersized MSDU will be further brok en do wn into se v eral b-MSDUs and the resulting b-MSDUs will inherit the IP header of the original MSDU. The b-MSDUs are k ept in the

PAGE 101

89 queues corresponding to their priorities. The dequeue controller operates according to the predened piggybacking rule, dequeues one or more b-MSDUs with the same ne xt hop and forms a MPDU satisfying the F T corresponding to the channel state. In order to reduce the o v erhead and the w ork to break a MSDU at the transmitter and to assembly the MSDU at the recei v er it is advisable to limit the pack et length at the netw ork layer to be no longer than F T m In order to a v oid further fragmentation of a b-MSDU to t the free space and assembly the b-MSDU, it is advisable to maintain the multiplicati v e relationship between the fragmentation threshold (FT) of the free-space-ef fecti v e state and F T m that is, the frame length for state k satises F T k = g k F T m where g k is a positi v e inte ger This can be achie v ed by properly partitioning the range of recei v ed SNR and adopting channeldependent modulation schemes. T o reduce the transmission time for a long frame, rate adapti v e transmission schemes may be used, so that the time for transmitting a frame does not v ary too much. T o a v oid making too man y modications to the MA C layer we prefer packing the b-MSDUs with the same ne xt hop in the routing table. T o f acilitate a recei v er in unpacking the bound pack ets, an unused bit [100] in the IP header of each b-MSDU is set to 1 at the transmitter to indicate that one bound b-MSDU is follo wed this b-MSDU, and the corresponding bit in the last b-MSDU is set to 0. At the recei v er the only thing it needs to do is to ackno wledge the recei v ed long frame, unpack the pack ed pack ets one by one according to the v alue of the unused bit. 4.5.3 Discussion Some properties of the CP. If we e xamine the destination of the bits in a single piggyback ed transmission, though these bits share the same ne xt hop, we nd out these bits may be destined to the ne xt hop or other nodes. Figure 4-8 showsthreescenariosfortheCourtesy Piggybacking. Consider that a mobile node A sends some bits (consisting of 2 b-MSDUs) to the ne xt hop B which ha v e three neighbors, including this mobile node A. Suppose that pack et 2 is piggyback ed by pack et 1, both pack ets should ha v e the same ne xt hop B. After the pack ets 1 and 2 arri v e at B, there are three cases at node B to process these tw o pack ets

PAGE 102

90 A B C D 1 2 1 2 A B C D 1 2 1 A B C D 1 2 C a s e 1 : E a c h h a s d i f f e r e n t n e x t h o p a t B C a s e 2 : B o t h d e s t i n e t o B C a s e 3 : O n e d e s t i n e s t o B t h e o t h e r n o t Figure 4: Three piggybacking cases. if we do not distinguish the dif ference between pack et 1 and pack et 2. Case 1 sho ws that pack et 1 and pack et 2 may be destined to dif ferent nodes and ha v e dif ferent ne xt hops at node B. Case 2 sho ws the case when both pack ets ha v e the same destination B. Case 3 sho ws that one pack et is destined to B while another one is destined to a node other than B. Since the probability that a pack et is piggyback ed lar gely depends on channel conditions and traf c pattern, the piggybacking may induce some delay jitter F or instance, due to our restriction that only pack ets sharing the same ne xt hop can be piggyback ed together one pack et that arri v es at one node later may lea v e the node earlier than some earlier arri ving pack ets. The signicance of piggybacking rules The piggybacking rules may play an impor tant role in allocating the bandwidth among the dif ferent prioritized traf c. Here, we w ant to discuss the design of a piggybacking rule based on a special case. When we ha v e plenty

PAGE 103

91 of dif ferent priority pack ets in the Tx queues w aiting for being serv ed, the design of piggybacking rule can be vie wed as an allocation problem. As discussed abo v e, by properly par titioning the range of recei v ed SNR, the fragmentation threshold of the free-space-ef fecti v e state k satises F T k = g k F T m where g k is a positi v e inte ger F or other non-free-spaceef fecti v e states, let g k = 1 Suppose we ha v e a total of K dif ferent channel states, and N dif ferent priority le v els. Let k j denote the number of b-MSDUs of j priority le v el to be pack ed when the channel is in state k W e should point out that in the non-free-spaceef fecti v e state, only the highest priority pack et is serv ed when there is plenty of traf c in the w aiting queue, thus k 0 = 1 and k j = 0 for 0 k m and 1 j N )Tj/T1_4 11.95509 Tf12.5201 0 Td(1 If we ne glect the MA C layer o v erhead, then the design problem can be reduced to choose a k j such that 8 > < > : 0 k j P j k j g k 0 k K )Tj/T1_4 11.95509 Tf11.95309 0 Td(1 ; 0 j N )Tj/T1_4 11.95509 Tf11.94411 0 Td(1 : (4.6) When the operation of packing pack ets with dif ferent priority le v els associates with certain objecti v e of f airness or prot, the design problem is a kind of knapsack problem which is considered to be NP-complete [72]. On the one hand, the resource (free space here) should ne fully utilized; On the other hand, the system attempts to maintain the best performance in terns of f airness among dif ferent classes of traf c. In f act, an upper bound for the e xpected v alue of throughput of priority le v el j at one node is can be written by P k R k p k k j ; 0 j N )Tj/T1_4 11.95509 Tf11.9621 0 Td(1 where p k is the probability that the channel is in state k and R k is the transmission rate in state k Some measures may impro v e CP. T o a v oid fragmentation of the b-MSDUs in a freespace-ef fecti v e state, in our piggyback scheme we should maintain the multiplicati v e relationship between the fragmentation threshold (FT) in the free-space-ef fecti v e state and the F T m Actually we can relax this requirement in the high traf c load case by allo wing fragmentation of the lo w-priority services at will to t into the free space the channel pro vides. Because at v ery hea vy traf c load, the piggybacking rule f a v oring high-priority

PAGE 104

92 services may still lead to bandwidth starv ation for lo w-priority services. By allo wing the fragmentation of the b-MSDUs from lo w-priority traf c, at least some lo w-priority traf c can be serv ed by piggybacking. Our Courtesy Piggybacking scheme does not preclude scheduling mechanisms; in f act, scheduling can still be emplo yed at higher layers to enhance the management of the heterogeneous traf c. F or e xample, we can use EDF (Earliest Deadline First) polic y to manage each priority queue, such that in a single queue, the pack et with early deadline may be put at the head of the queue and get transmitted earlier than those with later deadline. At the same time, the piggyback scheme can still tak e ef fect once the channel is in a freespace-ef fecti v e state. In our preliminary implementation of the piggybacking, the b-MSDUs are or g anized in the queues according to their priorities. When a transmitter w ants to pack more bits to the same recei v er e xhausti v e search is carried out to nd the proper bits in candidate queues according to the piggybacking rules. In addition, some processing time is also needed at the recei v er to unpack the pack ets. Thus, in the rst place we may e xpect some additional delay caused by courtesy piggybacking; ho we v er as can be seen from our performance e v aluation, the incurred delay is ne gligible and hence acceptable compared with the benet g ained from the Courtesy Piggybacking scheme. Furthermore, we ar gue that with more ef cient w ays to or g anize the b-MSDUs and to quickly acquire the proper b-MSDUs for packing, the benet from piggybacking scheme should be more visible. In our proposed Courtesy Piggybacking scheme, only the traf c sharing the same ne xt hop can be pack ed. One may w onder if it can be e xtended to the dif ferent ne xt hop scenarios. W e are cautions to mak e this mo v e. The main concern comes from the MA C layer The pre v alent MA C protocols including the IEEE 802.11 do not support multiple recei v ers at the same time. Undoubtedly it is a arduous task to coordinate the transmission and reception acti vities between multiple recei v ers. Especially when multiple recei v ers are in v olv ed in one transmission, the interference area should be much lar ger than that of one

PAGE 105

93 recei v er case and the coordination issue gets more complicated. Though a data ushing mechanism for multiple recei v er w as presented in [101], while without consideration of the enlar ged interference area, the benet may not outweigh the ne g ati v e impact of their approach. In our e x emplary piggybacking illustration (Fig.4), all the dequeued b-MSDUs are rst assembled as one MSDU and further encapsulated with only one MA C header and CRC. This e x emplary piggybacking method imposes no modication on current MA C protocol, e.g, IEEE 802.11, b ut it lacks e xibility F or e xample, when the recei v er detects something wrong with the recei v ed frame, it is not able to inform the transmitter which b-MSDU is damaged, and as a result, the transmitter has to retransmit the whole frame ag ain. In f act we can adopt an alternati v e method as the follo wing (Fig.4). Dif fer ent from the abo v e method, each b-MSDU will generate a separate MPDU with its o wn CRC and MA C header and the resulting MPDUs will nally be combined as an assembled MPDU and be sent out as one unit in one transmission. At the recei v er side, the recei v er can check the CRC for each b-MSDU and positi v ely or ne g ati v ely ackno wledge each bMSDU in a single short message, for e xample, A CK. W ith this alternati v e method, the transmitter can retransmit only those b-MSDUs f ailed CRC checking. In f act, these retransmitted b-MSDUs can be further bound with other fresh b-MSDUs from the queues to form another assembled MPDU. Compared the alterati v e one with our e x emplary one, the alterati v e method pro vides some de gree of e xibility b ut requires some modication on MA C protocol to f acilitate the abo v e communications. This modication can be possibly made similar to that in [101]. On the other hand, our e x emplary piggybacking method may incur less o v erhead of MA C header and CRC than the alternati v e one. In practice, some adapti v e piggyacking methods can be designed to accommodate the time-v arying channel condition, meanwhile, k eep the o v erhead reasonable.

PAGE 106

94 . A s s e m l e d M P D U M A C h e a d e r b M S D U C R C M A C h e a d e r b M S D U C R C M P D U Figure 4: An alternati v e piggybacking method. 4.6 Performance Analysis In this section, we present the performance analysis in order to theoretically sho w the ef fecti v eness of our proposed piggybacking scheme in alle viating the conict between throughput and f airness for dif ferent prioritized services. T o simplify the analysis, we consider the piggybacking at one node, and assume all the pack ets are destined for the same ne xt hop. Ag ain, we assume that there are a total of N dif ferent priority le v els, P 0 P 1 ..., P N )Tj/T1_5 7.97011 Tf6.5833 0 Td(1 whereby P i has higher priority than P j if i < j The arri v als of each priority P i service are P oisson processes with arri v al rate i The channel is modeled as FSMC as discussed in Section4.4: F T k = g k F T m where g k = 1 if k m4; otherwise g k is an inte ger greater than 1 and g i > g j if i > j m When the channel is in state j > m the channel adaptation scheme is used such that the time for transmitting F T j is almost the same as that for F T m Moreo v er we assume that the pack et length is P K max = F T m Under such assumptions, our Courtesy Piggybacking scheme can be modeled as a multiple-serv er queue system with non-preempti v e priority where the service rate is dependent on the channel state. More specically the service discipline is the follo wing if we limit the number of serv ers to 2. Serv er S R 0 operates in all channel states with a v erage service time X F or serv er S R 1 it operates when the channel is in the state S k ; k > m and does not w ork in an y state S k ; k m ; the service time for serv er S R 1 is X g k )Tj/T1_5 7.97011 Tf6.5833 0 Td(1 This means that when serv er S R 0 serv es one non-piggyback 4 Gi v en x ed transmitting po wer and modulation scheme, when the channel is in an y state j < m a fragmentation threshold that is smaller than F T m is required to maintain the same frame error rate. F or simplicity we assume F T m is also used in those states.

PAGE 107

95 pack et, serv er S R 1 can serv e g k )Tj/T1_5 11.95509 Tf12.5741 0 Td(1 piggyback ed pack ets. The non-w orking probability of serv er S R 1 is m P k =0 k where k is the steady-state probability of state S k Serv er S R 0 operates with non-preempti v e head-of-line priority service discipline, while serv er S R 1 operates according to the discipline dened in the piggybacking rules. F or e xample, it may rst serv e the traf c with the longest w aiting queue. W ithin a priority traf c, service is pro vided on a rst-come-rst-serv ed basis. Figure 4 showstheanalyticalmodel. F or the purpose of obtaining tractable analysis, we study the scheme in a simple scenario in which the traf c is of tw o priorities, and the channel has tw o states, with F T 1 = 2 F T 0 Accordingly the service time for each serv er is X As discussed abo v e, serv er S R 0 w orks all the time with non-preempti v e service discipline, whereas S R 1 w ork only when channel state is in S 1 and does not w ork otherwise. Then our piggybacking queue model can be reduced to an M/D/2 non-preempti v e priority queueing system, in which one serv er w orks all the time, while the other w orks with some probability This system seems to be simple, ho we v er it is hard to quantitati v ely analyze it. T o our kno wledge, there is no ready solution to this interesting queue system. Hence we seek to get some bounds on the performance metrics such as a v erage w aiting time (that is, the a v erage queueing delay) and queue size. W e only focus on the a v erage w aiting time hereafter as the a v erage w aiting time and the queue size lead to each other according to Little' s La w W e rst consider the upper bound of the a v erage w aiting time for the system. Ob viously if a transmitter does not ha v e an y kno wledge of the channel status and hence does not adopt an y rate adaptation based on the channel state, both traf c types, that is, the high priority traf c and the lo w priority traf c, will suf fer from longer delay than the y will when piggybacking is used. In this case, only one serv er that is, serv er S R 0 w orks. The system thus is an M/D/1 priority queue system [102], and it is easy to obtain the a v erage w aiting time for each type of traf c: 8 > < > : W 0 = ( 0 + 1 ) X 2 2(1 )Tj/T1_3 7.97011 Tf6.5833 0 Td( 0 X ) W 1 = ( 0 + 1 ) X 2 2(1 )Tj/T1_3 7.97011 Tf6.5833 0 Td( 0 X )(1 )Tj/T1_3 7.97011 Tf6.5833 0 Td( 0 X )Tj/T1_3 7.97011 Tf6.5833 0 Td( 1 X ) : (4.7)

PAGE 108

96 . . S R0 S R1 0 l 1 l 1 N l S 0 S 1 S k 1 t 0 1 t 1 0 t 1 2 t 2 1 t k 2 k 1 t k 1 k 2 t 0 0 t 1 1 t k 1 k 1 0 1 m 1 1 m P r i o r i t y Q u e u e s S e r v e r s S m t m 1 m t m m 1 t m m + 1 t m + 1 m t m mFigure 4: The queue model for piggybacking. Recall that serv er S R 0 w orks all the time and serv er S R 1 w orks only when channel state is in S 1 If the piggybacking rule is set such that when S R 1 operates, it al w ays serv es the high-priority traf c rst, the follo wing case will pro vide the lo wer bound for the a v erage w aiting time. In this case, the transmitter is a w are of the channel state and adopts the channel dependent transmission rate. Moreo v er the channel is al w ays in good state. Therefore, the system becomes a tw o-serv er queue system with non-preempti v e priority where the tw o serv ers, serv er S R 0 and S R 1 w ork all the time. According to the results in [103], we can obtain the a v erage w aiting time for each type of traf c: 8 > < > : W 0 = 4( 0 + 1 ) 2 X 3 3(2+( 0 + 1 ) X )(2 )Tj/T1_12 7.97011 Tf6.5833 0 Td( 0 X ) W 1 = 4( 0 + 1 ) 2 X 3 3(4 )Tj/T1_9 7.97011 Tf6.5833 0 Td(( 0 + 1 ) 2 X 2 )(2 )Tj/T1_12 7.97011 Tf6.5833 0 Td( 0 X ) : (4.8) T o v alidate the upper and lo wer bounds, we implement the abo v e queue model in OPNET [57] and present the analytical results and simulation results in Fig.4, where the service time X = 0.01s. It can be observ ed that the a v erage w aiting time for the piggybacking scheme in three dif ferent channel models as specied in T able4is completely

PAGE 109

97 6 5 7 0 7 5 8 0 8 5 9 0 9 5 1 0 0 1 03 1 02 1 01 1 00 A r r i v a l r a t e ( p a c k e t / s )A v e r a g e w a i t i n g t i m e f o r p r i o r i t y 1 ( s ) p0= 0 7 5 p1= 0 2 5 : s i m u l a t i o n p0= 0 5 0 p1= 0 5 0 : s i m u l a t i o n p0= 0 2 5 p1= 0 7 5 : s i m u l a t i o n L o w e r b o u n d : a n a l y s i s U p p e r b o u n d : a n a l y s i s (a) A v erage w aiting time of priority 1 service. 6 5 7 0 7 5 8 0 8 5 9 0 9 5 1 0 0 1 04 1 03 1 02 A r r i v a l r a t e ( p a c k e t / s )A v e r a g e w a i t i n g t i m e f o r p r i o r i t y 0 ( s ) p0= 0 7 5 p1= 0 2 5 : s i m u l a t i o n p0= 0 5 0 p1= 0 5 0 : s i m u l a t i o n p0= 0 2 5 p1= 0 7 5 : s i m u l a t i o n L o w e r b o u n d : a n a l y s i s U p p e r b o u n d : a n a l y s i s (b) A v erage w aiting time of priority 0 service. Figure 4: A v erage w aiting time. T able 4: Channel model statistics Setting State i P r ( i ) t 1 ;i t 0 ;i 1 0 0.75 0.0075 0.9975 1 0.25 0.9925 0.0025 2 0 0.5 0.002 0.998 1 0.5 0.998 0.002 3 0 0.25 0.0025 0.9925 1 0.75 0.9975 0.0075 bounded by the obtained bounds. Therefore, we kno w the g ain that can be accrued by adopting the piggybacking scheme as opposed to the case without piggybacking. 4.7 Performance Ev aluation In this section, we implement our proposed piggybacking scheme in OPNET with DSD V and the IEEE 802.11 as underlying routing and MA C protocols, and conduct e xtensi v e simulations to e v aluate the performance of the piggybacking scheme. The simulation results v alidate the ef fecti v eness and ef cienc y in impro ving the channel utilization and f airness thus alle viating the conict between throughput and f airness. 4.7.1 Simulation Setup In our simulation study we assume the ph ysical channel is a slo w f ading channel with only tw o states satisfying F T 1 = 2 F T 0 Three dif ferent channel settings are adopted

PAGE 110

98 in our studies. The channel statistics are listed in T able4. While setting 2 represents a neutral channel in the sense that the steady state probabilities of both states are identical, channel setting 1 and 3 respecti v ely represent a relati v ely bad channel and a relati v ely good channel. W ithout loss of generality we limit the number of priorities to 2. W e simulate an ad hoc netw ork consisting of 50 mobile nodes5, whose mobility follo ws the random w aypoint mobility model in a 1500 300 m 2 area. The transmission range of each node is 250 m Each node generates traf c according to a Poisson process with parameter and the destination for each generated pack et is randomly chosen among all other nodes. W e assume that the pack et length is 1024 bits that agrees with F T 0 and each pack et is a b-MSDU. The generated traf c is further assigned with 1 (lo w priority) or 0 (high priority) with probability 0.5. All pack ets are b uf fered in the queues according to their priorities. In our simulations, when the channel is in the state corresponding to F T 0 we use basic transmission rate 1Mbps, while for state corresponding to F T 1 we use 2Mbps, so that the transmission time for one fragment with channel-dependent length in tw o states does not change too much. Each simulation runs for 300 seconds. W e dene tw o piggybacking rules for comparison. Rule 1 f a v ors the high priority traf c. When the channel state is in F T 0 a priority-1 pack et can be serv ed only when no priority-0 pack et e xists in the queue. When the state is in F T 1 a priority-1 pack et is piggyback ed by priority-0 pack ets only when no priority-0 pack et is left in the queue (of course, the pack ets should share the same ne xt hop in the routing table). In contrast, rule 2 f a v ors the lo w priority traf c. In rule 2, when the channel is in F T 0 the stations act as in the rule 1. When the state is in F T 1 a priority-1 pack et is piggyback ed no matter ho w man y priority-0 pack ets are left in the queue. W e study the performance of the netw ork in four 5 T o minimize the possible unf airness between dif ferent nodes due to unequal channel access, we use such conguration and treat all the nodes equally for e xample, using the same traf c and mobility pattern.

PAGE 111

99 scenarios: case 1. the netw ork is una w are of channel states; case 2. the netw ork is a w are of channel states and thus adopts dynamic transmission rate; case 3. the netw ork emplo ys the Courtesy Piggybacking with rule 1; and case 4. the netw ork emplo ys the Courtesy Piggybacking with rule 2. W e choose tw o metrics to analyze and compare the performance of piggybacking: End-to-end delay and P ack et deli v ery ratio End-to-end delay measures the a v erage onew ay latenc y observ ed between the time instant that the pack et is generated at the source and the time instant the pack et is recei v ed at the destination. This metric should count in all the delays, such as propag ation delay queueing delay and transmission delay which the pack et has e xperienced during the whole process. P ack et deli v ery ratio measures the ratio of the total number of pack et successfully recei v ed by the destinations to the total number of pack ets generated at the sources. This metric reects the o v erall throughput and f airness of each prioritized traf c. T o observ e the ef fect of the channel characteristics, we rst disable node mobility and compare the performance of piggybacking under dif ferent channel settings and under dif ferent traf c loads. Then we x the channel setting as setting 2, and enable node mobility according to the random w aypoint mobility model described in the Section4.5with min speed = 1 m=s and max speed = 19 m=s in order to study the ef fect of traf c load and mobility on the performance of piggybacking. 4.7.2 Impact of Channel Characteristics Figure4 illustratesperformanceofthenetworkinthefourdifferentcases,when three channel settings are specied as in T able4. Since in case 1, the netw ork has no kno wledge of the channel, the performance is the same no matter what channel settings are used. In all the other three cases, the performance changes with the channel setting. More specically as the channel is getting better that is, the channel setting changes from 1, to 2, and to 3, for each traf c priority the end-to-end delay decreases and the pack et deli v ery ratio increases. This sho ws that the dynamic channel conditions are v aluable resources that

PAGE 112

100 1 2 3 0 1 2 3 4 5 e n d t o e n d d e l a y ( s )c h a n n a l s e t t i n g ( a ) p r i o r i t y 1 c a s e 1 c a s e 2 c a s e 3 c a s e 4 1 2 3 0 0 1 0 2 0 3 0 4 0 5 e n d t o e n d d e l a y ( s )c h a n n a l s e t t i n g ( b ) p r i o r i t y 0 1 2 3 0 0 2 0 4 0 6 0 8 1 p a c k e t d e l i v e r y r a t i oc h a n n a l s e t t i n g ( c ) p r i o r i t y 1 1 2 3 0 0 2 0 4 0 6 0 8 1 p a c k e t d e l i v e r y r a t i oc h a n n a l s e t t i n g ( d ) p r i o r i t y 0Figure 4: Simulation results with dif ferent channel settings. should be e xplored. More impro v ements in terms of both metrics can be observ ed when piggybacking is adopted, for e xample, case 3 and case 4, compared to case 2. Further our piggybacking scheme can achie v e a better f airness between these tw o prioritized traf c while still maintaining a higher pack et deli v ery ratio, thus higher throughput. This is because, in addition to taking adv antage of channel states, the piggybacking scheme reduces the time to contend for the channel, thereby further impro ving channel utilization. 4.7.3 Impact of T raf c Load The performance of the netw ork is studied under dif ferent traf c loads as well, as sho wn in Fig.4, in which the a v erage pack et inter -arri v al time of 0.3s represents the relati v ely light traf c load and the a v erage pack et inter -arri v al time of 0.25s represents relati v ely hea vy traf c load. From Fig.4, we can see the inter -class ef fects in the dif ferentiated services system, which becomes more pronounced in the case where the highpriority traf c load is high. The delay for priority 1 (7.89 second) is f ar greater than that for priority 0 (0.30 second), and the pack et deli v ery ratio for priority 1 is much smaller than that for priority 0. It also can be observ ed that case 3 and 4, where our piggybacking scheme is emplo yed, ha v e better performance than case 1 and 2. Under both hea vy traf c load and

PAGE 113

101 0 2 5 0 3 0 2 4 6 8 e n d t o e n d d e l a y ( s )p a c k e t a r r i v a l i n t e r v a l ( s ) ( a ) p r i o r i t y 1 0 2 5 0 3 0 0 1 0 2 0 3 0 4 e n d t o e n d d e l a y ( s )p a c k e t a r r i v a l i n t e r v a l ( s ) ( b ) p r i o r i t y 0 c a s e 1 c a s e 2 c a s e 3 c a s e 4 0 2 5 0 3 0 5 0 6 0 7 0 8 0 9 1 p a c k e t d e l i v e r y r a t i op a c k e t a r r i v a l i n t e r v a l ( s ) ( c ) p r i o r i t y 1 0 2 5 0 3 0 5 0 6 0 7 0 8 0 9 1 p a c k e t d e l i v e r y r a t i op a c k e t a r r i v a l i n t e r v a l ( s ) ( d ) p r i o r i t y 0Figure 4: Simulation results with dif ferent pack et arri v al rates. light traf c load, the piggybacking scheme can greatly reduce the end-to-end delay and impro v e the pack et deli v ery ratio for both priorities. It is note w orth y that although channel a w are mechanisms (case 2) can impro v e both metrics compared to case 1, our piggybacking scheme pro vides more benets that the channel a w are mechanism alone cannot achie v e for the reasons described abo v e. 4.7.4 Impact of Node Mobility Ne xt, we study the impact of mobility on the performance of the proposed Courtesy Piggybacking scheme when the a v erage pack et inter -arri v al time is 0.25s. In Fig.4, the lar ger the pause time the lo wer the mobility W e observ e that in all the cases the performance is sensiti v e to mobility and de grades as mobility increases. The reason for this is that high mobility may frequently cause route breakages and pack et losses, and hence increase delay and decrease pack et deli v ery ratio. In addition, we can clearly observ e that case 2, case 3, and case 4 ha v e better performance than case 1, which is consistent with our observ ations in pre vious simulations. In general, these three cases ha v e shorter end-to-end delay and higher pack et deli v ery ratio than case 1 for both priorities, especially for the lo w priority traf c, the disadv antageous traf c in the dif ferentiated services system. Since all

PAGE 114

102 0 5 0 1 0 0 1 5 0 2 0 0 2 5 0 3 0 0 0 5 1 0 1 5 p a u s e t i m e ( s )e n d t o e n d d e l a y ( s ) c a s e 1 c a s e 2 c a s e 3 c a s e 4 (a) A v erage end-to-end delay of priority-1 service. 0 5 0 1 0 0 1 5 0 2 0 0 2 5 0 3 0 0 0 0 5 1 1 5 2 2 5 3 3 5 p a u s e t i m e ( s )e n d t o e n d d e l a y ( s ) c a s e 1 c a s e 2 c a s e 3 c a s e 4 (b) A v erage end-to-end delay of priority-0 service. 0 5 0 1 0 0 1 5 0 2 0 0 2 5 0 3 0 0 0 4 0 5 0 6 0 7 0 8 0 9 1 p a u s e t i m e ( s )p a c k e t d e l i v e r y r a t i o c a s e 1 c a s e 2 c a s e 3 c a s e 4 (c) P ack et deli v ery ratio of Priority-1 service. 0 5 0 1 0 0 1 5 0 2 0 0 2 5 0 3 0 0 0 6 0 6 5 0 7 0 7 5 0 8 0 8 5 0 9 0 9 5 1 p a u s e t i m e ( s )p a c k e t d e l i v e r y r a t i o c a s e 1 c a s e 2 c a s e 3 c a s e 4 (d) P ack et deli v ery ratio of Priority-0 service. Figure 4: Simulation results with dif ferent pause time.

PAGE 115

103 the cases e xcept case 1 mak e use of the channel conditions and rate adaptation, ag ain, we v alidate that the dynamic channel conditions can be used to impro v e the channel utilization. W e further compare case 3 and case 4 as a group with case 2 to sho w the ef fecti v eness of our Courtesy Piggybacking scheme. From Fig.4, we can clearly see that our scheme can further shorten the end-to-end delay and impro v e the pack et deli v ery ratio for both types of traf c. Moreo v er the Courtesy Piggybacking scheme not only impro v es the performance of the priority-0 traf c, b ut also impro v es signicantly the performance of the priority-1 traf c. This v eries that our Courtesy Piggybacking scheme is capable of alle viating the conict between the dif ferent prioritized traf c. As discussed in Section4.5, all these g ains come from the Courtesy Piggybacking scheme. In case 2, the channel state information is e xploited only to some e xtent, b ut not fully harv ested in the sense that the free space cannot completely be utilized. Ho we v er our piggybacking scheme can mak e use of these system dynamics, not only the channel dynamics b ut also the traf c dynamics, so that the free space can be best e xploited without an y w aste. 4.7.5 Impact of Piggybacking Rules Finally we focus on case 3 and 4 and study the impact of the piggybacking rules. The piggybacking rule in case 3 f a v ors the high-priority traf c in the system, priority 0, while the rule in case 4 f a v ors the lo w-priority traf c, priority 1. Thus, there is no surprise that in Fig.4(a)and Fig.4(c), the end-to-end delay for the priority 1 in case 4 is generally shorter than that in case 3, and the pack et deli v ery ratio is generally greater than that in case 3. F or the priority-0 traf c, all the measured metrics generally ha v e better performance in case 3 than those in case 4. Compared with the performance in case 3, the Courtesy Piggybacking in case 4 sacrices the priority-0 traf c a little bit to piggyback the priority-1 traf c. In Fig.4(d), we also observ e some oscillations in the pack et deli v ery ratio when the mobility is high, for e xample, when the pause time is less than 60. The pack et deli v ery ratio of priority 0 in case 3 seems v ery sensiti v e to the high mobility and has w orse performance than that in the case 4, the one with piggybacking rule preferring

PAGE 116

104 the lo w priority This can be e xplained as follo ws. When the mobility is high, the pack et loss may primarily result from the mobility of nodes in v olv ed in the communications, not necessarily from the channel impairments due to other f actors. On the other hand, the high mobility prolongs pack et deli v ery and brings do wn the pack et deli v ery ratio, which further results in man y w aiting pack ets of both types in the queues. In case 3, since the piggybacking rule prefers the traf c of priority 0, quite often we may ha v e tw o priority-0 pack ets pack ed together for transmission to the ne xt hop when the channel is in state 1. If the recei v er does not recei v e them successfully due to high mobility in this case, then more pack ets of priority 0 will be dropped, leading to lo wer pack et deli v ery ratio, thus the pack et loss due to high mobility under piggybacking rule in case 3 may be amplied and accordingly de grades the performance further than in case 4 for high-priority traf c. On the contrary in case 4, instead of packing tw o priority-0 pack ets when possible, a sender packs one pack et of priority 1 with one pack et of priority 0. When the pack ed pack ets cannot be successfully recei v ed due to high mobility only one pack et of each priority is in v olv ed, hence the impact on the high-priority traf c is less se v ere. Thus the Courtesy Piggybacking with properly designed piggybacking rules may compensate for the ne g ati v e ef fect of high mobility 4.8 Summary In this chapter we propose a no v el Courtesy Piggybacking scheme to alle viate the conict between throughput and f airness for dif ferent prioritized services in mobile ad hoc netw orks. By making use of system dynamics, such as the v ariable channel quality and changing traf c conditions, it can harness the a v ailable residual bandwidth that w ould otherwise be w asted. Thereby it signicantly impro v es the end-to-end delay and pack et deli v ery ratio. More precisely when the traf c load is light, it can shorten the end-toend delay; when the traf c load is high, it can not only shorten the end-to-end delay b ut also impro v e the pack et deli v ery ratio for all prioritized services. W ith properly dened piggybacking rules, the piggybacking scheme can e xibly allocate the bandwidth among

PAGE 117

105 dif ferent types of traf c, thus achie v e good f airness between dif ferent priority services without using con v entional costly scheduling mechanisms. By deri ving the delay bounds, we e xplicitly specify the performance g ain the piggybacking scheme can achie v e in some simplied scenarios. Further e xtensi v e simulations v erify the performance of our proposed piggybacking scheme. Our scheme is also sho wn to be easily implemented in a distrib uted f ashion, and thus could be incorporated into man y scheduling schemes to pro vide better support of the dif ferentiated and heterogeneous services in both mobile ad hoc netw orks and traditional wireless netw orks.

PAGE 118

CHAPTER 5 R OB UST AND ENERGY -EFFICIENT D A T A DISSEMIN A TION IN WIRELESS SENSOR NETW ORKS 5.1 Introduction Recent adv ances in Micro-Electro-Mechanical Systems (MEMS) technology and wireless communications ha v e resulted in the emer gence of small, lo w-cost sensors with more and more po werful processing and netw orking capabilities, which mak es wireless sensor netw orks (WSNs) be identied as one of the most important emer ging technologies [104]. Especially wireless sensor netw orks can furnish us with ne-granular observ ation about the ph ysical w orld we are li ving in. Potential applications include the remote sensing in nuclear plants, mines, and other hazardous industrial v enues, real-time traf c monitoring, realtime weather monitoring, wild animal monitoring and tracking, disaster rescue, ener gy management, medical monitoring, logistics and in v entory management, and military reconnaissance. While much research has been focused on making sensor netw orks feasible and useful [105] [106], some important problems resulting from the error -prone and resource-constrained nature of WSNs ha v e not been well addressed yet. Notable are the issues associated with scalability reliability and ener gy ef cienc y F or instance, since a WSN may consist of hundreds or thousands or e v en millions of sensor nodes, an ef cient data dissemination technique should w ork well not only in small-scale sensor netw orks b ut also in lar ge-scale ones. In addition, it should be rob ust ag ainst the harsh en vironmental ef fects and temporal or permanent f ailures of sensors and wireless links in between them, so that the functionality of the WSN can be sustained without an y interruption. Moreo v er it should ha v e good ener gy ef cienc y in terms of both lo w a v erage ener gy consumption per observ ation report and balanced ener gy usage instead of o v erb urdening a small set of nodes in the netw ork. 106

PAGE 119

107 The research presented in this chapter tar gets at real-time and continuous monitor ing applications such as battleeld monitoring netw orks and v olcano monitoring netw orks, where sensors are deplo yed in an ad hoc manner and the aforementioned nice properties are desirable. Those sensors collaborati v ely accomplish the sensing task and forw ard the sensing data to the closest data processing centers or sink nodes through wireless links. T raditional routing protocols proposed for ad hoc netw orks may not be suitable for our tar get applications due to the substantial dif ferences between ad hoc netw orks and sensor netw orks [105]. In contrast, ooding, as a reacti v e technique with inbred reliability seems to be a good candidate for sensor netw orks because it does not in v olv e costly topology maintenance and comple x route disco v ery algorithms. Ho we v er the main problems with ooding are that it typically causes unproducti v e and often harmful bandwidth congestion, as well as inef cient use of node resources such as ener gy which is scarce in resource-constrained sensor netw orks. Though se v eral data dissemination schemes ha v e been proposed specically for sensor netw orks [105] [106], research on nding a scheme that can strik e a good balance among reliability scalability and ener gy ef cienc y is still lacking. In this chapter we propose a h ybrid data dissemination frame w ork for WSNs that features lo w o v erhead, high reliability good scalability and e xibility and preferable ener gy ef cienc y Our contrib utions are mainly three-fold. First of all, to the best of our kno wledge, this is the rst ef fort to study a wireless sensor netw ork from the point of vie w of supply chain management. W e introduce the notion of supply c hain into the design of sensor netw orks and conceptually partition a sensor eld into se v eral functional re gions according to the supply chain management methodology Secondly we apply dif ferent routing techniques to dif ferent re gions in order to pro vide better performance in terms of reliability and ener gy consumption. Lastly we propose a no v el zone ooding scheme which is a combination of con v entional ooding and geometric routing. Our rationale here is to of fer the desired reliability and routing simplicity with ooding and to mitig ate the decienc y of blind ooding with geometric routing.

PAGE 120

108 The remainder of this chapter is or g anized as follo ws. W e start with Section5.2describing the related w ork and discussing some basics of supply chains and the resemblance between supply chains and wireless sensor netw orks in Section5.3. Then, we detail our h ybrid data dissemination frame w ork in Section5.4. In Section5.5, simulation studies are carried out to e v aluate the performance of the proposed scheme. The concluding remarks are gi v en in Section5.6. 5.2 Related W ork Ho w to ef ciently deli v er the information in densely deplo yed sensor netw orks is a v ery challenging task. Directed dif fusion [107], and SPIN [108] are tw o e x emplary data dissemination paradigms for sensor netw orks. SPIN adopts meta-data ne gotiation to eliminate the redundant data transmission, and it is suitable for the scenarios where an individual sensor disseminates its observ ations to all the sensors in a netw ork. As a datacentric approach, directed dif fusion emplo ys lo w rate ooding to establish gradients and uses gradual reinforcement of better paths to accommodate certain le v els of netw ork and sink dynamics. Recently Heidemann et al. proposed tw o v ariations of directed dif fusion and the y adv ocated the importance of matching dissemination algorithms to application performance requirements [109]. Besides the abo v e schemes, the cluster -based LEA CH [110] and hierarchical-based TTDD [111] are also a v ailable in the literature. In LEA CH, nodes are or g anized into clusters, each of which has a randomly selected cluster header Sensor nodes send their observ ations to af liated cluster headers and request cluster headers to forw ard their reported data to a base station. As the authors mentioned, their scheme w as designed for applications where a sensor node is able to transmit data to a f ar a w ay base station with higher transmission po wer In TTDD, each data source rst b uilds a grid structure, based on which the sink' s query and sensing data are forw arded through this grid structure and further ooded in a local grid cell. TTDD can accommodate the e xistence of mobile sinks. W e notice that the abo v e approaches need lots of o v erhead to be w orkable, for e xample, the y need e xchange a lot of information with neighbors to b uild routing tables

PAGE 121

109 or to form and propag ate the cluster/grid structures. In addition, reliability in the presence of the error -prone wireless links is another constraint of the abo v e approaches because the y usually maintain only a single route for each source-sink pair Though the multipath e xtension of [107] w as reported [112], it is quite dif ferent from our approach. P articularly the alternati v e paths are only used when some node f ailures occur along the primary path. Moreo v er the multipath e xtension of directed dif fusion is still not able to deal with other transmission f ailures due to the wireless links. In contrast, our RRP inherits the inbred reliability of ooding and is rob ust ag ainst transmission f ailures coming from both node f ailures and wireless link impairment. As a reacti v e technique with inbred reliability ooding seems to be a good candidate for sensor netw orks because it does not in v olv e costly topology maintenance and comple x route disco v ery algorithms. In vie w of this, quite a fe w research ef forts ha v e been done to optimize the use of ooding in terms of reducing the unproducti v e and often harmful bandwidth congestion, as well as inef cient use of nodes resources caused by ooding. Gossiping [113], probabilistic-based ooding [114] [115], and other controlled-ooding approaches [116] belong to this cate gory In contrast, our zone ooding scheme utilizes the node location information instead of a probabilistic method to control the ooding range. Furthermore, based on the partition strate gy and applying dif ferent routing strate gies to different functional re gions, our data dissemination paradigm can well balance the reliability and the ener gy ef cienc y and achie v e good scalability and notable e xibility Geographic routing is another potential candidate for sensor netw orks. It uses nodes' locations as their addresses based on which pack ets are forw arded to w ards the destination in a greedy manner Se v eral approaches ha v e been proposed to impro v e the ef cienc y and accurac y of the geographic routing. In GPSR [117], a detour algorithm w as proposed to o v ercome the possible dead ends with the greedy method. In addition, LAR [64] uses restricted area ooding to reduce the cost of disco v ery Moreo v er in DREAM [63], each

PAGE 122

110 node obtains its geographic location through e xternal de vices such as GPS, and periodically transmits its location coordinates to other nodes in the netw ork. A source sends a data pack et to a subset of its neighbors in the direction of the destination. Unlik e the abo v e approaches that require a mass of location information e xchange among neighboring nodes or require a node to be a w are of the locations of all the other nodes, our zone ooding scheme only requires a node to learn the locations of its o wn and sink nodes, and to mak e a forw arding decision based on its o wn location and the ooding zone information carried in recei v ed pack ets. Therefore, our scheme demonstrates good scalability Besides, our scheme pro vides the option of multi-zone ooding and hence is more resilient to node or route f ailures. T rajectory based routing (TBF) [118] is another notable w ork in geographic routing. TBF embeds a trajectory in each pack et and intermediate nodes around the trajectory collaborati v ely forw ard a gi v en pack et. Dif ferent from TBF our zone ooding scheme uses tw o curv es to specify a ooding zone and all the nodes in the specied zone will par ticipate in forw arding a gi v en pack et in an autonomous w ay Hence our approach is more reliable than TBF More important, our approach is completely loop free, while trajectory based forw arding still risks possible routing loops in some scenarios lik e other geographic routing. Our data dissemination frame w ork demands the location information of nodes, which is indispensable for man y applications, such as tar get tracking and en vironment monitoring [119], and is fundamental for man y geographic routing schemes ([64,63,117]). There is a rich literature in ad hoc netw orks on ho w to retrie v e the location information as accurate as possible. F or e xample, se v eral approaches based on time dif ference of arri v al (TDO A) [120], angle of arri v al (A O A) [121], or signal strength [122] ha v e been proposed for the scenarios without GPS de vices. More recently tw o no v el approaches ha v e been proposed specically for sensor netw orks [123] [124]. W e belie v e that the location information should be ef fecti v ely utilized to pro vide better performance in wireless sensor netw orks, the more accurate the location information is, the more benet we can get.

PAGE 123

111 A B B C C D E F 2 1 3 p a r t s s h i p m e n t p r o d u c t s s u p p l i e r s p a r t s w a r e h o u s e s m a n u f a c t u r e r s p r o d u c t w a r e h o u s e r e t a i l e r s c u s t o m e r s v o l u m e s h i p m e n t A B D A B C D E B C F Figure 5: An e x emplary supply chain. 5.3 Modelling Sensor Netw orks as Supply Chains In this section, we rst introduce some basics of supply chains and then discuss wh y and ho w we can model a WSN as a supply chain. 5.3.1 Introduction to Supply Chains In the b usiness w orld, a supply c hain (SC) is a series of links and shared processes e xisting between suppliers and consumers, which in v olv e all acti vities from the acquisition of ra w materials to the deli v ery of nished goods to the end consumers [125]. Generally speaking, a supply chain consists of se v eral components: ra w material manuf acturers, inter mediate product manuf acturers, end product manuf acturers, wholesalers and distrib utors, and retailers. These components are connected by transportation and storage acti vities, and are inte grated through information, planning, and inte gration acti vities. Fig.5sho ws an e x emplary SC, where ra w materials or parts enter into a manuf acturing or g anization via a supply system and are transformed into nished goods. The nished goods are then supplied to consumers through a distrib ution system. Usually supply chains are operated with certain strate gy to coordinate all the components' functions and to smooth the material and information o ws in the supply chains. In the past 40 years, SC strate gies ha v e e v olv ed from push strate gies to pull strate gies and nally to push-pull strate gies [125]. In a push-type SC, production and distrib ution decisions are based on long-term forecasts. A manuf acturer uses orders recei v ed from retailer w arehouses to forecast demands. The decisions of consumers' orders are based on in v entory

PAGE 124

112 P u s h S t r a t e g y : M a k e t o s t o c k m o d e l P u l l S t r a t e g y : A s s e m b l y t o o r d e r m o d e l P u s hP u s h--P u l l B o u n d a r yP u l l B o u n d a r yP U S H S T R A T E G Y P U L L S T R A T E G YP u s h P u l l S t r a t e g y : H y b r i d m o d e ls u p p l i e r s A s s e m b l y C o n f i g u r a t i o ns u p p l i e r s A s s e m b l y C o n f i g u r a t i o n Figure 5: Supply chain strate gies. rather than consumers' demands. Bull whip ef fects such as e xcessi v e in v entory e xcessi v e production v ariability and poor service le v els, are v ery common in the push-type SC. In contrast, in a pull-type SC, production and distrib ution are demand-dri v en and are based on actual customer demands rather than forecasts. The rm no longer needs to hold an y in v entory and only responds to specic orders. Ho we v er in a pull-type SC, it is hard to le v erage economies of scale. These adv antages and disadv antages of push strate gies and pull strategies ha v e resulted in hybrid strate gies: push-pull strate gies. In push-pull strate gies, push strate gies are applied in the initial stages such as parts in v entory where production and distrib ution decisions are based on long-term demand forecasts by manuf acturers on the basis of orders recei v ed from retailers; and pull strate gies tak e ef fect in the nal stages such as product assembly where production and distrib ution are purely demand dri v en and rely on actual customer demands rather than forecasts. Usually a b uf fer in v entory such as a w arehouse, is located at the push-pull boundary Dell Computer a giant computer manuf acturer and retailer is an e xcellent e xample of the push-pull-type SC. Dell k eeps an in v entory of components and assembles only when there is an actual order Fig.5sho ws those three dif ferent strate gies. On the other hand, supply c hain mana g ement (SCM) is the act of optimizing all acti vities throughout the supply chain, so that products and services are supplied in the right quantity to the right location, at the right time, and at the optimal cost. One of the fundamental concepts in SCM is that all the autonomous entities in the SC may ha v e their o wn inner operations and management strate gies, b ut the y w ork in a cooperati v e f ashion

PAGE 125

113 to achie v e the management goal of the whole SC: satisfying the customer service requirements while minimizing the o v erall system cost and obtaining as much re v enue as possible from the cooperations. No w adays, SCM has been widely accepted in the b usiness w orld as a vital f actor for competiti v e adv antage and sustainable b usiness impro v ement. 5.3.2 Ho w Could Supply Chains Help Us Similar to supply chains, wireless sensor netw orks are designed to cooperati v ely transmit meaningful sensing data (products) to the sinks (retailers) with certain QoS requirements. F or a designed sensing task, the sensors (ra w material manuf acturers) sense the ra w data (ra w materials) as suppliers and send them to some nodes (product manuf actur ers) for further processing. These processed data (products or sometimes intermediate products) are deli v ered hop-by-hop to sinks with the help from intermediate sensor nodes (transporters). In f act, supply chains and wireless sensor netw orks ha v e man y k e y components or functions in common. F or e xample, the parts w arehouse in Fig.5is designed to consolidate ra w parts from dif ferent suppliers and it serv es as a multiple x er to decouple the need from the a v ailability Its counterpart in sensor netw orks is the desirable functionality of data aggre g ation, which is used to combine the data gradually at inter mediate nodes enroute from dif ferent sensor nodes to the sink node. The objecti v es of the data aggre g ation are eliminating redundanc y minimizing the number of transmissions, and thus sa ving ener gy [126]. Further the product w arehouse close to retailers, ensuring lo w in v entory reduced transportation costs, and quick replenishment capability acts a similar role as that of the mechanisms for reducing information implosion [105]. V ital to cost management of supply chains, transportation planning is equi v alent in functionality to routing in sensor netw orks, whose major objecti v e is to transfer sensing data from sensors to sink nodes as ef ciently as possible. Moreo v er both require all the entities in the system to w ork cooperati v ely for goals common in nature: pro viding good quality of service (QoS) and k eeping the o v erall system cost as lo w as possible. In T able5we list the analogous components between supply chains and sensor netw orks. It is a rather soft matching in

PAGE 126

114 T able 5: Analogue between supply chains and wireless sensor netw orks Items in supply chains Counter parts in sensor netw orks Ra w materials or parts Phenomena of interest, for e xample, sound, images, and mo v ements of observ ed objects Suppliers or manuf acturers Sensor nodes generating sensing data T ransportation netw ork Intermediate sensor nodes Distrib utors or retailers Sink nodes Finished products Data processed by sink nodes Consumers End-users of the data of fered by sink nodes that not e v ery sensor node can be absolutely or perfectly t into a matching component in a supply chain architecture. A sensor may act as multiple roles under dif ferent scenarios. F or e xample, one sensor may be a supplier of one communication, while be a tr ansporter of the other communication in the meantime. Intermediate nodes sometimes are also called semi-manuf acturers or parts w arehouses when performing data aggre g ation to reduce the data redundanc y What we emphasize here is the conceptual and technical moti v ations and the approaches leading to the viable solutions behind the tw o seemingly dif ferent systems. Interestingly we notice that tw o notable routing protocols for sensor netw orks, namely directed dif fusion [107] and SPIN [108], conform well with the SC methodology In particular directed dif fusion can be vie wed as a pull-type SC, in which the sink node propag ates its interests throughout the sensor netw ork and sensors possessing the data of interest respond with sensing data via intermediate sensors. Thus, directed dif fusion is a pull-type SC in that pr oduction is demand (inter est) driven In contrast, SPIN with AD V -REQ-D A T A handshaking is more lik e a push-type SC. It is designed for disseminating information to all nodes in a sensor netw ork, where the nodes generating the data can be re g arded as suppliers or manuf acturers in a SC. Thus, SPIN is a push-type SC where or der decisions (REQ) ar e based on in ventory (AD V) The abo v e resemblance between supply chains and wireless sensor netw orks moti v ates us to model a WSN as a SC, thus we can apply the sophisticated kno wledge of SCM in the b usiness w orld to impro v e the performance of the sensor netw orks. More specically we

PAGE 127

115 C o n s u m e r 1 C o n s u m e r 2 C o n s u m e r 3 C o n s u m e r 4 I a m i n t e r e s t e d i n t h e b e a k w h a t c o l o r ? h o w l o n g ? S e r v i c e a r e a W a r e h o u s e a r e a T r a n s p o r t a t i o n a r e a M a n u f a c t u r e a r e a C a n y o u s e n d m e t h e c h i r m o f t h e b i r d ? I c a n t e l l y o u t h e n a m e o f t h e b i r d b y i t s c h i r m L e t m e s e e t h e w h o l e p i c t u r e o f t h e b i r d I c o l l e c t p h o t o s o f f e e t o f a l l k i n d s o f b i r d s T h e y a r e b e a u t i f u l A B C D E F G H I J K L M N O P Q R S T w 4 w 5 w 6 w 7 w 8 w 9 R 3 R 2 w 1 w 2 w 3 Z o n e 1 Z o n e 2 R 1 R 4P u l l P u s h B o u n d a r yP U S H S T R A T E G Y P U L L S T R A T E G Y Figure 5: A system architecture for habitat monitoring kno w that it yields better performance and lo wer cost in SCM by partitioning the supply chain into se v eral dif ferent components ( partition str ate gy ), applying dif ferent management mechanisms to dif ferent components ( hybrid str ate gy ), and designing cooperations among dif ferent components ( cooper ation str ate gy ). Similar ideas can be introduced into sensor netw orks for solving reliability and ef cienc y problems. 5.4 A Hybrid Data Dissemination Frame w ork for W ireless Sensor Netw orks Inspired by SCM methodology we design a no v el h ybrid data dissemination framew ork for wireless sensor netw orks. In this section, we rst describe the system model and then detail the management strate gies or routing techniques applied to dif ferent functional components. 5.4.1 System Model In what follo ws, we use a WSN for habitat monitoring as an e xample to illustrate our scheme. As sho wn in Fig.5, for one specic sensing task, the whole sensor eld is conceptually partitioned into se v eral functional areas according to the aforementioned push-pull strate gies of supply chains (see Section5.3.1). In the manufactur e ar ea some nodes such as those from A to J are in v olv ed in generating the ra w data about the objects of interest, that is, birds in this case, while other nodes such as K L and P are responsible

PAGE 128

116 for data aggre g ation, that is, consolidating the ra w data and reducing possible information o v erlapping. The ltered data is fed into the tr ansportation ar ea to be relayed by inter mediate sensors to sink nodes. In addition, we introduce the war ehouse ar ea as a b uf fer area between the transportation area and the service area to reduce the possible traf c congestion and information implosion [105] at the sink nodes. The service area consists of sink nodes which can directly communicate with each other through f ast and reliable links, either wired or wireless. The sink nodes perform collaborati v e reception of sensing e v ents and of fer dif ferent data items to end-users or consumers with dif ferent interests. W e assume that each node has the kno wledge of its o wn position1and the positions of the sink nodes, which is a reasonable assumption for man y monitoring applications [105,106]. W e further assume that all the nodes including common sensor nodes and sink nodes are identied by their geographic information. W e should emphasize here that the sensor eld partition is only conceptual and application-dependent rather than a x ed one. Basically v arious sensing tasks may ha v e quite dif ferent partition instances. In particular for each specic sensing task, those nodes sensing the e v ents of interest can form a manuf acture area together with their neighboring nodes just for that task. Accordingly the transportation area lies in the forw ard direction from the manuf acture area to w ards the sink nodes. In this dissertation, we assume that the sinks are f ar enough a w ay from the manuf acture area, and thus the transportation area does e xist2. Moreo v er we dene the w arehouse area to be the area within the sink nodes' n -hop range, where n is a tunable design parameter T o form its w arehouse area, one sink node needs to simply broadcast a special request with the TTL v alue set to n An y node recei ving this request becomes a w arehouse member 1 Note that a node can obtain its location information at lo w cost from GPS or some localization system [120,122,121,123,124]. 2 F or those sensing tasks close to the sinks, a modied partition method can be used where transportation area may not be necessary .

PAGE 129

117 (w arehouse node) of the w arehouse area of the requesting sink node. In f act, the v alue of n can be application-dependent, and v arious sensing tasks may ha v e dif ferent v alues of n to reect the di v erse QoS requirements. In contrast to the abo v e three areas whose locations and sizes are closely related to sensing tasks, the service area is determinate because of the in v ariable locations of sink nodes. One of the no v el features of our h ybrid data dissemination paradigm is that we apply dif ferent data forw arding mechanisms in dif ferent functional areas. More specically local broadcasting in the manuf acture area, a unicasting-based routing in the w arehouse area, and a specially designed zone ooding, which is essentially a combination of con v entional ooding and geometric routing, in the transportation area, are used respecti v ely W ith this h ybrid data dissemination paradigm in place, a better balance between ener gy ef cienc y and reliability can be e xpected. In what follo ws, the h ybrid data dissemination paradigm is elaborated in more detail. 5.4.2 Manuf acture Area W e postulate that the nodes in the manuf acture area are a w are of their o wn missions3. Each mission might represent a sensing task of the sensor netw ork. In this e xample, the mission may be collecting the information of birds, such as the beak color the feet length, or e v en the bird chirm. Due to the limitation of sensors' capabilities, each sensor may only sense part of the interested e v ent so that the y might need to locally e xchange some sensing e v ents and select among themselv es one node as an aggre g ation center to fulll the data fusion task. F or e xample, nodes K L and P in Fig.5are selected as aggre g ation centers. Since aggre g ation centers, in most cases, are only se v eral hops a w ay from the sensing nodes, the simplest w ay to forw ard the sensed ra w data to aggre g ation centers is 3 Here we assume that certain mission allocation scheme is used during the deplo yment and initialization phases of sensor netw orks.

PAGE 130

118 to broadcast pack ets with limited TTL v alues. F or lack of space, we do not detail ho w to manage the sensing tasks and accomplish data aggre g ation in this disseratation. Besides data fusion, each aggre g ation center assumes a special role in our data dissemination frame w ork. It is also responsible for determining the transportation method for the ltered data by itself, that is, using single zone ooding or multi-zone ooding, and the proper transportation zone(s) through which the data will tra v el in the tr ansportation ar ea F or e xample, after nishing the aggre g ation of the ra w data from nodes E F and G node P mak es choice of using tw o ooding zones and then chops the ltered data into tw o parts, both of which are labelled with their respecti v e designated ooding zone. In f act, the operations in the manuf acture area e xhibit lots of e xibility during the selection of the transportation methods and the ooding zone(s). W ith proper selection, our scheme can strik e a good balance between reliability and ener gy ef cienc y F or e xample, if the en vironment is good, an aggre g ation center can choose a single ooding zone instead of multiple ooding zones which are usually used in the f ace of harsh en vironment to forw ard the data. Moreo v er if no w arehouse area is allocated and the ooding zone is the whole sensor eld, our zone ooding e xpands to blind ooding; on the other hand, if we can squash the ooding zone into an area containing only a single path, our zone ooding reduces to single path routing. F or a b ursty and b ulk y e v ent report, an aggre g ation center can choose multiple ooding zones and split the report into se v eral portions which are simultaneously deli v ered to multiple sinks through dif ferent ooding zones. Since sink nodes can communicate with each other through f ast and reliable links, the y can e xchange the recei v ed portions and easily reconstruct the original e v ent report. In this w ay a reduction in the end-to-end delay can be e xpected. Such multipath approaches are also well kno wn for their ef fecti v eness in increasing the reliability and security of data disseminations [127]. Moreo v er to combat the node or wireless link f ailures especially in a harsh en vironment, an aggre g ation center may introduce some redundanc y by sending duplicate reports to the sinks through multiple ooding zones. W e should note that, aggre g ation centers can v ary

PAGE 131

119 the sizes and the locations of ooding zones with time in order to distrib ute the traf c load more e v enly among sensors and thus a v oid o v erb urdening a small set of nodes. Besides, if proper scheduling is a v ailable, nodes can enter into w ak e or sleep modes zone by zone to sa v e ener gy In what follo ws, we shall discuss ho w an aggre g ation center chooses appropriate ooding zones, what data structure a data pack et manufactur ed by an aggre g ation center may use, and ho w a node in the w arehouse area processes a zone-ooded pack et. 5.4.3 T ransportation Area Sensor nodes in the transportation area undertak e the task of relaying data to possible multiple sink nodes. T o a v oid costly topology maintenance and comple x route disco v ery algorithms, we propose a no v el zone ooding scheme, which is a combination of con v entional ooding and geometric routing techniques. The basic idea is as follo ws: a zone containing the source and the destination is specied by some geometric means; then, instead of netw ork-wide omnidirectional ooding, the zone ooding is guided along the direction from the source to the destination and is restricted in the designated zone. Once a node recei v es a pack et carrying parameters that identify a ooding zone, it rst checks whether or not it is in the indicated zone through some tri vial calculations based on its o wn location information and the recei v ed zone parameters. Only when located in the ooding zone w ould it rebroadcast the pack et. Figure 5-3 showsanexampleofthezoneoodingscheme,inwhichellipses 4 areused to specify the ooding zones. Suppose one of the aggre g ation centers, say node P has the coordinates ( x 1 ; y 1 ) in the cartesian plane of Fig.5, and the intended sink node R 3 has the coordinates ( x 2 ; y 2 ). W e should mention that all the required information for zone ooding are embedded in the transmitted data pack ets rather than the control pack ets. 4 As we will see, ellipse can be represented with lo w o v erhead. More important, it is easy to manipulate.

PAGE 132

120 P ( x1, y1) R3 ( x2, y2) U ( x3, y3) X Y f ( x y ) g ( x y ) F1 fF2 fF1 gF2 gO O Y X 1 2 3 4 I I I I I I U U U U C a s e 1 : b 1 b 2 ( b 1 < b 2 ) C r i t e r i a : y 3 = ( y 3 ( y 1 + y 2 ) / 2 ) > 0 a n d L ( b 1 ) < 2 a < L ( b 2 ) C a s e 2 : b 1 b 2 ( b 1 < b 2 ) C r i t e r i a : y 3 = ( y 3 ( y 1 + y 2 ) / 2 ) < 0 a n d L ( b 1 ) < 2 a < L ( b 2 ) C a s e 3 : b 1 b 2 ( b 1 < b 2 ) C r i t e r i a : y 3 = ( y 3 ( y 1 + y 2 ) / 2 ) > = 0 a n d L ( b 2 ) < 2 a o r y 3 = ( y 3 ( y 1 + y 2 ) / 2 ) < 0 a n d L ( b 1 ) < 2 a C a s e 4 : b 1 b 2 ( b 1 < b 2 ) C r i t e r i a : y 3 = ( y 3 ( y 1 + y 2 ) / 2 ) > = 0 a n d L ( b 1 ) < 2 a o r y 3 = ( y 3 ( y 1 + y 2 ) / 2 ) < 0 a n d L ( b 2 ) < 2 a .P R3P R3P R3P R3Y Y Y Y Y X X X X O O O O Figure 5: The forw arding-decision-making process of nodes in the transportation area. A C l o c a t i o n S i n k l o c a t i o n I n n e r S e m i m i n o r A x i s P a y l o a d I n t e r e s t ( e v e n t ) D e s c r i p t i o n ( l o c a t i o n t i m e t y p e ) O t h e r C o n t r o l F i e l d s ( m u l t i p l e z o n e f l o o d i n g r e d u n d a n c y . ) W a r e h o u s e F l a g O u t e r S e m i m i n o r A x i sFigure 5: An e x emplary pack et structure.

PAGE 133

121 Figure 5-5 showsthetypicalstructureofadatapacketmanufacturedbyanaggregation node. Besides the ltered data, each data pack et sent from node P will carry four e xtra zone parameters: A C location indicating the coordinates of the aggre g ation node P Sink location indicating the coordinates of the sink node R 3 and Inner -SemiminorAxis and Outer -SemiminorAxis indicating the semiminors of the inner and outer ellipses5of the desired ooding zone respecti v ely The eld W arehouse Flag is used to inform intermediate nodes whether this pack et has e v er been processed by a w arehouse node. The usage of this eld will be discussed in Section5.4.4. In addition, based on the elds Interest (e v ent) Description and Other Control Fields, intermediate nodes and sinks can determine whether one recei v ed pack et has already been processed, or it needs to assemble those partitioned pack ets belonging to the same interest, or to remo v e the possible redundanc y added. The concrete use of these elds depends on specic applications, which is be yond the scope of this dissertation. From the ellipse geometry we kno w that when the tw o endpoints of the major axis are x ed (that is, A C location of P and Sink location of R 3 ), a v alue of the semiminor axis can uniquely determine an ellipse. And tw o dif ferent semiminor v alues (that is, Inner SemiminorAxis b 1 and outer -SemiminorAxis b 2 ) will determine tw o ellipses as sho wn in Fig.5, f ( x; y ) and g ( x; y ) Such ellipses can be used to specify multiple ooding zones. F or e xample, the tw o ellipses with the same endpoints of the major axis can jointly deter mine three non-o v erlapping ooding zones Zone I between curv e 1 and curv e 2, Zone II between curv e 2 and curv e 3, and Zone III between curv e 3 and curv e 4. In f act, an y tw o of the four curv es can specify a ooding zone, for e xample, a bigger zone determined by curv e 1 and curv e 3. W e already kno w that, with the tw o endpoints of the major axis x ed, the semiminor axis uniquely determines an ellipse. But we need a means to dif ferentiate 5 Here the inner ellipse means the ellipse with smaller semiminor axis, while the outer ellipse means the ellipse with bigger semiminor axis.

PAGE 134

122 the tw o curv es constituting the ellipse the upper half (abo v e the X 0 axis, for e xample, curv e 1 and 2) and the lo wer half (belo w the X 0 axis, for e xample, curv e 3 and 4). In our design, we adopt a simple rule we use the positi v e v alue, for e xample, b 1 or b 2 to denote the upper half of an ellipse while using the ne g ati v e v alue to denote the lo wer half curv e of the ellipse in the shifted coordinate plane. F or e xample, b 1 together with b 2 determines zone I, )Tj/T1_1 11.95509 Tf9.28912 0 Td(b 1 and )Tj/T1_1 11.95509 Tf9.28912 0 Td(b 2 jointly specify zone III, and )Tj/T1_1 11.95509 Tf9.28912 0 Td(b 1 and b 2 select zone I+II. In particular Inner -SemiminorAxis b 1 can be 0 to denote the X 0 axis as a special elliptic curv e. Thus, a f (+ = )Tj/T1_6 11.95509 Tf9.28912 0 Td() b 1 (+ = )Tj/T1_6 11.95509 Tf9.28912 0 Td() b 2 g pair of real numbers can uniquely determine a ooding zone. Moreo v er by v arying the v alues of semiminor axis, we can easily get ph ysically separated or interlea v ed multiple paths (multiple ooding zones) without incurring an y signicant additional costs. As discussed in Section5.4.2, the multipath routing is a po werful technique to impro v e the reliability and security among other system performance f actors. When one node, say U with coordinates ( x 3 ; y 3 ), recei v es a data pack et containing the abo v e-mentioned zone parameters f ( x 1 ; y 1 ) ; ( x 2 ; y 2 ) ; (+ = )Tj/T1_6 11.95509 Tf9.28912 0 Td() b 1 ; (+ = )Tj/T1_6 11.95509 Tf9.28912 0 Td() b 2 g the question whether or not it should rebroadcast the pack et is reduced to a simple geometric problem: whether or not point U lies between the tw o elliptic curv es specied by the embedded parameters? Suppose the semiminor axis of an ellipse with tw o major -axis endpoints P and R 3 is b Then the sum of the distance from point U to tw o x ed points F 1 and F 2 (the foci) can be e xpressed as L ( b ) = D 1 + D 2 [128], where D 1 = r ( q ( x 1 )Tj/T1_10 7.97011 Tf6.5833 0 Td(x 2 ) 2 +( y 1 )Tj/T1_10 7.97011 Tf6.5833 0 Td(y 2 ) 2 4 )Tj/T1_7 10.9091 Tf10.89729 0 Td(b 2 + ( x 3 )Tj/T1_10 7.97011 Tf12.105 4.49101 Td(x 1 + x 2 2 )) 2 + ( y 3 )Tj/T1_10 7.97011 Tf12.105 4.93201 Td(y 1 + y 2 2 ) 2 and D 2 = r ( q ( x 1 )Tj/T1_10 7.97011 Tf6.5833 0 Td(x 2 ) 2 +( y 1 )Tj/T1_10 7.97011 Tf6.5833 0 Td(y 2 ) 2 4 )Tj/T1_7 10.9091 Tf10.9063 0 Td(b 2 )Tj/T1_8 10.9091 Tf10.89729 0 Td(( x 3 )Tj/T1_10 7.97011 Tf12.105 4.49098 Td(x 1 + x 2 2 )) 2 + ( y 3 )Tj/T1_10 7.97011 Tf12.105 4.93198 Td(y 1 + y 2 2 ) 2 : T o determine whether a node itself is in the specied ooding zone, it needs to compare the distance to the foci of each ellipse with 2 a where a = p ( x 1 )Tj/T1_1 11.95509 Tf11.95309 0 Td(x 2 ) 2 + ( y 1 )Tj/T1_1 11.95509 Tf11.94411 0 Td(y 2 ) 2 = 2 is the semimajor axis of the ellipse with major -axis endpoints P and R 3 In Fig.5we

PAGE 135

123 gi v e four possible cases and the corresponding decision criteria that a node may lie in the specied ooding zone. Therefore, for Case 1 where b 1 < b 2 6= 0 in our e xample (Fig.5), node U needs to rebroadcast the pack et if ( y 3 )Tj/T1_6 7.97011 Tf12.85201 5.25604 Td(y 1 + y 2 2 ) > 0 and L ( b 1 ) < 2 a < L ( b 2 ) Otherwise, it will simply ignore the pack et because it is not in the designated ooding zone for that pack et. F ollo wing the abo v e procedures, sensor nodes ( tr ansporter s ) in the transportation area can nally relay a data pack et to the w arehouse area through multi-hop wireless links. In our e xample, since we use tw o elliptic curv es to specify one ooding zone, the Flooding Zone P arameters eld only needs to include tw o v alues for the inner and the outer semiminor ax es, respecti v ely Although an y tw o noncrossing curv es sharing the same tw o ends could be used to specify a ooding zone, we should cautiously choose those curv es that not only can be represented with as fe w bytes as possible to reduce the communication o v erhead, b ut also can simplify the forw arding-decision-making processes of intermediate nodes. In this sense, arcs and ellipses are tw o promising candidates. Moreo v er a ooding zone specied by tw o curv es should be wide enough to ha v e suf cient nodes to forw ard the pack ets while maintaining high ener gy ef cienc y in the meantime. T o accomplish this, the aggre g ation center in our e xample, say node P should properly choose the v alues of the semiminors of the tw o ellipses. Besides, to balance the nodal usage in the transportation area, aggre g ation centers should v ary the ooding zones by using dif ferent and alternating ne g ati v e and positi v e v alues for semiminor ax es. By doing so, our scheme could achie v e e v en load balance and f air ener gy usage without incurring an y signicant additional costs. 5.4.4 W arehouse Area and Service Area In the e x emplary SC sho wn in Fig.5, the w arehouse near the retailers creates a break point in the mo v ement of the products and acts as a b uf fer to reduce the cost of stock at the retailers, hence impro ving the e xibility of the retailers. The w arehouse frequently updates its in ventory list to the retailers, and the retailers can quickly get the products out

PAGE 136

124 W a r e h o u s e A r e a S e r v i c e a r e a w 4 w 5 w 6 w 7 w 8 w 9 R 3 R 2 w 1 w 2 w 3 A D V R E Q D A T A R 1 R 4 Figure 5: The routing process in the w arehouse area. of stock in the stores replenished from the w arehouse. In addition, the w arehouse may consolidate small shipments into a lar ger shipment to the same retailer to sa v e transportation costs. W e notice that such a w arehouse component is also needed in sensor netw orks for realtime and continuous monitoring applications. In these applications, b ursty and b ulk y traf c may be simultaneously transmitted to sink nodes, as a result of which notorious traf c congestion may happen frequently in the vicinity of sink nodes and thus cause the unf a v orable loss of information and the w aste of scarce netw ork resources. Moreo v er the redundant pack ets ooded to w ards sinks may result in the information implosion problem as well. Thus, the introduction of the w arehouse area as a b uf fer area can decouple need ( inter ests ) from a v ailability ( r edundant e vent r eports ) and hence help mitig ate the abo v e information implosion problem and possible traf c congestion. F or the w arehouse area, we use a modied SPIN [108] instead of zone ooding as the underlying routing protocol. Dif ferent from the AD V -REQ-D A T A e xchange in SPIN that happens between neighbors, the AD V -REQ-D A T A e xchange in our scheme is between a w arehouse node and a sink node se v eral hops a w ay and is realized by end-to-end unicast. The unicasting path could be established using the Destination-Sequenced Distance-V ector Routing protocol (DSD V) [7]. W e note that only those w arehouse nodes may participate in

PAGE 137

125 the routing maintenance acti vities. W ith a limited w arehouse size, small o v erhead of routing maintenance can be e xpected. Of course, other routing schemes are applicable to this area as well. Fig.5illustrates our routing strate gy used in the war ehouse ar ea Once recei ving data pack ets from outside the w arehouse area (by e xamining the eld W arehouse Field), a w arehouse member of the desired sink of the pack ets (cf. Section5.4.1), say W 7 lying on the boundary of the w arehouse area, rst sets the eld W arehouse Flag and temporarily stores the pack ets. Afterw ards, W 7 the data holder will unicast an AD V message, essentially an in ventory containing the descriptors of stored pack ets, to the tar geted sink node R 3 either on a per -pack et basis or periodically or when the number of stored pack ets e xceeds a threshold. F ollo wing the reception of the AD V message, R 3 will send a REQ message requesting the interested data. Upon recei ving the REQ message, W 7 can unicast the requested data to R 3 via a D A T A message. In case that R 3 does not recei v e the requested data in time after sending out a REQ, it can resend a REQ to the same data holder W 7 or another data holder who also sent to it an AD V containing the descriptors of the same data. Note that, just as what the w arehouse does in the b usiness SC, if a w arehouse node has enough storage space, it can periodically consolidate se v eral stored pack ets and send them in a v olume shipment to the same sink, in which w ay ener gy sa vings can be e xpected. Here we w ant to e xplain ho w the w arehouse area can help reduce the information implosion. Suppose pack ets describing the same e v ent for sink R 3 arri v e at both W 4 and W 7 which are located in the same ooding zone and the w arehouse area of R 3 When no AD V -REQ-D A T A e xchange is used, depending on the routing protocol, both W 4 and W 7 will broadcast or unicast the pack ets for the same e v ent to the sink R 3 which will lead to information implosion and ener gy w aste o wing to redundant pack ets, especially when numerous nodes send the same e v ent report to the sink. Ho we v er using the proposed AD V REQ-D A T A procedure, both W 4 and W 7 will send short AD V messages rather than relati v e long data pack ets to R 3 It is up to R 3 to mak e a decision on which data holder it should

PAGE 138

126 pull the data from based on certain criteria such as hop count or delay Suppose W 7 is chosen, R 3 will send a REQ to W 7 and accordingly W 7 can unicast the requested data in a D A T A message to R 3 After a certain period, W 4 may delete the stored stale data. From this e xample, we can see that redundant pack ets can be successfully eliminated by the means of AD V -REQ-D A T A e xchange. Moreo v er sink nodes in the service area perform collaborati v e reception in the sense that the y could communicate with each other through f ast and reliable means, for e xample, wired links or separate wireless channels. F or e xample, if sink R 2 recei v es an AD V message from node W 1 it can contact other sink nodes f ar from W 2 to see if the y need the pro vided data, though R 2 itself may not need it in some cases. Suppose R 4 needs the data, R 2 can help obtain the data from W 1 and send it to R 4 Such collaboration also helps a w arehouse node deal with cases when the w arehouse node has no unicasting route to the desired sink of the report due to node f ailures or other reasons. In such a case, the w arehouse node may choose to send the AD V message to a nearest sink. F or instance, in Fig.5, W 1 is a common w arehouse member of both sinks R 2 and R 3 F or some reason, W 1 temporarily has no unicasting route to the reports' destination R 3 and thus it unicasts an AD V to R 2 instead. After recei ving the AD V from R 2 R 3 may request R 2 to help collect the other portions of the information which are destined to R 3 After collecting al the portions, R 3 can reconstruct the original information manuf actured by node P and serv e consumers later 5.4.5 Discussion Our h ybrid data dissemination frame w ork is an open frame w ork and allo ws dif ferent routing techniques to coe xist in the same netw ork. In f act, depending on the applications, a v ariety of routing protocols, not limited to the ones presented in this chapter can nd their possible applications in this frame w ork. In addition, similar to its support of scalability in database engineering [129], the partitioning strate gy enables our scheme to be more scalable and e xible, and reduces the dif culty of designing a feasible o v erall routing scheme.

PAGE 139

127 The decision on ho w to partition the sensor eld and what routing technique should be used for each functional area depends on the application requirements and sensor eld features. F or e xample, multipath routing could be used in the w arehouse area if it is in a v ery harsh en vironment. Further with more information about the sensor eld, the ooding zones can be specied in an ef cient and adapti v e w ay F or e xample, some irre gular curv es can be used to a v oid the ph ysical obstacles in the eld or to a v oid the dead zone with sparse sensor nodes or empty of sensor nodes. As an another e xample, a ooding zone can be specied in a smart w ay such that it co v ers a relati v e small area where communications are of good quality while co v ers a relati v e lar ge area where communications are of bad quality Moreo v er our zone ooding scheme does not e xclude other ooding enhancement mechanisms. Note that our zone ooding scheme attempts to impro v e the ener gy ef cienc y of ooding by restricting the ooding range in the spatial domain. W e can further impro v e the ener gy ef cienc y in the temporal domain. F or e xample, if wireless links are reliable enough, a redundanc y elimination technique can be enabled to further optimize the pack et ooding, which w orks as follo ws: sensor nodes k eep track of redundant pack ets recei v ed o v er a short time interv al, termed Random Assessment Delay (RAD), randomly chosen from a uniform distrib ution between 0 and T max seconds, where T max is the longest possible delay; each node needs to rebroadcast one gi v en pack et if not recei ving redundant ones during the RAD. This RAD method is designed to reduce the collisions among neighboring nodes and to eliminate unnecessary transmissions for one pack et. Besides, ener gy information can be incorporated into the routing decision. F or e xample, an ener gy-ef cient cost metric can be used in the w arehouse area to set up ener gyef cient paths. In the abo v e RAD enhancement, residual ener gy information can also be mapped into delays such that nodes with more ener gy can assume smaller delays than those

PAGE 140

128 nodes short of ener gy Actually ener gy ef cienc y can be further impro v ed in other dimensions. F or instance, if the ooding zone is properly chosen, we can mak e use of Gossipingbased ooding [113], probabilistic-based ooding [114] [115], and other controlled-ooding approaches to further impro v e the ener gy ef cienc y Considering the f act that some sensor nodes may ha v e much more po werful capabilities in terms of ener gy and other important resources than common sensor nodes, more enhancement schemes can be proposed to mak e use of such heterogeneity to operate and manage the whole sensor netw ork more ef fecti v ely and ef ciently In some cases, such po werful nodes may perform data fusion and other management acti vities. The y may e v en act as freight agencies as what UPS or Fede x does in the b usiness supply chains. The aggre g ation centers can send their data to these freight agencies and further these agencies multiple x the data and ship the data in a manner of mass transportation through some high speed data link. In some circumstances, some mobile sensor nodes may be deplo yed in the eld. In such cases, the mobile sensor nodes can perform as mobile data collectors, patrolling around some area and collecting data and carrying them to freight agencies or data processing agencies. These mobile nodes can greatly f acilitate the data forw arding process. Moreo v er in practice, it is possible to replace the depleted nodes near the sinks, or deplo y some nodes with solar po wer supply in the eld. 5.5 Performance Ev aluation In this section, we e v aluate the performance of our h ybrid data dissemination framew ork through simulations. W e rst describe the simulation congurations, the performance metrics, and our simulation methodology W e then study the impact of some en vironment and control parameters on the performance of our scheme. W e compare our scheme with ooding and directed dif fusion [107] in terms of ener gy ef cienc y and reliability 5.5.1 Methodology and Metrics T o v alidate the ef fecti v eness and ef cienc y of our proposed scheme (denoted by RRP), we ha v e de v eloped an e v aluation en vironment within GlomoSim [130] and implemented

PAGE 141

129 s r c 1 s r c 2 s r c 3 s i n k 1 s i n k 2 s i n k 3 1 2 3 4 5 Figure 5: The simulated sensor eld. our h ybrid data dissemination paradigm, including the zone ooding scheme designed for the transportation area and the unicast-based AD V -REQ-D A T A e xchange for the w arehouse area. W e simulate a sensor eld consisting of 606 sensor nodes. W e ha v e 3 independent equally-spaced sources at the left boundary of the sensor eld and 3 independent equallyspaced sinks at the right boundary of the sensor eld. The other 600 sensors are uniformly deplo yed in the sensor eld. The simulated netw ork is sho wn in Fig.5, where the sensor eld is composed of the transportation area and the w arehouse area only The manuf acture area and the service area are on the boundaries of the eld and omitted for simplicity Besides, each of the three sources generates a 128-byte data pack et (e v ent) destined to a randomly selected sink e v ery 1.5 seconds, 1.0 second, and 2.0 seconds, respecti v ely Since our purpose is to study the routing performance of the h ybrid data dissemination framew ork, with an emphasis on the zone ooding scheme, there is no data fusion and collaboration among sink nodes implemented in our simulation. Therefore, a source acts as both an e v ent observ er and a data aggre g ation center T able5lists the conguration parameters of our simulation, where the transmission/reception po wer consumption of sensors are in line with those of Motes [131]. In our simulation, we v ary the size of the w arehouse area, dened as the area within a sink node' n -hop range, to see its impact on RRP' s performance. Besides, each sourcesink pair uses equally-spaced elliptic curv es to partition the entire sensor eld into d (another control parameter) non-o v erlapping ooding zones. F or e xample, Fig.5sho ws the curv es used by source sr c 2 and sink 2 where d = 4 W e e v aluate the impact of the ooding zone size on RRP' s performance by v arying the v alue of d Since the eld size is

PAGE 142

130 T able 5: Simulation conguration Simulation Area 500 m 300 m Number of Nodes 606 T ransmission Range 40 m Initial Ener gy 60 J T ransmit Po wer 81 mW Recei v e/Idle Po wer 30 mW MA C Protocol IEEE 802.11 Radio Bandwidth 2 M bps Data P ack et 128 B y tes Directed Dif fusion Interests 36 B y tes AD V/REQ 12 B y tes Simulation time 15 min x ed, the bigger the v alue of d the narro wer the ooding zone is. Furthermore, we introduce an en vironment parameter namely pack et error rate ( PER ), to reect the error -prone natures of wireless links in WSNs. F or simplicity we assume the error properties of all radio transmissions are independent b ut ha v e the identical PERs during one simulation run. Therefore, by v arying the control parameters n and d and the en vironment parameter PER we can study the performance of our proposed RRP under dif ferent settings. Moreo v er we compare our RRP with ooding and directed dif fusion [107] in terms of ener gy ef cienc y and reliability In our implementation, directed dif fusion uses delay as the criterion to select preferred neighbors during reinforcement. F or our RRP we use the AD V -REQ-D A T A e xchanges (cf. Section5.4.4) and our proposed zone ooding (cf. Section5.4.3) as underlying routing techniques for w arehouse area and transportation area, respecti v ely There are four performance metrics of interest to us. The e v ent deli v ery ratio (EDR) reecting the reliability is dened as the ratio of the total number of e v ent reports (data pack ets) that are successfully deli v ered from the sources to the intended sinks o v er the total number of pack ets generated at the sources. The normalized ener gy consumption reecting the ener gy ef cienc y is dened as the ener gy consumption per pack et per node normalized by the ener gy consumption for one single pack et reception. The a v erage e v ent

PAGE 143

131 end-to-end delay is dened as the a v erage delay from when a pack et (e v ent) is generated and transmitted by the source till it is recei v ed by the sink. And the a v erage r outing o v erhead is dened as the a v erage number of routing pack ets generated per data pack et. 5.5.2 Simulation Results F or the rst set of gures (Fig.5Fig.5), we x the number of ooding zones to be 4, that is, d = 4 and study the performance of RRP under dif ferent en vironment conditions, that is, dif ferent v alues of pack et error rate ( PER ), and with dif ferent w arehouse size n 0 0 0 0 1 0 0 0 2 0 0 0 3 0 0 0 4 0 0 0 5 0 0 0 6 0 0 0 7 0 0 0 8 0 0 0 9 0 0 1 0 8 6 0 8 8 0 9 0 9 2 0 9 4 0 9 6 0 9 8 1 1 0 2 P a c k e t e r r o r r a t e ( P E R )E v e n t d e l i v e r y r a t i o ( E D R ) d i r e c t e d d i f f u s i o n f l o o d i n g R R P ( n = 1 ) R R P ( n = 2 ) R R P ( n = 3 ) R R P ( n = 4 ) R R P ( n = 5 ) Figure 5: Ev ent deli v ery ratio vs. pack et error rate. 0 0 0 0 1 0 0 0 2 0 0 0 3 0 0 0 4 0 0 0 5 0 0 0 6 0 0 0 7 0 0 0 8 0 0 0 9 0 0 1 0 2 4 6 8 1 0 1 2 1 4 1 6 1 8 P a c k e t e r r o r r a t e ( P E R )N o r m a l i z e d e n e r g y c o n s u m p t i o n d i r e c t e d d i f f u s i o n f l o o d i n g R R P ( n = 1 ) R R P ( n = 2 ) R R P ( n = 3 ) R R P ( n = 4 ) R R P ( n = 5 ) Figure 5: Normalized ener gy consumption vs. pack et error rate. Figure 5-8 compareseventdeliveryratioofRRP,pureooding,anddirecteddiffusion under dif ferent PERs. As we can see, since directed dif fusion maintains single path

PAGE 144

132 0 0 0 0 1 0 0 0 2 0 0 0 3 0 0 0 4 0 0 0 5 0 0 0 6 0 0 0 7 0 0 0 8 0 0 0 9 0 0 1 0 1 0 1 5 0 2 0 2 5 0 3 0 3 5 0 4 0 4 5 0 5 P a c k e t e r r o r r a t e ( P E R )E v e n t e n d t o e n d d e l a y ( s ) d i r e c t e d d i f f u s i o n f l o o d i n g R R P ( n = 1 ) R R P ( n = 2 ) R R P ( n = 3 ) R R P ( n = 4 ) R R P ( n = 5 ) Figure 5: A v erage e v ent end-to-end delay vs. pack et error rate. 0 0 0 0 1 0 0 0 2 0 0 0 3 0 0 0 4 0 0 0 5 0 0 0 6 0 0 0 7 0 0 0 8 0 0 0 9 0 0 1 0 5 1 0 1 5 2 0 2 5 3 0 3 5 P a c k e t e r r o r r a t e ( P E R )A v e r a g e r o u t i n g o v e r h e a d d i r e c t e d d i f f u s i o n f l o o d i n g R R P ( n = 1 ) R R P ( n = 2 ) R R P ( n = 3 ) R R P ( n = 4 ) R R P ( n = 5 ) Figure 5: A v erage routing o v erhead vs. pack et error rate. for each source-sink pair its EDR is v ery sensiti v e to the change of PER, dropping almost linearly from 99 % to 86 % with the increase of PERs. In contrast, the EDRs of ooding al w ays stabilize around 100 % This result is not surprising because of the inbred reliability of ooding techniques. Our RRP demonstrates a stable EDR greater than 99 % under all v e dif ferent w arehouse sizes. This result indicates that our RRP using zone ooding in the transportation area and unicasting in the w arehouse area, has achie v ed the reliability comparable to that of ooding, b ut superior to that of directed dif fusion. W e can also observ e the small de gradation of EDR with the increase of n in RRP which can be e xplained as follo ws: since our RRP uses unicasting in the w arehouse area, data pack ets tra v elling in the w arehouse area may suf fer from pack et dropping as in direct dif fusion. Intuiti v ely ,

PAGE 145

133 the greater n is, the more hops a pack et may tra v el in the w arehouse area, hence increasing the dropping probability of pack ets. F ortunately our AD V -REQ-D A T A e xchanges can compensate for such possible pack et droppings by allo wing sink nodes to recei v e AD Vs from multiple data holders and retransmit REQs to a data holder if the e xpected D A T A message is not recei v ed in time. Therefore, our RRP can still maintain a pretty high EDR comparable to that of ooding in the f ace of error -prone wireless links. Figure 5-9 showsnormalizedenergyconsumptionofourRRP,oodinganddirected dif fusion. Compared with the other tw o schemes, directed dif fusion demonstrates the minimum ener gy consumption because it uses lo w rate ooding for interest propag ation and unicasting for data pack ets. Our RRP outperforms ooding because of the use of zone ooding instead of netw ork-wide ooding. W e observ e that the greater n is, the less ener gy our RRP consumes. The reason is that zone ooding and unicasting are respecti v ely used in the transportation area and the w arehouse area. W ith the increase of the w arehouse size n more unicasting and less zone ooding will be in v olv ed and the former is kno wn to be more ener gy ef cient than the latter Figure 5-10 givesaverageeventend-to-enddelayofeachschemeunderdifferent PERs Since directed dif fusion adopts minimum-delay paths, it has the shortest delay among the three schemes. F or pure ooding, the netw ork-wide pure ooding of data packets may result in much more collisions than our zone ooding and pack ets belonging to the same source-sink pair may follo w dif ferent, quite unpredictable, and possibly v ery long routes. Therefore, pure ooding e xperiences longer a v erage e v ent delay than that of our RRP F or our RRP since an e v ent report rst tra v els through the transportation area and then the w arehouse area, there are tw o sources contrib uting to the delay One is the zone ooding in the transportation area, and the other is the AD V -REQ-D A T A e xchanges in the w arehouse area. The a v erage delay coming from zone ooding decreases with the increase of n while the a v erage delay coming from AD V -REQ-D A T A e xchanges increases with the

PAGE 146

134 0 0 0 0 1 0 0 0 2 0 0 0 3 0 0 0 4 0 0 0 5 0 0 0 6 0 0 0 7 0 0 0 8 0 0 0 9 0 0 1 0 8 6 0 8 8 0 9 0 9 2 0 9 4 0 9 6 0 9 8 1 1 0 2 P a c k e t e r r o r r a t e ( P E R )E v e n t d e l i v e r y r a t i o ( E D R ) d i r e c t e d d i f f u s i o n f l o o d i n g R R P ( d = 2 ) R R P ( d = 3 ) R R P ( d = 4 ) R R P ( d = 5 ) Figure 5: Ev ent deli v ery ratio vs. pack et error rate. increase of n An interesting observ ation here is that the delay performance is not monotone with re g ard to n Our RRP achie v es smallest delay when n = 1 follo wed by n = 5 and n = 2 the other tw o are v ery close and not easy to tell them apart. This observ ation suggests a tradeof f between zone ooding and w arehouse routing should be made to achie v e a desired latenc y Figure 5 presentsaverageroutingoverheadofeachscheme.Amongthethree schemes, directed dif fusion requires sinks to periodically ood inter ests to maintain the gradients, as a result of which it encounters the lar gest routing o v erhead. F or our RRP the routing o v erhead comes from the routing maintenance in the small w arehouse area. Therefore, its routing o v erhead is much smaller than that of directed dif fusion b ut lar ger than that of pure ooding which is supposed to ha v e zero routing o v erhead. Besides, since the o v erhead of RRP mainly comes from the routing maintenance in the w arehouse area, it is of no surprise that we observ e the o v erhead of RRP increases with the increase of n that is, the w arehouse size. In the second set of gures (Fig.5Fig.5), we present our results on the RRP performance with a x ed w arehouse size n = 3 b ut dif ferent PER s. W e also v ary the control parameter d to study the impact of dif ferent ooding zone sizes.

PAGE 147

135 0 5 1 0 1 5 2 0 d i r e c t e d d i f f u s i o n f l o o d i n g R R P ( d = 2 ) R R P ( d = 3 ) R R P ( d = 4 ) R R P ( d = 5 ) (a) Normalized ener gy consumption vs. d (PER=0.0005). 0 0 2 0 4 0 6 0 8 d i r e c t e d d i f f u s i o n f l o o d i n g R R P ( d = 2 ) R R P ( d = 3 ) R R P ( d = 4 ) R R P ( d = 5 ) (b) A v erage e v ent end-to-end delay vs. d (PER=0.0005). Figure 5: Ener gy consumption and e v ent end-to-end delay of RRP with dif ferent d

PAGE 148

136 Fig.5sho ws the e v ent deli v ery ratio with dif ferent d W e observ e that the size of ooding zone has signicant impact on the performance of RRP When d = 2 the EDR of RRP is nearly 98% b ut when d increases to 3 or 4, the EDR of RRP becomes better and is comparable to that of pure ooding. Ho we v er the lar ger d does not al w ays imply better EDR. As we can see, when d increases to 5, the EDR drops to 96% This observ ation can be interpreted as follo ws. When d is small, meaning a lar ge ooding zone, the collisions in the zone may occur quite often and result in relati v ely lo wer EDR. Though pure ooding suf fers from collisions as well, ooding has higher probability to recei v e a cop y of the original data pack et because it is ooded in the whole eld rather than in a zone. W ith the increase of d and thus the decrease of the ooding zone size, pack et collisions may decrease, leading to a better EDR. Ho we v er the ooding zone size should be reasonably wide to ha v e enough nodes to relay pack ets. Ho we v er our RRP outperforms directed dif fusion in most cases e v en when the ooding zone is small. Figure 5(a) showsnormalizedenergyconsumptionofourRRPwithregardto d where P E R = 0 : 0005 The results of pure ooding and directed dif fusion are sho wn for reference only W e observ e that the lar ger d is, the less ener gy each pack et consumes. This result is reasonable because the lar ger d is, the narro wer the zone is, the fe wer nodes are in v olv ed in the ooding, and hence the less the ener gy is consumed. Considering the impact of d on both EDR and ener gy consumption, it clearly indicates that the ooding zone size d is an important parameter in RRP and should be well chosen to strik e a good balance between reliability and ener gy ef cienc y Figure 5(b) showsaverageeventend-to-enddelayofourRRPwithregardto d with P E R equal to 0.0005. W e can see that the e v ent latenc y decreases with the increase of d When d is small, meaning a lar ge ooding zone, pack ets recei v ed by a sink may tra v el along unpredictable paths, leading to a longer a v erage delay In contrast, when d is lar ge, meaning a small ooding zone and fe wer nodes in v olv ed in the pack et forw arding, pack ets recei v ed by a sink may tra v el along more predictable paths limited by the small zone, which

PAGE 149

137 results in a shorter a v erage delay W e can further e xpect a small latenc y comparable to that of directed dif fusion if we carefully choose a small enough ooding zone. T o sum up, our simulation results sho w that ooding and our zone ooding techniques are less sensiti v e to the link f ailures thereby pro vide more rob ust pack et deli v ery service. Ho we v er pure ooding suf fers from inef cient ener gy usage and long end-to-end pack et latenc y By choosing appropriate w arehouse size and ooding zone size, we can adjust the ener gy consumption, the e v ent end-to-end latenc y and the routing o v erhead of RRP to competiti v e le v els while maintaining nearly perfect reliability Therefore, our RRP provides a good solution to balance the reliability and ener gy ef cienc y requirements in sensor netw orks. 5.6 Summary In this chapter we introduced the concept of supply chain into wireless sensor netw orks and propose a h ybrid data dissemination frame w ork based on the supply chain management methodology F or each sensing task, a whole sensor eld is conceptually partition into se v eral functional re gions, and v arious routing schemes are applied to dif ferent re gions in order to pro vide better performance in terms of reliability and ef cient ener gy usage. F or this purpose, we also proposed a no v el zone ooding technique which is a combination of geometric routing and ooding. On top of our scheme, ph ysically separated or interlaced multipath routing can be easily implemented without incurring an y signicant additional costs. Our h ybrid data dissemination frame w ork features lo w o v erhead, high reliability good scalability and notable e xibility The ef fecti v eness and ef cienc y of our scheme were v alidated through simulation studies. W e demonstrated that our scheme can strik e a good balance between reliability and ener gy ef cienc y with proper sizes of the w arehouse area and ooding zones.

PAGE 150

CHAPTER 6 CONCLUSIONS AND FUTURE W ORK 6.1 Conclusions In this dissertation, we study the issues of ener gy conserv ation and QoS pro visioning in resource-constrained wireless ad hoc netw orks. W e particularly identify the inherent heterogeneity in such netw orks, and classify it into tw o groups: de vice heterogeneity and en vironment heterogeneity W e propose to adopt the cross-layer design philosoph y to guide the protocol design. F or the issue of ener gy conserv ation. W e rst propose a de vice-ener gy-load a w are relaying frame w ork (DELAR) to utilize heterogeneity in terms of po wer supply to conserv e ener gy This frame w ork is a joint design of po wer control, scheduling and routing. The proposed A-MA C protocol also addresses the issue of reliable communications o v er unidirectional links. Further based on the DELAR frame w ork, a multiple-pack ets transmission scheme is proposed to impro v e the o v erall performance of DELAR. Through simulation studies, DELAR is sho wn to be ener gy-ef cient. W e also study ho w to utilize mobility to conserv e ener gy W e present a general mo v ement model and formulate the general mo v ement problem. W e particularly study the W aterhunter problem and reduce it to a path-constrained path-optimization routing problem. W e present a heuristic algorithm for this NP-complete routing problem. Through simulation, it sho ws that the resource-a w are mo v ement scheme can better utilize the P-nodes in the netw orks and lead to more ener gy sa vings. F or the issue of Quality of Service, we study ho w to use system dynamics to better support dif ferentiated services in wireless ad hoc netw orks. W e v alidate that there are unused resources, that is, bandwidth due to system dynamics. W e propose a cross-layer designed scheme called courtesy piggybacking to mak e use of such unused resources. This 138

PAGE 151

139 scheme is sho wn to be v ery ef fecti v e in alle viating the conict between throughput and f airness and the starv ation problem in a dif ferentiated services system. W e study the issues of reliability and ener gy conserv ation in wireless sensor netw orks. W e propose a supply-chain based wireless sensor netw ork model and propose a h ybrid data dissemination frame w ork. This frame w ork is scalable, reliable and ener gy-ef cient. 6.2 Future W ork F or the future research on the issue of ener gy conserv ation, there are se v eral aspects can be impro v ed and in v estig ated. Since the h ybrid scheduling has hea vy impacts on the system performance of DELAR. W e w ould lik e to seek a more dynamic transmission scheduling scheme to strik e a better balance between ener gy conserv ation and other system performance f actors. W e also w ant to in v estig ate the impact on ener gy conserv ation of node placement of P-nodes [75,132]. W e belie v e that the controlled and intelligent mo v ement and its inte gration with resource heterogeneity can bring lots of benets to mobile ad hoc netw orks. From the resource-a w are mo v ement scheme, it suggests that netw orks e xhibiting features of small-w orld tend to ha v e better o v erall performance, so we w ould lik e to rst study the Firehunter Mo v ement problem (cf. Section3.3.2) and stri v e to propose a unied solution for the general resource-a w are mo v ement problem. Further we w ould lik e to study ho w to b uild up such a small w orld in wireless ad hoc netw orks. As to the research on the issue of QoS pro visioning, we w ould lik e to apply this piggybacking scheme to more concrete applications to e v aluate its performance, and we are also interested in the inte gration of this scheme into the DELAR frame w ork. F or the wireless sensor netw orks, the supply chain model can be further impro v ed to t dif ferent application scenarios. W e w ould lik e to gi v e a formal formulation of this model and to carry out more strict analysis about the partition boundary between dif ferent

PAGE 152

140 functional re gions. W e are particularly interested in b uilding some real test beds to e v aluate this model and this frame w ork. F or wireless ad hoc netw orks as a whole, there are still lots of open problems such as security issues, and the interoperateability with other wireless and mobile netw orks. In the future, we will de v ote ourselv es to these interesting topics.

PAGE 153

REFERENCES [1] E. Biglieri and P T orino, Digital transmission in the 21st century: Conating modulation and coding, IEEE Communications Ma gazine v ol. 40, no. 5, pp. 128, May 2002. [2] S. K umar V Ragha v an, and J. Deng, Medium access control protocols for ad-hoc wireless netw orks: A surv e y Else vier Ad Hoc Networks J ournal 2005, to appear [3] V Bhar gha v an, A. Demers, S. Shenk er and L. Zhang, Maca w: A media access protocol for wireless lans, in Pr oc. of A CM SIGCOMM'94 1994. [4] C. Fullmer and J. Garcia-Luna-Ace v es, Floor acquisition multiple acess (f ama) for pack et radio netw orks, in Pr oc. of A CM SIGCOMM'95 1995. [5] P art 11: W ireless lan medium access control (mac) and ph ysical layer (ph y) specications, Standard, IEEE, IEEE 802.11WG, August 1999. [6] D. Lang, A comprehensi v e o v ervie w about selected ad hoc netw orking routing protocols, T echnische Uni v ersity Munchen, Department of Computer Science, T ech. Rep., March 2003. [Online]. A v ailable:http://wwwbib .informatik. tumuenchen.de/infberichte/2003/TUMI0311.pdf[7] C. Perkins and P Bhagw at, Highly dynamic destination-sequenced distance-v ector routing (dsdv) for mobile computers, in Pr oc. of A CM SIGCOMM'94 Sept. 1994, pp. 234. [8] Optimized link state routing protocol (olsr), IETF T ech. Rep., October 2003. [9] D. Johnson, D. Maltz, and Y .-C. Hu, The dynamic source routing protocol for mobile ad hoc netw orks, IETF Internet Draft, April 2003, draft-ietf-manet-dsr -09.txt. [10] C. Perkins, E. Belding-Ro yer and S. Das, Ad hoc on demand distance v ector (aodv) routing, RFC 3561, July 2003. [11] Z. Haas, M. Pearlman, and P Samar The zone routing protocol (zrp) for ad hoc netw orks, IETF Internet Draft, draft-ietf-manet-zone-zrp-04.txt, July 2002. [12] M. Gerla, X. Hong, L. Ma, and G. Pei, Landmark routing protocol (lanmar) for lar ge scale ad hoc netw orks, IETF Internet Draft, MANET W orking Group, draftietf-manet-lanmar -05.txt, No v ember 2002. 141

PAGE 154

142 [13] C. Jones, K. S. P Agra w al, and J.-C. Chen, A surv e y of ener gy ef cient netw ork protocols for wireless netw orks, A CM/Kluwer J ournal on W ir eless Networks (WINET) v ol. 7, no. 4, August 2001. [14] P Mohapatra, J. Li, and C. Gui, Qos in mobile ad hoc netw orks, IEEE W ir eless Communications Ma gazine v ol. 10, no. 3, pp. 44, June 2003. [15] X. Hannan, C. Chaing, and S. W inston, Quality of service models for ad hoc wireless netw orks, in The handbook of ad hoc wir eless networks Boca Raton, FL, USA: CRC Press, Inc., 2003, pp. 467. [16] G.-S. Ahn, A. Campbell, A. V eres, and L.-H. Sun, Supporting service dif ferentiation for real-time and best ef fort traf c in stateless wireless ad hoc netw orks (sw an), IEEE T r ans. on Mobile Computing v ol. 1, no. 3, pp. 192, July-September 2002. [17] D. Black, Dif ferentiated services and tunnels, RFC2983, 2000. [18] Dening cross-layer design for wirelss netw orking, in Pr oc. of IEEE ICC'03 2003, panel. [Online]. A v ailable:http://www .eas.asu.edu/ junshan/ICC03panel. html[19] S. Shakk ottai, T Rappaport, and P Karlsson, Cross-layer design for wireless netw orks, IEEE Commun. Ma g October 2003. [20] I. Ak yildiz, X. W ang, and W W ang, W ireless mesh netw orks: a surv e y Computer Networks 2005. [21] I. Chak eres and E. Belding-Ro yer Resource biased path selection in heterogeneous mobile netw orks, in UCSB T ec hnical Report 2003-18 July 2003. [22] M. Krunz, A. Muqattash, and S.-J. Lee, T ransmission po wer control in wireless ad hoc netw orks: challenges, solutions and open issues, IEEE Network v ol. 18, no. 5, pp. 8, 2004. [23] S. Singh and C. S. Ragha v endra, P amas-po wer a w are multi-access protocol with signaling for ad hoc netw orks, Computer Communication Re vie w v ol. 28, no. 3, pp. 5, 1998. [24] J. H. W Y e and D. Estrin, An ener gy-ef fcient mac protocol for wireless sensor netw orks, in Pr oc. of INFOCOM'02 2002. [25] Y .-C. Tseng, C.-S. Hsu, and T .-Y Hsieh, Po wer -sa ving protocols for ieee802.11based multi-hop ad hoc netw orks, in Pr oc. of INFOCOM'02 2002. [26] V Roodoplu and T Meng, Minimum ener gy mobile wireless netw orks, IEEE J Select. Ar eas Commun. v ol. 17, no. 8, pp. 1333, August 1999.

PAGE 155

143 [27] B. Chen, K. Jamieson, H. Balakrishnan, and R. Morris, Span: an ener gy-ef cient coordination algorithm for topology maintainence in ad hoc wireless netw orks, A CM/Kluwer J ournal on W ir eless Networks (WINET) v ol. 8, no. 5, Sept. 2002. [28] J. W u, F Dai, M. Gao, and I. Stojmoeno vic, On calculating po wer -a w are connected dominating sets for ef cient routing in ad hoc wireless netw orks, J ournal of Communications and Networks v ol. 4, no. 1, March 2002. [29] Y Xu, J. Heidemann, and D. Estrin, Geograph y-informed ener gy conserv ation for ad hoc routing, in Pr oc. of MobiCom'01 2001. [30] D. Blough, M. Leoncini, G. Resta, and P Santi, The k-neigh protocol for symmetric topology control in ad hoc netw orks, in Pr oc. of MobiHoc'03 2003. [31] L. Feene y A qos a w are po wer sa v e protocol for wireless ad hoc netw orks, in Pr oc. of INFOCOM'02 2002. [32] R. Ramanathan and R. Rosales-hain, T opology control of multihop wireless netw orks using transmit po wer adjustment, in Pr oc. of INFOCOM'00 2000. [33] J. Gomez and A. T Campbell, A case for v ariable-range transmission po wer control in wireless ad hoc netw orks, in Pr oc. of IEEE INFOCOM'04 Hongk ong, China, March 2004. [34] W .-Z. Song, Y W ang, X.-Y Li, and O. Frieder Localized algorithms for ener gy ef cient topology in wireless ad hoc netw orks, in Pr oc. of MobiHoc'04 2004. [35] A. Muqattash and M. Krunz, Po wmac: A single-channel po wer -control protocol for throughput enhancement in wireless ad hoc netw orks, IEEE J Select. Ar eas Commun. 2005, special Issue on Adv ances in Military W ireless Communications (to appear). [36] E.-S. Jung and N. V aidya, A po wer control mac protocol for ad hoc netw orks, in Pr oc. of MobiCom'02 2002. [37] S. Singh, M. W oo, and C. Ragha v endra, Po wer -a w are routing in mobile ad hoc netw orks, in Pr oc. of MobiCom'98 1998. [38] J.-H. Chang and L. T assiulas, Ener gy conserv ation routing in wireless ad-hoc netw orks, in Pr oc. of INFOCOM'00 2000. [39] C. T oh, Maximum battery life routing to support ubiquitous mobile computing in wireless ad hoc netw orks, IEEE Commun. Ma g pp. 138, June 2001. [40] S. Banerjee and A. Misra, Minimum ener gy paths for reliable communication in multi-hop wireless netw orks, in Pr oc. of MobiHoc'02 2002.

PAGE 156

144 [41] J. Gomez, A. Campbell, M. Naghshineh, and C. Bisdikian, P aro: Supporting dynamic po wer controlled routing in wireless ad hoc netw orks, A CM/Kluwer J ournal on W ir eless Networks (WINET) v ol. 9, no. 5, September 2003. [42] J.-H. Ryu and D. Cho, A ne w routing scheme concerning po wer -sa ving in mobile ad hoc netw orks, in Pr oc. of ICC'00 2000. [43] A. A vudainayag am, W Lou, and Y F ang, Dear: A de vice and ener gy a w are routing protocol for heterogeneous ad hoc netw orks, J ournal of P ar allel and Distrib uted Computing Feb 2003. [44] U. K ozat, I. K outsopoulos, and L. T assiulas, A frame w ork for cross-layer design of ener gy-ef cient communication with qos pro visioning in multi-hop wireless netw orks, in Pr oc. of IEEE INFOCOM'04 Hongk ong, China, March 2004. [45] R. Bhatia and M. K odialam, On po wer ef cient communication o v er multi-hop wireless netw orks: Joint routing, scheduling and po wer control, in Pr oc. of IEEE INFOCOM'04 Hongk ong, China, March 2004. [46] V Ka w adia and P R. K umar Principles and protocols for po wer control in ad hoc netw orks, IEEE J Select. Ar eas Commun. v ol. 23, no. 5, pp. 76, January 2005, special Issue on W ireless Ad Hoc Netw orks. [47] A. Leon-Garcia and I. W idjaja, Computer Networks: Fundmental Concepts and K e y Ar c hitectur es McGra w-Hill, 1999. [48] L. Bao and J. Garcia-Luna-Ace v es, Link-state routing in netw orks with unidirectional links, in Pr oc. of ICCCN'99 1999. [49] M. Marina and S. Das, Routing performance in the presence of unidirectional links in mulithop wireless netw orks, in Pr oc. of MobiHoc'02 Lausanne, Switzerland, June 2002. [50] R. Prakash, A routing algorithm for wireless ad hoc netw orks with unidirectional links, A CM/Kluwer J ournal on W ir eless Networks (WINET) v ol. 7, pp. 617, 2001. [51] S. Nesar gi and R. Prakash, A tunneling approach to routing with unidirectional links in mobile ad-hoc netw orks, in Pr oc. of ICCCN'00 2000. [52] V Ramasubramanian, R. Chandra, and D. Mosse, Pro viding a bidirectional abstraction for unidirectional ad hoc netw orks, in Pr oc. of INFOCOM'03 2003. [53] M.-S. Alouini, X. T ang, and A. Goldsmith, An adapti v e modulation scheme for simultaneous v oice and data transmission o v er f ading channels, IEEE J Select. Ar eas Commun. v ol. 17, no. 3, pp. 837, may 1999.

PAGE 157

145 [54] M. Pursle y and J. Shea, Nonuniform phase-shift-k e y modulation for multimedia multicast transmission in mobile wireless netw orks, IEEE J Select. Ar eas Commun. v ol. 17, no. 9, pp. 774, May 1999. [55] P V itthalade vuni and M.-S. Alouini, Exact ber computation of generalized hier archical psk constellations, IEEE T r ans. on W ir eless Comm. v ol. 51, no. 12, pp. 2030, December 2003. [56] M. Pearlman and Z. Haas, Determining the optimal conguration for the zone routing protocol, IEEE J Select. Ar eas Commun. v ol. 17, no. 8, pp. 1395, August 1999. [57] OPNET Making netw orks and applications perform, http://www .opnet.com. [58] J. Y oon, M. Liu, and B. Noble, Sound mobility models, in Pr oc. of A CM MobiCom'03 San Die go, CA, September 2003. [59] Proxim, Orinoco 11a/b/g combocard specications. [Online]. A v ailable:www proxim.com/learn/library/datasheet/11abgcombocard.pdf[60] C. Chiasserini and R. Rao, Ener gy ef cient battery management, in Pr oc. of INFOCOM'00 v ol. 2, 2000, pp. 396. [61] S. Jayashree, B. Manoj, and C. S. R. Murth y On using battery state for medium access control in ad hoc wireless netw orks, in Pr oc. of MobiCom'04 Sept. 2004. [62] M. Grossglauser and D. Tse, Mobility increase the capacity of ad-hoc wirelss netw orks, in Pr oceedings of IEEE INFOOCM'00 T el A vi v Israel, March 2000, pp. 1360. [63] S. Basagni, I. Chlamtac, and V Syrotiuk, A distance routing ef fect algorithm for mobility (dream), in Pr oc. of the F ifth Anuual A CM/IEEE International Confer ence on Mobile Computing and Networking (MobiCom'98) Dallas, TX, 1998. [64] Y .-B. K ou and N. V aidya, Location-aided routing (lar) in mobile ad hoc netw orks, Pr oc. of the F ifth Anuual A CM/IEEE International Confer ence on Mobile Computing and Networking (MobiCom'98) October 1998. [65] S. Chakraborty D. Y au, and J. Lui, On the ef fecti v eness of mo v ement prediction to reduce ener gy consumption in wireless sommunication (e xtended abstract), in Pr oc. of A CM SIGMETRICS'03 San Die go, CA, June 2003. [66] Q. Li and D. Rus, Sending messages to mobile users in disconnected ad-hoc wireless netw orks, in Pr oc. of A CM MobiCom'00 August 2000. [67] D. Glodenber g, J. Lin, A. Morse, B. Rosen, and Y Y ang, T o w ards mobility as a netw ork control primiti v e, in Pr oceedings of A CM MobiHoc'04 Roppongi, Japan, May 2004, pp. 163.

PAGE 158

146 [68] R. Shah, S. Ro y S. Jain, and W Brunette, Data mules: Modeling a three-tier architecture for sparse sensor netw orks, in Pr oc. of IEEE SNP A W orkshop 2003. [69] W Zhao, M. Ammar and E. Ae gura, A message ferry approach for data deli v ery in sparse mobile ad hoc netw orks, in Pr oc. of A CM MobiCom'04 26 Sep. 1 Oct. 2004, pp. 187. [70] D. Johnson and D. Maltz, Dynamic source routing in ad hoc wireless netw orks, in Mobile Computing Imielinski and K orth, Eds. Kluwer Academic Publishers, 1996, v ol. 353. [71] D. Ree v es and H. Salama, A distrib uted algorithm for delay-constrained unicast routing, IEEE/A CM T r ansaction on Networking v ol. 8, no. 2, pp. 230, 2000. [72] M. Gare y and D. Johnson, Computer and Intr actability: A Guide to the Theory of NP-completeness San Francisco, CA: Freeman, 1979. [73] W Liu, W Lou, and Y F ang, An ef cient qos routing algorithm for delay sensiti v e applications, Computer Networks v ol. 47, no. 1, pp. 87, 2005. [74] Draft supplement to part 11: W ireless lan medium access control (mac) and ph ysical layer (ph y) specications: Medium access control (mac) enhancements for quality of service (qos),ieee standard 802.11e/d4.3, Standard, IEEE, IEEE 802.11WG, May 2003. [75] Z. Y e, S. Krishnamurth y and S. K. T ripathi, A frame w ork for reliable routing in mobile ad hoc netw orks, in Pr oc. of INFOCOM'03 2003. [76] L. Feene y and M. Nilsson, In v estig ating the ener gy consumption of wireless netw ork interf ace in an ad hoc netw orking en vironment, in Pr oc. of IEEE INFOCOM'01 2001, pp. 1548. [77] R. Albert and A.-L. Barabsi, Statisctical mechanics of somple x netw orks, Re vie w of Modern Physics pp. 47, Jan. 2002. [78] M. Barry A. T Campbell, and A. V eres, Distrib uted control algorithm for service dif ferentiation in wireless pack et netw orks, in Pr oc. of IEEE INFOCOM'01 2001. [79] X. P allot and L. Miller Implementing message priority policies o v er an 802.11 based mobile ad hoc netw ork, in Pr oc. of IEEE MILCOM'01 2001. [80] Y Cao and V Li, Scheduling algorithms in broad-band wirelss netw orks, in Pr ocedings of the IEEE v ol. 89, no. 1, 2001, pp. 76. [81] J. W ang and K. Nahrstedt, Hop-by-hop routing algorithms for premium-class traf c in dif fserv netw orks, in Pr oc. of IEEE INFOCOM'02 2002. [82] A. Demers, S. K esha v and S. Shenk er Analysis and simulation of a f air queueing algorithm, in Pr oc. of A CM SIGCOMM'89 1989, pp. 3.

PAGE 159

147 [83] L. Zhang, V irtual clock: A ne w traf c control algorithm for pack et switching neto wrks, in Pr oc. of A CM SIGCOMM'90 1990, pp. 19. [84] S. Lu and V Bhar gha v an, F air scheduling in wireless pack et netw orks, IEEE/A CM T r ans. Networking v ol. 7, no. 4, pp. 473, 1999. [85] T Ng, I. Stoica, and H. Zhang, P ack et f air queueing algorithms for wireless netw orks with locaiton-dependent errors, in Pr oc. of IEEE INFOCOM'98 Mar 1998, pp. 1103. [86] P Bhagw at, A. Krishna, and S. T ripathi, Enhance throughput o v er wireless lans using channel state depentdent pack et scheduling, in Pr oc. of IEEE INFOCOM'96 Mar 1996, pp. 1133. [87] C. Fragouli, V Si v araman, and M. Sri v asta v a, Controlled multimedia wirelss link sharing via enhanced class-based queuing with channel-state dependent pack et scheduling, in Pr oc. of IEEE INFOCOM'98 1998, pp. 572. [88] S. Flo yd and V Jacobson, Link-sharing and resource management models for pack et netwroks, IEEE/A CM T r ans. Networking v ol. 3, pp. 365, August 1995. [89] J. Gomez and A. T Campbell, Ha v ana: Supporting application and channel dependent qos in wireless netw orks, A CM/Kluwer J ournal on W ir eless Networks (WINET) v ol. 9, no. 1, pp. 21, Jan. 2003. [90] I. K outsopoulous and L. T assiulas, Channel state-adpti v e techiniques for throughput enhancements in wireless broadband netw orks, in Pr oc. of IEEEINFOCOM'01 Apr 2001, pp. 757. [91] G. Holland, N. V aidya, and P Bahl, A rate-adpati v e mac protocol for multi-hop wireless netw orks, in Pr oc. of A CM MobiCom'01 2001. [92] P Che villat, J. Jelitto, A. Barreto, and H. T ruong, A dynamic link adaptation algorithm for ieee 802.11a wireless lans, in Pr oc. of IEEE ICC'03 2003. [93] B. Kim, Y F ang, T F W ong, and Y Kw on, Dynamic fragmentation scheme for rate-adapti v e wireless lan, in Pr oc. of IEEE PIMRC'03 Sept. 2003. [94] J. T ourrilhes, Dwell adapti v e fragmentation: Ho w to cope with short dwells required by multimedia wirelss lans, in Pr oc. of IEEE GLOBECOM'00 2000. [95] D. Qiao and S. Choi, Goodput enhancement of ieee 802.11a wireless lan via link adaptation, in Pr oc. of IEEE ICC'01 2001. [96] K. B. S. R. Kadaba and S. Nanda, Channel quality estimation and rate adaption for cellular mobile radio, IEEE J Select. Ar eas Commun. v ol. 17, no. 7, pp. 1244 1256, July 1999.

PAGE 160

148 [97] H. W ang and N. Moayeri, Finite-state mark o v channel-a useful model for radio communication channels, IEEE T r ans. V eh. T ec hnol. v ol. 44, no. 1, 1995. [98] E. Modiano, An adapti v e algorithm for optimizing the pack et size used in wireless arq protocols, A CM/Kluwer J ournal on W ir eless Networks (WINET) v ol. 5, no. 5, pp. 279, 1999. [99] I. Ada and C. Castelluccia, Dif ferentiation mechanisms for ieee 802.11, in Pr oc. of IEEE INFOCOM'01 2001. [100] A. S. T anenbaum, Computer Networks 3rd ed. Prentice Hall, 1996. [101] S.-T Sheu, T Chen, J. Chen, and F Y e, An impro v ed data ushing mac protocol for ieee 802.11 wireless ad hoc netw ork, in Pr oc. of IEEE VTC 2002 F all 2002. [102] D. Bertsekas and R. Gallager Data Networks 2nd ed. Prentice Hall, 1992. [103] K. Altink emer I. Bose, and R. P al, A v erage w aiting time of customers in an m/d/k queue with nonpreempti v e priorities, Computer s & Oper ations Resear c h v ol. 25, no. 4, pp. 317, 1998. [104] emer ging technologies that will change the w orld, T ec hnolo gy Re vie w v ol. 106, no. 1, pp. 33, Feb 2003. [105] I. Ak yildiz, W Su, Y Sankarasubramaniam, and E. Cayirci, A surv e y on sensor netw orks, IEEE Communication Ma gazine August 2002. [106] C.-Y Chong and S. P K umar Sensor netw orks: e v olution, opportunities, and challenges, Pr oceedings of the IEEE v ol. 91, no. 8, August 2003. [107] C. Intanagonwiw at, R. Go vindan, and D. Estrin, Directed dif fusion: a scalable and rob ust communication paradigm for sensor netw orks, in Pr oc. of the Sixth Anuual A CM/IEEE International Confer ence on Mobile Computing and Networking (MobiCom'00) Boston, MA, August 2000. [108] J. K ulik, W R. Heinzelman, and H. Balakrishnan, Ne gotiation-based protocols for disseminating information in wireless sensor netw orks, A CM W ir eless Networks v ol. 8, no. 2-3, pp. 169, 2002. [109] J. Heidemann, F Silv a, and D. Estrin, Matching data dissemination algorithms to application requirements, in Pr oc. of the F ir st International Confer ence on Embedded Network ed Sensor Systems (MobiSys'03) Los Angeles, CA, 2003, pp. 218. [110] W Heinzelman, A. Chandrakasan, and H. Balakrishnan, Ener gy-ef fcient rotuing protocols for wireless microsensor netw orks, in r oc. of the 33r d Hawaii Intenational Confer ence on System Sciences (HICSS'00) Ha w aii, Jan. 2000. [111] F Y e, H. Luo, J. Cheng, S. Lu, and L. Zhang, A tw o-tier data dissemination model for lar ge-scale wireless sensor netw orks, in Pr oc. of the Ninth Anuual A CM/IEEE

PAGE 161

149 International Confer ence on Mobile Computing and Networking (MobiCom'02) Atlanta, Geor gia, Spet. 2002. [112] D. Ganessan, R. Go vindan, S. Shenk er and D. Estrin, Highly-resilient,ener gyef cient multipath routing in wireless sensor netw orks, A CM MC 2 R v ol. 5, no. 4, October 2001. [113] J. H. Z. Hass and L. Li, Gossip-based ad hoc routing, in Pr oc. of IEEE INFOCOM'02 Ne w Y ork, June 2002. [114] Y Sasson, D. Ca vin, and A. Schiper Probabilistic broadcast for ooding in wireless mobile ad hoc netw orks, in Pr oc. of IEEE W ir eless Communications and Networking Confer ence (WCNC'03) Ne w Orleans, Louisiana, March 2003. [115] C. Barrett and S. E. L. Kroc, P arametric probabilistic sensor netw ork routing, in Pr oc. of the 2nd International W orkshop on W ir eless Sensor Networks and Applications (WSN A 2003) San Die go, CA, Spet. 2003. [116] S. T ilak, A. Murph y and W Heinzelman, Non-uniform information dissemination for sensor netw orks, in Pr oc. of the 11th International Confer ence on Network Pr otocols (ICNP'03) Atlanta, Geor gia, No v 2003. [117] B. Karp and H. K ung, Greedy perimeter stateless routing, in Pr oc. of the Se venth Anuual A CM/IEEE International Confer ence on Mobile Computing and Networking (MobiCom'00) Boston, MA, August 2000. [118] D. Niculescu and B. Nath, rajectory based forw arding and its applications, in Pr oc. of the T enth Anuual A CM/IEEE International Confer ence on Mobile Computing and Networking (MobiCom'03) San Die go, CA, Sept. 2003. [119] A. Mainw aring, J. Polastre, R. Sze wczyk, D. Culler and J. Anderson, W ireless sensor netw orks for habitat monitoring, in Pr oc. of the 1st International W orkshop on W ir eless Sensor Networks and Applications (WSN A 2002) Atlanta,Geogia,USA, 2002. [120] N. B. Priyantha, A. Chakraborty and H. Balakrishnan, The crick et location-support system, in Pr oc. of the Se venth Anuual A CM/IEEE International Confer ence on Mobile Computing and Networking (MobiCom'00) Boston, MA, August 2000. [121] D. Niculescu and B. Nath, Ad hoc position system (aps) using aoa, in Pr oc. of IEEE INFOCOM'03 San Fransciso, CA, April 2003. [122] P Bahl and V P admanabhan, Radar: an inb uiding rf-based user location and tracking system, in Pr oc. of IEEE INFOCOM'00 T el A vi v Israel, March 2000. [123] A. Rao, S. Ratnasamy C. papadimitriou, S. Shenk er and I. Stoica, Geograhpic routing without location information, in Pr oc. of the T enth Anuual A CM/IEEE International Confer ence on Mobile Computing and Networking (MobiCom'03) San Die go, CA, Sept. 2003.

PAGE 162

150 [124] T He, C. Huang, B. Blum, J. Stank o vic, and T Abdelzaher Range-free location schemes for lar ge scale sensor netw orks, in Pr oc. of the T enth Anuual A CM/IEEE International Confer ence on Mobile Computing and Networking (MobiCom'03) San Die go, CA, Spet. 2003. [125] L. D. Fredenhall and E. Hill, Basics of supply c hain mana g ement St. Lucie Press, 2001. [126] B. Krishnamachari, D. Estrin, and S. W ick er Impact of data aggre g ation in wireless sensor netw orks, in Pr oc. of International W orkshop on Distrib uted Event-Based Systems V ienna, Austria, July 2002. [127] W Lou, W Liu, and Y F ang, Spread: enhancing data condentiality in mobile ad hoc netw orks, in Pr oc. of IEEE INFOCOM'04 HongK ong, China, March 2004. [128] E. W eisstein, Ellipse: From mathw orlda w olfram web resource, http://mathw orld.w olfram.com/Ellipse.html. [129] T Ozsu and P V alduriez, Principles of distrib uted database mana g ement systems 2nd ed. Prentice Hall, 1999. [130] L. Bajaj, M. T akai, R. Ahuja, K. T ang, R. Bagrodia, and M. Gerla, Glomosim: A scalable netw ork simulation en vironment, UCLA Computer Science Department, T ech. Rep. 990027, May 1999. [131] Crossbo w Crossbo w technology inertial & gyro systems, wireless sensor netw orks, smart dust, adv anced sensors, http://www .xbo w .com. [132] M. Y arvis, N. K ushalnag ar H. Singh, A. Rang arajan, Y Liu, and S. Singh, Exploiting heterogeneity in sensor netw orks, in Pr oc. of IEEE INFOCOM'05 2005.

PAGE 163

BIOGRAPHICAL SKETCH W ei Liu recei v ed the B.E. and M.E. de grees in electrical engineeringfromHuazhong Uni v ersity of Science and T echnology W uhan, China, in 1998 and 2001, respecti v ely He is currently pursuing the PhD de gree in the Department of Electrical and Computer Engineering,Uni v ersity of Florida, Gainesville, where he is a research assistant in the W ireless Netw orks Laboratory (WINET). His research interest includes QoS, Secure, and Po wer Ef cient Routing, and MA C protocols in Mobile Ad Hoc Netw orks and W ireless Sensor Netw orks. 151

PAGE 164

I certify that I ha v e read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope and quality as a disser tation for the de gree of Doctor of Philosoph y Y uguang F ang, Chair Associate Professor of Electrical and Computer Engineering I certify that I ha v e read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope and quality as a disser tation for the de gree of Doctor of Philosoph y Shig ang Chen Assistant Professor of Computer and Information Science and Engineering I certify that I ha v e read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope and quality as a disser tation for the de gree of Doctor of Philosoph y Sartaj Sahni Professor of Conputer and Information Science Engineering I certify that I ha v e read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope and quality as a disser tation for the de gree of Doctor of Philosoph y John M. Shea Assistant Professor of Electrical and Computer Engineering I certify that I ha v e read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope and quality as a disser tation for the de gree of Doctor of Philosoph y Dapeng W u Assistant Professor of Electrical and Computer Engineering

PAGE 165

This dissertation w as submitted to the Graduate F aculty of the Colle ge of Engineering and to the Graduate School and w as accepted as partial fulllment of the requirements for the de gree of Doctor of Philosoph y. August 2005 Pramod P Khar gonekar Dean, Colle ge of Engineering K enneth Gerhardt Interim Dean, Graduate School

PAGE 166

CR OSS-LA YER DESIGN OF RESOURCE A W ARE PR O T OCOLS FOR HETER OGENEOUS WIRELESS AD HOC NETW ORKS W ei Liu (352) 392–2746 Department of Electrical and Computer Engineering Chair: Y uguang F ang De gree: Doctor of Philosoph y Graduation Date: August 2005 W ireless ad hoc netw orks are v ery attracti v e in military and ci vil applications for which x ed infrastructure are una v ailable or unreliable. In this dissertation we in v estig ate the protocol design in resource-constrained wireless ad hoc netw orks, for the purpose of supporting communications with high ener gy-ef cienc y and desired quality-of-service (QoS). W e propose se v eral schemes that can utilize inherent heterogeneity in the netw orks to better use the precious resources. The nds in this dissertation are no v el. The research on ener gy conserv ation is the rst w ork that comprehensi v ely addresses this challenging issue. The research on QoS can be applied to man y other wireless and mobile netw orks and systems. The research on wireless sensor netw orks pro vides a ne w w ay to ef ciently manage wireless sensor netw orks. All of the proposed schemes are v ery ef fecti v e and easy to implement, and can be readily applied to real applications.