Citation
Control schemes for real-time communication in high-speed networks

Material Information

Title:
Control schemes for real-time communication in high-speed networks
Added title page title:
Control schemes in high speed real-time communication networks
Creator:
Shen, Li-Tao
Publication Date:
Language:
English
Physical Description:
vi, 118 leaves : ill. ; 29 cm.

Subjects

Subjects / Keywords:
Bandwidth ( jstor )
Communication models ( jstor )
Communication systems ( jstor )
Communications industries ( jstor )
Deadlines ( jstor )
Employment statistics ( jstor )
Modeling ( jstor )
Scheduling ( jstor )
Simulations ( jstor )
Traffic delay ( jstor )
Communication -- Network analysis ( lcsh )
Computer and Information Sciences thesis Ph. D
Computer network protocols ( lcsh )
Dissertations, Academic -- Computer and Information Sciences -- UF
Real-time data processing ( lcsh )
Genre:
bibliography ( marcgt )
non-fiction ( marcgt )

Notes

Thesis:
Thesis (Ph. D.)--University of Florida, 1992.
Bibliography:
Includes bibliographical references (leaves 111-117).
General Note:
Typescript.
General Note:
Vita.
Statement of Responsibility:
by Li-Tao Shen.

Record Information

Source Institution:
University of Florida
Holding Location:
University of Florida
Rights Management:
The University of Florida George A. Smathers Libraries respect the intellectual property rights of others and do not claim any copyright interest in this item. This item may be protected by copyright but is made available here under a claim of fair use (17 U.S.C. §107) for non-profit research and educational purposes. Users of this work have responsibility for determining copyright status prior to reusing, publishing or reproducing this item for purposes other than what is allowed by fair use or other copyright exemptions. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder. The Smathers Libraries would like to learn more about this item and invite individuals or organizations to contact the RDS coordinator (ufdissertations@uflib.ufl.edu) with any additional information they can provide.
Resource Identifier:
029758948 ( ALEPH )
29373990 ( OCLC )

Downloads

This item has the following downloads:


Full Text









CONTROL SCHEMES IN HIGH SPEED
REAL-TIME COMMUNICATION NETWORKS








BY

LI-TAO SHEN


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


1992















ACKNOWLEDGEMENTS


I would like to express my sincere thanks to my advisor, supervisory committee chairman, and friend, Dr. Yann-flang Lee, for his guidance and encouragement. Without his support, this work would never have been accomplished. During the years of this relationship, he enlightened me not only by his knowledge, but also with his way of exploring truth.

I also would like to express my appreciation to the other members of my supervisory committee, Dr. Randy Chow, Dr. Richard Newman-Wolfe, Dr. Panos Livadas, and Dr. Haniph Latchman, for their commitment and service on this committee.

Special thanks go to my wife, Min Pu, for her understanding and continuous support for this work.















TABLE OF CONTENTS

page

ACKNOW LEDGEMENTS .................................................... ii

A B ST R A CT .................................................................. v

CHAPTERS

1 INTRODUCTION .............................................. 1

1.1 Problem Statement ...................................... 1
1.2 Background and Related Works ......................... 4
1.3 Dissertation Outline ..................................... 8

2 REAL-TIME COMMUNICATIONS AND CONTROLS ........ 11

2.1 Real-Time Communication Service ..................... 11
2.1.1 Applications and Requirements ....................... 11
2.1.2 Time-Constrained Performance Measures ............. 14
2.2 Real-Time Communication Support .................... 16
2.2.1 Networking Architectures ............................. 16
2.2.2 Traffic Characterization .............................. 19
2.2.3 Control Enforcement Structure ....................... 20
2.3 General Control Approaches ............................ 21
2.3.1 Priority-Driven Mechanism ........................... 21
2.3.2 Dynamic Scheduling Control Approach ............... 23
2.3.3 Message Level Processing Approach ................... 25

3 MEDIA ACCESS CONTROLS IN LANS AND MANS ......... 27

3.1 Multiple Channel Token Ring (MCTR) ................. 27
3.1.1 Background .......................................... 27
3.1.2 System Architecture Model ........................... 30
3.1.3 WUT Control Strategy and Mechanisms .............. 32
3.1.4 Performance Evaluations ............................. 37
3.2 Distributed Queue Dual Bus (DQDB) .................. 47
3.2.1 Background .......................................... 47









3.2.2 Dynamic Control Model and Strategies ............... 51
3.2.3 Priority Assignment Algorithm ....................... 54
3.2.4 Simulations and Evaluations .......................... 56

4 MESSAGE LEVEL PROCESSING SCHEMES ................. 65

4.1 Introduction ........................................... 65
4.2 Message Level Processing .............................. 66
4.3 Analysis Model and Evaluations ........................ 69
4.3.1 Service Time Distribution ............................ 69
4.3.2 Waiting and Delay Time .............................. 70
4.3.3 Num erical Results ................................... 71
4.4 Transmission Control Schemes .......................... 73
4.4.1 Multiple-Slot Reservation Protocol .................... 73
4.4.2 Enhanced Multiple-Slot Protocol ..................... 76
4.5 Performace Evaluations ................................ 78

5 INTEGRATED SCHEMES IN WIDE-AREA NETWORKS .... 83

5.1 Introduction ........................................... 83
5.2 Control Structure Model ............................... 85
5.2.1 Two-Level Control Model ............................. 85
5.2.2 Challenges ........................................... 88
5.3 Integrated Control Scheme ............................. 89
5.3.1 Algorithm for Connection Setup Control .............. 90
5.3.2 Algorithm for Switching Control ...................... 94
5.4 Evaluations and Analysis ............................... 99
5.4.1 Simulation M odel ................................... 100
5.4.2 Admission Controls ................................. 102
5.4.3 Swtching Controls ................................... 105

6 CONCLUSIONS ... .......................................... 112

R EFEREN CES ............................................................. 116

BIOGRAPHICAL SKETCH ................................................ 129















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 CONTROL SCHEMES IN HIGH SPEED REAL-TIME COMMUNICATION NETWORKS By

LI-TAO SHEN

November 1992


Chairman: Dr. Yann-Hang Lee
Major Department: Computer and Information Sciences

Recently rapid engineering advances and new applications of distributed computing technology impose stringent requirements on computer communication networks in supporting a wide variety of services. In additional to fast data delivery, features such as fault-tolerance, multicasting and security, are all in demand. Most importantly, as the distributed applications move to real-time domains, many timing requirements have to be enforced within the communication systems. A real-time communication network is such a system which is able to provide its users the ability to specify the timing requirements and to obtain guarantees about the satisfaction of those requirements. The predictable operation and a high degree of schedulability are two of the most desirable properties of a real-time communication network.

Timing correctness, as well as the traditional functional correctness, is ex-








tremely vital in various distributed real-time applications. However, the current networking systems employed in various time-constrained applications are generally not real-time communication systems. The main problem here is that there is no explicit real-time control schemes employed in these communication systems to guarantee the timing requirements.

While the multiple priority has been identified as one of the most commonly used mechanisms for current real-time communications, it is not very clear how to assign the proper priority consistently in a complex and dynamic networking environment. Several dynamic scheduling approaches are proposed and studied in detail in this dissertation. The main idea of these approaches is to schedule the messages to traverse through the network in a dynamic way such that the user specified timing requirements can be maximumly satisfied. By dynamic, it is meant that the priority should be time-variant and decided by both the deadline requirement and current network operating status. By scheduling, it is meant that the message transportation in real-time communication networks should be emphasized to minimize the lost percentage. A set of generic control schemes are proposed based on the general approaches and also evaluated extensively to show their promising performance for real-time communication applications.
















Chapter 1




INTRODUCTION





1.1 Problem Statement


Real-time communication networks are distinguished from the normal network systems with the introduction of time constraints. They are used to insure on-time delivery of messages and to support distributed real-time computations. The performance measures of such networks differ from those of the conventional networks. The principal performance considerations for the conventional network control protocols are to maximize the system throughput and to minimize the average delay. In real-time communication systems, however, the main performance consideration is to maximize the percentage of messages that are delivered within the given time constraints [50, 83]. The different performance metrics, reliability requirements, and performance trade-offs suggest that the control protocols previously developed for traditional communication networks may no longer be suitable for time-constrained communications.

According to the well-known open system interconnection (ISO/OSI) reference









model, the communication architecture can be identified as a seven-layered processing stack. Two main subsystems can be further defined:

" end-host processing system

" networking communication system

The former one is totally transparent from networking details. It is built upon a transport system meeting certain quality of service (QOS) requirements, and provides the end-host users with a unified application service. The functions of this interoperating system actually accomplish a kernel distributed operating system, which is usually corresponding to the Session layer, the Presentation layer, and the Application layer of the OSI model. The second part of the architecture is the transport system whose basic functions and services can be viewed similar as those of the Physical layer, the Data Link layer, the Network layer, and the Transport layer of the OSI model. While the end-host processing system is very important for providing application users various powerful and efficient services, its functional capability would mainly depend upon the networking communication system.

It can be obviously observed that there is a close relationship between the endhost processing system and the networking communication system. This is especially true when we deal with real-time distributed systems. However, there are some fundamental differences between real-time communication networks and traditional communication networks. This is majorly because a new performance dimension is introduced into the traditional model when the additional timeliness requirement should be considered and enforced. It is well recognized that real-time communications will be the backbone for the next generation networks where predictable communication services are the basis for further application development.








We believe that, a solid understanding of real-time communication networks and a developing methodology of such networks can not be overlooked to address the real-time distributed operating systems and applications.

Generally, the current networking systems employed in various time-constrained applications are not real-time communication systems. Some main problems include:

* lack of explicit timing representation and operation mechanism;

* undetermined communication services due to the network routing and delay;

* local and global priority inversion across different processing domain;

* the improper mapping of message's timing attributes;

* weak dynamic control mechanisms to adopt to various application environments;

The key problem here is that there is no explicit control support employed in the current communication systems to guarantee the real-time requirements, which is actually quite essential to many distributed processing environments.

It is the real-time communication system that must do whatever is necessary to bring the quality of service provided by vast variety networking architecture up to the level required by the communication service users. It is easy to see that this is an integrated control system with a substantial amount of processing efforts and various decision choices. Therefore, the real-time networking communication system is an integrated multilayered system, which not only provides the principal functions of conventional OSI lower layers but also, of the same importance, provides integrated real-time control mechanisms at each individual layer. While the









conventional protocols are not real-time oriented, a new design methodology and a set of new protocols are to be developed. The objective is to provide application users a predictable and guaranteed real-time communication service interface. The efficiency of the system is to be achieved by selecting the most suitable control function for the requests given by application users.

This research is intend to investigate the effective control schemes as well as the proper architecture model for real-time communication networks in a distributed environment. The main objectives of this dissertation research are:

e to investigate the communication architectures with guaranteed performance

in a high speed real-time environment;

9 to design, analyze, and evaluate various effective control schemes in real-time

communication systems,


Besides a better and more detailed understanding of the increasing real-time communication systems, this research work is intended to devise a number of generic control schemes and algorithms which can be applied in practical environment to enhance the real-time performance. It also provides significant values to the fundamental control of real-time communication in distributed systems and meets the increasing demands of real-time communications.



1.2 Background and Related Works


The OSI Basic Reference Model [42] evolved out of early work to describe the communications infrastructure required by applications such as banking, airline reservations and ticketing, and other industries requiring distributed access to large









data bases. The emphasis in such systems was on information accuracy and system reliability [19, 20, 31, 36, 71]. Real-time response was not considered as a prime design goal. The network system was deemed as a success if almost all queries were answered within a few seconds [66]. As a result, most standards that developed under the OSI model ignored the needs of industrial applications, where timeliness of messaging is paramount. Issues such as message priorities, bounded messaging delay, and efficient multi-node communication via selective broadcast either were ignored, or were relegated to the bonepile of unresolved standards problems by designating them "for future study".

With the advent of high-speed networking technology, the existing communication architecture and control schemes are no longer appropriate. New design and analysis methods are introduced for very high-speed networking architectures [3, 13, 22, 61]. The requirement for small delays and low processing overheads has also brought about the development of so called light-weight protocols [27]. Some of the representative systems are Xpress Transport Protocol (XTP) developed at University of Virginia [75], Versatile Message Transaction Protocol (VMTP) developed at Stanford University [11], Network Bulk Transfer (NETBLT) developed at MIT [16], and GAM-T-103 Transport System developed in France [59]. All these systems are dedicated to the single transport layer with no supporting control schemes in underlying layers or specific integrated control schemes. The main effort of these systems is to minimize the processing overhead for data communications.

Real-time computing systems is expected to be the backbone of the next generation system. The system requirements and concepts were well discussed by Ferrari [29, 30], Krishna and Lee [49], Kurose and et.al [50], and Stankovic [83, 84]. Several real-time system model as well as design strategies have also been proposed.









The ARTS is a distributed real-time kernel system developed by Tokuda and et.al [94]. The Spring is a real-time operating system kernel developed by Stankovic and Ramamritham [85]. The HARTOS is a distributed real-time system developed by Kandlur and et.al [44]. In addition, the CHAOS is an operating system developed for real-time applications by Gopinath and Schwan [33]. While most of these systems assume a guaranteed communication service and emphasize on operating system supports for real-time applications, the issues of real-time communication have not been well addressed.

There is a lot of research being undertaken in developing various media access control (MAC) schemes due to the wide installation and application of the local and metropolitan area networks. Surveys of media access control mechanisms in highspeed network systems were made by several researchers [3, 13, 25, 73]. However, real-time requirements were not considered in the study of most of the local and metropolitan area networks. IEEE 802.5 Token Ring network was well discussed and analyzed by many researchers [7, 8, 9, 40, 65, 78]. New technology like optic fiber has brought about a lot of attention on both the ANSI FDDI networks [4, 5, 72, 77, 79, 96] and IEEE 802.6 DQDB networks [17, 21, 41, 43, 63, 76]. Various analytical model and evaluations were made to these medium access control protocols [6, 34, 69, 70, 89, 95, 98].

Some enhanced and new protocols were also developed recently in the context of real-time communication requirements. Shin and Hou [81] evaluated three contention protocols used for real-time communications. In their model, the probability of missing message deadlines was taken into account and some analytical results were obtained. Virtual time CSMA protocols were proposed by Zhao and et.al [100, 101] for hard real-time communications. Strosnider and et.al [87, 881 developed a de-









ferrable service algorithm for periodic message scheduling. The algorithm not only guarantees the deadlines of periodic messages, but also substantially reduces the response time of aperiodic messages by assigning the highest priority to the aperiodic messages up to the point where periodic messages would start to miss their deadlines. The consideration of applying this algorithm to the IEEE 802.5 token ring protocol were discussed in [87].

The general wide area networks with any networking topology introduce more complicated problems and challenges [14, 15, 23, 28, 29, 45, 74], though they have the potential for higher performance and reliability than common bus or ring structures. Maxemchuk and Zarki [57] made a good survey on routing and flow control in high-speed wide-area networks as well as local area and metropolitan area networks. Since the information is delivered across multiple hops along the network, the issue to be addressed is not the control schemes for medium access but the message scheduling along the network with delivery time constraints. Ferrari [29, 30] specified several real-time performance measures from the point of user application requirements, and presented a real-time channel scheme which was claimed to be able to provide guaranteed performance. A multi-hop network model and scheduling scheme were further developed by Kandlur and et.al [45] to provide predictable inter-process communication in real-time multiple stage networks. Cidon and et.al [15] discussed their proposed methods for bandwidth management and congestion control in very high-speed networks and their experience from the plaNET wide area network [32]. While the connection establishment management is very important in wide area networks, the proper queueing and scheduling schemes at switching components are also crucial to the system performance [51, 74]. A priority assignment control scheme with quality of service constraints was proposed by Takagi








and et.al [90] for line buffers in ATM networks. Fair queueing algorithm was first proposed by Nagle [62], and was studied by several researchers [24]. The reasoning of these proposed algorithms was to prevent a source from arbitrarily increasing its share of the bandwidth or causing the delay to other sources. However, the time constraints were not the principal consideration despite the fact that the improper queueing discipline and scheduling could cost unnecessary percentage of out of deadline messages.



1.3 Dissertation Outline


Real-time communication can be conceived as a complex system that satisfies multiple classes of transmissions under various time-constraints imposed by application users. Simple and effective control mechanisms are the key components to guarantee the service requirements even if the communication system is under high utilization.

In the next Chapter, the general characteristics of real-time communication and control mechanisms are first discussed. By presenting an application example, the real-time service requirements are identified in detail. Dedicated real-time performance measures are also proposed, which are significantly different from the traditional networking performance measures. The supporting structure for real-time service is then presented. Several control strategies and mechanisms are proposed and discussed in general for real-time communications.

In Chapter 3, the effective media access control schemes are discussed in detail in the context of providing real-time services. In both the local area networks (LANs) and metropolitan area networks (MANs), the media access control (MAC) plays a key issue to provide real-time performance. The implication of deadline









and/or priority information to the design of MAC layer protocols is carefully investigated. Since MACs are closely network related, special implementation considerations should be given to each type of networks. Two examples have been used to examine the control schemes of different MAC protocols in high-speed local and metropolitan area networks. A wait-until (WUT) control scheme is first proposed for multiple-channel token ring networks. Then, a dynamic priority assignment control scheme is designed for distributed queue dual bus (DQDB) networks. It has been showed that the performance gain achieved could be substantial in real-time communications.

The message level processing (MLP) strategy and control schemes are discussed in Chapter 4. Specifically, several message level processing schemes have been studied in the DQDB network. The standard DQDB protocol is based on segment level operations and is not very efficient from the application users' point of view. A class of multiple-slot transmission schemes have been developed to reduce the message level transmission delay. The main idea of the proposed schemes is to perform the service reservation request at the message level instead of at the segment level. The proposed multiple-slot reservation protocols contribute to the DQDB network with processing efficiency, better bandwidth usage, and improved message delay performance. In additional to these, they are highly compatible with the original standard and are able to suppress the unfairness problem.

In Chapter 5, the design methodology for real-time control schemes in high-speed wide area networks (WANs) is discussed. Due to the additional multi-hop switching system, guaranteeing the required quality of service in WAN brings even more challenges for real-time communications. An integrated control scheme has been proposed and studied, which includes the connection setup control at networking








access point and dynamic scheduling control at switching nodes. The connection setup procedure not only checks the availability of the network resources for specified requirements, but also assigns the delay bound vector along the route. While the switching system is fundamental in wide area network environment, effective queueing discipline and dynamic scheduling schemes are carefully designed to accommodate the requirements of various QOS classes and to achieve good real-time communication performance.

Finally, a brief summary of the dissertation work as well as some further research works are given in Chapter 6.
















Chapter 2




REAL-TIME



COMMUNICATIONS AND



CONTROLS





2.1 Real-time Communication Service


2.1.1 Applications and Requirements

Real-time communications are driven by many distributed applications which need guaranteed real-time services at the communication interface. While the distributed systems become even more popular, the variety of the user requirements can be so vast. Besides the usual functional requirements such as transmitted messages should be free of error and in order, there are some other important requirements such as reliability and security. As the applications go to the real-time domain, timing requirements must be considered and guaranteed. These timing requirements include messages should be transmitted before their deadlines and the percentage of mes-

















NETWORKS


Figure 2.1: Service Requirements


sages missing deadlines should be minimized. All these various requirements can be shown as in Fig 2.1 and should be met at the interface of communication networks. In fact, the real-time communication is just one kind of networks which is able to enforce the timing requirements and provide predictable communication services. Within real-time communication networks, not only the functional correctness but also the timeliness correctness must be guaranteed.

The real-time communications and timing requirements are not something new and can be easily observed in our everyday life. There are enormous applications which have various real-time communication requirements. These applications cover almost every aspects, including bank transactions, airport scheduling control, multimedia telecommunications, and military networking management. One realistic example can be found in the Distributed Interactive Simulation (DIS) networking environment [103], which is sponsored by the Defense Advanced Research Project


Timing Requirement
* before deadline
* loss rate








Agency (DARPA) in partnership with the United States Army. The goal of this simulated battlefield program is to provide, for the first time, an opportunity for fully-manned platoon-, company-, and battalion-level units to fight force-on-force engagements against an opposing unit of similar composition. A description of a vehicle's appearance passes through several hands from the time it is first expressed as a protocol data unit (PDU), to the time the vehicle is displayed by an observer's simulator. The steps include processing by the software that provides communication services, transmission across a network, and perhaps queueing within the receiving simulator. The magnitude of this discrepancy is proportional to the speed of the vehicle described by the PDU, and to the magnitude of the network delay. Therefore, this effect is expected to be most evident and critical in certain situations such as when aircraft flying at high speed are able to observe each other closely while being simulated by widely separated simulators. It is quite obvious that efficient and fast are relative terms and are not sufficient when dealing with real-time requirements. In this case, timing correctness, as well as functional correctness (received correctly and in order), is extremely important for application requirements.

Certainly, a specific priority field can be added into PDUs to represent the criticalness of the message. The problem of real-time communication is unfortunately not so easy and far from being solved. It is not the real-time requirement or the user interface format, but the underlying mechanisms employed in existing networks that must provide explicit or enough supports for real-time communications.








2.1.2 Time-Constrained Performance Measures

Interests of new timing performance are recognized for real-time communications. Several performance metrics have been proposed and defined. Instead of the traditional performance measures like average delay, throughput, and fairness, real-time communication systems emphasize on the guarantees of the timing requirements and system utilization.

The most important performance measures, such as transport deadline, message lost percentage and channel utilization, can be defined as:

" Transport Deadline Duration (D):

The transport deadline duration is the period from the time a message entering

the system until the user specified expiration time.

D = expired time - arrival time (2.1)


" Message Lost Percentage (L):

A message is considered to be lost if it cannot be delivered before its deadline

is expired. The lost percentage is defined as follow.

L = no. of no. of messages lost
(no. of messages lost + no. of messages sent out)

" Channel Utilization (U):

The channel idle time is the time spent when a node has nothing to send.

This measure is limited to the on-time delivery.

= message transport time
(message transport time + channel idle time) (2.3) The increasing real-time data communication requires integrated networking transmission service for multi-class message traffic [74, 831. One important issue is








to satisfy all those multiple classes of transmissions so that the various service delay requirements are guaranteed even the communication system is under high utilization. Different classes of transmissions can be classified by their different requirements in time-dependent constraints (deadlines). With smaller deadline constraint, this class of transmission is often considered to be more urgent and certain service preference is needed to guarantee its real-time requirement. Priority queueing [70] is one of the mechanisms often used for real-time communication since it provides an effective means to give preference to individual message class.

In complex real-time communication systems, various types of traffic are stochastically multiplexed to efficiently utilize the network resources. However, the excessive use of the bandwidth may cause traffic-dependent quality of service (QOS) deteriorations. Generally, the service quality in communication context refers to accuracy and speed of information delivery, and to the absence of certain impairments such as excessive delay, excessive variance of delay, transmission error and out of sequence delivery. In additional to these, the quality of service (QOS) in real-time communication also includes

* the required lost percentages for each class of message,

* the bounded transport delay for eact, message class,

* the high system utilization.

There are two possible common ways of QOS management in real-time communications: (1) to define a single QOS class and manage the transmitting information equitably; (2) to define multiple QOS classes and manage each different classified class individually. Although the control scheme could be much simpler in the first method, the real-time performance is very hard to be ensured for a wide variety of








user requirements. The advantages of the second method are the bandwidth efficiency can be improved by suitable control such as priority scheme and the users are able to select one of the classes most appropriate to their requirements.

The multiple QOS classes strategy is of more interest in this research. Whenever a user intends to set up a network connection, the required QOS class of the sending messages should be clearly specified at the network entrance points. It is worth reminding that while the connection is class-oriented, the traffic on any particular link could come from various sources and belongs to multiple classes. Therefore, the control schemes need to be designed to make efficient use of the transmission bandwidth while satisfying the QOS requirements for all the classes.



2.2 Real-time Communication Support


2.2.1 Networking Architecture

Networking architecture can have significant effect on the design of effective control schemes. Different architecture not only invokes different networking protocols, but also has its exclusive advantages and special problems.

For local and metropolitan area networks, the architecture is usually very simple. The most common ones are star, ring, and bus as shown in Fig 2.2. There is only one networking access point and no intermediate switching node. The media access control (MAC), thus, is the key for cooperating all the nodes and supporting required communication services. The standard message transmission has two phases:


* compete for the media access control




























Figure 2.2: Architecture for LANs/MANs


9 transmit the message through the media

While the first phase is governed strictly by the MAC protocols such as CSMA/CD (IEEE 802.3) and Token Ring (IEEE 802.5), the second phase usually has the determined execution period. In these networks, therefore, the major challenge for real-time communications is how to make the best use of the special architecture and design the effective MAC protocols so that specified services can be provided.

For wide-area networks, there are basically two major different ways in dealing with the integrated data communications. In the circuit-switched (synchronous transfer mode) networks, sufficient resources are allocated to each call request to handle its maximum utilization. This guarantees that the user will get the quality of service required, but, on the other hand, may be wasteful of system resources. In the packet-switched (asynchronous transfer mode) networks, call requests from all sources are packetized, and statistical multiplexing techniques are used to combine


STAR


0 6 *








all network traffic through multihop switching fabrics. This allows higher network utilization. However, more complex and proper control schemes are inevitable to ensure the required service, especially for real-time communications.

The kind of wide area networks to be considered is a class of networks based on Asynchronous Transfer Mode (ATM) design principle and architecture. While there are various components in a wide area network, such as monitor and auditor, our study will focus on two primary components: interfacing nodes and switching nodes.

* Interfacing nodes

Interfacing nodes are those attached to external components in either network entrance point or exit point. They are responsible for the connection setup

and management.

e Switching nodes

Switching nodes are collectively referred as the switching subsystem which is the major part of a wide area network. A switching node can be logically

regarded as three parts: input buffer, switch, and output buffer.

It should be noticed that control schemes in wide-area networks usually invoke more challenges due to the intermediate switching systems. Two levels of control are needed in wide-area networks: connection setup control at the network access point and switching control within the network. Various schemes have been developed and employed at each level to provide specified communication services.









2.2.2 Traffic Characterization

It is believed that the guaranteed real-time communication service could only be accomplished by making use of certain types of resource reservation. The system needs detailed characteristics of the required guarantee service so that it can reserve the corresponding resources and avoid possible interference from other contending requests. However, the non-time-constrained traffic obviously does not have to reserve the bandwidth and compete with the time-constrained traffic. This traffic can use the gaps in the bandwidth usage of reserved ones, and its behavior will not affect the quality of service given to the reserved traffic.

To characterize the traffic features for each connection, it is necessary to select an appropriate model to specify the traffic in terms of known parameters. The two-state fluid flow (STFF) model is adopted and recognized as a proper means to capture the features of a wide range of connections. In this model, the source is either in an idle state transmitting at zero bit rate, or in a burst state transmitting at its peak rate. The advantage of using such a model is that it is flexible as well as simple. It can represent connections ranging from burst to continuous bit stream, or even as the approximation of more complex sources

Based on this model, the idle and burst periods are defined as the times during which the source is idle or active respectively. The peak rate of a connection and the distributions of idle and burst periods completely identify the traffics statistics of the connection. Specifically, connection i is represented by a request vector (Rpak,i, pi, bi), where Rpak,i is the peak rate at which the source generates data, pi is the utilization or fraction of time the source is active and transmitting at Rpeak,i, and the bi is the average duration of an active period. From these three













b


R peak


Figure 2.3: Two-State Fluid Flow (STFF) Model


basic parameters, we can also derive the mean bit rate mi and its variance oi.


ni = pRpeki (2.4) = mi(Rpok,i - Mi) (2.5)


It is worth mention that the mean burst period bi gives the information of how data is being generated by the source. Two sources, with identical mean and peak bit rates but different burst periods, have different impacts on the network.


2.2.3 Control Enforcement Structure

The whole control structure for real-time communication systems consist of both the system users and system providers. The users are responsible for the specification of the specific real-time requirements such as bounded delay and lost percentage. The users are also required to have an effective way to describe the expected traffic pattern to go through the network. Given these, the communication system provider should have the ability to satisfy the user requirements transparently. The layered structure and interface can be shown as in Fig 2.4. The users will be notified if the








Application Service

Communication Service
(Traffic Pattern
Deadline
Lost Percentage
QOS Class
. 0 S )


HS-NETWORK


Figure 2.4: Control in Real-Time Communications


provider cannot guarantee the required service at the current network operating status. The user can either withdraw the request and try later, or continue the request at the best-effort manner. As stated before, the media access control (MAC) protocol has been identified as the key challenge for local and metropolitan area networks. For wide-area networks, the challenges of real-time control come from two closely related parts: connection setup control and intermediate switching control.



2.3 General Control Approaches


2.3.1 Priority-Driven Mechanisms

The multiple priority mechanism is the most commonly employed scheme in attacking real-time communication problems. Examing carefully the existing communication systems, it is not difficult to observe that these systems actually do not








have consistent mechanisms explicitly in supporting real-time application requirements. Although the media access control protocols in some local and metropolitan area networks, like Token-Ring, DQDB, and FDDI, provide priority mechanisms for possible differential service, some others like CSMA/CD even do not have priority mechanisms. In these latter ones, priority inversion problem is common where higher priority (critical) messages may be transmitted after lower priority ones. This is a very undesirable property, especially for real-time communication systems where timing requirement is emphasized. Even for those networks providing priority mechanisms, there also exist problems such as the static priority assignment, inefficient enforcement scheme, and insufficient priority levels.

The problem in existing transport layer protocols is even worse. The expedited service of OSI transport class 4, for example, is weak in definition and undefined in implementation. Furthermore, it is unclear how such priorities are mapped to lower layer services, even if the priority is preserved. TCP uses a called Urgent field in its TPDU structure to indicate that some number of bytes are special and should be processed out of order [92]. It is still not clear to the users how to make use of this facility instead of sending interrupt messages. Some recently developed systems like Xpress Transport Protocol (XTP) and Versatile Message Transaction Protocol (VMTP) are expected to be more appropriate for high speed and real-time communications. While they concerned more about high speed processing, the realtime mechanism is not well addressed. The priority mechanism in XTP intends to support both incoming and outgoing messages [75]. For outgoing messages, the priority level is encoded into a 4-byte integer and placed into the SORT field before transmission. When the message arrives at the remote receiver, the SORT field is examined, and the message is enqueued according to its priority. However, this









scheme is of static type, in that the priority level remains constant as the message travels through the network.

It can be observed that the priority mechanisms in existing networks are either not sufficient or even not available system-wide. Even they are available through the whole protocol stack, inconsistency is another big problem. The priority in these systems is generally an intra-layer concept rather than a system-wide control mechanism. On the other side, it is anxious to know how to use this distributed mechanism effectively in real networking practice. It is argued that the application users should not take care of the priority assignment. Since the users are never able to track the detailed operations of the communication service provider, they often feel quite at a loss at priority assignment except following a static policy. The user should give the specific requirement like deadline, and it is the responsibility of the system provider to decide what and which is the most suitable mechanism, say priority, to satisfy the requirement.


2.3.2 Dynamic Scheduling Control Approach

While the multiple priority mechanism has been identified as one of the most commonly used control schemes for current real-time communications, it is not very clear who and how to assign the proper priority consistently in a complex and dynamic networking environment. A dynamic scheduling control approach has been developed as a general approach and investigated in detailed in various specific networks.

The main idea of this approach is to schedule the messages to traverse through the network in a dynamic way such that the user specified real-time requirements can be maximumly satisfied.








e By dynamic, it is meant that the priority should be time-variant and decided

by both the deadline requirement and current networking operating status

* By scheduling, it is meant that the message transportation in real-time communication network should be more emphasized as a scheduling issue to minimize the chance of missing message deadlines

Instead of static control mechanism like user specified priority, the priority should be time-variant for effective real-time controls. The priority is more likely in the mechanism domain which is employed by the service provider and transparent to the users. Users do not have to worry about the detailed mechanisms like priority but simply specify the requirement and policy. It is the provider's responsibility to provide the required service by optimizing the networking operating status and user requirements. The slack time, from the current time up to the expiration time, is widely used to reflect the dynamic property of each message. While the smaller slack time usually implies more attention for this message, various QOS classes and current network operating status have also to be considered to make the best use of the system and provide the required services.

The prime objective of real-time communications is to provide predictable service instead of fast transportation. As long as the message can be transmitted within its deadline, there is no reward for early delivery. In another word, the message transportation should be carefully scheduled so that the message delivery can be guaranteed and also system can be highly utilized. The interaction and ordering of various QOS classes have to be resolved dynamically in response to changing traffic flow. It is also very important, for the control schemes to be distributed. Efficiency, flexibility, and predictability are some other desirable properties have to








be considered.


2.3.3 Message Level Processing Approach

Message level processing approach is quite natural since the timeliness is an attribute belongs to the message level. Since the processing is now driven by each individual message rather than the segmented packets, it is a message-driven model with the relevant control schemes integrated as a single processing object. All the objects are interacted each other via their relative time constraints. The more urgent an object is, the more processing effort should be considered. This model actually provides a system-wide virtual priority processing mechanism in which the priority assigned to each message is dynamic and time-dependent. This is because that the message criticality is not only staticly determined by the real-time application but may also be dynamically changed as the message keeps remaining in the system. A time-constrained message with a very long deadline could also be very urgent when this message has been blocked for enough time. Therefore, there is no static priority imposed upon each message. All the scheduling and queueing algorithms are designed and "driven" by the time requirement at the message level. While this model naturally introduces the messages as parallel processing objects, some parallel processing architecture can be employed to provide potential much more fast processing.

Since segment delay does not make any sense to the application users, and only the complete message level performance is what they are really interested. It can be observed that this model is controlled at semantic level rather than the syntax format. The basic data unit for manipulation and control management is the message instead of the packet. Some benefits can be obtained including more






26


effective transmission and better real-time performance. For instance, when certain portion of a message fails to meet its deadline, then the rest of this message should be aborted, since there is no way to satisfy the requirement of this message. Also, endto-end time fence checking mechanism based on message boundary is quite effective to minimize the dependence of one message on the timing characteristics of other messages. Based upon the delay bound provided by network, this mechanism can quickly isolate a timing violation. By requiring explicit timing information about each message, a time encapsulation at message level can be achieved.














Chapter 3




MEDIA ACCESS CONTROL IN



LANS AND MANS




3.1 Multiple Channel Token Ring (MCTR)


3.1.1 Background

Token ring network systems have been extensively studied and also widely used during the past decade [7]. The basic single token ring network consists of a number of stations N attached on a ring and a control token rotates around the ring, station by station. If the station receiving the free token has message packets to transmit, it converts the free token into a connector and then follows the connector with its sending message. If the station has no message waiting for transmission, it simply passes the free token to the next station. The source station has the responsibility to remove the packet from the ring and to generate a new free token after removing the packet and then passes it to the next station.

Compared to other structures, the ring networks have the good properties of







bounded transfer delay and better channel utilization [6, 781. However, the transfer delay will be considerably high under moderate and high load. It is even worse when the messages are critical packets with real-time requirements. This is due to the fact that most current protocols aim to minimize the average transmission delay and have no mechanism to favor the critical packets.

For these real-time network systems, it is obvious that the round-robin methods resulted from token rotation are not appropriate. The priority based methods, which aim to favor messages according to their priorities, are the currently prevailing techniques. A real-time scheduling method for prioritized messages has been proposed and investigated by Strosnider [88]. Using IEEE 802.5 token ring protocol, it has been shown that a better real-time performance can be achieved by choosing a proper packet size and operating the token in priority mode.

Currently a high speed optical fiber ring network known as FDDI is being developed and standardized. The key characteristics of FDDI include optical transmission medium, fair and robust control protocol, a datarate of 100 Mbit/s, a ring length of up to 100 km and up to 500 stations on the ring. A fully decentralized priority mechanism is used in FDDI supporting synchronous and asynchronous transmission modes. Synchronous transmission is real-time sensitive and the delay of synchronous transmission is limited by reserving an appropriate bandwidth at ring initialization. The remaining available bandwidth can be used for asynchronous operation. Performance analysis has shown that the throughput and the real-time response of the FDDI cannot be optimized simultaneously, especially for long ring lengths [77].

Though the priority based and bandwidth allocation methods are much better than the conventional methods in real-time applications, there are still some







problems:

" the introduction of a high bandwidth channel may be accompanied by only insignificant increase in system capacity, and the increase of channel bandwidth

can only be partially utilized;

* the distribution of time-constrained messages is usually not predictable, and

a static allocation of bandwidth is hard to meet the general real-time requirements;

" to preempt the normal token rotation, a lot of time is wasted in doing token

reservations and hence the system utilization is low;

" the starvation and unnecessary delay of low priority messages is quite common; " the total percentage of messages that miss their time constraints may still be

quite high though that of the most urgent ones is decreased;

For a high speed transmission medium, it is attractive to partition the medium into multiple channels. It was also shown that for a given system bandwidth, the system capacity can be increased by bandwidth subdivision [13]. These channels can form a multiple ring network and have two main advantages: (i) they increase network capacity by operating on several slower channels so that the propagation delay and other penalties become a smaller fraction of the packet transmission time, and (ii) they can be easily implemented by expanding the existing interface technologies based on medium speed.

Several approaches have been suggested for multiple ring networks [9]: (1) separate queues with simultaneous transmissions, (2) single queue with simultaneous








transmissions, (3) single queue with single transmission. The analysis and simulation results showed that the single queue with simultaneous transmissions protocol has a better performance than the other two protocols but its interface design is more complex than the others.

In this section, a new protocol is presented which is designed for real-time communication in multiple token ring networks. With proper channel allocations and priority reservations, the protocol can reduce the percentage of messages that miss their time constraints and also maintain a high channel utilization. It is a dynamic control policy, flexible to any kind of system load, and easy to implement.


3.1.2 System Architecture Model The real-time communication model can be specified as follows: for each packet Pi, i = 1,2,... , m, there are two principal parameters A, and E. Ai is the arrival time of P and Ej is its expire time before which P must be transmitted and received. We define the deadline duration for P, Di, as the difference between Ai and Ej, as shown in Figure 3.1. The packets that miss their deadline are considered to be useless and lost no matter whether they are received or not.

It is assumed that the network considered consists of a physical ring which is "divided" into k channels and is operated as k token rings, C1,C2,...,Ck. This way, the ratio of propagation delay and packet transmission time in each channel can be decreased so that a high utilization can be achieved. Then each two rings Ci and C,+1 are grouped as a pair < Ci, Ui+1 >, where i is an odd number. To allocate packets to the two rings of a pair, we classify the packets into two groups according





31





Current Time

Arrival Time Expire Time
9/

Packet Slack Time T,-.
Packet Deadline Duration


Figure 3.1: Time Attributes of Packet


to their deadline values and a predefined threshold DL, tight deadline packet if Di < DL loose deadline packet if Di > DL In the next subsection, a dynamic load control protocol will be presented which aims to utilize this multiple token ring network with the consideration of timing constraints. The most important concern is the opportunity provided in this architecture such that the input load can be allocated dynamically into different rings according to different system requirements and the potential performance can be achieved in real-time applications.

To design the network controller in high-speed networks, the effort is to have a simple interface design. After the division of physical ring into channels, medium or low speed interface can be incorporated into each channel. Control mechanisms then can be associated with each ring and arriving packets can be routed with simple scheme. Thus, we eliminate the consideration of an integrated network controller for multiple rings in each station.




























ARCHITECTURE MODEL CHANNELS Figure 3.2: Architecture Model


3.1.3 WUT Control Strategy and Mechanisms Based on the architecture given before, a dynamic control protocol is proposed to make token ring networks achieve a high performance in distributed real-time applications. The proposed control protocol is based upon the following idea. The critical packets with tight time constraints need not necessarily request for token reservation immediately after they arrive in to the system. They can still wait for the chance of transmission like all the other packets until their slack time passes certain limit, that they are going to fail to meet their deadlines. These special packets are called alert packets. The alert packets have the privilege to do reservation at once to preempt the normal packet transmission until all alert packets have been sent out. Since each token reservation wastes certain rotation time, the preemption due to alert packets should be minimized. This can be attained through a proper








allocation. The protocol consists of four parts as follows:



A. Allocation of Incoming Packets

When packets arrive at a station, they are routed to consecutive pairs of the network in a round robin manner. In each pair, < Ci, C+j >, all the tight deadline packets are routed to Ci and all the loose deadline packets are routed to C+1. So now we have a multiple channel token ring network with k/2 subsystems of the same property.



B. Wait-Until (WUT) Control Policy and Switching

Within each pair of rings, < Ci, C+1 >, the wait-until (WUT) control policy is adopted to switch an alert packet from C, to Ci+,. Then, the alert packets will preempt the loose deadline packets in C-7+1 so that the on-time delivery can be guaranteed. The first ring of a pair Cj is operated in non-priority mode and allows no token reservation. When the free token arrives, it send out a tight deadline packet if its waiting queue is not empty, otherwise it passes the free token to the next station. At the same time, it checks and switches those alert packets, that are going to fail to meet their deadlines, to the next channel C+j which is allocated to those loose deadline packets. The checking process is done based on the following formula:

Tslack < Tgw,;ch (3.2) where Tot,, is the packet slack time. The value Tswitch classifies alert packets and functions as a maximum sliding window to control the flow of switching as shown in Figure 3.3.








On the other hand, the second ring of a pair, Ci+1, works in a reservation mode and an alert packet switched from Ci can reserve the token and preempt the normal token rotation. This will guarantee that all alert packets can be sent out immediately before the loose deadline packets are delivered.

The benefit of this method is that it guarantees real-time requirements while still keeping relatively high channel utilization. By using preemption, the real-time requirements can be met. The channel utilization is still kept high because the transmission of alert packets has no interference with the tight deadline packets and it only affects those loose deadline packets. Hence only the channel C,+. within a subsystem < Ci, C+j > is degraded by reservation and the lost percentage of loose deadline packets may not be increased due to the nature of their time constraints.



C. Migration of Loose Deadline Packets

With preemption and switching, the loads in two rings of a pair may be unbalanced. This is due to the fact the channel assigned to loose time packet has the inclination to be overloaded while the channel assigned to tight deadline packet is possibly under utilized.

This problem can be alleviated through a proper classification of tight and loose deadline packets by the threshold DL. However, this may still have a load fluctuation from time to time. A dynamic load- bal ance policy called migration is suggested. The key operations of this policy are: whenever a station at ring Ci gets a free token, it sends out a tight deadline packet if it has a nonempty waiting queue. Instead of passing the token immediately if its waiting queue is empty, it will fetch a loose deadline packet from channel C+1 and send it out in channel Ci. This migration of loose deadline packet from Cji+ to Ci should only be done when Ci is under-utilized.











Current Time
Arrival Time [Expire Time
*'"- T.-' -4

Packet Slack Time
Packet Deadline Duration


Figure 3.3: Switching of Packets


To identify the utilization of Ci, a checking mechanism is embedded within each pair. The detection is based upon investigating the token cycling time (Tci~,g) in each station.
1
T~c <_ -M Aze (3.3)
A

where Tei,ng is the period between the time that a station releases the free token and the time it receives the free token again. Tcv , g can be easily observed by Ci when it receives the free token. The 1/A is the packet interarrival time at each channel which can be measured over a long interval, and Mze is a constant that defines the migration window size. As it can be shown, when the above inequality is valid, C, has a low utilization and the average number of packets arrived during the last token rotation is less than A,.J,,.

So, a migration of packets from C+j to C, occurs whenever the latter one is under utilized and thus the load is dynamically balanced within such system. In addition, for each receipt of the free token in C,, a packet (either a tight deadline packet or a loose deadline packet migrated) can be transmitted. This is especially effective in high speed rings due to the high ratio of propagation delay and packet








transmission time so that a high utilization can be achieved.



D. Multiple Packet Transmission

There are normally two different packet service policies in ring networks: exhaustive service and non-exhaustive service. In an exhaustive service system, whenever a station get a free token, it sends out all the packets in its waiting queue. In a nonexhaustive service, every station only sends out one packet whenever it gets a free token. The later scheme prevents any station from monopolizing the service and is useful for real-time applications. But in high speed network system, the packet transmission time is much smaller and hence exhaustive service is more suitable if the physical ring length is longer than the packet length. However, the token may be held by a busy station and the packets with short deadlines in other stations may not get the chance to be transmitted. Thus, transmission of multiple packets per token receipt can only be adopted under an underutilized load.

In a paired ring < Ci, C+1 >, both rings can be in the multiple packet mode. While ring C,+j is working in a reservation mode, some loose deadline packets can possibly follow the switched alert packet (from channel C) to make full use of channel bandwidth. While channel Ci is working in a migration mode, more than one loose time packets can possibly be fetched from channel Cj+ so that the channel bandwidth is not wasted. However, the maximum number of multiple packets N..z has to be limited, it is defined as the following: Nm.Z T'.k" if Dmn >T di,9 (3.4) Nmaz = ckeT,., ifD,,>

if Dmin
Where Dmin is the minimum deadline value, and Tp,eg is the average packet transmission time on a ring. Based on the detection of the last token cycle delay on the








ring, N,,,._ is determined by means of the packet transmission delay to maximize the exiting channel utilization.

The operation of the proposed protocol is illustrated in Figure 3.2. In summary, the multiple ring network system works in the following way: all the packets arriving to a station are distributed into paired rings by a round-robin manner. Within each pair, there are two separate waiting queues attached on two rings, and tight deadline packets and loose deadline packets are queued and transmitted in their own channel respectively. A dynamic control protocol allows the switching of alert packets to preempt the loose deadline packets and at the same time allows the migration of loose deadline packet whenever the tight deadline packet channel is under utilized. In this way, every paired channel and thus the whole system is guaranteed to be working in a balanced state and to not only reduce the lost percentage of message packets but also increase the system utilization. The dynamic controlling policies like switching and migration in both directions are the key points that make such a kind of system achieve a remarkable performance improvement.


3.1.4 Performance Evaluation

Though some analytical models and comparisons of throughput-delay characteristics of ring networks have been given, they are usually only applied to simple cases. It is widely believed that studies of performance behavior using complex control policies are beyond analytical methods and therefore extensive simulations are required.


A. System Assumptions and Performance Measures

The simulation model is set up based upon the following assumptions:








* Arrival packets should be sent out and/or received within a certain period

of time (deadline), otherwise they will be meaningless no matter they can be received or not. Those packets who have passed their deadlines will be viewed

as lost-packets and will be discarded.

e All packets are classified into different classes based on their deadline durations. The deadline durations are uniformly distributed over a constant range.

Thus, the loads incurred by tight deadline packets and loose deadline packets

in a paired ring depend upon the value of the threshold DL.

* All packets arrival into each station in the network follow a Poission process

with a constant arrival rate.

* All stations are equally distanced on a ring;

* The lengths of all the packets are exponentially distributed with a constant

mean.

Various system parameters are defined as in the following table:

SYSTEM PARAMETERS IN EXPERIMENT network length 20 km transmission speed 200,000 km/s bitrate / channel 10 Mbps delay per station 3 bits overhead per packet 97 bits station number 30 packet length (mean value) 512 bits


Table 3.1. System Parameters








With different arrival rate to each channel A, we vary the network utilization from 60% to 80% under the conventional token ring protocol. The packet deadline durations are distributed uniformly over the range of 5T, ott and 25T,,otae where T,otjt, is the free token rotation time. A Ithough a real communication load may not follow the above assumption, it is our purpose to stress the network with a heavy load and equally distributed deadline durations. This will enable us to examine the important performance measures such as packet lost percentage and channel utilization as defined previously.



B. Simulated Control Protocols

It is obvious that different system performance can be achieved by using different medium access protocols. To evaluate the protocol proposed in the previous section, the performance of the following protocols have been investigated and compared.



(1). FCFS

The simplest way to run this kind of system is by using the first-come first-serve policy (FCFS) at each ring. Packets arriving at a station will routed to rings in a round robin manner. Thus, the total load is distributed equally among the multiple rings. In each station, there is an independent waiting queue connected to each channel Ci, i = 1,2,... ,K. When a ring Ci catches a free token, it sends out the earliest arrived packet in its waiting queue if the queue is not empty; otherwise the free token is passed to the next station. This scheme actually functions as a network with several independent single token rings with IEEE 802.5 protocol.


(2). Priority Transmission (PR)








This approach is to consider the different priorities between all incoming packets. There are various policies to decide which packet in a channel should be the next one to send out. To meet the real-time requirements, it is quite natural to set a high priority to those packets with tight deadline constraints. So, choose those "urgent" packets and assign high priority to them based on their slack time so that they can be send out as early as possible.

As in FCFS, all channels are treated as the same and independently. To allow tight deadline packet to use token reservation, two queues for tight and loose deadline packets respectively are associated with each channel. Packets arrived will be inserted into a queue after comparing its deadline duration with the slack time of existing packets. Tight deadline packets have a high priority and can reserve the channel by using reservation bit of the token. The channel can be reset to allow loose deadline packet's transmission until all tight deadline packets have been processed (sent out or thrown away)



(3). Wait-Until Protocol (WUT)

This one is called Wait-UnTil. Different from former ones, WUT groups each two channels as a structured pair < C;, C,+, >. The initial packet allocation, the wait-until policy and switching described in the previous section are applied.



(4). WUT-MIG

In addition to WUT, WUT-MIG (Wait-Util with MIGration) allow loose deadline packets to migrate from their channel C'+j to channel Ci whenever the latter is under utilized. In WUT-MIG, Miz, which functions as a flow control window, is set up to limit the migration load.












40.0
- FCFS
---_ PR
.-. WUT
- WUT-MIG . ---MWUT
30.0 -- WUT- MIG






20.0





10.0





0.0
400.0 500.0 600.0 700.0 ARRIVAL RATE


Figure 3.4: Lost Percentage for Tight Deadline Packets








(5). MWUT


Unlike WUT protocol, MWUT (Multiple packet WUT) adopts Single-TokenMultiple-Packet policy. Based on WUT, NIWUT allows more than one packets to be sent out whenever a free token arrives. The maximum number of packets per transmission Naz, is set dynamically and is dependent of the current network load.






(6). MWUT-MIG


Based on MWUT protocol, the MWIJT-MIG protocol allows migration of loose deadline packets from the second channel to the first channel in a pair so that the loads can be dynamically balanced and to have a full utilization of all the resources.


C. Simulation Results and Evaluation





42

80.0
________FCFS
- PR
* WUT1
50.0 WUTr-MIC
-- MWUT
MWUT-MIG
40.0 30.0

20.0

1 0.0,

0.0
400.0 500.0 600.0 700.0 ARRIVAL RATE

Figure 3.5: Lost Percentage for Loose Deadline Packets


Extensive performance simulations are developed based on those protocols discussed above. Figure 3.4 and 3.5 show the lost percentages of tight and loose deadline packets with different packet arrival rates. In Figure 3.6 and 3.7, the channel utilizations are illustrated. The classification threshold DL is set to equally partition packets into two classes. The results show that the lost percentage of tight deadline packets in FCFS protocol is very high while that in priority-based protocol is relatively low. This is because the former one has no mechanism to favor real-time requirements. However, since the PR protocol uses reservation extensively for tight deadline packets, a lot of channel bandwidth is wasted by favoring the most urgent packets. This leads to a much high lost percentage of loose deadline packets. WUT protocol aims to overcome this weak point of priority-based protocol and has decreased the total lost percentage significantly. With T,gch = 2Troje, the simulation results show a much robust utilization. However, since only tight deadline packets can be switched to the second channel, the loads at paired rings are not balanced.

















90.0 80.0 70.0 60.0




50.0 40.0


.30.0,
400.0 500.0 600.0
ARRIVAL RATE

Figure 3.6: Utilization of Channel Ci in (Ci, Ci+1)


90.0 80.0 70.0 60.0



50.0 40.0


30.0 1
400.0


500.0 600.0 700.0


ARRIVAL RATE

Figure 3.7: Utilization of Channel Ci+l in (Ci, Ci+1)


700.0











With the addition of migration, we can make a dynamic load balance among the paired rings. The lost percentage of loose deadline packet is reduced significantly with a slight impact on tight deadline packets. The curves of MWUT in Figure 3.4 and 3.5, represent the cases where migration window M,,, = 0.2. Further performance improvements can be obtained by adopting multiple packet transmission policy. Note that this multiple packet transmission is dynamically controlled through token cycling time and the limit Nmo.

Table 3.2, 3.3 and 3.4 show the impacts of different T, ,i, Moi,. and DL. The performance of the network is relative insensitive to the choice of the threshold DL once T,,,,i and Mi,. are set properly. Since T.,ih and Miz control the migration of packets among two channels of a pair dynamically according the network load. The initial classification become less crucial. Note that T ,i, and Mi, have different purposes: T,,,i aims to switch tight deadline packets such that they can preempt the normal transmission of loose deadline packets and meet with their real-time requirements, whereas Mi, intend to use the underutilized resources.


Lost percentage (tight-loose) v.s.T.,,

control protocols 1.5Totatc j 2Tro tae 2.5T, o,4, 3Troate WUT-MIG 6.8-32.0 3.5-27.5 2.2-26.8 1.8-28.3

MWUT-MIG 2.9-12.6 1.9-10.2 1.2-7.8 0.9-8.5


Table 3.2. The Lost Percentages with different T,,tch









Lost percentage (tight-loose) v.s. M,

control protocols 0.2 0.4 0.6 0.8 WUT-MIG 3.5-27.5 8.0-18.7 11.8-18.3 12.7-17.0

MWUT-MIG 1.9-10.2 4.6-5.8 6.5-4.4 6.6-3.7


Table 3.3. The Lost Percentages with different Mmzt


Lost Percentage (tight-loose) v.s. Threshold (DL)

control protocols 13TototI 14Trotat 15Totate 16Ttate 17Totat, 18Totate WUT 1.2-33.4 1.8-32.5 2.4-32.5 3.6-34.4 5.2-34.5 6.4-34.5

WUT-MIG 7.1-18.8 7.4-19.5 8.6-22.1 8.7-24.7 9.4-24.4 10.3-26.2

MWUT 0.9-15.4 1.2-13.2 1.3-11.4 2.0-11.9 2.7-13.7 3.1-13.8

MWUT-MIG 2.0-7.6 1.8-8.3 2.0-8.9 2.4-9.9 2.4-10.4 3.3-11.2

Table 3.4. The Lost Percentages with different threshold DL


The following conclusions are drawn from the performance of the protocols studied:

* As expected, FCFS performs badly concerning about the real-time requirements. FCFS has the best performance in channel utilization, but the performance of this protocol is not acceptable. The studies of this protocol, thus,

demonstrate the importance of prioritization.

e The performance of priority-based protocol (PR) gives a better performance

in trying to meet the time-constraints of tight deadline messages. However, as a total lost percentage, it is still unacceptable. The main problem of this








kind of protocols is the low utilization due to the reservation schemes. This

performance is even worse in the high-speed communication systems.

* As may be noted, the protocols based on wait-until scheme (WUT) can achieve

a better performance comparing to the PR protocols. It can increase the system utilization significantly while still guarantee the time-constrained requirements. That is because WUT policy allows all the packets to be sent out immediately if they are not time-out, and dynamically preempt the loose deadline packets only when necessary to guarantee the real-time requirements.

9 The combined dynamic protocol employing multiple-packet policy (MWUTMIG) performs the best on all accounts. The main reason for the better performance is that it dynamically balances the traffic load among available channel resources and cleverly schedules the transmission order. Therefore, it minimizes the lost percentages of all kinds of messages and still maximizes

the system utilization.

From the studies, it can be seen that a network control protocol plays a very important role in system design and system performance. In addition to its traditional role as an arbiter of channel accessing and sharing, a control protocol also serves as a distributed scheduling mechanism by imposing an implicit or explicit transmission order. This scheduling function can critically affect the distribution of packet transmission delays and thus the real-time performance of the protocol.








3.2 Distributed Queue Dual Bus (DQDB)


3.2.1 Background

The Distributed Queue Dual Bus (DQDB) is a recently developed LAN/MAN networking architecture for very high-speed and high-quality transmission services. It is defined by IEEE 802.6 working group with a high compatibility with the switching concepts used by promising B-ISDN standard. However, the standard DQDB protocol is not suitable for real-time communications. Studies have shown [95] that DQDB architecture inherits the serious problem of service unfairness. The position of the transmission station has a great effect upon the system performance. Since transmission delay is critical to real-time communications and the urgent packet in downstream stations can not afford to bear long queueing delay, this unfairness of DQDB networks is obviously undesirable for real-time multiclass message transmission. DQDB standard also provides a multiple priority mechanism with up to four level priorities in its distributed queue processing. It is hoped from theory that the multiple priority mechanism can improve the system performance by giving proper preference to certain urgent message class. Like most of the other systems, however, it is anxious to know how to use this distributed mechanism effectively in real networking practice.

The media access control (MAC) protocol of the DQDB standard is based on its unique physical architecture, which consists of two unidirectional buses and a multiplicity of stations along the buses. These two buses, supporting communication in opposite directions, allow full duplex communication between any pair of stations within the network. All the networking stations cooperate with each other in sharing the network bandwidth via a global distributed queue mechanism f60, 63]. Each








station maintains several counters to reflect the network states. A Request-Bit of request field on reverse bus is used to indicate that a segment of downstream station has been queued for access the bus. The request (RQ) counter in every station is incremented by one every time a nonzero Request-Bit is detected. This way, every station keeps a current state record of the number of segments waiting for access downstream on the bus. Stations with no segments queued to a specific bus decrement their RQ counter by one for each empty slot passing on that bus. Any station wishing to send a segment on a specific bus writes a request into the next free Request-Bit on the reverse bus. At the same time, it loads the current value of its RQ counter into a counterdown (CD) counter. This counter indicates the number of requests for access to this bus which have to be satisfied before the segment at the station may be sent, and is also decremented by one for each passing empty slot. When CD counter reaches zero, the station writes the segment into next empty slot. It should be noted that the operations of writing requests and sending segments are independent. The DQDB standard also provides a global multiple priority mechanism which allows up to four different priorities for segment level transmission. The priority mechanism is absolute in that segments with a higher priority will always gain access ahead of segments at all lower levels. This is achieved by having dedicated counters to each priority and operating separate distributed queues for each level of priority.

The DQDB standard can be simplified as a model in Fig 3.8. Within each station, an arriving message from upper user layer is first processed by Local Allocation Mechanism (LAM), which divides each message into several segments and allocates each segment into one of four local queues according to its priority and is assigned statically. The priority is usually a function of message deadline duration. The








smaller the deadline duration, the higher the priority. The local queues employ conventional First-Come-First-Served (FCFS) queueing discipline. There is a media access Bloking (B) mechanism in each station to control local load to flow into the distributed queue. It is the DQDB standard that requires each station to defer next segment request until the current one has been served. Therefore, only one representative segment from each station is allowed to join the distributed queue for any priority level.

It is obvious that the multiple priority scheme is expected to improves the performance by giving preference to higher priority messages. However, the local queues are of static type and can not reflect the dynamic nature of operating network. The relative criticality of message as time goes by is not considered. A long-waiting lower priority message may become critical because the deadline duration may be expired soon. Therefore one important issue is to develop a control policy which should be capable of adjusting the local queues according to the message waiting time and the up-to-date networking operating status.

It is also worth noting that even a good control policy is included, the current multiple priority mechanism is still not enough to provide guaranteed real-time performance. This is because there is no effective Priority Assignment Mechanism (PAM) enforced by the DQDB communication system. A proper priority assignment scheme is critically important in those distributed networks to ensure the required services. This is mainly due to the fact that the MAC protocol of DQDB networks does not provide individual stations a global view of priority operations. As it is indicated in Fig 3.8, the message in each station is directly passed from local queue to the distributed queue whenever it is permitted by the media access blocking mechanism. That is to say, each station actually decides the global pri-







Distributed Queue


.-.-...-.-----. .................


LAM: Lo= M ocado Mechmim


B: Ac*" Blocking Gate
S: Di.8: Queue Server
Figure 3.8: DQDB Control Model


ority of its messages according to its local users' requirements without concerning the relative criticalness with that of other stations. This certainly results in an inconsistent priority mapping in the distributed queue and may seriously affect the system performance. The analysis of DQDB protocol has shown that the relative traffic intensity of different priority distributions has great effect on the global system performance[70]. The overloaded high priority traffic will block and cause a significant transmission delay to lower priority classes. Since this undesirable situation can be artificially introduced by improper mapping from the local priority (deadline) requirement to the global priority mechanism, an important issue is to develop a scheme which is able to enforce all the communication stations to follow the same priority assignment criteria even though they are only cooperated with each other via a distributed queue.








3.2.2 Dynamic Control Model and Strategies To overcome the problems the standard DQDB protocol has for the real-time communications, a dynamic control model has been proposed as in Fig 3.9. In addition to the original LAM module, a new Priority Assignment Mechanism (PAM) is included in our model. Some different control strategies are employed in these two modules to provide much better performance for real-time communications.

As that in DQDB model, the LAM is responsible for message segmentation and local queue allocation. However, the LAM in our model will not use the conventional FCFS queueing discipline, this is because that the more important performance objective now is to guarantee the timing requirement of transmitting messages. In real-time communication networks, the time criticalness not only comes from the original semantic attribute of the message but also is caused by the limitation of the various network resources. As time goes by, the slack time ST(m) for one segment of message m changes dynamically according to the follow: ST(m) = AR(m) + DL(m) - t, (3.5) where AR(m) is the arrival time for message m, DL(m) is the predefined deadline duration constraint for message m, and tc stands for the current time point. As we are more concerned about transmission delay at message level, we treat each segment from one specific message to be of the same type. It is noted that a message with larger deadline duration could still have a smaller slack time ST(m) than that of another message with a smaller deadline duration. As time constraints are more concerned in real-time communication systems, the Local Allocation Mechanism (LAM) in our model employs the following allocation scheme: all the segments are queued with regards to their dynamic slack time ST(m) instead of static deadline






1itrbulad Quew

...................... ................................... .. ..............----- f
Phot"
swaos-1















LAM: Locad A o~o Mehs B: Accem Blocking
PAM: I' odty Aipment Memaa S: Diatributad Quetae Server
Figure 3.9: Dynamic Priority Control Model


duration DL(m). There is only one local queue instead of the previous four local queues and this queue is updated whenever a new message arrives. The message priority is actually not considered at this processing stage. The global priority assignment is delayed until the local blocking (B) mechanism permits next segment to join the distributed queue.
It is the additional Priority Assignment Mechanism (PAM) that dynamically decides which segment in local queue is the best candidate for the current allowable priority level to join the distributed queue. The decision should not only consider the local users' requirement, but also keep the DQDB multiple priority mechanism fully utilized. A dynamic priority assignment scheme is desirable which should be able to map the local user requirements properly to the existing global multiple priority mechanism. To apply the dynamic assignment scheme within the whole distributed system, the PAM in each station should first convert the continuous timing requests








to discrete ones so as to make use of the 4-level priority mechanism. Since it should be done in real time, the up-to-date network operating status could be very informative and helpful. It is observed that we can make a wise local decision if the latest operating information for each priority level is provided. Suppose we have estimated the delay bounds of segment transmissions at different global priorities, then we can decide the best candidate for a particular priority by considering all the segments whose slack times are within corresponding delay bounds and choosing the one with the smallest slack time. Let TB(j) denote the specific delay bound for station i and priority level j, where the larger level number j indicates higher priority. This measure gives the information that the next possible transmission for priority j segment at station i is about TB!j) time units away from now. It should be noted that TB(4) < TB(3) < TB(') < TB(') and they are dynamically determined during network operations at each station. If PR[S(m)] is the priority value to be assigned to a segment from message m at station i, it can be decided as follow:

PR[Si(m)] = if STi)n) TB!') (3.6) {,if TB~j) < ST,(m) TBf'1')
It should be noted the candidate for priority j at station i is always the segment with locally minimum slack time among all segments which have slack time within the range [TB('), TB(-j')]. Actually this method provides an approximate global priority decision mechanism, which is based on message slack time, so that every user can follow the same criteria to decide segment priority and participate the competition of medium access.

Besides the function of dynamic priority assignment, another effective strategy employed in PAM is the segment rejection. Since we are more interested in message








transmission, the rejection of a segment of one specific message means that all the remaining untransmitted segments of that message should also be rejected. By rejecting those waiting segments, the lost percentage of other messages can be reduced. This is due to the fact that this kind of rejection can release the buffer blocking problem and making the best use of local resource.


3.2.3 Priority Assignment Algorithm

In order to implement the above proposed method in practice, A simple and effective dynamic priority assignment algorithm to be employed by PAM has been developed as shown in Fig 3.9. The key point here is to properly determine the delay bounds TBW"0 for each priority level j(= 1,2, 3,4) and operational station i. These bounds have to be determined locally while the global network status should also be considered so that the best real-time performance can be achieved.

Motivated by the standard nonpreemptive M/G/1 priority queueing model, A method which is able to derive the approximate media access delay dynamically is presented in following. The mean waiting time of a priority j segment at station i can be expressed by

W~j) =WOi
[ - +J[ - +) (3.7) where U4j) = = p , which actually is the accumulated traffic intensity of all no lower than priority j segments. The WOi is the service residual time. It is noticed that W!) does not depend on the segments from lower priority groups except for their contribution to the numerator WOi. Since only interested in the media access part, we can consider the partial distributed waiting queue in which segments from all downstream stations (including station i) are accumulated for their turns of








transmission. There is a different partial distributed queue for each active station i along the network. For any station i, a standard nonpreemptive priority queueing discipline model can be employed to analysis this queue. The server here is the next available empty slot with fixed length, and the queueing objects are all the segments from downstream stations which have been permitted to join the distributed queue.

Let WOi(t) denote the last two consective empty slot passing time, which can be observed by each active station i in real network operating time. The RQi(j) is the RQ counter value at station i, which is provided by the standard DQDB protocol. It is noted that the RQi(j) actually counts the number of requests whose priority is no less than level j for all the downstream stations, and there are total (N - i + 1) downstream stations from station i. Taking these as a means to measure the priority-j traffic for all downstream stations from station i, then the following approximations are derived.

WO, = WO,(t) (3.8) U~) = RQi(j) (3.9) (5-j) (N - i + 1)
U(j+1) = RQi(j + 1)
a (4 - j) * (N- i + 1) (3.10) and finally the expected media access delays for any priority level j.

w0,(t)
I -RQi(j)/(5-j)(N-i+ I)]i-RQ,(j+I)/(4-j)(N-i+:)J Wi(j) _ ifj<4 (3.11) WO,(t)
[I -RQ, (j)/(s-j)(N -i+ 1)]

if j=4

Since all the values in right hand can be calculated at each station, the above approximate access delay is time-dependent and of the feedback nature so that it can reflect the dynamic operations of the network. Taking these delay bounds to be








the TBi0' discussed in our dynamic control model, the following priority assignment algorithm can be formulated.

{ 1 if ST,(m) > i() (3.12) j if W~il) < ST(m) < WiThus the segment priority can be dynamically assigned by service provider at the medium access request time rather than statically decided by application users at the message arrival time. In this way, the priority distribution are dynamically adjusted in accordance with the network operating status so that the optimal real-time performance for the whole system can be achieved. Inspecting this scheme carefully, it can be seen that the scheme actually approximates to the Smallest-Slack-TimeFirst (SSTF) scheduling scheme. It has been show [741 that SSTF minimizes the total lost percentage of all classes of transmissions and thus provides the optimal real-time communication service. The proposed time-dependent priority assignment scheme can be regarded as a practical implementation of SSTF scheduling strategy in the DQDB networks.


3.2.4 Simulation and Evaluation

Because of the dynamical nature of the real-time communication systems, the investigation and evaluation of the proposed protocol have been done in certain details through extensive simulations. There are three different MAC protocols to be investigated. The Ideal Static Priority (ISP) protocol is the ideal case where all active stations are able to follow exactly the same static mapping schemes. Though it is not possible in real network, this protocol is valuable for analysis and comparison. The second one is the Static Priority (SP) protocol, which is the real networking practice that employs a static mapping from the user's deadline to the segment's








priority at each station independently. The SP protocol under consideration is one of the typical unbalanced cases due to the distributed environment. For the simplicity, this protocol has the structure that the most upstream half of all the stations assign their segments with lowest priority while the rest downstream stations assign their segments with highest priority. The Dynamic Priority (DP) protocol, however, employs the proposed dynamic priority assignment protocol. It provides a dynamic mapping from the user's deadline to the segment's priority, depending on the message waiting time and operating status of the network. All three protocols under considerations employs segment rejection scheme for effective message transmission. It is our purpose to investigate the different priority assignment algorithms and the effectiveness of our proposed dynamic control scheme.

Since the two unidirectional buses are identical in a DQDB network, the study will concentrate on bus-A. The traffic load is uniformly distributed, that is to say, each station has the same amount of transmission requests. However, considering only bus-A, the arrival rate for each station will depend on its physical location since the station can access both buses in two different directions.

The network parameters under the simulation experiment environment are set according to the following table:









NETWORK PARAMETERS

Bandwidth of each bus (B) = 150 Mbps

Number of stations (N) = 20

Slot size = Segment length (S) = 750 bits

Distance between adjacent stations (D) = 2 slots

Local Buffer size (K) = 120 messages

Segments per message = 2

Destinations are uniformly distributed

Some other system parameters can be easily calculated.

Slot Unit (A) = S/B (3.13) Bus Length (L) = (D + 1) x N - D (3.14) The basic simulation load pattern is that there are four predefined classes of messages with different deadline durations. There is a deadline seed parameter dloeed, and the four deadline durations are generated as dl, = l*dl .ed, d2 = 3*dled,, d3 = 5 * dizeed, d14 = 7 * dl,,,d. The ISP protocol is the ideal static control case which assign the priorities according to the criticalness of the message's deadline duration in the whole system. That is, the segment's priority of one message is always j if that message has deadline dl(s-i),j = 1,2,3,4, no matter where this station is located. The message arrival rate A is also a changeable parameter so that the protocols can be investigated under different traffic loads. The unspecified parameters in later discussions correspond the default ones that A=0.0005 (message/msec), dl. ed=35.0

(A), where A is the slot unit time.

Let F(i) denote the number of failed messages and T(i) denote the number of successfully transmitted messages in station i respectively. Considering all the









40 1 1 , I , - . I I I
-DIP Protocol A.----.- ISp Protocol ISP Protocol
30






20



1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 15 17 18 19 Station Index I

Figure 3.10: Lost Percentage

priority segment transmissions, let P(i)(%) be the lost percentage value in station i and P be the total message lost percentages of any protocol considering all the networking stations.

F(i)
P(i) = F(i) + T(i) X 100% (3.15) P EN EXY, F(i) 1 (3.16) V=1 F(i) + Ev= T(i)

Fig 3.10 shows the lost percentage of messages for each individual station along the network. It can be seen the location position has a great effect on the service quality in DQDB networks. The downstream stations usually bear a higher lost percentage than the upstream stations. This is due to the special structure of the DQDB standard. However, the proposed protocol (DP) is able to significantly reduce this undesired property by using dynamic control strategies. While it shows the general lost percentage for all classes of transmitting messages, Fig 3.11 and Fig 3.12 show the lost percentages of messages for different priority classes under DP and ISP










































2 3 4 5 6 7


8 9 10 11 12 13 14 15 16 17 18 19


Station Index I




Figure 3.11: Class Lost Percentage Under DP


StaUon Index i


Figure 3.12: Class Lost Percentage Under ISP


20 18

16

14 0.12


LO


,------h Cke-2
* .. -,aCkm-2
---0 CIoa-3










Clo -4


12 .10









protocols, the priority level here is the static one which users under the ISP protocol.


is decided originally by the


Lost Percentages P(%)

[(10-3)I ]ISP I DP

0.40 35.0 1.1362 0.8932 0.6897 0.45 35.0 2.3457 2.0356 1.1764 0.50 35.0 5.0928 4.2453 1.8293 0.55 35.0 10.5006 8.0267 3.7152 0.60 35.0 20.0528 15.8923 9.2287 0.50 31.0 8.1022 5.5552 2.0751 0.50 33.0 5.8399 4.6504 1.9803 0.50 35.0 5.0928 4.2453 1.8293 0.50 37.0 4.5301 3.8139 1.8162


Table 3.5. Total Lost Percentage

Table 3.5 shows the total lost percentages of three investigated protocols under different traffic loads and timing requirements. As load increases, all three protocols tend to increase the lost percentage of message transmission. However, the proposed DP protocol has a much smaller lost percentage than the other two protocols. This good property can also be observed when the message deadlines change from loose to tight. In Fig 3.13 and Fig 3.14, the total lost percentages of all classes of messages are given under various arrival rate A and deadline duration seed dled. In general, the lost percentages of SP and ISP protocol are quite high for time-constrained message transmissions. This is because they never consider the dynamic nature of the local queue and the tight deadline messages tend to block the loose deadline









messages no matter how long they have been in the waiting queue. Also, they bypass the user requirement directly to the multiple priority mechanism of DQDB without any control which in most cases will result in an improper mapping and spoil the system's global performance. On the other hand, the proposed DP protocol achieves much better system performance. This advanced MAC protocol is dynamic in the sense that it can adjust the load of different priority levels according to the status of network operations. It should be noted that under the DP protocol, the load of each priority level is able to change from time to time no matter what the original local user requirement is. Therefore the distributed queue processing is optimized to achieve expected good performance.

In Fig 3.15, the average message transmission delay Dme,(i) for each station i is given. It is measured from time instant the first segment of a message attempts to access the medium to the time instant the last segment of that message successfully finishes its transmission. It is worth mention that this delay is an average measure which considers all four different priority levels at each station and only one unidirectional bus of the DQDB network is considered. The DP protocol has shown a significant smaller transmission delay than the standard DQDB static protocol. This is because the latter does not have the mechanism to provide message level service and all transmission is always undertaken at the segment level. The message transmission delay also reflects the chance of accessing the bus from station i and indicates the fairness property of the undertaking MAC protocol. Therefore it shows the DP protocol can provide a much better fairness performance than the conventional static DQDB protocol.

The proposed dynamic priority assignment scheme achieves a high quality realtime performance by properly making use of the multiple priority mechanism in








63



30.0 �
28.0 -OP Protocol
28.0 & . ......... A ISP Protocol

26.0 -' - SP Protocol

24.0 22.0

. 20.0 18.0
16.0 14.0
12.0

10.0
8.0 l

6.0
4.0........

2.0 0.0
40.0 45.0 50.0 55.0 60.0 Arrival Rate



Figure 3.13: Lost Percentage Over Arrival Rate



DQDB standard. It emphases on dynamic local message scheduling at message


level which is especially important in real-time communications. By forcing all the


stations to follow the same priority assignment criteria, the total lost percentage


of all kinds of messages due to the deadline constraints is minimized. The simulation results have validated the effectiveness of our control strategies and shown the


proposed MAC protocol significantly outperforms conventional MAC protocols for


real-time communications. It is believed that the future real-time protocols should


be able to function as a wise load scheduler as well as the conventional media access


arbitrator.


















16.0


14.0 12.0 10.0


8.0


0.0 ' 31.


0


35.0


Deodllne Duroton Seed





Figure 3.14: Lost Percentage Over Deadline


1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 Stoton Index I


Figure 3.15: Message Transmission Delay


-OP Protocol
---------- ISP Protocol
- SP Protocol


















..... .. ....


37.0
















Chapter 4




MESSAGE LEVEL



PROCESSING SCHEMES





4.1 Introduction


In real-time communication systems, the message level delay and processing is much more meanful and important to the application users. The flaw in conventional segment or slot level processing schemes is that it is generally an intra-layer concept rather than a global control mechanism. This is because all the messages have to be divided into segments to suit the fixed size formats when they go through the networking media. While it is unclear how effective such schemes in a lower layer can be mapped from upper layer services, they are usually undesirable from a user's point of view. The real-time application users are only interested in the performance at the complete message level, and the performance measure at network-specific segment level really does not make any sense to them. The emphasis of networking operations must be on the message level instead of viewing them independently at








the segment level. Some benefits can be obtained including more effective transmission and better real-time performance. For instance, when certain portion of a message fail to meet its deadline, then the rest of this message should be aborted, since there is no way to satisfy the requirement of this message.

We intend to investigate the proper processing level for control schemes in realtime communication networks. It is believed that the transmission delay should be considered at the message level, which is obviously not optimized at the current segment level processing schemes. The standard DQDB protocol is one of the segment level processing schemes, in which each request is asking for one empty slot to serve the current waiting segment and the next segment request will not be issued until the current segment has been transmitted. Therefore all the messages should be first divided into segments to suit the fixed size slot format. It has been found that this scheme is very undesirable for real-time communication service since segment level processing is not efficient and the segment delay does not make much sense to the application users. More importantly, this mechanism does not necessarily result in good message delay performance.



4.2 Message Level Processing


A processing model is presented whose scheduling and queueing algorithms are build upon the individual message timing requirement instead of the complex mapping mechanism. A reservation control strategy has been proposed. Whenever an incoming message arrives, the control scheme has to consider the media access duration at the message level although the real transmission is performed at the segment format. The access request must be notified to other stations with the information








at the message level such as the number of segments this message has. Under this strategy, it is the messages not the segments are interacted with each other and compete the media access control. The proper order of transmission is decided by the message criticality. It can be observed that the tight deadline messages are unlikely to be blocked by loose deadline messages under the proposed strategy, and there is no partial transmission success at the message level. Thus, this reservation control strategy is able to provide much better real-time performance as well as the processing efficiency.

The standard DQDB protocol is employed as an example to elaborate the proposed strategy in detail. Fig 4.1 gives the DQDB network architecture and its slot format[41]. A slot is the basic unit for data transfer. The access control field contains the bits that control access to slots. The BUSY bit indicates whether the slot contains information or is available. The SLTYPE bit indicates the type of slot and the PSR bit indicates whether the segment in the previous slot may be cleared or not. Finally, the REQUEST field consists of four bits, usually used in the operation of the multiple priority distributed queues. Messages at each station have to be divided into segments to fit in the fixed length slots. The access to slots is controlled by a distributed queue which allows the formation and operation of a queue of asynchronous segments generated across the network. Each station maintains certain dedicated counters for each bus so that the distributed queue can perform as a centralized queue. As discussed in Section 3.2, the DQDB protocol is based on segment level processing. Each request asks for one empty slot to serve the current waiting segment and the next segment request will not be issued until the current segment has been transmitted. The message level service is never considered. Therefore, the DQDB protocol has been found being not so efficient and



















Acc.e. Control Field gogint I



(I bit) (1 bit) (1 bit) 11 bit) (4 bit)

Figure 4.1: DQDB Network and Slot Format



effective when it is employed by real-time communication users.

To overcome these shortages, a class of new control protocols is proposed based upon multiple-slot reservation scheme. These new protocols are carefully designed to not only provide a better real-time performance in an efficient way but also be highly compatible with the standard DQDB protocol. The main idea of our proposed protocols is to do the service request at the message level instead of at the original segment level. That is to say, all the corresponding counters now operate in multiple-value mode rather than the simple increment-by-one and decrementby-one mode of the original DQDB standard. In this way, the reservation is done more infrequently and effectively. Rather than reserving empty slots one by one, the proposed protocol try to reserve all required empty slots for the current message all at once by making use of multiple value request in each slot.








4.3 Analysis Model and Evaluations


Due to the fact that the two DQDB buses are identical in nature, our model and corresponding analysis are based on one single bus, say bus-A.


4.3.1 Service Time Distribution

Since we are only interested in the media access part, a special partial distributed waiting queue is considered at each station i. This is a queue in which arrival segments come from all downstream stations (including station i) and are accumulated for their turns of transmission. There is a different partial distributed queue for each active station i along the network. For any station i, a standard preemptive resume

(PR) and feedback M/G/1 queueing discipline model can be employed to analysis this queue. The server here is the next available empty slot with fixed length T, ot, and the queueing objects are all the segments from downstream stations which have been permitted to join the distributed queue.

N-I N
A, = E E Ad (4.1) s=i d=s+l
i-I N
F, = Z A~d (4.2) s=1 d=s+l

where Ai is actually the accumulated traffic intensity of all downstream stations and Ai is the accumulated traffic intensity of all upstream stations.

For this special queue in each station i, the service time is actually the empty slot intervals. We can describe approximately the interval between empty slots seen from station i with the following geometric distribution:


Pr{x = k- slots} = p-'(I -pi) k = 1,2,... (4.3)








where pi = Fi * Ttot. This distribution is known as geometric distribution at one and the corresponding moments of service time x can be easily obtained. One important property of this model can be observed and is worth mention. The mean service time increases and the arrival traffic intensity decreases when the number i goes from upstream stations to downstream stations.
1
xi - (4.4) T2 -+Pi
- 1+pi (4.5) pi= Ai* s (4.6)


4.3.2 Waiting and Delay Time Processing-Sharing (PS) Discipline

This is a scheduling discipline employed by DQDB media access control protocol in the sense of ideal analytical case. Newly arriving customers join the single waiting queue, work their way up to the head of this queue in a first-come-first-serve fashion, and then finally receive a quantum of service. When that quantum expires and if they need more service, they then return to the tail of that same queue and repeat the cycle. It is clear in this system that a customer is required to make an infinite number of cycles each infinitely quickly and each time receiving infinitesimal service, until finally his attained service equals his required service, at which time he departs.

Based on the results from queueing theory and previous equations, we finally obtain the following message average waiting time Wi and transmission delay time Ti for this processing-sharing (PS) model.

W(PS) = (l-p,) (4.7) Ti(PS) = W,(PS) + Y. (4.8)








First-Come-First-Serve (FCFS) Discipline

Under this scheduling discipline, the server always selects for service that customer at the head of the waiting queue and offers a quantum od service to this customer. The difference, however, is the place the customer returns to this same queue waiting for another service quantum. The First-Come-First-Serve discipline feeds those customers that ejected from service due to the termination of their quanta directly back to the head of the single queue and thereby immediately takes them back into service until fulfill their required service.

Similarly, we can get the following message average waiting time Wi and transmission delay time 7i for this first-come-first-serve (FCFS) model.


W(FCFS) = 4 *.

7i(FCFS) = W(FCFS) + T (4.10)


4.3.3 Numerical Results

The analytical results are shown together with the corresponding simulation results as in Fig 4.2, While the PS.ANA and PS.SIM represent the analytical and simulation results for the processing-sharing model, the FCFS.ANA and FCFS.SIM denote the corresponding results for first-come-first-serve model. In this experiment, each message has the fixed length of two segments. It can be noticed that the analytical curves always have a large delay comparing with the simulation curves. This is because the networking propagation delay is never considered in our analytical model. In fact, the analytical models here overestimate the real traffic load and thus enlarge the transmission delay.

The Processing Sharing (PS) model has a flat curve for its analytical result and
















20 - I . I - I - I . I . I . I . I . I I . I - I - I - I ,
PS.ANA



.- . .....SS.....











O f a 4 5 6 7 8 9t0 11 1 I IS 14 15 19I' 1 to Sta.-an Ind"
Figure 4.2: Message Transmission Delay


shows a perfect fairness among all stations. However, the propagation delay is an essential part of any high speed network in reality. That is why the DQDB network suffers a serious unfairness as shown in the simulation experiments. While it is widely believed that the exact analysis of DQDB network is extremely difficult or intractable, the above approximation analysis do give us insights of some important features. More important, the analysis shows the better performance and promising of the proposed multiple-slot scheme, which is based on the First-Come-First-Serve (FCFS) model at the message level. Based on these observations, we will further investigate the multiple-slot schemes in more general cases where the messages have not to be restricted as fixed length. Since exact analysis may not be applied, simulation experiments are used in later sections to verify and demonstrate the better performance provided by the multiple-slot schemes.









4.4 Transmission Control Schemes


4.4.1 Multiple-Slot Reservation Protocol

Following the previous discussion and the message level processing model, further investigation of the proper control schemes in real-time networks is studied. In this section, a simple message level processing scheme is presented based on the multipleslot reservation concept. In order to be highly compatible with DQDB standard, a 4-bit Request field of the access control field (ACF) is used to represent different message length requests. It should be noted that it does not necessarily have to use these particular four bits to implement our strategy and algorithms. In general, one additional access control field can be added into the slot format dedicating to the specification of the message length and it will not affect our later discussion of algorithms and results. However, our intend in this research is to investigate the effectiveness of multiple-slot control schemes and to avoid complicated format discussion hampering the proposed ideas. Therefore we restrict the message length with a maximum value of 15 segments where value 0 means no request. By doing like this, no change is made at all of the DQDB standard slot format which is a very desirable design concern.

One way for a detailed description of a protocol is its state transition diagram (STD). It should be noted that the proposed control strategy does not conflict with the priority mechanism in principle although the priority field is used as one way of implementation. As long as additional control bits are provided, both priority mechanism and multiple-slot reservation scheme can be accommodated. To be compatible with the standard DQDB standard, the priority attribute in the corresponding STDs is still included. Fig 4.3 and Fig 4.4 show the detailed STD of












IDLE


RQ_I - RQJ . I








IN - SEL-PRQJ & J > RQI - R_. A





IN - EPMPTY-QA if RQ_. 1>0 then R(2_1 - RQI-


QA-DATA request
CDI - RQI
RQI = 0










IN - FMPTY-QA & CDI -0 Transmit Next Segment


COUNTDOWN

IN . REBQJ > I

CDI - CD + 1



I- - REQZJ & I - I

RQ.. - RQJ + I




IN - EMPKY-QA & CD > 0

CD-I - CD.-. - I



IN = SL-REQJ & J> I

CDI - CD_1 *4I


Figure 4.3: Standard DQDB Protocol





standard DQDB protocol and proposed protocol respectively. As in DQDB standard, the REQJ and SEL-REQJ denote the priority-J requests issued from downstream stations and local station respectively. The RQ_I and CDJ represent the request counter and countdown counter with priority I. While EMPTYQA is the empty slot passing by, the local data waiting for transmission is QA-DATA with length LEN(QA-DATA). It can be observed that two new variables are introduced. With LENJ represents the message length of the corresponding requests REQJ and SEL-REQJ, another variable W-NO denotes the remaining segment numbers of a message waiting for transmission which is used to guarantee the continuous transmitting of a message. Different from that in DQDB, all the counters now are operating according to the specific message length variable LENJ (J=0,1,2,3). Also, the COUNTDOWN state will not be switched back to IDLE state until all the segments of a message have been transmitted, which actually accomplishes a










IDLE COUNTDOWN
IN = REQ.J & J>
IN = REQJ & J , I a_ = C)_. + LENJ

RQJ = RQJ + LENJ r=IN=RFJ&= QA-DATA request
RQ._ = RQ.J + LENJ
CDJ = RQJ
RQJ. = 0
IN = SEL-REQJ & J > I W-NO = LEN(QA-DATA) IN = EMPrY-QA & CD > 0 RQJ = RQ. + LEN.- CDI = Cu -1

IN = SEL-REQJ &J > I
a,_I = aDi +. LEN..)
IN - E14PTY-QA
IN = EMPTY-QA & CD)_ = 0 & W-NO = I If RQ.) >0 IN = EM -YQA & CDJ = 0 & W-NO > I then RQ.) = RQ. - 1 Transmit Next Segment ~Trnmit Next Segnmnt W-NO = W-NO - I



Figure 4.4: Multiple-Slot Protocol


message level transmission.

It can be observed that in the distributed queue, the proposed protocol employs the scheme of the First-Come-First-Served (FCFS) model considering the message as a processing unit, the conventional DQDB protocol is, however, a segment interrelated system which can be considered as the Processing Sharing (PS) model at the message level. From the results of queueing theory, it can be expected that the conventional DQDB protocol always bears longer mean message delay comparing to the proposed protocol under the situation that the variation of message length distribution is not so large. The operations of the proposed protocol are very simple and highly compatible with those of the original DQDB protocol. It is also worth mention that the proposed protocol alleviates the request-bus traffic by efficiently employing multiple-slot reservation scheme and thus improves system bandwidth utilization.








4.4.2 Enhanced Multiple-Slot Protocol

In practical distributed system, one station may have some very long messages for transmission and thus cause unfair access by overusing the available bandwidth. More specifically, this undesirable situation will happen when the variation of message length distribution becomes greater. Motivated by the queueing analysis, an enhanced MAC protocol is further presented based upon the Shortest-Message-First control scheme, which is a direct variation of the Shortest-Job-First (SJF) scheme. Under this queueing discipline, both short and long messages can be properly scheduled to achieve a much better overall system performance. To accomplish this, some additional information is needed whenever a transmission decision is made locally. In each station, additional records of message length can be obtained for every active downstream message (which has entered distributed queue) through slot transmission on reverse bus. For station i, it is observed that there are at most 15 different message length types since 4-bit Request field is used to specify the message length. Therefore, any station i only needs 15 length counters to store the number of message requests with different length.

Let NL,(k), NL2(k), ..., NL1s(k) denote all the length counters in station k. It should be noted that NLI(k) = m means, viewing from station k, there are m requests of 1 segments from downstream stations have already entered the distributed queue. Instead of using CD (counterdown) counter to decide the order for transmission which provides FCFS service, station k decides next candidate by considering its pending message and all the downstream active messages and picking the message with smallest length. The corresponding STD is described in detail in Fig 4.5. Our algorithm can be simply defined by the rule that, whenever a message finishes

















IN - REQJ &) I RQJ = RQj + LEN_J





IN = SEL-REQ J & J > I RQJ - RQJ + LXKJ






IN = EbPIY-QA (RQJ> 0 then RQI = RQ.., - I


QA-DATA request


W-NO = LEN(QA-DATA)












IN - EMIPY-QA & MN" = t & W-NO =


COUNTDOWN

IN . RQJ & J >" I RQ_I = RQ.. + LNJ



IN = EMPTY-QA & Me = f

RQ_ = RQJ - I j



IN = SEL-REQ.J i > I

RQJ = ROW + LENJ


IN - EMPJY-QA & MIN* = t & W-NO> I Transmit the M - SegmentP Transmit Nea Segmnt W-NO = W-NO - I




Figure 4.5: Enhanced Multiple-Slot Protocol


transmission, the next candidate is the active message with smallest message length value. Let L(k) denotes the candidate message length at station k, which is actually the initial value of variable W-NO in the STDs. The MIN* operation in the state transition diagram can be specified in detail as follows.




MIN* - J t(rue) if L(k) < MINI=1,2 ... 1{I NLI(k) # 0) (4.11) f(alse) otherwise



The length counters NLI(k) are maintained by each station k. Upon receiving any request with length 1, the station k increases the corresponding counter NL1(k) by one. When the station passes empty slots to downstream stations, it decreases the counter NLt(k), which has the smallest length value I and is greater than zero, by one if it passed 1 empty slots. It can be seen that this is actually an approximate Shortest-Message-First scheme undertaken for the distributed DQDB networks. From the result of queueing theory, this SJF-based protocol is expected








to achieve near optimal performance for our distributed queue system.



4.5 Performance Evaluations The mathematical modeling and performance analysis of the DQDB network is known to be a very difficult problem. This is mainly because that high degree of interactions among a plethora of processes makes an exact analysis of the distributed network almost impossible. In order to verify our analysis and compare all these discussed protocols in detail, a simulation model is set up and extensive experiments are undertaken. There are three MAC protocols to be investigated. While Single Slot Processing Sharing (SS-PS) protocol is essentially the standard DQDB protocol, Multiple Slot First-Come-First-Served (MS-FCFS) protocol and Multiple Slot Shortest-Job-First (MS-SJF) protocol are the two multiple slot reservation protocols proposed in Section 3 and Section 4 respectively.

Since the two unidirectional buses are identical in a DQDB network, the concentration will be on the study of bus-A. The traffic load is uniformly distributed, that is to say, each station has the same amount of transmission requests. However, considering only bus-A, the number of unit arrival rate for each station will depend on its physical location since the station can access both buses in two different directions. The major network parameters are set according to the following table.









NETWORK PARAMETERS

Bandwidth of each bus (B) = 150 Mbps

Number of stations (N) = 20

Slot size = 750 bits

Distance between two adjacent stations (D) = 2 slots

Local Buffer size (K) = 120 messages Destinations are uniformly distributed

Some other system parameters can be easily calculated.

Slot unit time (A) = S/B = 5 (microseconds) (4.12)

Bus length (L) = (D + 1) x N - D = 25 (slots) (4.13)

The major performance concern here is the message level transmission delay. Let D,,(i) denotes the mean message transmission delay which is measured from the instant the first segment request is issued until the last segment of the message has been successfully transmitted. The intention is to investigate this measure obtained in each station on both buses and under various traffic patterns. The traffic load is uniformly distributed among all stations, and the message arrival is of the Poisson pattern.

As discussed early in this chapter, the minor change in protocol only comes when the user message requests have a very large variance in length distribution so that some message lengthes will exceed 15. In that case, several reservation requests are sent out by this station with each one having the maximum lengthes except the last request. In Fig 4.6, the message length varies and follows a Poisson distribution with a mean length M = 3. In this case, the MS-FCFS has a quite close mean delay to that of SS-PS protocol although the former has a relative flat curve. This

















. ........... a &ss-P
180 . Ms'-Flips f ro MS-SJP 160
150 140 180


110 100 80 80


50.. ..
40 SO


10 I 4 * 4 5 8 7 8 8 to It is 15 14 15 to 17 18 1* 0



Figure 4.6: Message Delay When Arrival Rate Per Station = 0.2


............ SS-PS
too - MS-FCPS I O i-. MS-SJ?


150
140

ISO

Ito



0
so .....' "A.



*0


I.
90 .. . ...- p
40




1 X 8 4 5 9 7 8 9 10 ItIS IS 14 15 1o 17 18 1950o Sta~ft hdm4

Figure 4.7: Message Delay When Arrival Rate Per Station = 0.25








is due to the fact that the coefficient of variation of the message length is now equal to one. Fig 4.7 shows the same measure when the traffic load is increased. The multiple-slot reservation protocols have shown to be able to outperform standard DQDB protocol quite significantly as the system is in high utilization. It should be noted that although all these protocols are based upon certain queueing model, their operational characteristics do not exactly follow the theoretical results. This is due to the fact that there are so many factors, including distributed queueing operation and networking delay, influence DQDB behavior. However, these analytic queueing model can be used as a guider to design more effective media access control protocols. As expected, when the variation of message length distribution becomes quite large the simple multiple-slot reservation protocol based on FCFS scheme has shown to provide a quite large message delay, which may even worse than that of the original segment-based DQDB protocol as analyzed before. However, the MS-SJF achieves the overall best message transmission delay performance under all situations. The proposed control strategy also shows a very efficient bandwidth usage on reverse bus to send requests and thus achieves a much better system utilization. This is because the multiple value request operations enable the system to send fewer number of requests and shorten the request transmission delay.

In real-time communication networks, the transmission delay at the message level is a much more important and proper performance measure. It is argued here that the MAC protocol in real-time MANs should be designed and evaluated from the application users' point of view (message level) instead of from the network designer' point of view (segment level) so that much meaningful and better service performance can be achieved. The proposed multiple-slot reservation protocols contribute to the DQDB standard with processing efficiency, better bandwidth usage,









and improved message delay performance. Of all the discussed protocols, the SJFbased multiple slot protocol has shown to be the best in the sense that it minimizes the average message delay under all circumstances. Also, it is worth mention that the proposed protocols have shown a relative more flat delay curve than that of the single slot protocol (DQDB standard). This indicates that the multiple slot scheme is able to suppress the unfairness problem. In additional to these, the proposed schemes require no or minor changes to the slot format and are highly compatible with the DQDB standard protocol. As discussed early, the minor change in format only comes when the user message requests have a very large variance in length distribution. In that case, one additional field dedicating to the length specification can be compatibly added into the slot format and all the presented results and discussions remain the same in nature. The proposed strategy is surely a promising candidate of MAC protocols for the integrate-service applications in high speed real-time communication system.
















Chapter 5




INTEGRATED SCHEMES IN



WIDE-AREA NETWORKS





5.1 Introduction

The wide area networking model is defined as an environment in which there are lots of processing hosts or local area networks connected by high speed broadband trucks though intermediate routing nodes. It should be noted that the underlying networking medium is usually of very high speed and possible large bandwidth. This kind of communication system is believed to be the backbone of the next generation distributed real-time processing systems. Different from local and metropolitan area networks [22], the user connection establishment and the effective network buffering have to be considered in this wide-area communication model. Also, flow control and congestion control [82] needs to be addressed so that more efficient and more predictable performance can be achieved.

Most of the current communication systems are not real real-time oriented in



























Figure 5.1: Wide-area Network Model


the sense that the timing constraint is not addressed directly and explicitly. One apparent substitution for real-time communication is high speed networking techniques. This is based on the wrong argument that if a communication system is so much faster than the most time-constrained messages ever needed, then no scheduling or queueing algorithm need to be used to ensure the time responsiveness. In fact, the real-time requirement not necessarily means the existence of the shortest deadline. Nevertheless the service quality control to each individual class must be provided. The real-time communication system must be able to ensure that the delivery of messages is accomplished with the quality specified, and it happens within the deadline imposed by the application. Also, due to the complex multi-hop structure of WAN [45], proper buffer management and switching control under the time constraints invokes even more challenges.

A complete model of the wide area network can be shown in Fig 5.1. There are two levels of control in the network: connection setup control at the network ac-








cess point and switching control within the network. The connection setup control is responsible for the end-to-end resource reservation and allocation, and notifies the user whether this connection can be accepted or not. The switching control is to schedule the order of transmission to all those accepted messages at each intermediate nodes. Effective schemes need to be developed and employed at both levels to ensure the required real-time communication services. In this chapter, the effective control structure and schemes are going to be examined for real-time communications in wide-area networks. The emphasis will be on the effective resource allocation and scheduling schemes. They have been recognized to be very important and should be highly integrated to effectively satisfy the expected real-time services.



5.2 Control Structure Model


5.2.1 Two-Level Control Model

As pointed out in Chapter 2, there are usually two control levels due to the architecture of the wide-area network. We shall discuss briefly the existing control schemes in the following and then show the proposed integrated scheme for real-time communication in next Section.



A. Connection Setup Control

The bandwidth allocation is one major part in connection setup phase. The decision should be based on the information that how much additional bandwidth needs to be reserved on links over which the requested connection is to be routed. However, due to the statistical multiplexing of connections within the network, the








- Connection Setup Control --------------Switch
Control









End-System
Figure 5.2: Control Structure in WAN


exact measure of this reservation requirements is not so easily obtainable. The traffic demand should be based on some aggregate statistical measures rather than on the maximal demand per connection. A reasonable scheme is carefully developed from the so called Equivalent Bandwidth (EB) which is introduced by Cidon and et.al [15]. This bandwidth represents the equivalent amount of link capacity that is to be consumed by the request connection. It is a function of both the characteristics of individual connections and their interaction within the link. The quality of service can be met only if, at all links, the aggregate equivalent bandwidth remains below the available network capacity. The proposed approach relies on simple approximations that estimate the bandwidth requirements or equivalent capacities. The validity of this approach has been verified by comparison to both exact computations and simulation results [35].














0 0 0


. . .


Figure 5.3: Two-Level Control and Schemes


B. Switching Control

The message switching control determines the message transmission order within the network. It is a run-time scheme and should conform with the guarantees computed at the virtual connection setup time. The control should also consider the effects of early arrivals and develop effective mechanisms to avoid possible problems such as priority inversion and congestion. It has been well recognized that the conventional packet switching data networks with window-based flow control and first-come-first-served discipline can not provide services with strict performance guarantees. Several new control strategies and schemes have been proposed and analyzed recently [14, 24, 45]. The rate-based service discipline provides a client with a minimum service rate independent of the traffic characteristics of other clients. Such a discipline, operating at switching component, manages the bandwidth, service priority, and buffer space.








5.2.2 Challenges

It is not so difficulty to observe that the problems existing in wide area networks, especially when additional time constraints need to be considered, are far from being solved.

First, it should be recognized that the guaranteeing of bandwidth requirements does not necessarily guarantees the real-time requirements due to the statistically multiplexing traffic. In fact, the time-constraints imposed by application users have not been explicitly addressed at the connection setup phase.

Second, most of the effective switching control schemes, such as smallest-slacktime-first (SSTF) based schemes [74], assume some deadline information is available at all the intermediate switching node. But it is not always possible to obtain the accurate local deadline due to the dynamic nature of the network operating and the distributed control mechanisms.

Third, the control schemes employed in the connection setup and switch mechanisms are usually unrelated and working independently. This may cause the inconsistence and degrade system performance at the user interface.

Real-time communication requires the explicit processing efforts to individual message classes which are with different degree of service quality. After investigating various control structures in current networking systems, it is identified that the control schemes for real-time communications should be closely integrated so that the timing constraints can be enforced system-wide. While the connection setup control at the interfacing components and switching control at the switching components play crucial roles in wide area networks, they should not be separated and independent. A consistent and integrated control management is critical to satisfy









the timing constraints all the way through the network.



5.3 An Integrated Control Scheme


Based upon the previous discussion and the proposed models, an integrated control scheme has been developed which consists of two levels. The first level is the end-to-end network access level, which focuses on the guaranteed connection set up and bandwidth allocation. At this level, the control schemes address the global network scheduling problem to achieve the maximum system utilization with required quality of service for each message classes. The second level is, however, a sort of micro-tune processing within the network. Rather than considering the global context, this level addresses the scheduling problems locally at individual switching node along the network. The control schemes at this level actually enforce the global scheduling policies, and tune the system to meet timing requirement individually. Fig 5.4 shows the structure of our proposed integrated control scheme. At the connection setup control level, a request has to go through a delay bound checking procedure in additional to the conventional bandwidth checking procedure. The new delay bound checking procedure is to ensure the user required timing requirement can be satisfied. The acceptance is only issued when the request can pass both checking procedures so that not only bandwidth is allocated but also the delay can be bounded. As a by-product, a bound vector is computed by the delay bound checking procedure. This bound vector then feeds into each intermediate switching nodes as a deadline reference. So at the switching control level, the switching scheduling algorithm reorders the incoming packets and sends them along the outgoing link. Wise decisions can be performed by effectively using the informative







Reject




d


Connection Request



Setup
Control
Delay Bound Checking Pr
-- - - -- - - -


Accept
I


[-Bound Vector


Incoming Packets
i


Switch Control


ch S


cheduling Algorithm


Outgoing Packet

Figure 5.4: Integrated Control Structure


bound vector. It can be seen these two control levels are highly integrated in our scheme and thus able to achieve much better real-time performance.



5.3.1 Algorithm for Setup Control

At connection setup phase, the decision to accept a new request is usually based on whether the quality of service can be maintained under the present network load along the calculated route. The simplest scheme is to assign the peak bit rate required by all source users. Therefore the new request will be accepted only when the sum of peak rates of all current, including the new, requests on every link composing the route does not exceed the bit rate of that link. It is worth mention that different routes may have different results. Anyway, this is appropriate only in the environment where smooth and constant rate traffic is expected. Since the practical real-time communication traffic is often stochastic and hard to predict, it can be


0


| r


cedure


edure








easily seen that this scheme cannot reflect the real-time processing requirement and cannot achieve significant bandwidth efficiency.

To start determining the proper amount of bandwidth required by a connection, the characteristics of each connection request have to be clearly specified. Recall from the previous model discussed and the results of [15], the bandwidth required by an individual connection with vector (Rpeak,i, P,, bi) can be estimated using a simple fluid-flow model, and is given by:

EB = flRPeak - X + V[3Rpeak - x]2 + 4xp/3Rpk (5.1) 20

where P = ab(1 - p), x represents the available buffer space, and a = ln (1/e) with c being the required quality of service. This expression provides a simple and reasonably accurate estimate of the load request on network links. It is the basis of our further discussions on bandwidth allocation control and quality of service control.

A proposed connection setup control procedure can be described as following major steps.

1. Choose a route vector (L1, ..., Lk)c for the connection request c which starts at

link L1 and ends at link Lk

2. Calculate the equivalent bandwidth EBL,., for each link Li of this connection

c along the route

3. For each link Li, assign the new connection a highest priority level such that

the link delay for this connection WYL,,c is minimized. Note that, since there are existing connections at link Li, the priority assignment should keep the

delay of existing connections within their bounds (BL,,c,..., BL,,c,).








* The above calculations of the link delay WL,,c are based on the M/G/l

priority model and will be described later.

4. If the A = TD,-Es=I WLi,, > 0, the connection is established with guaranteed

service to the bound vector (BL,,C, ..., BL,,C)

* The bound vector is obtained by allocating the A to link delay WL,,, along

the route

Our scheme is under the assumption that there is a routing topology database similar to the one in ARPANET. Every node maintains a routing topology database with link weights representing the traffic over that link and the updating to each node is accomplished by a broadcast algorithm.

As shown before, the calculation of WL,,c is the key part for the whole procedure. Based on the standard M/G/1 priority queue, the non-preemptive priority scheme only influences the waiting time experienced by randomly arriving packets of different classes. The expected delay for a class-k packet can be written as Wk = J=1 E3(t2)/2 (5.2)
Ej=k- P1 _ j=k j
(1 - ZJk= - Ey=k )
Where Ej(t2) is the second moment of the class-j service time distribution and the smaller index value indicates the higher priority level.

While the packet transmission time T can be a close upper bound of the service residue time WO, the traffic intensity can be estimated by making use of the precalculated equivalent bandwidth. The computations of the accumulated traffic intensity and utilization of connection c can be obtained respectively.
k k
ak(L,) = >Zpj(L,)-= PL,,c (5.3) j=l c=1
PLi,c = EBL,,, (5.4) PL1,c BL,,max








Following is the procedure of finding the possible smallest link delay for the new connection request while not affecting the other already established connections.

Procedure FindtheSmallestDelay:

position = 0;

flag = ON;

While ((position < theConNo) and (flag == ON))
{

position + +; flag = OFF;

/ * check all the existing connections * /

Pt = Ph = 0.0;
for(i = 1; i < theConNo; i + +)
{

/ * get the i - th smallest connection � /

pi = Ph + linkfi].EB;

if(i == position)
{

P1+ = Pnew; Ph+ = Pne;


at pi/theMaxCapacity;

arh = Ph/theMaxCapacity;

newDelay = T/(l - at) * (1 - ah);

/ * feasibility check * /

if(newDelay > BD[i])










/ � failed and try next higher index * /

flag = ON;



Ph = pI;


}/* while* /



It is worth mention that here the priority is not really assigned to each connection but works as a means to obtaining the delay bound vector. Actually, we do not intend to assign a static priority to each connection which is usually not so effective when the traffic is multiplexing and unpredictable. What we really want is the more accurate local link delay (through the delay bound vector) for each connection and transforming the global timing constraints into a distributed control system. The exact transmission order at each switching node is decided locally by considering both the local delay bound and current traffic status at that node, which is the job to be accomplished by the next level control called dynamic switching control. Under this integrated control scheme, the ene-to-end real-time services can be guaranteed as long as the local delays along the network are enforced.


5.3.2 Algorithm for Switching Control

As it has been widely accepted that the overflow error instead of bit-error has become the major factor in high speed real-time communications, the role of switching nodes has become more and more important. One primary task of the switching nodes is to schedule various incoming messages and to determine their service pref-




Full Text

PAGE 1

CONTROL SCHEMES IN HIGH SPEED REAL-TIME COMMUNICATION NETWORKS BY LI-TAO SHEN 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 1992

PAGE 2

ACKNOWLEDGEMENTS I would like to express my sincere thanks to my advisor, supervisory committee chairman, and friend, Dr. Yann-IIang Lee, for his guidance and encouragement. Without his support, this work would never have been accomplished. During the years of this relationship, he enlightened me not only by his knowledge, but also with his way of exploring truth. I also would like to express my appreciation to the other members of my supervisory committee, Dr. Randy Chow, Dr. Richard NewmanWolfe, Dr. Panos Livadas, and Dr. Haniph Latchman, for their commitment and service on this committee. Special thanks go to my wife, Min Pu, for her understanding and continuous support for this work. ii

PAGE 3

TABLE OF CONTENTS page ACKNOWLEDGEMENTS ii ABSTRACT v CHAPTERS 1 INTRODUCTION 1 1.1 Problem Statement 1 1.2 Background and Related Works 4 1.3 Dissertation Outline 8 2 REAL-TIME COMMUNICATIONS AND CONTROLS 11 2.1 Real-Time Communication Service 11 2.1.1 Applications and Requirements 11 2.1.2 TimeConstrained Performance Measures 14 2.2 Real-Time Communication Support 16 2.2.1 Networking Architectures 16 2.2.2 Traffic Characterization 19 2.2.3 Control Enforcement Structure 20 2.3 General Control Approaches 21 2.3.1 Priority-Driven Mechanism 21 2.3.2 Dynamic Scheduling Control Approach 23 2.3.3 Message Level Processing Approach 25 3 MEDIA ACCESS CONTROLS IN LANS AND MANS 27 3.1 Multiple Channel Token Ring (MCTR) 27 3.1.1 Background 27 3.1.2 System Architecture Model 30 3.1.3 WUT Control Strategy and Mechanisms 32 3.1.4 Performance Evaluations 37 3.2 Distributed Queue Dual Bus (DQDB) 47 3.2.1 Background 47 iii

PAGE 4

3.2.2 Dynamic Control Model and Strategies 51 3.2.3 Priority Assignment Algorithm 54 3.2.4 Simulations and Evaluations 56 4 MESSAGE LEVEL PROCESSING SCHEMES 65 4.1 Introduction 65 4.2 Message Level Processing 66 4.3 Analysis Model and Evaluations 69 4.3.1 Service Time Distribution 69 4.3.2 Waiting and Delay Time 70 4.3.3 Numerical Results 71 4.4 Transmission Control Schemes 73 4.4.1 Multiple-Slot Reservation Protocol 73 4.4.2 Enhanced Multiple-Slot Protocol 76 4.5 Performace Evaluations 78 5 INTEGRATED SCHEMES IN WIDE-AREA NETWORKS .... 83 5.1 Introduction 83 5.2 Control Structure Model 85 5.2.1 TwoLevel Control Model 85 5.2.2 Challenges 88 5.3 Integrated Control Scheme 89 5.3.1 Algorithm for Connection Setup Control 90 5.3.2 Algorithm for Switching Control 94 5.4 Evaluations and Analysis 99 5.4.1 Simulation Model 100 5.4.2 Admission Controls 102 5.4.3 Swtching Controls 105 6 CONCLUSIONS 112 REFERENCES 116 BIOGRAPHICAL SKETCH 129 iv

PAGE 5

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 CONTROL SCHEMES IN HIGH SPEED REAL-TIME COMMUNICATION NETWORKS By LI-TAO SHEN November 1992 Chairman: Dr. Yann-Hang Lee Major Department: Computer and Information Sciences Recently rapid engineering advances and new applications of distributed computing technology impose stringent requirements on computer communication networks in supporting a wide variety of services. In additional to fast data delivery, features such as fault-tolerance, multicasting and security, are all in demand. Most importantly, as the distributed applications move to real-time domains, many timing requirements have to be enforced within the communication systems. A real-time communication network is such a system which is able to provide its users the ability to specify the timing requirements and to obtain guarantees about the satisfaction of those requirements. The predictable operation and a high degree of schedulability are two of the most desirable properties of a real-time communication network. Timing correctness, as well as the traditional functional correctness, is ex-

PAGE 6

tremely vital in various distributed real-time applications. However, the current networking systems employed in various time-constrained applications are generally not real-time communication systems. The main problem here is that there is no explicit real-time control schemes employed in these communication systems to guarantee the timing requirements. While the multiple priority has been identified as one of the most commonly used mechanisms for current real-time communications, it is not very clear how to assign the proper priority consistently in a complex and dynamic networking environment. Several dynamic scheduling approaches are proposed and studied in detail in this dissertation. The main idea of these approaches is to schedule the messages to traverse through the network in a dynamic way such that the user specified timing requirements can be maximumly satisfied. By dynamic, it is meant that the priority should be time-variant and decided by both the deadline requirement and current network operating status. By scheduling, it is meant that the message transportation in real-time communication networks should be emphasized to minimize the lost percentage. A set of generic control schemes are proposed based on the general approaches and also evaluated extensively to show their promising performance for real-time communication applications. vi

PAGE 7

Chapter 1 INTRODUCTION 1.1 Problem Statement Real-time communication networks are distinguished from the normal network systems with the introduction of time constraints. They are used to insure on-time delivery of messages and to support distributed real-time computations. The performance measures of such networks differ from those of the conventional networks. The principal performance considerations for the conventional network control protocols are to maximize the system throughput and to minimize the average delay. In real-time communication systems, however, the main performance consideration is to maximize the percentage of messages that are delivered within the given time constraints [50, 83]. The different performance metrics, reliability requirements, and performance trade-offs suggest that the control protocols previously developed for traditional communication networks may no longer be suitable for time-constrained communications. According to the well-known open system interconnection (ISO/OSI) reference 1

PAGE 8

2 model, the communication architecture can be identified as a seven-layered processing stack. Two main subsystems can be further defined: • end-host processing system • networking communication system The former one is totally transparent from networking details. It is built upon a transport system meeting certain quality of service (QOS) requirements, and provides the end-host users with a unified application service. The functions of this interoperating system actually accomplish a kernel distributed operating system, which is usually corresponding to the Session layer, the Presentation layer, and the Application layer of the OSI model. The second part of the architecture is the transport system whose basic functions and services can be viewed similar as those of the Physical layer, the Data Link layer, the Network layer, and the Transport layer of the OSI model. While the end-host processing system is very important for providing application users various powerful and efficient services, its functional capability would mainly depend upon the networking communication system. It can be obviously observed that there is a close relationship between the endhost processing system and the networking communication system. This is especially true when we deal with real-time distributed systems. However, there are some fundamental differences between real-time communication networks and traditional communication networks. This is majorly because a new performance dimension is introduced into the traditional model when the additional timeliness requirement should be considered and enforced. It is well recognized that real-time communications will be the backbone for the next generation networks where predictable communication services are the basis for further application development.

PAGE 9

We believe that, a solid understanding of real-time communication networks and a developing methodology of such networks can not be overlooked to address the real-time distributed operating systems and applications. Generally, the current networking systems employed in various time-constrained applications are not real-time communication systems. Some main problems include: • lack of explicit timing representation and operation mechanism; • undetermined communication services due to the network routing and delay; • local and global priority inversion across different processing domain; • the improper mapping of message's timing attributes; • weak dynamic control mechanisms to adopt to various application environments; The key problem here is that there is no explicit control support employed in the current communication systems to guarantee the real-time requirements, which is actually quite essential to many distributed processing environments. It is the real-time communication system that must do whatever is necessary to bring the quality of service provided by vast variety networking architecture up to the level required by the communication service users. It is easy to see that this is an integrated control system with a substantial amount of processing efforts and various decision choices. Therefore, the real-time networking communication system is an integrated multilayered system, which not only provides the principal functions of conventional OSI lower layers but also, of the same importance, provides integrated real-time control mechanisms at each individual layer. While the

PAGE 10

4 conventional protocols are not real-time oriented, a new design methodology and a set of new protocols are to be developed. The objective is to provide application users a predictable and guaranteed real-time communication service interface. The efficiency of the system is to be achieved by selecting the most suitable control function for the requests given by application users. This research is intend to investigate the effective control schemes as well as the proper architecture model for real-time communication networks in a distributed environment. The main objectives of this dissertation research are: • to investigate the communication architectures with guaranteed performance in a high speed real-time environment; • to design, analyze, and evaluate various effective control schemes in real-time communication systems, Besides a better and more detailed understanding of the increasing real-time communication systems, this research work is intended to devise a number of generic control schemes and algorithms which can be applied in practical environment to enhance the real-time performance. It also provides significant values to the fundamental control of real-time communication in distributed systems and meets the increasing demands of real-time communications. 1.2 Background and Related Works The OSI Basic Reference Model [42] evolved out of early work to describe the communications infrastructure required by applications such as banking, airline reservations and ticketing, and other industries requiring distributed access to large

PAGE 11

5 data bases. The emphasis in such systems was on information accuracy and system reliability [19, 20, 31, 36, 71]. Real-time response was not considered as a prime design goal. The network system was deemed as a success if almost all queries were answered within a few seconds [66]. As a result, most standards that developed under the OSI model ignored the needs of industrial applications, where timeliness of messaging is paramount. Issues such as message priorities, bounded messaging delay, and efficient multi-node communication via selective broadcast either were ignored, or were relegated to the bonepile of unresolved standards problems by designating them "for future study" . With the advent of high-speed networking technology, the existing communication architecture and control schemes are no longer appropriate. New design and analysis methods are introduced for very high-speed networking architectures [3, 13, 22, 61]. The requirement for small delays and low processing overheads has also brought about the development of so called light-weight protocols [27]. Some of the representative systems are Xpress Transport Protocol (XTP) developed at University of Virginia [75], Versatile Message Transaction Protocol (VMTP) developed at Stanford University [11], Network Bulk Transfer (NETBLT) developed at MIT [16], and GAM-T-103 Transport System developed in France [59]. All these systems are dedicated to the single transport layer with no supporting control schemes in underlying layers or specific integrated control schemes. The main effort of these systems is to minimize the processing overhead for data communications. Real-time computing systems is expected to be the backbone of the next generation system. The system requirements and concepts were well discussed by Ferrari [29, 30], Krishna and Lee [49], Kurose and et.al [50], and Stankovic [83, 84]. Several real-time system model as well as design strategies have also been proposed.

PAGE 12

6 The ARTS is a distributed real-time kernel system developed by Tokuda and et.al [94]. The Spring is a real-time operating system kernel developed by Stankovic and Ramamritham [85]. The HARTOS is a distributed real-time system developed by Kandlur and et.al [44]. In addition, the CHAOS is an operating system developed for real-time applications by Gopinath and Schwan [33]. While most of these systems assume a guaranteed communication service and emphasize on operating system supports for real-time applications, the issues of real-time communication have not been well addressed. There is a lot of research being undertaken in developing various media access control (MAC) schemes due to the wide installation and application of the local and metropolitan area networks. Surveys of media access control mechanisms in highspeed network systems were made by several researchers [3, 13, 25, 73]. However, real-time requirements were not considered in the study of most of the local and metropolitan area networks. IEEE 802.5 Token Ring network was well discussed and analyzed by many researchers [7, 8, 9, 40, 65, 78]. New technology like optic fiber has brought about a lot of attention on both the ANSI FDDI networks [4, 5, 72, 77, 79, 96] and IEEE 802.6 DQDB networks [17, 21, 41, 43, 63, 76]. Various analytical model and evaluations were made to these medium access control protocols [6, 34, 69, 70, 89, 95, 98]. Some enhanced and new protocols were also developed recently in the context of real-time communication requirements. Shin and Hou [81] evaluated three contention protocols used for real-time communications. In their model, the probability of missing message deadlines was taken into account and some analytical results were obtained. Virtual time CSMA protocols were proposed by Zhao and et.al [100, 101] for hard real-time communications. Strosnider and et.al [87, 88] developed a de-

PAGE 13

7 ferrable service algorithm for periodic message scheduling. The algorithm not only guarantees the deadlines of periodic messages, but also substantially reduces the response time of aperiodic messages by assigning the highest priority to the aperiodic messages up to the point where periodic messages would start to miss their deadlines. The consideration of applying this algorithm to the IEEE 802.5 token ring protocol were discussed in [87]. The general wide area networks with any networking topology introduce more complicated problems and challenges [14, 15, 23, 28, 29, 45, 74], though they have the potential for higher performance and reliability than common bus or ring structures. Maxemchuk and Zarki [57] made a good survey on routing and flow control in high-speed wide-area networks as well as local area and metropolitan area networks. Since the information is delivered across multiple hops along the network, the issue to be addressed is not the control schemes for medium access but the message scheduling along the network with delivery time constraints. Ferrari [29, 30] specified several real-time performance measures from the point of user application requirements, and presented a real-time channel scheme which was claimed to be able to provide guaranteed performance. A multi-hop network model and scheduling scheme were further developed by Kaudlur and et.al [45] to provide predictable inter-process communication in real-time multiple stage networks. Cidon and et.al [15] discussed their proposed methods for bandwidth management and congestion control in very high-speed networks and their experience from the plaNET wide area network [32]. While the connection establishment management is very important in wide area networks, the proper queueing and scheduling schemes at switching components are also crucial to the system performance [51, 74]. A priority assignment control scheme with quality of service constraints was proposed by Takagi

PAGE 14

8 and et.al [90] for line buffers in ATM networks. Fair queueing algorithm was first proposed by Nagle [62], and was studied by several researchers [24]. The reasoning of these proposed algorithms was to prevent a source from arbitrarily increasing its share of the bandwidth or causing the delay to other sources. However, the time constraints were not the principal consideration despite the fact that the improper queueing discipline and scheduling could cost unnecessary percentage of out of deadline messages. 1.3 Dissertation Outline Real-time communication can be conceived as a complex system that satisfies multiple classes of transmissions under various t ime-constraints imposed by application users. Simple and effective control mechanisms are the key components to guarantee the service requirements even if the communication system is under high utilization. In the next Chapter, the general characteristics of real-time communication and control mechanisms are first discussed. By presenting an application example, the real-time service requirements are identified in detail. Dedicated real-time performance measures are also proposed, which are significantly different from the traditional networking performance measures. The supporting structure for real-time service is then presented. Several control st rategies and mechanisms are proposed and discussed in general for real-time communications. In Chapter 3, the effective media access control schemes are discussed in detail in the context of providing real-time services. In both the local area networks (LANs) and metropolitan area networks (MANs), the media access control (MAC) plays a key issue to provide real-time performance. The implication of deadline

PAGE 15

9 and/or priority information to the design of MAC layer protocols is carefully investigated. Since MACs are closely network related, special implementation considerations should be given to each type of networks. Two examples have been used to examine the control schemes of different MAC protocols in high-speed local and metropolitan area networks. A wait-until (WUT) control scheme is first proposed for multiple-channel token ring networks. Then, a dynamic priority assignment control scheme is designed for distributed queue dual bus (DQDB) networks. It has been showed that the performance gain achieved could be substantial in real-time communications . The message level processing (MLP) strategy and control schemes are discussed in Chapter 4. Specifically, several message level processing schemes have been studied in the DQDB network. The standard DQDB protocol is based on segment level operations and is not very efficient from the application users' point of view. A class of multiple-slot transmission schemes have been developed to reduce the message level transmission delay. The main idea of the proposed schemes is to perform the service reservation request at the message level instead of at the segment level. The proposed multiple-slot reservation protocols contribute to the DQDB network with processing efficiency, better bandwidth usage, and improved message delay performance. In additional to these, they are highly compatible with the original standard and are able to suppress the unfairness problem. In Chapter 5, the design methodology for real-time control schemes in high-speed wide area networks (WANs) is discussed. Due to the additional multi-hop switching system, guaranteeing the required quality of service in WAN brings even more challenges for real-time communications. An integrated control scheme has been proposed and studied, which includes the connection setup control at networking

PAGE 16

10 access point and dynamic scheduling control at switching nodes. The connection setup procedure not only checks the availability of the network resources for specified requirements, but also assigns the delay bound vector along the route. While the switching system is fundamental in wide area network environment, effective queueing discipline and dynamic scheduling schemes are carefully designed to accommodate the requirements of various QOS classes and to achieve good real-time communication performance. Finally, a brief summary of the dissertation work as well as some further research works are given in Chapter 6.

PAGE 17

Chapter 2 REAL-TIME COMMUNICATIONS AND CONTROLS 2.1 Real-time Communication Service 2.1.1 Applications and Requirements Real-time communications are driven by many distributed applications which need guaranteed real-time services at the communication interface. While the distributed systems become even more popular, the variety of the user requirements can be so vast. Besides the usual functional requirements such as transmitted messages should be free of error and in order, there are some other important requirements such as reliability and security. As the applications go to the real-time domain, timing requirements must be considered and guaranteed. These timing requirements include messages should be transmitted before their deadlines and the percentage of mes11

PAGE 18

12 Figure 2.1: Service Requirements sages missing deadlines should be minimized. All these various requirements can be shown as in Fig 2.1 and should be met at the interface of communication networks. In fact, the real-time communication is just one kind of networks which is able to enforce the timing requirements and provide predictable communication services. Within real-time communication networks, not only the functional correctness but also the timeliness correctness must be guaranteed. The real-time communications and timing requirements are not something new and can be easily observed in our everyday life. There are enormous applications which have various real-time communicat ion requirements. These applications cover almost every aspects, including bank transactions, airport scheduling control, multimedia telecommunications, and military networking management. One realistic example can be found in the Distributed Interactive Simulation (DIS) networking environment [103], which is sponsored by the Defense Advanced Research Project

PAGE 19

13 Agency (DARPA) in partnership with the United States Army. The goal of this simulated battlefield program is to provide, for the first time, an opportunity for fully-manned platoon-, company-, and battalion-level units to fight force-onforce engagements against an opposing unit of similar composition. A description of a vehicle's appearance passes through several hands from the time it is first expressed as a protocol data unit (PDU), to the time the vehicle is displayed by an observer's simulator. The steps include processing by the software that provides communication services, transmission across a network, and perhaps queueing within the receiving simulator. The magnitude of this discrepancy is proportional to the speed of the vehicle described by the PDU, and to the magnitude of the network delay. Therefore, this effect is expected to be most evident and critical in certain situations such as when aircraft flying at high speed are able to observe each other closely while being simulated by widely separated simulators. It is quite obvious that efficient and fast are relative terms and are not sufficient when dealing with real-time requirements. In this case, timing correctness, as well as functional correctness (received correctly and in order), is extremely important for application requirements. Certainly, a specific priority field can be added into PDUs to represent the criticalness of the message. The problem of real-time communication is unfortunately not so easy and far from being solved. It is not the real-time requirement or the user interface format, but the underlying mechanisms employed in existing networks that must provide explicit or enough supports for real-time communications.

PAGE 20

14 2.1.2 TimeConstrained Performance Measures Interests of new timing performance are recognized for real-time communications. Several performance metrics have been proposed and denned. Instead of the traditional performance measures like average delay, throughput, and fairness, real-time communication systems emphasize on the guarantees of the timing requirements and system utilization. The most important performance measures, such as transport deadline, message lost percentage and channel utilization, can be defined as: • Transport Deadline Duration (D): The transport deadline duration is the period from the time a message entering the system until the user specified expiration time. D = expired time — arrival time (2.1) • Message Lost Percentage (L): A message is considered to be lost if it cannot be delivered before its deadline is expired. The lost percentage is defined as follow. no. of messages lost = 7 7 i r r \^"*") [no. of messages lost + no. of messages sent out) • Channel Utilization (U): The channel idle time is the time spent when a node has nothing to send. This measure is limited to the on-time delivery. y _ message transport time (message transport time + channel idle time) The increasing real-time data communication requires integrated networking transmission service for multi-class message traffic [74, 83]. One important issue is

PAGE 21

to satisfy all those multiple classes of transmissions so that the various service delay requirements are guaranteed even the communication system is under high utilization. Different classes of transmissions can be classified by their different requirements in time-dependent constraints (deadlines). With smaller deadline constraint, this class of transmission is often considered to be more urgent and certain service preference is needed to guarantee its real-time requirement. Priority queueing [70] is one of the mechanisms often used for real-time communication since it provides an effective means to give preference to individual message class. In complex real-time communication systems, various types of traffic are stochastically multiplexed to efficiently utilize the network resources. However, the excessive use of the bandwidth may cause traffic-dependent quality of service (QOS) deteriorations. Generally, the service quality in communication context refers to accuracy and speed of information delivery, and to the absence of certain impairments such as excessive delay, excessive variance of delay, transmission error and out of sequence delivery. In additional to these, the quality of service (QOS) in real-time communication also includes • the required lost percentages for each class of message, • the bounded transport delay for each message class, • the high system utilization. There are two possible common ways of QOS management in real-time communications: (1) to define a single QOS class and manage the transmitting information equitably; (2) to define multiple QOS classes and manage each different classified class individually. Although the control scheme could be much simpler in the first method, the real-time performance is very hard to be ensured for a wide variety of

PAGE 22

16 user requirements. The advantages of the second method are the bandwidth efficiency can be improved by suitable control such as priority scheme and the users are able to select one of the classes most appropriate to their requirements. The multiple QOS classes strategy is of more interest in this research. Whenever a user intends to set up a network connection, the required QOS class of the sending messages should be clearly specified at the network entrance points. It is worth reminding that while the connection is class-oriented, the traffic on any particular link could come from various sources and belongs to multiple classes. Therefore, the control schemes need to be designed to make efficient use of the transmission bandwidth while satisfying the QOS requirements for all the classes. 2.2 Real-time Communication Support 2.2.1 Networking Architecture Networking architecture can have significant, effect on the design of effective control schemes. Different architecture not only invokes different networking protocols, but also has its exclusive advantages and special problems. For local and metropolitan area networks, the architecture is usually very simple. The most common ones are star, ring, and bus as shown in Fig 2.2. There is only one networking access point and no intermediate switching node. The media access control (MAC), thus, is the key for cooperating all the nodes and supporting required communication services. The standard message transmission has two phases: • compete for the media access control

PAGE 23

17 L-J BUS 1 RING 1 Figure 2.2: Architecture for LANs/MANs • transmit the message through the media While the first phase is governed strictly by the MAC protocols such as CSMA/CD (IEEE 802.3) and Token Ring (IEEE 802.5), the second phase usually has the determined execution period. In these networks, therefore, the major challenge for real-time communications is how to make the best use of the special architecture and design the effective MAC protocols so that specified services can be provided. For wide-area networks, there are basically two major different ways in dealing with the integrated data communications. In the circuit-switched (synchronous transfer mode) networks, sufficient resources are allocated to each call request to handle its maximum utilization. This guarantees that the user will get the quality of service required, but, on the other hand, may be wasteful of system resources. In the packet-switched (asynchronous transfer mode) networks, call requests from all sources are packetized, and statistical multiplexing techniques are used to combine

PAGE 24

18 all network traffic through multihop switching fabrics. This allows higher network utilization. However, more complex and proper control schemes are inevitable to ensure the required service, especially for real-time communications. The kind of wide area networks to be considered is a class of networks based on Asynchronous Transfer Mode (ATM) design principle and architecture. While there are various components in a wide area network, such as monitor and auditor, our study will focus on two primary components: interfacing nodes and switching nodes. • Interfacing nodes Interfacing nodes are those attached to external components in either network entrance point or exit point. They are responsible for the connection setup and management. • Switching nodes Switching nodes are collectively referred as the switching subsystem which is the major part of a wide area net work. A switching node can be logically regarded as three parts: input buffer, switch, and output buffer. It should be noticed that control schemes in wide-area networks usually invoke more challenges due to the intermediate switching systems. Two levels of control are needed in wide-area networks: connection setup control at the network access point and switching control within the network. Various schemes have been developed and employed at each level to provide specified communication services.

PAGE 25

19 2.2.2 Trafiic Characterization It is believed that the guaranteed real-time communication service could only be accomplished by making use of certain types of resource reservation. The system needs detailed characteristics of the required guarantee service so that it can reserve the corresponding resources and avoid possible interference from other contending requests. However, the non-time-constrained traffic obviously does not have to reserve the bandwidth and compete with the time-constrained traffic. This traffic can use the gaps in the bandwidth usage of reserved ones, and its behavior will not affect the quality of service given to the reserved traffic. To characterize the traffic features for each connection, it is necessary to select an appropriate model to specify the traffic in terms of known parameters. The two-state fluid flow (STFF) model is adopted and recognized as a proper means to capture the features of a wide range of connections. In this model, the source is either in an idle state transmitting at zero bit rate, or in a burst state transmitting at its peak rate. The advantage of using such a model is that it is flexible as well as simple. It can represent connections ranging from burst to continuous bit stream, or even as the approximation of more complex sources Based on this model, the idle and burst periods are defined as the times during which the source is idle or active respectively. The peak rate of a connection and the distributions of idle and burst periods completely identify the traffics statistics of the connection. Specifically, connection i is represented by a request vector {Rpeak,iiPiybi)i where Rj,eak,i is the peak rate at which the source generates data, Pi is the utilization or fraction of time the source is active and transmitting at Rp ea k,ii and the 6, is the average duration of an active period. From these three

PAGE 26

20 b j | |I|| 1 ||1| | ||| 1 | I |---TR peak ! I I «... ! I i : Unit Period — { ! I Figure 2.3: Two-State Fluid Flow (STFF) Model basic parameters, we can also derive the mean bit rate m, and its variance of. m, = pR pea k,i (2.4) a\ = m,(Rp Cak ,i m.) (2.5) It is worth mention that the mean burst period 6, gives the information of how data is being generated by the source. Two sources, with identical mean and peak bit rates but different burst periods, have different impacts on the network. 2.2.3 Control Enforcement Structure The whole control structure for real-time communication systems consist of both the system users and system providers. The users are responsible for the specification of the specific real-time requirements such as bounded delay and lost percentage. The users are also required to have an effective way to describe the expected traffic pattern to go through the network. Given these, the communication system provider should have the ability to satisfy the user requirements transparently. The layered structure and interface can be shown as in Fig 2.4. The users will be notified if the

PAGE 27

21 Application Service Figure 2.4: Control in Real-Time Communications provider cannot guarantee the required service at the current network operating status. The user can either withdraw the request and try later, or continue the request at the best-effort manner. As stated before, the media access control (MAC) protocol has been identified as the key challenge for local and metropolitan area networks. For wide-area networks, the challenges of real-time control come from two closely related parts: connection setup control and intermediate switching control. 2.3 General Control Approaches 2.3.1 Priority-Driven Mechanisms The multiple priority mechanism is the most commonly employed scheme in attacking real-time communication problems. Examing carefully the existing communication systems, it is not difficult to observe that these systems actually do not

PAGE 28

22 have consistent mechanisms explicitly in supporting real-time application requirements. Although the media access control protocols in some local and metropolitan area networks, like Token-Ring, DQDB, and FDDI, provide priority mechanisms for possible differential service, some others like CSMA/CD even do not have priority mechanisms. In these latter ones, priority inversion problem is common where higher priority (critical) messages may be transmitted after lower priority ones. This is a very undesirable property, especially for real-time communication systems where timing requirement is emphasized. Even for those networks providing priority mechanisms, there also exist problems such as the static priority assignment, inefficient enforcement scheme, and insufficient priority levels. The problem in existing transport layer protocols is even worse. The expedited service of OSI transport class 4, for example, is weak in definition and undefined in implementation. Furthermore, it is unclear how such priorities are mapped to lower layer services, even if the priority is preserved. TCP uses a called Urgent field in its TPDU structure to indicate that some number of bytes are special and should be processed out of order [92]. It is still not clear to the users how to make use of this facility instead of sending interrupt messages. Some recently developed systems like Xpress Transport Protocol (XTP) and Versatile Message Transaction Protocol (VMTP) are expected to be more appropriate for high speed and real-time communications. While they concerned more about high speed processing, the realtime mechanism is not well addressed. The priority mechanism in XTP intends to support both incoming and outgoing messages [75]. For outgoing messages, the priority level is encoded into a 4-byte integer and placed into the SORT field before transmission. When the message arrives at the remote receiver, the SORT field is examined, and the message is enqueued according to its priority. However, this

PAGE 29

23 scheme is of static type, in that the priority level remains constant as the message travels through the network. It can be observed that the priority mechanisms in existing networks are either not sufficient or even not available system-wide. Even they are available through the whole protocol stack, inconsistency is another big problem. The priority in these systems is generally an intra-layer concept rather than a system-wide control mechanism. On the other side, it is anxious to know how to use this distributed mechanism effectively in real networking practice. It is argued that the application users should not take care of the priority assignment. Since the users are never able to track the detailed operations of the communication service provider, they often feel quite at a loss at priority assignment except following a static policy. The user should give the specific requirement, like deadline, and it is the responsibility of the system provider to decide what and which is the most suitable mechanism, say priority, to satisfy the requirement. 2.3.2 Dynamic Scheduling Control Approach While the multiple priority mechanism has been identified as one of the most commonly used control schemes for current real-time communications, it is not very clear who and how to assign the proper priority consistently in a complex and dynamic networking environment. A dynamic scheduling control approach has been developed as a general approach and investigated in detailed in various specific networks. The main idea of this approach is to schedule the messages to traverse through the network in a dynamic way such that the user specified real-time requirements can be maximumly satisfied.

PAGE 30

24 • By dynamic, it is meant that the priority should be time-variant and decided by both the deadline requirement and current networking operating status • By scheduling, it is meant that the message transportation in real-time communication network should be more emphasized as a scheduling issue to minimize the chance of missing message deadlines Instead of static control mechanism like user specified priority, the priority should be timevariant for effective real-time controls. The priority is more likely in the mechanism domain which is employed by the service provider and transparent to the users. Users do not have to worry about the detailed mechanisms like priority but simply specify the requirement and policy. It is the provider's responsibility to provide the required service by optimizing the networking operating status and user requirements. The slack time, from the current time up to the expiration time, is widely used to reflect the dynamic property of each message. While the smaller slack time usually implies more attention for this message, various QOS classes and current network operating status have also to be considered to make the best use of the system and provide the required services. The prime objective of real-time communications is to provide predictable service instead of fast transportation. As long as the message can be transmitted within its deadline, there is no reward for early delivery. In another word, the message transportation should be carefully scheduled so that the message delivery can be guaranteed and also system can be highly utilized. The interaction and ordering of various QOS classes have to be resolved dynamically in response to changing traffic flow. It is also very important for the control schemes to be distributed. Efficiency, flexibility, and predictability are some other desirable properties have to

PAGE 31

25 be considered. 2.3.3 Message Level Processing Approach Message level processing approach is quite natural since the timeliness is an attribute belongs to the message level. Since the processing is now driven by each individual message rather than the segmented packets, it is a message-driven model with the relevant control schemes integrated as a single processing object. All the objects are interacted each other via their relative time constraints. The more urgent an object is, the more processing effort should be considered. This model actually provides a system-wide virtual priority processing mechanism in which the priority assigned to each message is dynamic and time-dependent. This is because that the message criticality is not only staticly determined by the real-time application but may also be dynamically changed as the message keeps remaining in the system. A time-constrained message with a very long deadline could also be very urgent when this message has been blocked for enough time. Therefore, there is no static priority imposed upon each message. All the scheduling and queueing algorithms are designed and "driven" by the time requirement at the message level. While this model naturally introduces the messages as parallel processing objects, some parallel processing architecture can be employed to provide potential much more fast processing. Since segment delay does not make any sense to the application users, and only the complete message level performance is what they are really interested. It can be observed that this model is controlled at semantic level rather than the syntax format. The basic data unit for manipulation and control management is the message instead of the packet. Some benefits can be obtained including more

PAGE 32

26 effective transmission and better real-time performance. For instance, when certain portion of a message fails to meet its deadline, then the rest of this message should be aborted, since there is no way to satisfy the requirement of this message. Also, endto-end time fence checking mechanism based on message boundary is quite effective to minimize the dependence of one message on the timing characteristics of other messages. Based upon the delay bound provided by network, this mechanism can quickly isolate a timing violation. By requiring explicit timing information about each message, a time encapsulation at message level can be achieved.

PAGE 33

Chapter 3 MEDIA ACCESS CONTROL IN LANS AND MANS 3.1 Multiple Channel Token Ring (MCTR) 3.1.1 Background Token ring network systems have been extensively studied and also widely used during the past decade [7J. The basic single token ring network consists of a number of stations N attached on a ring and a control token rotates around the ring, station by station. If the station receiving the free token has message packets to transmit, it converts the free token into a connector and then follows the connector with its sending message. If the station has no message waiting for transmission, it simply passes the free token to the next station. The source station has the responsibility to remove the packet from the ring and to generate a new free token after removing the packet and then passes it to the next station. Compared to other structures, the ring networks have the good properties of 27

PAGE 34

28 bounded transfer delay and better channel utilization [6, 78]. However, the transfer delay will be considerably high under moderate and high load. It is even worse when the messages are critical packets with real-time requirements. This is due to the fact that most current protocols aim to minimize the average transmission delay and have no mechanism to favor the critical packets. For these real-time network systems, it is obvious that the round-robin methods resulted from token rotation are not appropriate. The priority based methods, which aim to favor messages according to their priorities, are the currently prevailing techniques. A real-time scheduling method for prioritized messages has been proposed and investigated by Strosnider [88]. Using IEEE 802.5 token ring protocol, it has been shown that a better real-time performance can be achieved by choosing a proper packet size and operating the token in priority mode. Currently a high speed optical fiber ring network known as FDDI is being developed and standardized. The key characteristics of FDDI include optical transmission medium, fair and robust control protocol, a datarate of 100 Mbit/s, a ring length of up to 100 km and up to 500 stations on the ring. A fully decentralized priority mechanism is used in FDDI supporting synchronous and asynchronous transmission modes. Synchronous transmission is real-time sensitive and the delay of synchronous transmission is limited by reserving an appropriate bandwidth at ring initialization. The remaining available bandwidth can be used for asynchronous operation. Performance analysis has shown that the throughput and the real-time response of the FDDI cannot be optimized simultaneously, especially for long ring lengths [77]. Though the priority based and bandwidth allocation methods are much better than the conventional methods in real-time applications, there are still some

PAGE 35

29 problems: • the introduction of a high bandwidth channel may be accompanied by only insignificant increase in system capacity, and the increase of channel bandwidth can only be partially utilized; • the distribution of time-constrained messages is usually not predictable, and a static allocation of bandwidth is hard to meet the general real-time requirements; • to preempt the normal token rotation, a lot of time is wasted in doing token reservations and hence the system utilization is low; • the starvation and unnecessary delay of low priority messages is quite common; • the total percentage of messages that miss their time constraints may still be quite high though that of the most urgent ones is decreased; For a high speed transmission medium, it is attractive to partition the medium into multiple channels. It was also shown that for a given system bandwidth, the system capacity can be increased by bandwidth subdivision [13]. These channels can form a multiple ring network and have two main advantages: (i) they increase network capacity by operating on several slower channels so that the propagation delay and other penalties become a smaller fraction of the packet transmission time, and (ii) they can be easily implemented by expanding the existing interface technologies based on medium speed. Several approaches have been suggested for multiple ring networks [9]: (1) separate queues with simultaneous transmissions, (2) single queue with simultaneous

PAGE 36

30 transmissions, (3) single queue with single transmission. The analysis and simulation results showed that the single queue with simultaneous transmissions protocol has a better performance than the other two protocols but its interface design is more complex than the others. In this section, a new protocol is presented which is designed for real-time communication in multiple token ring networks. With proper channel allocations and priority reservations, the protocol can reduce the percentage of messages that miss their time constraints and also maintain a high channel utilization. It is a dynamic control policy, flexible to any kind of system load, and easy to implement. 3.1.2 System Architecture Model The real-time communication model can be specified as follows: for each packet P t , i = 1, 2, . . . , m, there are two principal parameters A { and E { . A { is the arrival time of Pi and Ei is its expire time before which Pi must be transmitted and received. We define the deadline duration for P„ D„ as the difference between A; and as shown in Figure 3.1. The packets that miss their deadline are considered to be useless and lost no matter whether they are received or not. It is assumed that the network considered consists of a physical ring which is "divided" into k channels and is operated as k token rings, C\, C 2 , • , CfcThis way, the ratio of propagation delay and packet transmission time in each channel can be decreased so that a high utilization can be achieved. Then each two rings C, and C.+i are grouped as a pair < Cj,C,+i >, where t is an odd number. To allocate packets to the two rings of a pair, we classify the packets into two groups according

PAGE 37

31 Arrival Time Current Time Expire Time T Packet Slack Time Td«* Packet Deadline Duration (3-1) Figure 3.1: Time Attributes of Packet to their deadline values and a predefined threshold DL, tight deadline packet if D% < DL loose deadline packet if A > DL In the next subsection, a dynamic load control protocol will be presented which aims to utilize this multiple token ring network with the consideration of timing constraints. The most important concern is the opportunity provided in this architecture such that the input load can be allocated dynamically into different rings according to different system requirements and the potential performance can be achieved in real-time applications. To design the network controller in high-speed networks, the effort is to have a simple interface design. After the division of physical ring into channels, medium or low speed interface can be incorporated int o each channel. Control mechanisms then can be associated with each ring and arriving packets can be routed with simple scheme. Thus, we eliminate the consideration of an integrated network controller for multiple rings in each station.

PAGE 38

32 ARCHITECTURE MODEL Figure 3.2: Architecture Model CHANNELS 3.1.3 WUT Control Strategy and Mechanisms Based on the architecture given before, a dynamic control protocol is proposed to make token ring networks achieve a high performance in distributed real-time applications. The proposed control protocol is based upon the following idea. The critical packets with tight time constraints need not necessarily request for token reservation immediately after they arrive in to the system. They can still wait for the chance of transmission like all the other packets until their slack time passes certain limit, that they are going to fail to meet their deadlines. These special packets are called alert packets. The alert packets have the privilege to do reservation at once to preempt the normal packet transmission until all alert packets have been sent out. Since each token reservation wastes certain rotation time, the preemption due to alert packets should be minimized. This can be attained through a proper

PAGE 39

33 allocation. The protocol consists of four parts as follows: A. Allocation of Incoming Packets When packets arrive at a station, they are routed to consecutive pairs of the network in a round robin manner. In each pair, < <7,,C, + i >, all the tight deadline packets are routed to C, and all the loose deadline packets are routed to C <+1 . So now we have a multiple channel token ring network with k/2 subsystems of the same property. B. Wait-Until (WUT) Control Policy and Switching Within each pair of rings, < C,,C t+1 >, the wait-until (WUT) control policy is adopted to switch an alert packet from C, to Then, the alert packets will preempt the loose deadline packets in C i+i so that the on-time delivery can be guaranteed. The first ring of a pair C, is operated in non-priority mode and allows no token reservation. When the free token arrives, it send out a tight deadline packet if its waiting queue is not empty, otherwise it passes the free token to the next station. At the same time, it checks and switches those alert packets, that are going to fail to meet their deadlines, to the next channel C,+i which is allocated to those loose deadline packets. The checking process is done based on the following formula: T 3 l ac k < T, u ,itch (3.2) where T,i ac k is the packet slack time. The value T sw i tc h classifies alert packets and functions as a maximum sliding window to control the flow of switching as shown in Figure 3.3.

PAGE 40

34 On the other hand, the second ring of a pair, C, + i, works in a reservation mode and an alert packet switched from C{ can reserve the token and preempt the normal token rotation. This will guarantee that all alert packets can be sent out immediately before the loose deadline packets are delivered. The benefit of this method is that it guarantees real-time requirements while still keeping relatively high channel utilization. By using preemption, the real-time requirements can be met. The channel utilization is still kept high because the transmission of alert packets has no interference with the tight deadline packets and it only affects those loose deadline packets. Hence only the channel C, + i within a subsystem < C,, Ci+i > is degraded by reservation and the lost percentage of loose deadline packets may not be increased due to the nature of their time constraints. C. Migration of Loose Deadline Packets With preemption and switching, the loads in two rings of a pair may be unbalanced. This is due to the fact the channel assigned to loose time packet has the inclination to be overloaded while the channel assigned to tight deadline packet is possibly under utilized. This problem can be alleviated through a proper classification of tight and loose deadline packets by the threshold DL. However, this may still have a load fluctuation from time to time. A dynamic load-balance policy called migration is suggested. The key operations of this policy are: whenever a station at ring C, gets a free token, it sends out a tight deadline packet if it has a nonempty waiting queue. Instead of passing the token immediately if its waiting queue is empty, it will fetch a loose deadline packet from channel Ci+i and send it out in channel C%. This migration of loose deadline packet from C l+ i to C, should only be done when C,is under-utilized.

PAGE 41

35 Current Time Arrival Time Expire Time ? TiwMch Packet Slack Time Packet Deadline Duration Figure 3.3: Switching of Packets To identify the utilization of C,, a checking mechanism is embedded within each pair. The detection is based upon investigating the token cycling time (Tcyding) in each station. Tcyding < T Mme ( 3 3 ) where Tiding is the period between the time that a station releases the free token and the time it receives the free token again. Tcyding can be easily observed by C, when it receives the free token. The 1/A is the packet interarrival time at each channel which can be measured over a long interval, and A/„« is a constant that defines the migration window size. As it can be shown, when the above inequality is valid, C, has a low utilization and the average number of packets arrived during the last token rotation is less than M SiZ eSo, a migration of packets from C,+i to C, occurs whenever the latter one is under utilized and thus the load is dynamically balanced within such system. In addition, for each receipt of the free token in C,, a packet (either a tight deadline packet or a loose deadline packet migrated) can be transmitted. This is especially effective in high speed rings due to the high ratio of propagation delay and packet

PAGE 42

36 transmission time so that a high utilization can be achieved. D. Multiple Packet Transmission There are normally two different packet service policies in ring networks: exhaustive service and non-exhaustive service. In an exhaustive service system, whenever a station get a free token, it sends out all the packets in its waiting queue. In a nonexhaustive service, every station only sends out one packet whenever it gets a free token. The later scheme prevents any station from monopolizing the service and is useful for real-time applications. But in high speed network system, the packet transmission time is much smaller and hence exhaustive service is more suitable if the physical ring length is longer than the packet length. However, the token may be held by a busy station and the packets with short deadlines in other stations may not get the chance to be transmitted. Thus, transmission of multiple packets per token receipt can only be adopted under an underutilized load. In a paired ring < C.C+i >, both rings can be in the multiple packet mode. While ring C,+i is working in a reservation mode, some loose deadline packets can possibly follow the switched alert packet (from channel C.) to make full use of channel bandwidth. While channel C, is working in a migration mode, more than one loose time packets can possibly be fetched from channel Cj+i so that the channel bandwidth is not wasted. However, the maximum number of multiple packets N max has to be limited, it is defined as the following: if Dmin — Tcyding Where D mfn is the minimum deadline value, and T pac k ei is the average packet transmission time on a ring. Based on the detection of the last token cycle delay on the

PAGE 43

37 ring, N max is determined by means of the packet transmission delay to maximize the exiting channel utilization. The operation of the proposed protocol is illustrated in Figure 3.2. In summary, the multiple ring network system works in the following way: all the packets arriving to a station are distributed into paired rings by a round-robin manner. Within each pair, there are two separate waiting queues attached on two rings, and tight deadline packets and loose deadline packets are queued and transmitted in their own channel respectively. A dynamic control protocol allows the switching of alert packets to preempt the loose deadline packets and at the same time allows the migration of loose deadline packet whenever the tight deadline packet channel is under utilized. In this way, every paired channel and thus the whole system is guaranteed to be working in a balanced state and to not only reduce the lost percentage of message packets but also increase the system utilization. The dynamic controlling policies like switching and migration in both directions are the key points that make such a kind of system achieve a remarkable performance improvement. 3.1.4 Performance Evaluation Though some analytical models and comparisons of throughput-delay characteristics of ring networks have been given, they are usually only applied to simple cases. It is widely believed that studies of performance behavior using complex control policies are beyond analytical methods and therefore extensive simulations are required. A. System Assumptions and Performance Measures The simulation model is set up based upon the following assumptions:

PAGE 44

38 • Arrival packets should be sent out and/or received within a certain period of time (deadline), otherwise they will be meaningless no matter they can be received or not. Those packets who have passed their deadlines will be viewed as lost-packets and will be discarded. • All packets are classified into different classes based on their deadline durations. The deadline durations are uniformly distributed over a constant range. Thus, the loads incurred by tight deadline packets and loose deadline packets in a paired ring depend upon the value of the threshold DL. • All packets arrival into each station in the network follow a Poission process with a constant arrival rate. • All stations are equally distanced on a ring; • The lengths of all the packets are exponentially distributed with a constant mean. Various system parameters are denned as in the following table: SYSTEM PARAMETERS IN EXPERIMENT network length 20 km transmission speed 200,000 km/s bitrate / channel 10 Mbps delay per station 3 bits overhead per packet 97 bits station number 30 packet length (mean value) 512 bits Table 3.1. System Parameters

PAGE 45

39 With different arrival rate to each channel A, we vary the network utilization from 60% to 80% under the conventional token ring protocol. The packet deadline durations are distributed uniformly over the range of 5T ro
PAGE 46

40 This approach is to consider the different priorities between all incoming packets. There are various policies to decide which packet in a channel should be the next one to send out. To meet the real-time requirements, it is quite natural to set a high priority to those packets with tight deadline constraints. So, choose those "urgent" packets and assign high priority to them based on their slack time so that they can be send out as early as possible. As in FCFS, all channels are treated as the same and independently. To allow tight deadline packet to use token reservation, two queues for tight and loose deadline packets respectively are associated with each channel. Packets arrived will be inserted into a queue after comparing its deadline duration with the slack time of existing packets. Tight deadline packets have a high priority and can reserve the channel by using reservation bit of the token. The channel can be reset to allow loose deadline packet's transmission until all tight deadline packets have been processed (sent out or thrown away) (3) . Wait-Until Protocol (WUT) This one is called Wait-UnTil. Different from former ones, WUT groups each two channels as a structured pair < Ci,C l+i >. The initial packet allocation, the wait-until policy and switching described in the previous section are applied. (4) . WUT-MIG In addition to WUT, WUT-MIG (Wait-Util with MIGration) allow loose deadline packets to migrate from their channel C l+ i to channel C, whenever the latter is under utilized. In WUT-MIG, A/„ ze , which functions as a flow control window, is set up to limit the migration load.

PAGE 47

41 o.o 30.0 20.0 1 0.0 -m FCFS PR — WUT — ». WOT — MIG — MWUT -» MWUT— MIG 700.0 Figure 3.4: Lost Percentage for Tight Deadline Packets (5) . MWUT Unlike WUT protocol, MWUT (Multiple packet WUT) adopts Single-TokenMultiple-Packet policy. Based on WUT, MWUT allows more than one packets to be sent out whenever a free token arrives. The maximum number of packets per transmission N max , is set dynamically and is dependent of the current network load. (6) . MWUTMIG Based on MWUT protocol, the MWUT-MIG protocol allows migration of loose deadline packets from the second channel to the first channel in a pair so that the loads can be dynamically balanced and to have a full utilization of all the resources. C. Simulation Results and Evaluation

PAGE 48

42 so.o o ARRIVAL RATE Figure 3.5: Lost Percentage for Loose Deadline Packets Extensive performance simulations are developed based on those protocols discussed above. Figure 3.4 and 3.5 show the lost percentages of tight and loose deadline packets with different packet arrival rates. In Figure 3.6 and 3.7, the channel utilizations are illustrated. The classification threshold DL is set to equally partition packets into two classes. The results show that the lost percentage of tight deadline packets in FCFS protocol is very high while that in priority-based protocol is relatively low. This is because the former one has no mechanism to favor real-time requirements. However, since the PR protocol uses reservation extensively for tight deadline packets, a lot of channel bandwidth is wasted by favoring the most urgent packets. This leads to a much high lost percentage of loose deadline packets. WUT protocol aims to overcome this weak point of priority-based protocol and has decreased the total lost percentage significantly. With T sw , tc h = 2T rotote , the simulation results show a much robust utilization. However, since only t ight deadline packets can be switched to the second channel, the loads at paired rings are not balanced.

PAGE 49

43 30.0 1 1 ' 1 1 400.0 500.0 600.0 700.0 ARRIVAL RATE Figure 3.7: Utilization of Channel C,+i in (C,, C, + i)

PAGE 50

44 With the addition of migration, we can make a dynamic load balance among the paired rings. The lost percentage of loose deadline packet is reduced significantly with a slight impact on tight deadline packets. The curves of MWUT in Figure 3.4 and 3.5, represent the cases where migration window M siie = 0.2. Further performance improvements can be obtained by adopting multiple packet transmission policy. Note that this multiple packet transmission is dynamically controlled through token cycling time and the limit N max . Table 3.2, 3.3 and 3.4 show the impacts of different Twitch, M tixe and DL. The performance of the network is relative insensitive to the choice of the threshold DL once T sw itch and Af„« are set properly. Since T tw itch and Af„« control the migration of packets among two channels of a pair dynamically according the network load. The initial classification become less crucial. Note that T^tch and A/„je have different purposes: T lw itch aims to switch tight deadline packets such that they can preempt the normal transmission of loose deadline packets and meet with their real-time requirements, whereas M,{ ze intend to use the underutilized resources. Lost percentage (tight-loose) v.s. T^teh control protocols 1.5T rotate 2T' r0 j a t e 2.5T rotate WUT-MIG 6.8-32.0 3.5-27.5 2.2-26.8 1.8-28.3 MWUT-MIG 2.9-12.6 1.9-10.2 1.2-7.8 0.9-8.5 Table 3.2. The Lost Percentages with different Tnnteh

PAGE 51

45 Lost percentage (tight-loose) v.s. M„, e control protocols 0.2 0.4 0.6 0.8 WUT-MIG 3.5-27.5 8.0-18.7 11.8-18.3 12.7-17.0 MWUT-MIG 1.9-10.2 4.6-5.8 6.5-4.4 6.6-3.7 Table 3.3. The Lost Percentages with different Af„« Lost Percentage (tight-loose) v.s. Threshold (DL) control protocols lZT TO tate 14T TO tate 15T ro < 0 t e 16T TO tate 177V 0 t a fe lST rot ate WUT 1.2-33.4 1.8-32.5 2.4-32.5 3.6-34.4 5.2-34.5 6.4-34.5 WUT-MIG 7.1-18.8 7.4-19.5 8.6-22.1 8.7-24.7 9.4-24.4 10.3-26.2 MWUT 0.9-15.4 1.2-13.2 1.3-11.4 2.0-11.9 2.7-13.7 3.1-13.8 MWUT-MIG 2.0-7.6 1.8-8.3 2.0-8.9 2.4-9.9 2.4-10.4 3.3-11.2 Table 3.4. The Lost Percentages with different threshold DL The following conclusions are drawn from the performance of the protocols studied: • As expected, FCFS performs badly concerning about the real-time requirements. FCFS has the best performance in channel utilization, but the performance of this protocol is not acceptable. The studies of this protocol, thus, demonstrate the importance of prioritization. • The performance of priority-based protocol (PR) gives a better performance in trying to meet the time-constraints of tight deadline messages. However, as a total lost percentage, it is still unacceptable. The main problem of this

PAGE 52

kind of protocols is the low utilization due to the reservation schemes. This performance is even worse in the high-speed communication systems. • As may be noted, the protocols based on wait-until scheme (WUT) can achieve a better performance comparing to the PR protocols. It can increase the system utilization significantly while still guarantee the time-constrained requirements. That is because WUT policy allows all the packets to be sent out immediately if they are not time-out, and dynamically preempt the loose deadline packets only when necessary to guarantee the real-time requirements. • The combined dynamic protocol employing multiple-packet policy (MWUTMIG) performs the best on all accounts. The main reason for the better performance is that it dynamically balances the traffic load among available channel resources and cleverly schedules the transmission order. Therefore, it minimizes the lost percentages of all kinds of messages and still maximizes the system utilization. From the studies, it can be seen that a network control protocol plays a very important role in system design and system performance. In addition to its traditional role as an arbiter of channel accessing and sharing, a control protocol also serves as a distributed scheduling mechanism by imposing an implicit or explicit transmission order. This scheduling function can critically affect the distribution of packet transmission delays and thus the real-time performance of the protocol.

PAGE 53

47 3.2 Distributed Queue Dual Bus (DQDB) 3.2.1 Background The Distributed Queue Dual Bus (DQDB) is a recently developed LAN/MAN networking architecture for very high-speed and high-quality transmission services. It is defined by IEEE 802.6 working group with a high compatibility with the switching concepts used by promising B-ISDN standard. However, the standard DQDB protocol is not suitable for real-time communications. Studies have shown [95] that DQDB architecture inherits the serious problem of service unfairness. The position of the transmission station has a great effect upon the system performance. Since transmission delay is critical to real-time communications and the urgent packet in downstream stations can not afford to bear long queueing delay, this unfairness of DQDB networks is obviously undesirable for real-time multiclass message transmission. DQDB standard also provides a multiple priority mechanism with up to four level priorities in its distributed queue processing. It is hoped from theory that the multiple priority mechanism can improve the system performance by giving proper preference to certain urgent message class. Like most of the other systems, however, it is anxious to know how to use this distributed mechanism effectively in real networking practice. The media access control (MAC) protocol of the DQDB standard is based on its unique physical architecture, which consists of two unidirectional buses and a multiplicity of stations along the buses. These two buses, supporting communication in opposite directions, allow full duplex communication between any pair of stations within the network. All the networking stations cooperate with each other in sharing the network bandwidth via a global distributed queue mechanism [60, 63]. Each

PAGE 54

48 station maintains several counters to reflect the network states. A Request-Bit of request field on reverse bus is used to indicate that a segment of downstream station has been queued for access the bus. The request (RQ) counter in every station is incremented by one every time a nonzero Request-Bit is detected. This way, every station keeps a current state record of the number of segments waiting for access downstream on the bus. Stations with no segments queued to a specific bus decrement their RQ counter by one for each empty slot passing on that bus. Any station wishing to send a segment on a specific bus writes a request into the next free Request-Bit on the reverse bus. At the same time, it loads the current value of its RQ counter into a counterdown (CD) counter. This counter indicates the number of requests for access to this bus which have to be satisfied before the segment at the station may be sent, and is also decremented by one for each passing empty slot. When CD counter reaches zero, the station writes the segment into next empty slot. It should be noted that the operations of writing requests and sending segments are independent. The DQDB standard also provides a global multiple priority mechanism which allows up to four different priorities for segment level transmission. The priority mechanism is absolute in that segments with a higher priority will always gain access ahead of segments at all lower levels. This is achieved by having dedicated counters to each priority and operating separate distributed queues for each level of priority. The DQDB standard can be simplified as a model in Fig 3.8. Within each station, an arriving message from upper user layer is first processed by Local Allocation Mechanism (LAM), which divides each message into several segments and allocates each segment into one of four local queues according to its priority and is assigned statically. The priority is usually a function of message deadline duration. The

PAGE 55

49 smaller the deadline duration, the higher the priority. The local queues employ conventional First-Come-First-Served (FCFS) queueing discipline. There is a media access Bloking (B) mechanism in each station to control local load to flow into the distributed queue. It is the DQDB standard that requires each station to defer next segment request until the current one has been served. Therefore, only one representative segment from each station is allowed to join the distributed queue for any priority level. It is obvious that the multiple priority scheme is expected to improves the performance by giving preference to higher priority messages. However, the local queues are of static type and can not reflect the dynamic nature of operating network. The relative criticality of message as time goes by is not considered. A long-waiting lower priority message may become critical because the deadline duration may be expired soon. Therefore one important issue is to develop a control policy which should be capable of adjusting the local queues according to the message waiting time and the up-to-date networking operating status. It is also worth noting that even a good control policy is included, the current multiple priority mechanism is still not enough to provide guaranteed real-time performance. This is because there is no effective Priority Assignment Mechanism (PAM) enforced by the DQDB communication system. A proper priority assignment scheme is critically important in those distributed networks to ensure the required services. This is mainly due to the fact that the MAC protocol of DQDB networks does not provide individual stations a global view of priority operations. As it is indicated in Fig 3.8, the message in each station is directly passed from local queue to the distributed queue whenever it is permitted by the media access blocking mechanism. That is to say, each station actually decides the global pri-

PAGE 56

50 Dutributcd Queut LAM: Local Allocation Mechanism B Acceu Blockill( Gau S: Distributed Queue Server Figure 3.8: DQDB Control Model ority of its messages according to its local users' requirements without concerning the relative criticalness with that of other stations. This certainly results in an inconsistent priority mapping in the distributed queue and may seriously affect the system performance. The analysis of DQDB protocol has shown that the relative traffic intensity of different priority distributions has great effect on the global system performance[70]. The overloaded high priority traffic will block and cause a significant transmission delay to lower priority classes. Since this undesirable situation can be artificially introduced by improper mapping from the local priority (deadline) requirement to the global priority mechanism, an important issue is to develop a scheme which is able to enforce all the communication stations to follow the same priority assignment criteria even though they are only cooperated with each other via a distributed queue.

PAGE 57

51 3.2.2 Dynamic Control Model and Strategies To overcome the problems the standard DQDB protocol has for the real-time communications, a dynamic control model has been proposed as in Fig 3.9. In addition to the original LAM module, a new Priority Assignment Mechanism (PAM) is included in our model. Some different control strategies are employed in these two modules to provide much better performance for real-time communications. As that in DQDB model, the LAM is responsible for message segmentation and local queue allocation. However, the LAM in our model will not use the conventional FCFS queueing discipline, this is because that the more important performance objective now is to guarantee the timing requirement of transmitting messages. In real-time communication networks, the time criticalness not only comes from the original semantic attribute of the message but also is caused by the limitation of the various network resources. As time goes by, the slack time ST(m) for one segment of message m changes dynamically according to the follow: ST(m) = AR(m) + DL(m) t c (3.5) where AR(m) is the arrival time for message m, DL(m) is the predefined deadline duration constraint for message m, and t c stands for the current time point. As we are more concerned about transmission delay at message level, we treat each segment from one specific message to be of the same type. It is noted that a message with larger deadline duration could still have a smaller slack time ST(m) than that of another message with a smaller deadline duration. As time constraints are more concerned in real-time communication systems, the Local Allocation Mechanism (LAM) in our model employs the following allocation scheme: all the segments are queued with regards to their dynamic slack time ST(m) instead of static deadline

PAGE 58

52 Distributed Queue LAM: Local AUocstioa Meduuutm PAM: Priority Assignment Mechanism B: Access Blocking S: Distributed Queue Server Figure 3.9: Dynamic Priority Control Model duration DL(m). There is only one local queue instead of the previous four local queues and this queue is updated whenever a new message arrives. The message priority is actually not considered at this processing stage. The global priority assignment is delayed until the local blocking (B) mechanism permits next segment to join the distributed queue. It is the additional Priority Assignment Mechanism (PAM) that dynamically decides which segment in local queue is the best candidate for the current allowable priority level to join the distributed queue. The decision should not only consider the local users' requirement, but also keep the DQDB multiple priority mechanism fully utilized. A dynamic priority assignment scheme is desirable which should be able to map the local user requirements properly to the existing global multiple priority mechanism. To apply the dynamic assignment scheme within the whole distributed system, the PAM in each station should first convert the continuous timing requests

PAGE 59

53 to discrete ones so as to make use of the 4-level priority mechanism. Since it should be done in real time, the up-to-date network operating status could be very informative and helpful. It is observed that we can make a wise local decision if the latest operating information for each priority level is provided. Suppose we have estimated the delay bounds of segment transmissions at different global priorities, then we can decide the best candidate for a particular priority by considering all the segments whose slack times are within corresponding delay bounds and choosing the one with the smallest slack time. Let TB^ denote the specific delay bound for station t and priority level where the larger level number ; indicates higher priority. This measure gives the information that the next possible transmission for priority ;' segment at station t is about TB\ }) time units away from now. It should be noted that TB\ A) < TB\ 3) < TB\ 2) < TB\ X) and they are dynamically determined during network operations at each station. If PR[Si(m)] is the priority value to be assigned to a segment from message m at station t, it can be decided as follow: [ 1 if STi(m) > TB\ l) t x PR[Si(m)] = (3.6) [ j if TB l t j) < 5T,(m) < TB\ } x) It should be noted the candidate for priority j at station i is always the segment with locally minimum slack time among all segments which have slack time within the range [T B\ 3) ,T B\ 3 ~ 1) ). Actually this method provides an approximate global priority decision mechanism, which is based on message slack time, so that every user can follow the same criteria to decide segment priority and participate the competition of medium access. Besides the function of dynamic priority assignment, another effective strategy employed in PAM is the segment rejection. Since we are more interested in message

PAGE 60

54 transmission, the rejection of a segment of one specific message means that all the remaining untransmitted segments of that message should also be rejected. By rejecting those waiting segments, the lost percentage of other messages can be reduced. This is due to the fact that this kind of rejection can release the buffer blocking problem and making the best use of local resource. 3.2.3 Priority Assignment Algorithm In order to implement the above proposed method in practice, A simple and effective dynamic priority assignment algorithm to be employed by PAM has been developed as shown in Fig 3.9. The key point here is to properly determine the delay bounds TB^ for each priority level j(= 1,2,3,4) and operational station t*. These bounds have to be determined locally while the global network status should also be considered so that the best real-time performance can be achieved. Motivated by the standard nonpreemptive M/G/l priority queueing model, A method which is able to derive the approximate media access delay dynamically is presented in following. The mean waiting time of a priority j segment at station i can be expressed by where <7 f = Ylt=j Pi > which actually is the accumulated traffic intensity of all no lower than priority j segments. The WO, is the service residual time. It is noticed that Wf' does not depend on the segments from lower priority groups except for their contribution to the numerator WO, . Since only interested in the media access part, we can consider the partial distributed waiting queue in which segments from all downstream stations (including station t) are accumulated for their turns of

PAGE 61

55 transmission. There is a different partial distributed queue for each active station i along the network. For any station i, a standard nonpreemptive priority queueing discipline model can be employed to analysis this queue. The server here is the next available empty slot with fixed length, and the queueing objects are all the segments from downstream stations which have been permitted to join the distributed queue. Let W0,(t) denote the last two consective empty slot passing time, which can be observed by each active station t in real network operating time. The RQi(j) is the RQ counter value at station t, which is provided by the standard DQDB protocol. It is noted that the RQi{j) actually counts the number of requests whose priority is no less than level ; for all the downstream stations, and there are total + downstream stations from station i. Taking these as a means to measure the priorityj traffic for all downstream stations from station t, then the following approximations are derived. = — (3.9) WO, = W0,(t) (3.8) RQiU) (5-j)*(N-i + l) a 0+» = RQtij + l) (3.10) and finally the expected media access delays for any priority level j. tvo.(t) , „ [l-«(3.(»/(5-j)(N-i + l)J|l-flg.(i+l)/(4-;)(N-.+l)] if 7 < 4 wo.(t) (3.11) li-fl<3,0)/(5-»(^-'+>)J if j=4 Since all the values in right hand can be calculated at each station, the above approximate access delay is time-dependent and of the feedback nature so that it can reflect the dynamic operations of the network. Taking these delay bounds to be

PAGE 62

56 the TB\^ discussed in our dynamic control model, the following priority assignment algorithm can be formulated. 1 if STi(m) > W t (l) t PR[ Si (m)) = { ' ' (3-12) j if Wl j) < ST t (m) < Wj 3 1} Thus the segment priority can be dynamically assigned by service provider at the medium access request time rather than statically decided by application users at the message arrival time. In this way, the priority distribution are dynamically adjusted in accordance with the network operating status so that the optimal real-time performance for the whole system can be achieved. Inspecting this scheme carefully, it can be seen that the scheme actually approximates to the Smallest-Slack-TimeFirst (SSTF) scheduling scheme. It has been show [74] that SSTF minimizes the total lost percentage of all classes of transmissions and thus provides the optimal real-time communication service. The proposed timedependent priority assignment scheme can be regarded as a practical implementation of SSTF scheduling strategy in the DQDB networks. 3.2.4 Simulation and Evaluation Because of the dynamical nature of the real-time communication systems, the investigation and evaluation of the proposed protocol have been done in certain details through extensive simulations. There are three different MAC protocols to be investigated. The Ideal Static Priority (ISP) protocol is the ideal case where all active stations are able to follow exactly the same static mapping schemes. Though it is not possible in real network, this protocol is valuable for analysis and comparison. The second one is the Static Priority (SP) protocol, which is the real networking practice that employs a static mapping from the user's deadline to the segment's

PAGE 63

57 priority at each station independently. The SP protocol under consideration is one of the typical unbalanced cases due to the distributed environment. For the simplicity, this protocol has the structure that the most upstream half of all the stations assign their segments with lowest priority while the rest downstream stations assign their segments with highest priority. The Dynamic Priority (DP) protocol, however, employs the proposed dynamic priority assignment protocol. It provides a dynamic mapping from the user's deadline to the segment's priority, depending on the message waiting time and operating status of the network. All three protocols under considerations employs segment rejection scheme for effective message transmission. It is our purpose to investigate the different priority assignment algorithms and the effectiveness of our proposed dynamic control scheme. Since the two unidirectional buses are identical in a DQDB network, the study will concentrate on bus-A. The traffic load is uniformly distributed, that is to say, each station has the same amount of transmission requests. However, considering only bus-A, the arrival rate for each station will depend on its physical location since the station can access both buses in two different directions. The network parameters under the simulation experiment environment are set according to the following table:

PAGE 64

58 NETWORK PARAMETERS Bandwidth of each bus (B) = 150 Mbps Number of stations (N) = 20 Slot size = Segment length (S) = 750 bits Distance between adjacent stations (D) = 2 slots Local Buffer size (K) = 120 messages Segments per message = 2 Destinations are uniformly distributed Some other system parameters can be easily calculated. Slot Unit (A) = S/B (3.13) Bus Length (L) = (D + 1) x N D (3.14) The basic simulation load pattern is that there are four predefined classes of messages with different deadline durations. There is a deadline seed parameter dl, ee j, and the four deadline durations are generated as dl\ = l*dl Ke j,
PAGE 65

59 40 I ' ' I DP Protocol I i I ' I I I l ' l 1 ' Figure 3.10: Lost Percentage priority segment transmissions, let P(i)(%) be the lost percentage value in station t and P be the total message lost percentages of any protocol considering all the networking stations. P{i) = P = F[i) x 100% F(i) + T(i) E," , F(i) — % (3.15) (3.16) Fig 3.10 shows the lost percentage of messages for each individual station along the network. It can be seen the location position has a great effect on the service quality in DQDB networks. The downstream stations usually bear a higher lost percentage than the upstream stations. This is due to the special structure of the DQDB standard. However, the proposed protocol (DP) is able to significantly reduce this undesired property by using dynamic control strategies. While it shows the general lost percentage for all classes of transmitting messages, Fig 3.11 and Fig 3.12 show the lost percentages of messages for different priority classes under DP and ISP

PAGE 66

60 20 18 16 14 'l2 t 10 8 6 4 2 I Clara1 ~* Clara -2 Oora-3 4 Clora-4 10 11 12 13 14 15 16 17 18 19 Station Index Figure 3.11: Class Lost. Percentage Under DP 20 18 16 14 12 „ 10 a. 1 ' 6 4 2 0 I I • i I CkJM-1 -* Clara— 2 < CkJM-3 4 1Ckjra-4 Figure 3.12: Class Lost Percentage Under ISP

PAGE 67

61 protocols, the priority level here is the static one which is decided originally by the users under the ISP protocol. Lost Percentages P{%) A (103 ) deseed SP ISP DP 0 40 3^ 0 1.1362 0.8932 0.6897 35 0 2.3457 2.0356 1.1764 35 0 5 0928 4.2453 1.8293 0.55 35.0 10.5006 8.0267 3.7152 0.60 35.0 20.0528 15.8923 9.2287 0.50 31.0 8.1022 5.5552 2.0751 0.50 33.0 5.8399 4.6504 1.9803 0.50 35.0 5.0928 4.2453 1.8293 0.50 37.0 4.5301 3.8139 1.8162 Table 3.5. Total Lost Percentage Table 3.5 shows the total lost percentages of three investigated protocols under different traffic loads and timing requirements. As load increases, all three protocols tend to increase the lost percentage of message transmission. However, the proposed DP protocol has a much smaller lost percentage than the other two protocols. This good property can also be observed when the message deadlines change from loose to tight. In Fig 3.13 and Fig 3.14, the total lost percentages of all classes of messages are given under various arrival rate A and deadline duration seed dl 3ee( i. In general, the lost percentages of SP and ISP protocol are quite high for time-constrained message transmissions. This is because they never consider the dynamic nature of the local queue and the tight deadline messages tend to block the loose deadline

PAGE 68

62 messages no matter how long they have been in the waiting queue. Also, they bypass the user requirement directly to the multiple priority mechanism of DQDB without any control which in most cases will result in an improper mapping and spoil the system's global performance. On the other hand, the proposed DP protocol achieves much better system performance. This advanced MAC protocol is dynamic in the sense that it can adjust the load of different priority levels according to the status of network operations. It should be noted that under the DP protocol, the load of each priority level is able to change from time to time no matter what the original local user requirement is. Therefore the distributed queue processing is optimized to achieve expected good performance. In Fig 3.15, the average message transmission delay D met (i) for each station i is given. It is measured from time instant the first segment of a message attempts to access the medium to the time instant the last segment of that message successfully finishes its transmission. It is worth mention that this delay is an average measure which considers all four different priority levels at each station and only one unidirectional bus of the DQDB network is considered. The DP protocol has shown a significant smaller transmission delay than the standard DQDB static protocol. This is because the latter does not have the mechanism to provide message level service and all transmission is always undertaken at the segment level. The message transmission delay also reflects the chance of accessing the bus from station t and indicates the fairness property of the undertaking MAC protocol. Therefore it shows the DP protocol can provide a much better fairness performance than the conventional static DQDB protocol. The proposed dynamic priority assignment scheme achieves a high quality realtime performance by properly making use of the multiple priority mechanism in

PAGE 69

63 OP Protocol CP Protocol SP Protocol 55.0 60.0 Figure 3.13: Lost Percentage Over Arrival Rate DQDB standard. It emphases on dynamic local message scheduling at message level which is especially important in real-time communications. By forcing all the stations to follow the same priority assignment criteria, the total lost percentage of all kinds of messages due to the deadline constraints is minimized. The simulation results have validated the effectiveness of our control strategies and shown the proposed MAC protocol significantly outperforms conventional MAC protocols for real-time communications. It is believed that the future real-time protocols should be able to function as a wise load scheduler as well as the conventional media access arbitrator.

PAGE 70

64 20.0 16.0 16.0 14.0 0. §" 12.0 l £ 10.0 0.0 31 — a DP Protocol — » ISP Protocol — • SP Protocol 33.0 35.0 D«odlln« Duration S«»d 37.0 Figure 3.14: Lost Percentage Over Deadline 200 190 180 170 160 150 J 130 g 120 8 110 | 100 E I 90 60 70 60 50 0 30 20 10 0 1 ' I ' I ' I ' I — « DP Protocol ISP Protocol — SP Protocol _i_ _i_ 8 9 10 11 12 13 14 15 16 17 16 19 Station lnd« I Figure 3.15: Message Transmission Delay

PAGE 71

Chapter 4 MESSAGE LEVEL PROCESSING SCHEMES 4.1 Introduction In real-time communication systems, the message level delay and processing is much more meanful and important to the application users. The flaw in conventional segment or slot level processing schemes is that it is generally an intra-layer concept rather than a global control mechanism. This is because all the messages have to be divided into segments to suit the fixed size formats when they go through the networking media. While it is unclear how effective such schemes in a lower layer can be mapped from upper layer services, they are usually undesirable from a user's point of view. The real-time application users are only interested in the performance at the complete message level, and the performance measure at network-specific segment level really does not make any sense to them. The emphasis of networking operations must be on the message level instead of viewing them independently at 65

PAGE 72

66 the segment level. Some benefits can be obtained including more effective transmission and better real-time performance. For instance, when certain portion of a message fail to meet its deadline, then the rest of this message should be aborted, since there is no way to satisfy the requirement of this message. We intend to investigate the proper processing level for control schemes in realtime communication networks. It is believed that the transmission delay should be considered at the message level, which is obviously not optimized at the current segment level processing schemes. The standard DQDB protocol is one of the segment level processing schemes, in which each request is asking for one empty slot to serve the current waiting segment and the next segment request will not be issued until the current segment has been transmitted. Therefore all the messages should be first divided into segments to suit the fixed size slot format. It has been found that this scheme is very undesirable for real-time communication service since segment level processing is not efficient and the segment delay does not make much sense to the application users. More importantly, this mechanism does not necessarily result in good message delay performance. 4.2 Message Level Processing A processing model is presented whose scheduling and queueing algorithms are build upon the individual message timing requirement instead of the complex mapping mechanism. A reservation control strategy has been proposed. Whenever an incoming message arrives, the control scheme has to consider the media access duration at the message level although the real transmission is performed at the segment format. The access request must be notified to other stations with the information

PAGE 73

67 at the message level such as the number of segments this message has. Under this strategy, it is the messages not the segments are interacted with each other and compete the media access control. The proper order of transmission is decided by the message criticality. It can be observed that the tight deadline messages are unlikely to be blocked by loose deadline messages under the proposed strategy, and there is no partial transmission success at the message level. Thus, this reservation control strategy is able to provide much better real-time performance as well as the processing efficiency. The standard DQDB protocol is employed as an example to elaborate the proposed strategy in detail. Fig 4.1 gives the DQDB network architecture and its slot format[41]. A slot is the basic unit for data transfer. The access control field contains the bits that control access to slots. The BUSY bit indicates whether the slot contains information or is available. The SL.TYPE bit indicates the type of slot and the PSR bit indicates whether the segment in the previous slot may be cleared or not. Finally, the REQUEST field consists of four bits, usually used in the operation of the multiple priority distributed queues. Messages at each station have to be divided into segments to fit in the fixed length slots. The access to slots is controlled by a distributed queue which allows the formation and operation of a queue of asynchronous segments generated across the network. Each station maintains certain dedicated counters for each bus so that the distributed queue can perform as a centralized queue. As discussed in Section 3.2, the DQDB protocol is based on segment level processing. Each request asks for one empty slot to serve the current waiting segment and the next segment request will not be issued until the current segment has been transmitted. The message level service is never considered. Therefore, the DQDB protocol has been found being not so efficient and

PAGE 74

68 000 o o o o © 0 Accui Control Flald BDST RSSBRVSD PSR REQUEST (1 bit) (1 bit) (1 bit) (1 bit) (4 bit) Figure 4.1: DQDB Network and Slot Format effective when it is employed by real-time communication users. To overcome these shortages, a class of new control protocols is proposed based upon multiple-slot reservation scheme. These new protocols are carefully designed to not only provide a better real-time performance in an efficient way but also be highly compatible with the standard DQDB protocol. The main idea of our proposed protocols is to do the service request at the message level instead of at the original segment level. That is to say, all the corresponding counters now operate in multiple-value mode rather than the simple increment-by-one and decrementby-one mode of the original DQDB standard. In this way, the reservation is done more infrequently and effectively. Rather than reserving empty slots one by one, the proposed protocol try to reserve all required empty slots for the current message all at once by making use of multiple value request in each slot.

PAGE 75

69 4.3 Analysis Model and Evaluations Due to the fact that the two DQDB buses are identical in nature, our model and corresponding analysis are based on one single bus, say bus-A. 4.3.1 Service Time Distribution Since we are only interested in the media access part, a special partial distributed waiting queue is considered at each station i. This is a queue in which arrival segments come from all downstream stations (including station i) and are accumulated for their turns of transmission. There is a different partial distributed queue for each active station i along the network. For any station t, a standard preemptive resume (PR) and feedback M/G/l queueing discipline model can be employed to analysis this queue. The server here is the next available empty slot with fixed length T s i ot , and the queueing objects are all the segments from downstream stations which have been permitted to join the distributed queue. N-l N A. i = E £ ^ (4-1) 5=1 d=s + l r. = £ £ x ><* (4-2) #=1 d=3+\ where A, is actually the accumulated traffic intensity of all downstream stations and A,is the accumulated traffic intensity of all upstream stations. For this special queue in each station i, the service time is actually the empty slot intervals. We can describe approximately the interval between empty slots seen from station i with the following geometric distribution: Pr{z = kslots] =rfl {l Pi ) * = 1,2,... (4.3)

PAGE 76

70 where p,= Ti*T s i ot . This distribution is known as geometric distribution at one and the corresponding moments of service time x can be easily obtained. One important property of this model can be observed and is worth mention. The mean service time increases and the arrival traffic intensity decreases when the number i goes from upstream stations to downstream stations. xl = y^— (4.4) 1 pi x? = l+Pi (4.5) pi = A, * x~i (4.6) 4.3.2 Waiting and Delay Time Processing-Sharing (PS) Discipline This is a scheduling discipline employed by DQDB media access control protocol in the sense of ideal analytical case. Newly arriving customers join the single waiting queue, work their way up to the head of this queue in a first-come-first-serve fashion, and then finally receive a quantum of service. When that quantum expires and if they need more service, they then return to the tail of that same queue and repeat the cycle. It is clear in this system that a customer is required to make an infinite number of cycles each infinitely quickly and each time receiving infinitesimal service, until finally his attained service equals his required service, at which time he departs. Based on the results from queueing theory and previous equations, we finally obtain the following message average waiting time W, and transmission delay time Ti for this processing-sharing (PS) model. WPS) = j£« (4.7) Ti(PS) = Wi(PS) + x(4.8)

PAGE 77

71 First-Come-First-Serve (FCFS) Discipline Under this scheduling discipline, the server always selects for service that customer at the head of the waiting queue and offers a quantum od service to this customer. The difference, however, is the place the customer returns to this same queue waiting for another service quantum. The First-Come-First-Serve discipline feeds those customers that ejected from service due to the termination of their quanta directly back to the head of the single queue and thereby immediately takes them back into service until fulfill their required service. Similarly, we can get the following message average waiting time W% and transmission delay time T, for this first-come-first-serve (FCFS) model. Wi{FCFS) = (4.9) 2 * (1 pi) Ti(FCFS) = W t {FCFS) + xl (4.10) 4.3.3 Numerical Results The analytical results are shown together with the corresponding simulation results as in Fig 4.2, While the PS. ANA and PS. SIM represent the analytical and simulation results for the processing-sharing model, the FCFS. ANA and FCFS. SIM denote the corresponding results for first-come-first-serve model. In this experiment, each message has the fixed length of two segments. It can be noticed that the analytical curves always have a large delay comparing with the simulation curves. This is because the networking propagation delay is never considered in our analytical model. In fact, the analytical models here overestimate the real traffic load and thus enlarge the transmission delay. The Processing Sharing (PS) model has a flat curve for its analytical result and

PAGE 78

72 30 T" T PSJLNA rdrsjLNA ps.snt rdrs.siM — i— — i— • b 10 11 iz is 14 is it tr ia io Station tndam Figure 4.2: Message Transmission Delay shows a perfect fairness among all stations. However, the propagation delay is an essential part of any high speed network in reality. That is why the DQDB network suffers a serious unfairness as shown in the simulation experiments. While it is widely believed that the exact analysis of DQDB network is extremely difficult or intractable, the above approximation analysis do give us insights of some important features. More important, the analysis shows the better performance and promising of the proposed multiple-slot scheme, which is based on the First-Come-First-Serve (FCFS) model at the message level. Based on these observations, we will further investigate the multiple-slot schemes in more general cases where the messages have not to be restricted as fixed length. Since exact analysis may not be applied, simulation experiments are used in later sections to verify and demonstrate the better performance provided by the multiple-slot schemes.

PAGE 79

73 4.4 Transmission Control Schemes 4.4.1 Multiple-Slot Reservation Protocol Following the previous discussion and the message level processing model, further investigation of the proper control schemes in real-time networks is studied. In this section, a simple message level processing scheme is presented based on the multipleslot reservation concept. In order to be highly compatible with DQDB standard, a 4-bit Request field of the access control field (ACF) is used to represent different message length requests. It should be noted that it does not necessarily have to use these particular four bits to implement our strategy and algorithms. In general, one additional access control field can be added into the slot format dedicating to the specification of the message length and it will not affect our later discussion of algorithms and results. However, our intend in this research is to investigate the effectiveness of multiple-slot control schemes and to avoid complicated format discussion hampering the proposed ideas. Therefore we restrict the message length with a maximum value of 15 segments where value 0 means no request. By doing like this, no change is made at all of the DQDB standard slot format which is a very desirable design concern. One way for a detailed description of a protocol is its state transition diagram (STD). It should be noted that the proposed control strategy does not conflict with the priority mechanism in principle although the priority field is used as one way of implementation. As long as additional control bits are provided, both priority mechanism and multiple-slot reservation scheme can be accommodated. To be compatible with the standard DQDB standard, the priority attribute in the corresponding STDs is still included. Fig 4.3 and Fig 4.4 show the detailed STD of

PAGE 80

74 IDLE IN REQJ Sc J : RQ I RQJ + 1 IN SEL-REQJ & J > I RQJ RQ_I + 1 IN EMPTY -Q A If RQ_I > O then RQ_I RQ_I 1 COUNTDOWN IN REQ_J * J > I QA-DATA request CDJ RQ_I RQ I 0 IN-EMPTY-QA Sc CDJ 0 Transmit Next Segment CDJ CDJ + 1 IN REQJ * J I RQ_I RQ_I + 1 IN EMPTY -OA * CD > 0 CDJ CDJ 1 IN SEL-REQJ Sc J > I CDJ CDJ + 1 Figure 4.3: Standard DQDB Protocol standard DQDB protocol and proposed protocol respectively. As in DQDB standard, the REQ_J and SEL-REQ.J denote the priorityJ requests issued from downstream stations and local station respectively. The RQJ and CDJ represent the request counter and countdown counter with priority /. While EMPTY.QA is the empty slot passing by, the local data waiting for transmission is QA-DATA with length LEN(QA-DATA). It can be observed that two new variables are introduced. With LEN.J represents the message length of the corresponding requests REQ J and SEL-REQ.J, another variable W-NO denotes the remaining segment numbers of a message waiting for transmission which is used to guarantee the continuous transmitting of a message. Different from that in DQDB, all the counters now are operating according to the specific message length variable LENJf (J=0, 1,2,3). Also, the COUNTDOWN state will not be switched back to IDLE state until all the segments of a message have been transmitted, which actually accomplishes a

PAGE 81

75 IDLE IN = REQ J & J x= I RQJ = RQJ + LENJ IN = SEL-REQJ A J > I RQJ = RQJ + LENJ IN = EMPTY-QA if RQJ>0 then RQJ = RQJ 1 COUNTDOWN IN = REQJ * J>I QA-DATA request CD J = RQJ RQJ = 0 W-NO = LEN(QA-DATA) IN = EMPTY-QA & CD J = 04 W-NO 1 Tranimil Next Segment CD J = CD J + LENJ IN = REQJ A J = I RQJ = RQJ + LENJ IN = EMPTY-QA & CD > 0 CD J = CD J -1 IN « S ELREQJ A J > I CD J = CD J + LENJ IN = EMPTY-QA & CD J = 0 & W-NO > 1 Transmit Next Segment W-NO = W-NO 1 Figure 4.4: Multiple-Slot Protocol message level transmission. It can be observed that in the distributed queue, the proposed protocol employs the scheme of the First-Come-First-Served (FCFS) model considering the message as a processing unit, the conventional DQDB protocol is, however, a segment interrelated system which can be considered as the Processing Sharing (PS) model at the message level. From the results of queueing theory, it can be expected that the conventional DQDB protocol always bears longer mean message delay comparing to the proposed protocol under the situation that the variation of message length distribution is not so large. The operations of the proposed protocol are very simple and highly compatible with those of the original DQDB protocol. It is also worth mention that the proposed protocol alleviates the request-bus traffic by efficiently employing multiple-slot reservation scheme and thus improves system bandwidth utilization.

PAGE 82

76 4.4.2 Enhanced Multiple-Slot Protocol In practical distributed system, one station may have some very long messages for transmission and thus cause unfair access by overusing the available bandwidth. More specifically, this undesirable situation will happen when the variation of message length distribution becomes greater. Motivated by the queueing analysis, an enhanced MAC protocol is further presented based upon the Shortest-Message-First control scheme, which is a direct variation of the ShortestJob-First (SJF) scheme. Under this queueing discipline, both short and long messages can be properly scheduled to achieve a much better overall system performance. To accomplish this, some additional information is needed whenever a transmission decision is made locally. In each station, additional records of message length can be obtained for every active downstream message (which has entered distributed queue) through slot transmission on reverse bus. For station i, it is observed that there are at most 15 different message length types since 4-bit Request field is used to specify the message length. Therefore, any station t only needs 15 length counters to store the number of message requests with different length. Let NLi(k),NL,2(k),...,NLis(k) denote all the length counters in station k. It should be noted that NLi(k) = m means, viewing from station k, there are m requests of / segments from downstream stations have already entered the distributed queue. Instead of using CD (counterdown) counter to decide the order for transmission which provides FCFS service, station k decides next candidate by considering its pending message and all the downstream active messages and picking the message with smallest length. The corresponding STD is described in detail in Fig 4.5. Our algorithm can be simply defined by the rule that, whenever a message finishes

PAGE 83

77 COUNTDOWN IN = REQJ * J >«= I QA-DATA request IN » REQ_J & J > = 1 RQJ = RQJ + LENJ RQJ = RQJ + Lcrt_f IN = EMPTY-QA & MIN* IN-SEL-REQJ*J>I W-NO LEN(QA-DATA) IN « EMPTY-QA & MIN* = t & W-NO = 1 RQJ = RQJ 1 RQJ « RQJ LENJ IN = SEL-REQJ * I > I RQJ = RQ_I + LENJ IN = EMPTY-QA IN = EMPTY-QA & MIN* IfRQJX) then RQJ = RQJ 1 Truicmit the Last Segment Transmit Next Segment W-NO W-NO 1 Figure 4.5: Enhanced Multiple-Slot Protocol transmission, the next candidate is the active message with smallest message length value. Let L{k) denotes the candidate message length at station k, which is actually the initial value of variable W-NO in the STDs. The MIN* operation in the state transition diagram can be specified in detail as follows. { t(rue) if L(k) < MIN M fl n{l\NL t (k) ^0} MIN* = I (4.11) I f(alse) otherwise The length counters NLi(k) are maintained by each station k. Upon receiving any request with length /, the station k increases the corresponding counter NLi(k) by one. When the station passes empty slots to downstream stations, it decreases the counter NLi(k), which has the smallest length value / and is greater than zero, by one if it passed / empty slots. It can be seen that this is actually an approximate Shortest-Message-First scheme undertaken for the distributed DQDB networks. From the result of queueing theory, this SJF-based protocol is expected

PAGE 84

78 to achieve near optimal performance for our distributed queue system. 4.5 Performance Evaluations The mathematical modeling and performance analysis of the DQDB network is known to be a very difficult problem. This is mainly because that high degree of interactions among a plethora of processes makes an exact analysis of the distributed network almost impossible. In order to verify our analysis and compare all these discussed protocols in detail, a simulation model is set up and extensive experiments are undertaken. There are three MAC protocols to be investigated. While Single Slot Processing Sharing (SS-PS) protocol is essentially the standard DQDB protocol, Multiple Slot First-Come-First-Served (MS-FCFS) protocol and Multiple Slot ShortestJobFirst (MS-SJF) protocol are the two multiple slot reservation protocols proposed in Section 3 and Section 4 respectively. Since the two unidirectional buses are identical in a DQDB network, the concentration will be on the study of busA. The traffic load is uniformly distributed, that is to say, each station has the same amount of transmission requests. However, considering only bus-A, the number of unit arrival rate for each station will depend on its physical location since the station can access both buses in two different directions. The major network parameters are set according to the following table.

PAGE 85

79 NETWORK PARAMETERS Bandwidth of each bus (B) = 150 Mbps Number of stations (N) = 20 Slot size = 750 bits Distance between two adjacent stations (D) = 2 slots Local Buffer size (K) = 120 messages Destinations are uniformly distributed Some other system parameters can be easily calculated. Slot unit time (A) = S/B = 5 (microseconds) (4-12) Bus length (L) = (D + 1) x N D = 25 (slots) (4.13) The major performance concern here is the message level transmission delay. Let D me3 (i) denotes the mean message transmission delay which is measured from the instant the first segment request is issued until the last segment of the message has been successfully transmitted. The intention is to investigate this measure obtained in each station on both buses and under various traffic patterns. The traffic load is uniformly distributed among all stations, and the message arrival is of the Poisson pattern. As discussed early in this chapter, the minor change in protocol only comes when the user message requests have a very large variance in length distribution so that some message lengthes will exceed 15. In that case, several reservation requests are sent out by this station with each one having the maximum lengthes except the last request. In Fig 4.6, the message length varies and follows a Poisson distribution with a mean length M = 3. In this case, the MS-FCFS has a quite close mean delay to that of SS-PS protocol although the former has a relative flat curve. This

PAGE 86

» too too tao 'tro h 180 ISO 14C ISO 120 no too \00 BO 70 eo 60 40 SO zo 10 o SS-PS -m Ms-rdrs — •• us sjr -1_l-i-1_l_ 9 10 11 IX 13 M IS 16 17 18 10 MO Station tnOmm Figure 4.6: Message Delay When Arrival Rate Per Station = 0.2 MOO i — i — i — i — i — i — i — — i — i — i — • — i — — i — i — i — i — i — • — i — — i — — i — — i — • — i — • — i — — i — • — i — — r 1 * a 4 s a r a o io ii it is 1* is ie 17 ia 19 to Station IniMm Figure 4.7: Message Delay When Arrival Rate Per Station = 0.25

PAGE 87

81 is due to the fact that the coefficient of variation of the message length is now equal to one. Fig 4.7 shows the same measure when the traffic load is increased. The multiple-slot reservation protocols have shown to be able to outperform standard DQDB protocol quite significantly as the system is in high utilization. It should be noted that although all these protocols are based upon certain queueing model, their operational characteristics do not exactly follow the theoretical results. This is due to the fact that there are so many factors, including distributed queueing operation and networking delay, influence DQDB behavior. However, these analytic queueing model can be used as a guider to design more effective media access control protocols. As expected, when the variation of message length distribution becomes quite large the simple multiple-slot reservation protocol based on FCFS scheme has shown to provide a quite large message delay, which may even worse than that of the original segment-based DQDB protocol as analyzed before. However, the MS-SJF achieves the overall best message transmission delay performance under all situations. The proposed control strategy also shows a very efficient bandwidth usage on reverse bus to send requests and thus achieves a much better system utilization. This is because the multiple value request operations enable the system to send fewer number of requests and shorten the request transmission delay. In real-time communication networks, the transmission delay at the message level is a much more important and proper performance measure. It is argued here that the MAC protocol in real-time MANs should be designed and evaluated from the application users' point of view (message level) instead of from the network designer' point of view (segment level) so that much meaningful and better service performance can be achieved. The proposed multiple-slot reservation protocols contribute to the DQDB standard with processing efficiency, better bandwidth usage,

PAGE 88

82 and improved message delay performance. Of all the discussed protocols, the SJFbased multiple slot protocol has shown to be the best in the sense that it minimizes the average message delay under all circumstances. Also, it is worth mention that the proposed protocols have shown a relative more flat delay curve than that of the single slot protocol (DQDB standard). This indicates that the multiple slot scheme is able to suppress the unfairness problem. In additional to these, the proposed schemes require no or minor changes to the slot format and are highly compatible with the DQDB standard protocol. As discussed early, the minor change in format only comes when the user message requests have a very large variance in length distribution. In that case, one additional field dedicating to the length specification can be compatibly added into the slot format and all the presented results and discussions remain the same in nature. The proposed strategy is surely a promising candidate of MAC protocols for the integrate-service applications in high speed real-time communication system.

PAGE 89

Chapter 5 INTEGRATED SCHEMES IN WIDEAREA NETWORKS 5.1 Introduction The wide area networking model is defined as an environment in which there are lots of processing hosts or local area networks connected by high speed broadband trucks though intermediate routing nodes. It should be noted that the underlying networking medium is usually of very high speed and possible large bandwidth. This kind of communication system is believed to be the backbone of the next generation distributed real-time processing systems. Different from local and metropolitan area networks [22], the user connection establishment and the effective network buffering have to be considered in this wide-area communication model. Also, flow control and congestion control [82] needs to be addressed so that more efficient and more predictable performance can be achieved. Most of the current communication systems are not real real-time oriented in 83

PAGE 90

84 Figure 5.1: Wide-area Network Model the sense that the timing constraint is not addressed directly and explicitly. One apparent substitution for real-time communication is high speed networking techniques. This is based on the wrong argument that if a communication system is so much faster than the most time-constrained messages ever needed, then no scheduling or queueing algorithm need to be used to ensure the time responsiveness. In fact, the real-time requirement not necessarily means the existence of the shortest deadline. Nevertheless the service quality control to each individual class must be provided. The real-time communication system must be able to ensure that the delivery of messages is accomplished with the quality specified, and it happens within the deadline imposed by the application. Also, due to the complex multi-hop structure of WAN [45], proper buffer management and switching control under the time constraints invokes even more challenges. A complete model of the wide area network can be shown in Fig 5.1. There are two levels of control in the network: connection setup control at the network ac-

PAGE 91

85 cess point and switching control within the network. The connection setup control is responsible for the end-to-end resource reservation and allocation, and notifies the user whether this connection can be accepted or not. The switching control is to schedule the order of transmission to all those accepted messages at each intermediate nodes. Effective schemes need to be developed and employed at both levels to ensure the required real-time communication services. In this chapter, the effective control structure and schemes are going to be examined for real-time communications in wide-area networks. The emphasis will be on the effective resource allocation and scheduling schemes. They have been recognized to be very important and should be highly integrated to effectively satisfy the expected real-time services. 5.2 Control Structure Model 5.2.1 Two-Level Control Model As pointed out in Chapter 2, there are usually two control levels due to the architecture of the wide-area network. We shall discuss briefly the existing control schemes in the following and then show the proposed integrated scheme for real-time communication in next Section. A. Connection Setup Control The bandwidth allocation is one major part in connection setup phase. The decision should be based on the information that how much additional bandwidth needs to be reserved on links over which the requested connection is to be routed. However, due to the statistical multiplexing of connections within the network, the

PAGE 92

86 End-System Figure 5.2: Control Structure in WAN exact measure of this reservation requirements is not so easily obtainable. The traffic demand should be based on some aggregate statistical measures rather than on the maximal demand per connection. A reasonable scheme is carefully developed from the so called Equivalent Bandwidth (EB) which is introduced by Cidon and et.al [15]. This bandwidth represents the equivalent amount of link capacity that is to be consumed by the request connection. It is a function of both the characteristics of individual connections and their interaction within the link. The quality of service can be met only if, at all links, the aggregate equivalent bandwidth remains below the available network capacity. The proposed approach relies on simple approximations that estimate the bandwidth requirements or equivalent capacities. The validity of this approach has been verified by comparison to both exact computations and simulation results [35].

PAGE 93

87 Setup Control No W-C A-C Control Control Control Switch Control FCFS SP SSTF Global Local Figure 5.3: Two-Level Control and Schemes B. Switching Control The message switching control determines the message transmission order within the network. It is a run-time scheme and should conform with the guarantees computed at the virtual connection setup time. The control should also consider the effects of early arrivals and develop effective mechanisms to avoid possible problems such as priority inversion and congestion. It has been well recognized that the conventional packet switching data networks with window-based flow control and first-come-first-served discipline can not provide services with strict performance guarantees. Several new control strategies and schemes have been proposed and analyzed recently [14, 24, 45]. The rate-based service discipline provides a client with a minimum service rate independent of the traffic characteristics of other clients. Such a discipline, operating at switching component, manages the bandwidth, service priority, and buffer space.

PAGE 94

88 5.2.2 Challenges It is not so difficulty to observe that the problems existing in wide area networks, especially when additional time constraints need to be considered, are far from being solved. First, it should be recognized that the guaranteeing of bandwidth requirements does not necessarily guarantees the real-time requirements due to the statistically multiplexing traffic. In fact, the time-constraints imposed by application users have not been explicitly addressed at the connection setup phase. Second, most of the effective switching control schemes, such as smallest-slacktime-first (SSTF) based schemes [74], assume some deadline information is available at all the intermediate switching node. But it is not always possible to obtain the accurate local deadline due to the dynamic nature of the network operating and the distributed control mechanisms. Third, the control schemes employed in the connection setup and switch mechanisms are usually unrelated and working independently. This may cause the inconsistence and degrade system performance at the user interface. Real-time communication requires the explicit processing efforts to individual message classes which are with different degree of service quality. After investigating various control structures in current networking systems, it is identified that the control schemes for real-time communications should be closely integrated so that the timing constraints can be enforced system-wide. While the connection setup control at the interfacing components and switching control at the switching components play crucial roles in wide area networks, they should not be separated and independent. A consistent and integrated control management is critical to satisfy

PAGE 95

89 the timing constraints all the way through the network. 5.3 An Integrated Control Scheme Based upon the previous discussion and the proposed models, an integrated control scheme has been developed which consists of two levels. The first level is the end-to-end network access level, which focuses on the guaranteed connection set up and bandwidth allocation. At this level, the control schemes address the global network scheduling problem to achieve the maximum system utilization with required quality of service for each message classes. The second level is, however, a sort of micro-tune processing within the network. Rather than considering the global context, this level addresses the scheduling problems locally at individual switching node along the network. The control schemes at this level actually enforce the global scheduling policies, and tune the system to meet timing requirement individually. Fig 5.4 shows the structure of our proposed integrated control scheme. At the connection setup control level, a request has to go through a delay bound checking procedure in additional to the conventional bandwidth checking procedure. The new delay bound checking procedure is to ensure the user required timing requirement can be satisfied. The acceptance is only issued when the request can pass both checking procedures so that not only bandwidth is allocated but also the delay can be bounded. As a by-product, a bound vector is computed by the delay bound checking procedure. This bound vector then feeds into each intermediate switching nodes as a deadline reference. So at the switching control level, the switching scheduling algorithm reorders the incoming packets and sends them along the outgoing link. Wise decisions can be performed by effectively using the informative

PAGE 96

90 Setup Control Switch Control Connection Request Incoming Packets _i Switch Scheduling Algorithm T Bandwidth Cheching Procedure * Delay Bound Checking Procedure Reject Accept A Bound Vector Outgoing Packet Figure 5.4: Integrated Control Structure bound vector. It can be seen these two control levels are highly integrated in our scheme and thus able to achieve much better real-time performance. 5.3.1 Algorithm for Setup Control At connection setup phase, the decision to accept a new request is usually based on whether the quality of service can be maintained under the present network load along the calculated route. The simplest scheme is to assign the peak bit rate required by all source users. Therefore the new request will be accepted only when the sum of peak rates of all current, including the new, requests on every link composing the route does not exceed the bit rate of that link. It is worth mention that different routes may have different results. Anyway, this is appropriate only in the environment where smooth and constant rate traffic is expected. Since the practical real-time communication traffic is often stochastic and hard to predict, it can be

PAGE 97

91 easily seen that this scheme cannot reflect the real-time processing requirement and cannot achieve significant bandwidth efficiency. To start determining the proper amount of bandwidth required by a connection, the characteristics of each connection request have to be clearly specified. Recall from the previous model discussed and the results of [15], the bandwidth required by an individual connection with vector (i2p ea *,., Puk) can be estimated using a simple fluid-flow model, and is given by: EB = ^ (5-1) where /? = ab(l — p), x represents the available buffer space, and a = ln(l/e) with e being the required quality of service. This expression provides a simple and reasonably accurate estimate of the load request on network links. It is the basis of our further discussions on bandwidth allocation control and quality of service control. A proposed connection setup control procedure can be described as following major steps. 1. Choose a route vector (L\, ...,Lk) c for the connection request c which starts at link L\ and ends at link Lk 2. Calculate the equivalent bandwidth EBi iiC for each link Li of this connection c along the route 3. For each link Li, assign the new connection a highest priority level such that the link delay for this connection Wl,, c is minimized. Note that, since there are existing connections at link Z,,, the priority assignment should keep the delay of existing connections within their bounds {Bi, iiCl , Bi i Cn ).

PAGE 98

92 • The above calculations of the link delay W LttC are based on the M/G/l priority model and will be described later. 4. If the A = TD C -Y^i=i ^Li,e > 0, the connection is established with guaranteed service to the bound vector (-6/, liC , Bi ktC ) • The bound vector is obtained by allocating the A to link delay W Li , c along the route Our scheme is under the assumption that there is a routing topology database similar to the one in ARPANET. Every node maintains a routing topology database with link weights representing the traffic over that link and the updating to each node is accomplished by a broadcast algorithm. As shown before, the calculation of Wi l>c is the key part for the whole procedure. Based on the standard M/G/l priority queue, the non-preemptive priority scheme only influences the waiting time experienced by randomly arriving packets of different classes. The expected delay for a class-k packet can be written as _ Ef = i ^Ej(t 2 )/2 (i-£j=i Pj)(i EUi Pi) Where Ej(t 2 ) is the second moment of the class-j service time distribution and the smaller index value indicates the higher priority level. While the packet transmission time T can be a close upper bound of the service residue time WO, the traffic intensity can be estimated by making use of the precalculated equivalent bandwidth. The computations of the accumulated traffic intensity and utilization of connection c can be obtained respectively. k k ,(£,) = £/>l„c (5-3) ;=1 c=l EB LliC ^L, ,max PL ifi = j-^ (5.4)

PAGE 99

93 Following is the procedure of finding the possible smallest link delay for the new connection request while not affecting the other already established connections. Procedure FindtheSmallestDelay: position = 0; flag = ON; While ((position < theConNo) and (flag == ON)) { position + +; flag = OFF] I * check all the existing connections * / Pl = ph = 0.0; for(i = 1; i < theConNo; i + +) { / * get the i — th smallest connection * / Pi = Ph + Hnk[i].EB; if(i == position) { Pl+ = Pnew\ Ph+ = Pnew] } o~i = pi /theMaxC opacity; °~h = Ph/theMaxC opacity; newDelay = T/{1a,) * (1 a k ); I * feasibility check * / if (newDelay > BD[i))

PAGE 100

94 { / * failed and try next higher index * / flag = ON; } Ph = pi; }/*for*/ }/ * while * I It is worth mention that here the priority is not really assigned to each connection but works as a means to obtaining the delay bound vector. Actually, we do not intend to assign a static priority to each connection which is usually not so effective when the traffic is multiplexing and unpredictable. What we really want is the more accurate local link delay (through the delay bound vector) for each connection and transforming the global timing constraints into a distributed control system. The exact transmission order at each switching node is decided locally by considering both the local delay bound and current traffic status at that node, which is the job to be accomplished by the next level control called dynamic switching control. Under this integrated control scheme, the ene-to-end real-time services can be guaranteed as long as the local delays along the network are enforced. 5.3.2 Algorithm for Switching Control As it has been widely accepted that the overflow error instead of bit-error has become the major factor in high speed real-time communications, the role of switching nodes has become more and more important. One primary task of the switching nodes is to schedule various incoming messages and to determine their service pref-

PAGE 101

95 Figure 5.5: Dynamic Scheduling Scheme Model erence. While the classic internetworking routers only deal with data at the packet level, it is impossible for them to provide flow and congestion control. A switching control model has been developed which enables the flow and congestion control be enforced at the switching nodes so that a better and more efficient real-time communication system is expected. Queueing schemes, which control the order in which packets are sent and the usage of the router's buffer space, do not affect congestion directly, in that they do not change the total traffic on the router's outgoing line. However, they do determine the way in which packets from different sources interact with each other which, in turn, affects the collective behavior of flow control algorithms. It should be noted that this effect, which is often ignored, makes queueing schemes a crucial component in effective congestion control over wide-area networks. More discriminating queueing schemes must be used in conjunction with source flow control algorithms to control congestion effectively in aggregately multiplexing environments.

PAGE 102

96 To overcome these disadvantages and provide real-time performance, some other control schemes are worth investigation. A QOS control scheme needs to be developed which allots the total loss probability and transmission delay into individual classes so that a feasible region can be identified for each particular individual message class. One idea to accomplish that quality control is to consider not only the slack time of a message but also the proper bandwidth allocation and scheduling among all the interacted message classes. A class-oriented control schemes based on the dynamic transmission region (DTR) idea have been developed. The intuition of this scheme is that different transmission region is assigned to each message class with particular guaranteed quality of service. In order to describe our control scheme more specificly, the slack time ST(c, Li) of a class c message at link L,should be first defined as follow. ST(c, Li) = £ B L]iC CurT ArrT (5.5) i=i where B(Lj,c) is the delay bound of this class c message on link Lj. CurT and ArrT are the current time and the message arrival time respectively. The accumulated delay bound along the route gives a precise deadline duration requirement the message should meet at this particular hop. It can be seen that the slack time represents the amount of time for which the server may remain idle, or serve other packets, and still be able to serve this packet by its deadline. This is actually the schedulability of this packet. As it can be observed, more schedulability can be achieved if more packets have a loose slack time at a time. The dynamic transmission region (DTR) scheduling algorithm can be conceptually described as follows: 1. Choose a control period (CP) as the scheduling region for all the available

PAGE 103

97 connections. On each link Z,,, the slack time ST(c, Z-,) of a class c message can be derived from bound vector as in equation (5.5). 2. For each class, decide the number of messages whose slack time will expire before CP. The transmission region TR(c) for class c can be obtained by checking equation (5.7) periodically. 3. The transmission order is prioritized by the criticality of each class. While still schedulable, all the non-guaranteed traffic can be transmitted within this control period CP to maximize the utilization. The DTR scheme decides a control period CP, and assigns the transmission region TR(i) or duration as the possible minimum to satisfy the QOS requirement of class t in the order from the highest class to the lowest class. Also, this control is enforced dynamically in corresponding to the traffic load of various classes so that the high system utilization can be achieved. Procedure DynamicTx: position = 1; min = MAX; /* update the control period */ While (sim t tme > period) period-r = CP; I * find the candidate among the local queue * / (5.6) 772(d) + TR(c 2 ) + ... + TR(c k ) + ... < CP (5.7)

PAGE 104

98 for(i = 1; i< QLEN; i + +) { / * obtain slack time * / slJLime = BV[i] durJime; I * check the validation * / remJLime = period — simJime; if(slJime < remJime) { I * within period, and decide the order * / if(ORDER[con] < min) { position = t; min = ORDER[con]; } }/ * for * I As it cam be observed, the delay bound vector is used to obtain the proper local deadline in our algorithm. Therefore, this is an integrated part of the whole control scheme. Whenever a new request coming, the equivalent bandwidth and delay bound vector are calculated according to the coming traffic characteristics and the current networking status. If it is accepted, the local deadlines are computed as discussed in the first level control and assigned to each switching node along the route. The dynamic switching control then enforces the timing constraints at each node independently.

PAGE 105

99 Besides the transmission delay deadline, the packet lost percentage is another major QOS requirement measure in real-time communication systems. Although the two issues are closely related with each other, it can be observed that the proposed DTR scheme mainly deals with the control of the packet transmission delay in order to meet the deadlines for each class. The packet lost percentage requirement of each class, however, can also be used in deciding which class of packets should be discarded when the congestion has caused the overflow situation. Our control scheme always discards the packets in the order whose class has a loose requirement (high constraint value) on lost percentage. Therefore, there are two independent control schemes employed in each switching node along the network. One is responsible for the proper order of packet transmission at outgoing links and the other controls the order of packet discarding when congestion is too severe. Both control schemes are, however, based on and served to the particular QOS requirement for each message classes. 5.4 Evaluations Due to the nature of traffic unpredictability, the analytical model of the real-time wide area networks is believed to be very difficult if not impossible. Several simulation models have been developed and extensive experiments have been taken under various settings for the purpose of performance evaluations. In this section, one example of the simulation model and corresponding experiment results are presented.

PAGE 106

100 Figure 5.6: Simulated Network and Workload 5.4.1 Network Model The performance evaluation of the proposed control schemes is conducted in a 9node 13-link network as shown in Fig 5.6. The distribution of this network workload is defined in the following way. There are totally 12 predefined routing pathes denoted as route.1, route J2. The exact links passing through by each route are listed as a n-tuple where n is the route length. Whenever a new request comes, a route is randomly assigned to this connection.

PAGE 107

101 Parameters Peak Rate Rpeak 4.0 Mbps Utilization p 0.4 Burstiness b 0.1 ms Link Capacity ,max 20 Mbps Delay Domain [BD m i n ,BD ma x] [4.0,10.0] T LP Domain [LPjnini ^B max] [lo, me, hi] Connection Duration T -* con 40UUU 1 Train Interval INT idle 13 T Control Period CP 5 T Simulation Time T 100000 T Table 5.1. The System Parameters Routing Table ID Route Vector ID Route Vector 1 (L1,L2,L4) 7 (L5,L6,L7) 2 (L1,L2,L9) 8 (L8,L10,L11) 3 (L2,L4,L7) 9 (L9,L10,L12) 4 (L2,L5,L6) 10 (L11,L6,L13) 5 (L3,L4,L13) 11 (L3,L4,L13) 6 (L3,L9,L10) 12 (L8,L10,L12) Table 5.2. The Routing Table For each coming request, additional parameters are needed to describe the user traffic load pattern and required quality of services. While the vector (Rp ea k,i, pi, &«)

PAGE 108

102 is used as the bandwidth requirement, BD and LP are the user required end-to-end bounded delay and packet lost percentage. A detailed system parameters and their values are illustrated in Table 5.1, where the packet transmission time T is used as the system unit time. 5.4.2 Setup Control For the connection setup control which is responsible for the admission management, the following four schemes are investigated. • NOC No admission control and accepts any incoming connections • W-C Worst-case bandwidth admission control • EB Pure equivalent bandwidth admission control • EB-D Combined equivalent bandwidth and delay bound admission control Control Schemes Load NOC W-C EB EB-D EB=3.12 Mbps 120 40 60 57 EB=2.79 Mbps 120 42 67 60 EB=2.47 Mbps 120 51 73 67 EB=2.15 Mbps 120 51 81 77 Table 5.3. The Number of Accepted Connections All the connection request follows a global arrival pattern no matter which accessing point it actually wants to enter the network. The Poisson distribution arrival pattern is employed in the simulation for all the incoming requests. The connection

PAGE 109

103 5 140.0 130.0 120.0 1 10.0 100.0 ©O.O 80.0 70.0 60.0 SO.O 40.0 30.0 20.0 10.0 O.O EB A. EB-D 1 2.2 2.3 2.4 2.S 2.8 2.7 2.8 2.8 3.0 3.1 Connection Load Figure 5.7: Established Connections 3.2 setup control decides whether each request can be accepted or rejected according to network status and user requirements. The accepted connection will depart from the network whenever its duration time expires. Fig 5.7 shows the number of accepted connections when the four different setup control schemes are employed. Different load can be obtained by changing the peak rate of each connection request. The connection load here is measured by the calculated equivalent bandwidth EB requested from each incoming connection. It should be noted that the total traffic load the network can be accepted is kind of fixed. In another word, the larger the bandwidth requirement of each connection, the smaller the number of connection can be established. So the number of connection decreases when the measured connection load increases along the x-axis. As it can be seen, the NOC scheme always accepts the maximum number of connections whereas the W-C scheme allows the smallest number of connections going into the network. The

PAGE 110

104 control schemes based on equivalent bandwidth computation compromise these two extreme cases and accept a more reasonable amount of connections at the networking access interface. However, the EB-D scheme may block some connections when certain or all the connections have a tighter delay bound requirement. Connection in Par an LP oters Route Vector Connection ID BD Par an LP leters Route Vector 1 9.00 low fx i T 0 T 41 11 7.00 high (L11,L6,L13) 2 6.00 nigh 12 5.00 low (L1.L2.L4) 3 8.00 high (L1.L2.L9) 13 9.00 median (L3.L4.L13) 4 7.00 high (L3,L4,L13) 14 9.00 low (L5,L6,L7) 5 8.00 low (L2,L5,L6) 15 8.00 high (L3.L4.L13) 6 6.00 median (L8,L10,L11) 16 6.00 low (L3.L4.L13) 7 9.00 high (L8,L10,L12) 17 4.00 low (L2,L4,L7) 8 4.00 high (L3,L4,L13) 18 5.00 low (L1,L2,L9) 9 8.00 high (L11.L6.L13) 19 8.00 low (L1,L2,L9) 10 4.00 high (L9,L10,L12) 20 6.00 high (L11,L6,L13) Table 5.4. The Generated Connections While there are so many connections within our network, 20 connections are arbitrarily chosen among them to study in detail. Table 5.4 gives one set of such 20 connections with their EB = 2.47. This set of connections is also elaborated by their detailed information, including BD, LP, and route vector in our simulation run. Since real-time connections have timing requirement, the proposed EB-D scheme may accommodate fewer connections than the EB scheme as in Fig 5.7. However, the EB-D scheme guarantees the accepted connections delay bounded when they go through the network and the lost percentage can be minimized. While the NOC

PAGE 111

105 scheme bears a large lost percentage due to the amount of connections it accepts, it has been found the EB scheme also has to tolerant the lost percentage especially for those critical classes. This is because the the EB scheme only guarantees the bandwidth bound and has no checking procedure for delay bound. Although the W-C scheme has been observed with no lost percentage in our simulation setting, it has the danger of accepting extremely tight timing requests due to the lack of delay bound checking procedure. It also has the problem of limiting the system utilization. Thus, the proposed EB-D scheme has shown to be a more powerful and effective admission control mechanism, especially for the real-time communication networks. In addition, the EB-D scheme provides a delay bound vector which can be integrated into the switching control and make the communication service more predictable. 5.4.3 Switching Control While the previous tables show the connection characteristics and the number of connection being accepted at networking access interface, the switching control is responsible for the exact order of transmission at each switching node. Six different control schemes are compared and evaluated. The first three are based on no priority or static priority. For all these schemes, message will be discarded at any switching node if it has been detected that it cannot meet the local delay bound. • FCFS First-Come-First-Served • Min-BD Static priority scheme with the smallest bounded delay sent first • Min-LP Static priority scheme with the smallest lost percentage sent first • SSTF Smallest-Slack-Time-First

PAGE 112

106 1 30.0 25. 0 20.0 1S.O 1 o.o s.o o.o € MlnBD O MlnLP 2.1 2.2 2.3 2.4 2.S 2.B 2.7 2.S 2.0 3.0 3.1 3.2 Connection Load Figure 5.8: Lost Percentage for Static Schemes • DTR-BD Dynamic transmission region scheme with the smallest bounded delay sent first • DTR-LP Dynamic transmission region scheme with the smallest lost percentage sent first Fig 5.8 and Fig 5.9 show the lost percentages for three static schemes and three dynamic schemes. The connection setup control scheme employed here is the EB-D although the three static schemes never use the delay bound vector. As discussed before, there is almost no lost percentage for connections under EB-D scheme when the number of connections is very small. In this experiment, we purposely increase the real traffic load to exceed the requested load which is measured by estimated equivalent bandwidth EB. The load overplay magnitude is of 8% in this experiment. It is our intention to investigate the proposed the EB-D scheme under over estimated load so that lost percentage can be observed to show the effectiveness of various

PAGE 113

107 30.0 2S.O 20.0 1S.O 10.0 s.o °'°2.1 2.2 2.3 2.4 2.S 2.S 2.7 2.S 2.8 3.0 3.1 3.2 ConneoHon Load Figure 5.9: Lost Percentage for Dynamic Schemes switching control schemes. This situation may also happen in real practice when the user cannot specify exact traffic load in advance or when arrival is so dynamic that the traffic model is inaccurate. While the lost percentages of three static schemes in Fig 5.8 are generally larger than those of three dynamic schemes in Fig 5.9, there are two distinguishable groups of control schemes. The better performance of the three dynamic control schemes can contribute to the effective usage of delay bound vector at switching control.

PAGE 114

108 Connection Control Schemes Connection Control Schemes ID FCFS MinBD MinLP ID FCFS MinBD MinLP 1 0.00 83.36 0.00 11 12.62 23.13 38.77 2 85.26 0.00 2.18 12 57.75 0.00 0.00 3 32.49 69.88 74.00 13 0.00 71.85 18.67 4 37.84 4.09 62.13 14 0.00 43.04 0.00 5 1.62 0.93 0.00 15 0.00 2.55 27.38 6 26.77 0.00 0.00 16 6.43 0.00 0.00 7 0.00 14.34 0.00 17 29.96 0.00 0.00 8 38.25 0.00 53.22 18 60.42 0.00 0.00 9 0.00 0.00 7.60 19 11.37 21.18 0.00 10 4.57 0.00 49.32 20 0.00 57.79 99.00 Table 5.5. The Lost Percentage for Static Schemes Connection Control Schemes Connection Control Schemes ID SSTF DTR-BD DTR-LP ID SSTF DTR-BD DTR-LP 1 17.13 39.31 6.32 11 9.16 2.18 22.46 2 0.00 0.00 23.07 12 28.76 0.00 0.00 3 -rej-rej-rej13 0.00 21.12 14.60 4 -rej-rej-rej14 0.00 15.87 0.00 5 0.00 0.36 0.00 15 0.00 3.61 3.16 6 -rej-rej-rej16 0.00 11.09 0.00 7 0.00 2.55 2.85 17 26.84 0.00 0.00 8 23.22 0.00 9.98 18 28.03 0.00 0.00 9 0.90 1.47 2.88 19 10.91 24.09 0.00 10 3.66 0.00 11.93 20 0.00 52.72 69.06 Table 5.6. The Lost Percent age for Dynamic Schemes

PAGE 115

109 Table 5.5 and Table 5.6 give the lost percentages of three static schemes and three dynamic schemes respectively for 20 selected connections. It can be observed the three dynamic schemes generally provide much better lost percentage performance than the three static ones. This is due to the fact the former ones have the additional control information, local deadline measures. The bound vector (Bl 1iC , Bi kiC ) obtained during setup phase in our integrated control scheme is very informative and plays a key role in such a dynamic environment. Among the three dynamic schemes, we also found that while the SSTF reduces lost percentage significantly, the DTRBD and DTR-LP provide more powerful policy to the control mechanism without sacrificing the low lost percentage. Actually, the DTR-BD tends to accumulate lost percentage on connections with a larger DD value like connections cl, cl3 and cl4, and the DTR-LP accumulates on connections with a larger LP value like connection c2, clO, ell, and c20. These latter two schemes suggest a set of control schemes can be derived from different performance interest. The control power can be enforced into each QOS classes rather than at the general level. While the proposed dynamic control schemes are able to provide more powerful and effective control in real-time communication networks, it has been found these benefits can be obtained with little cost. Table 5.7 gives the link utilization for each of six control schemes. It can be noticed there is not much difference in average system utilization even though the the latter three may reject some connections. Fig 5.10 shows more clearly the average link utilization for all six schemes. Again, we can see there are two groups corresponding to static schemes and dynamic schemes. However, the utilization difference is negligible in our simulation experiments, which means the utilization is not sacrificed to achieve better real-time performance.

PAGE 116

110 Link Control Schemes ID FCFS MinrJD MinLP SSTF DTR-BD DTR-LP LI 62.43 60.91 60.09 62.59 62.59 R1 1 1 01. 11 L2 81.63 80.43 79.88 79.31 78.85 77 KC\ ( 1 .OU L3 66.09 65.75 64.91 62.07 62.07 bi.y7 L4 79.48 79.97 81.21 80.26 80.69 on ac oU.4o L5 40.63 40.58 34.34 37.12 36.50 O/J 1 A oo.l4 L6 74.97 72.19 65.90 73.84 72.78 71.3b L7 32.55 33.25 34.02 32.79 33.94 o a ^ i 34.41 L8 38.89 38.89 38.89 34.27 34.27 1 A 07 34.27 L9 66.37 67.75 65.97 70.20 69.67 68.13 L10 76.20 75.81 73.20 72.47 72.58 ry 1 f\r\ 71.90 T 1 1 M 41 50.91 48.19 48.17 48.19 L12 52.98 46.39 48.63 49.02 47.86 47.99 L13 78.75 72.11 66.96 78.20 76.31 74.79 Table 5.7. The Link Utilization for Various Schemes In this chapter, the wide-area communication networks for providing timeconstrained services have been investigated. Proper networking structure and control enforcement strategies are discussed in identifying the lack of explicit real-time control schemes in current communication systems. An integrated control scheme is presented with the main idea of enforcing the time-constraints and lost percentage along the network. The detailed control algorithms for connection setup control and switching control are given. By employing the dynamic scheduling strategy, it is hoped that the system utilization can be significantly improved as well as the real-

PAGE 117

Ill 70.0 65. 0 60.0 55. 0 BO.O AS. O 40.0 35.0 30.0 2S.O 20. 0 -O FCFS -E3 MlnBD -<> MlnLP — ISSTF -»< DTR-BD -* DTR-LP 2.1 2.2 2.3 2.4 2.S 2.B 2.7 2.8 2.8 3.0 3.1 3.2 Connsotlon Load Figure 5.10: Link Utilizations time services can be ensured. Our extensive simulation experiments and evaluation results show the effectiveness and feasibility of the scheme. It contributes to the communication systems with its simplicity, effectiveness, and predictable services for complex real-time applications.

PAGE 118

Chapter 6 CONCLUSIONS The media access control schemes play the key roles for real-time communications in local and metropolitan area networks. The proposed multiple channel token ring (MCTR) architecture and control protocols have been developed and evaluated to show several advantages in ring networks. The Wait-Until (WUT) based schemes can minimize the lost percentage of those critical message packets while still maintain a high channel utilization. They are easy for the implementation by possibly expanding the existing ring interface technologies. Also it is worth noted that these dynamic control schemes are so flexible to various application situations that it can also be applied to conventional communications as well as time-constrained communications. In order to use the global multiple priority mechanism more effectively in the bus networks, a dynamic priority control scheme is presented for Distributed Queue Dual Bus (DQDB) network. The proposed scheme emphases on dynamic local message scheduling at individual level which is especially important in real-time communications. The dynamic priority assignment scheme achieves a high quality real-time performance by properly making use of the multiple priority mechanism 112

PAGE 119

113 in DQDB standard. The priority access control is absolute in that immediate access can be guaranteed without any priority inversion delay for those smallest slack time first (SSTF) segments. By forcing all the stations to follow the same distributed media access criteria, the total lost percentage of all kinds of messages due to the deadline constraints is minimized. The simulation results have validated the effectiveness of our control strategies and shown the proposed MAC protocol significantly outperforms conventional MAC protocols for real-time communications. It is believed that the future real-time protocols should be able to function as a wise load scheduler as well as the conventional media access arbitrator. The transmission delay at the message level is a much more important and proper performance measure for real-time communications. It is argued that the MAC protocol in real-time MANs should be designed and evaluated from the application users' point of view (message level) instead of from the network designer' point of view (segment level) so that much meaningful and better service performance can be achieved. Message level processing (MLP) is natural due to the fact that time-constraint is an attribute at the message level instead of slot level. The proposed multiple-slot reservation protocol and enhanced multiple-slot protocol contribute to the DQDB standard with processing efficiency, better bandwidth usage, and improved message delay performance. Of all the discussed protocols, the SJF-based multiple slot protocol has shown to be the best in the sense that it minimizes the average message delay under all circumstances. Also, it is worth mention that the proposed protocols have shown a relative more flat delay curve than that of the single slot protocol (DQDB standard). This indicates that the multiple slot scheme is able to suppress the unfairness problem. In additional to these, the proposed schemes require no or minor changes to the slot format and

PAGE 120

114 are highly compatible with the DQDB standard protocol. The message level processing strategy is surely a promising candidate of effective control schemes for the integrate-service applications in high speed real-time communication systems. While the communication domain changes from local and metropolitan area networks to wide-area networks, the control and support for real-time services bring more challenges. The wide-area communication networks for providing timeconstrained services have been investigated, especially for the effect the additional complex switching system incurs. Proper networking structure and control enforcement strategies are discussed in identifying the lack of explicit real-time control schemes in current communication systems. An integrated control scheme is presented with the main idea of enforcing the time-constraints along the network. The detailed control algorithms for connection setup control and switching control are given. By employing the dynamic scheduling strategy, it is hoped that the system utilization can be significantly improved as well as the real-time services can be guaranteed. Our extensive simulation experiments and evaluation results show the effectiveness and feasibility of the scheme. It contributes to the communication systems with its simplicity, effectiveness, and predictable services for real-time applications in complex wide-area networks. Real-time communication systems are complex and dynamic in nature. As currently in its early stage, the research in real-time communications needs great efforts in better understanding and is far from being solved. The capability of providing predictable services brings a rewarding challenge to both researchers and application users. Fast connection establishment is critical to certain real-time application users. For general establishment, a connection request usually takes a round trip through

PAGE 121

all the intermediate nodes. The context and resources are negotiated and prepared by all the nodes involved. The exact data transmission is normally started after this establishment phase. One way for fast establishment is to allow the data to be sent together with the request along the network without waiting for a connection establishment to be confirmed by the destination. Efficient mechanisms have to be developed and are useful for best-effort real-time applications. The traffic pattern in current networks is so unpredictable and hard to model accurately by the available theory. This is due to the fact that the amount of traffic is so large and diverse. It is also because the network speed could be so fast that it is able to change the flow pattern within the network from the original source. More suitable and efficient models are needed to capture the complex characterization of traffic in high speed real-time communications. Simple and effective approximation models may more likely be approached and comparison methods are necessary. What is the basis for a more predictable performance analysis ? We are now faced with the difficult scientific challenge of creating a set of unified theories and technologies which will allow us to reason about the correctness, timeliness, and reliability for complex integrated systems. The development of effective control schemes for real-time communication have brought about challenges in a wide range of science and engineering disciplines that are not being met. The task is very formidable and success will require a concerted effort on the part of many different participants in the computer and communication community.

PAGE 122

Bibliography [1] Aho, A.V., Hopcroft, J.E. and Ullman, J.D., "The Design and Analysis of Computer Algorithms," AddisionWesley Publishing Company, 1974. [2] Alluisi, E.A., "The Promise of Interactive Networking: New Levels of Training and Readiness in Peacetime," Proceedings Interactive Networked Simulation for Training, April, 1989. [3] Abeysundara, B.W. and Kamal, A.E., "High-Speed Local Area Networks and Their Performance: A Survey," A CM Computing Surveys, Vol.23, No. 2, June 1991, pp.221-264. [4] American National Standard, "FDD I Token Ring Media Access Control (MAC)," ANSI X3.139, 1987. [5] American National Standard, "FDDI Token Ring Physical Layer Protocol (PHY)," ANSI X3.148, 1988. [6] Boxma, O.J. and Meister, B., "Waiting Times in Multi-Queue Systems with Cyclic Service," IBM Zurich Research Lab, 8803 Ruschlikon, Switzerland, Jan.21, 1985. llfi

PAGE 123

117 [7] Bux, W., "Local Area Subnetworks: A Performance Comparison," IEEE transactions on Communications, Vol.29, No.10, Oct. 1981, pp. 1465-1473. [8] Bux, W., "Token-Ring Local-Area Networks and Their Performance," Proceedings of the IEEE, Vol.77, No.2, February 1989, pp.238-256. [9] Chen, C.H. and Bhuyan, L.N., "Design and Analysis of Multiple Token Ring Networks," INFOCOM 88, 1988, pp.477-486. [10] Chen, T.M., Walrand, J. and Messerschmitt, D.G., "Dynamic Priority Protocols for Packet Voice", IEEE Journal on Selected Area in Communications, Vol.7, No.5, June 1989, pp.632-643. [11] Cheriton, D.R., "The V Kernel: A Software Base for Distributed Systems," IEEE Software, Vol.1, No.2, April, 1981, pp.19-43. [12] Cheriton, D.R., "VMTP: A Transport Protocol for the Next Generation of Communication Systems," IEEE SICCOM'86, 1986, pp.406-415. [13] Chlamtac, I. and Ganz, A., "Design and Analysis of Very High-Speed Network Architectures," IEEE Transactions on Communications, Vol.36, No. 3, Mar. 1988, pp.252-262. [14] Cidon, I., et.al, "Real-Time Packet Switching: A Performance Analysis," IEEE Journal on Selected Areas in Communications, Vol.6, No.9, December 1988, pp.1576-1586. [15] Cidon, I., Gopal, I. and Guerin, R., "Bandwidth Management and Congestion Control in plaNET," IEEE Communications Magazine, October 1991, pp. 5464.

PAGE 124

118 [16] Clark,D.D., Lambert,M.L. and Zhang,L, "NETBLT: A High Throughput Transport Protocol," IEEE SIGCOMW, 1987, pp.353-359. [17] Conti, M., Gregori, E. and Lenzini, L., "A Methodological Approach to an Extensive Analysis of DQDB Performance and Fairness," IEEE Journal on Selected Areas in Communications, Vol.9, No.l, 1991. [18] Conti, M., Gregori, E. and Lenzini, L., "DCP: A Distributed-Control Polling MAC Protocol: Specifications and Comparison with DQDB," IEEE Journal on Selected Area in Communications, Vol.9, No.2, February 1991, pp.241-247. [19] Conway, A.E., "A Perspective on the Analytical Performance Evaluation of Multilayered Communication Protocol Architectures," IEEE Journal on Selected Area in Communications, Vol.9, No.l, January 1991. [20] Danthine, A., "10 Years with OSI," Information Network and Data Communication, III, Elsevier Science Publishers B.V. (North-Holland), IFIP, 1990, pp.1-16. [21] Davids, P. and Martini, P., "Performance Analysis of DQDB," IEEE INFOCOM'90, 1990. [22] Davids, P. and Martini, P., "High Speed LANs and MANs Protocols and Problems," Information Network and Data Communication, III, Elsevier Science Publishers B.V. (North-Holland), IFIP, 1990, pp.119-132. [23] Decina, M. and Toniatti, T., "On Bandwidth Allocation to Bursty Virtual Connections in ATM Networks," GLOBECOM, 1990.

PAGE 125

119 [24] Demers, A, Keshav, S. and Shenker, S., "Analysis and Simulation of a Fair Queueing Algorithm," ACM SIGCOM '89, 1989, pp.1-12. [25] de Moraes, Luis F.M., "Priority Scheduling in Multiaccess Communication," Stochastic Analysis of Computer and Communication Systems, Elsevier Science Publishers B.V. (North-Holland), IFIP, 1990. [26] Dionysios, K. and Alberto, L.G., "Performance of Integrated Packet Voice/Data Token-Passing Rings," IEEE Journal on Selected Areas in Communications, Vol.4, No.6, September 1986, pp. 823-832. [27] Doeringer, W., et.al, "A Survey of LightWeight Transport Protocols for HighSpeed Networks," IEEE Transactions on Communications, Vol.38, No. 11, November 1990, pp.2025-2039. [28] Fraser, A.G., "Designing a Public Data Network," IEEE Communications Magazine, October 1991, pp. 31-35. [29] Ferrari, D., "Guaranteeing Performance for Real-Time Communication in WideArea Networks," University of California, Berkeley, TR-89-001, January, 1989. [30] Ferrari, D. and Verma, D.C., "A Scheme for Real-Time Channel Establishment in Wide-Area Networks," IEEE Journal on Selected Area in Communications, Vol.8, No.3, April 1990, pp.368-379. [31] Gihr, 0. and Kuehn, P.J., "Comparison of Communication Services With ConnectionOriented and Connectionless Data Transmission," Computer Networking and Performance Evaluation, Elsevier Science Publishers B.V. (NorthHolland), IFIP, 1986, pp.173-186.

PAGE 126

120 [32] Gopal, I.S. and Guerin, R.A., "Network Transparency: The plaNET Approach," IBM Research Report, 1991. [33] Gopinath, P. and Schwan, K., "CHAOS: Why One Cannot Have Only An Operating System for Real-Time Applications," Operating System Review, July 1989, pp.106-125. [34] Grillo, D., "Polling Mechanism Models in Communication Systems Some Application Examples," Stochastic Analysis of Computer and Communication Systems, Elsevier Science Publishers H.V. (NorthHolland), IFIP, 1990. [35] Guerin, R., Ahmadi, H., and Naghshineh, M., "Equivalent Capacity and Its Application to Bandwidth Allocation in High-Speed Networks," IEEE Journal on Selected Areas in Communications, Vol.9, No.7, September 1991, pp. 968981. [36] Hayes, J.F., "Modeling and Analysis of Computer Communication Networks," Plenum Press, 1984. [37] Hyman, J.M., Lazar, A. A., and PacinV.i, G., "Real-Time Scheduling with Quality of Service Constraints," IEEE Journal on Selected Areas in Communications, Vol.9, No.7, September 1991, pp. 1052-1063. [38] IEEE 802.3 Local Area Network Standard, Carrier Sense Multiple Access with Collision Detection (CSMA/CD) Access Method and Physical Layer Specifications, IEEE Computer Society, 1985. [39] IEEE 802.4 Local Area Network Standard, TokenPassing Bus Access Method and Physical Layer Specifications, IEEE Computer Society, 1985.

PAGE 127

121 [40] IEEE 802.5 Local Area Network St andard, Token Ring Access Method, IEEE Computer Society, Silver Spring, Maryland, 1985. [41] IEEE P802.6/D11, "Project 802-Loral ans Metropolitan Area Networks, Proposed Standard: Distributed Queue Dual Bus (DQDB) Metropolitan Area Network (MAN)," November 1989. [42] ISO/IEC 7498, Data Communications Open Systems Interconnection — Basic Reference Model, 1984. [43] Jeon, W., Kim, C. and Kim, D.K., "Design of a Distributed Isochronous Channel Management Protocol for IEEE 802.fi MAN QPSX," IEEE INFOCOM'89, 1989, pp.268-275. [44] Kandlur, D.D., Kiskis, D.L. and Shin, K.G., "HARTOS; A Distributed RealTime Operating System," Operating System Review, July 1989, pp. 72-89. [45] Kandlur, D.D., Shin, K.G. and Ferrari, D., "Real-Time Communication in Multi-hop Networks," 11th Int. Conference on Distributed Computing Systems, 1991. [46] Kleinrock, L., "Queueing Systems, Vol.1: Theory," Wiley, New York, 1975. [47] Kleinrock, L., "Queueing Systems, Vol .11: Computer Applications," Wiley, New York, 1976. [48] Klessig, R.W., "Overview of Metropolitan Area Networks," IEEE Communications Magazine, Vol.24, No.l , January 1986, pp. 9-15. [49] Krishna, CM. and Lee, Y.H., "Guest Editors' Introduction: Real-Time Systems," IEEE Computer, May 1991, pp.10 11.

PAGE 128

122 [50] Kurose, J.F., Schwartz, M. and Yemini, Y., "Multiple-access Protocols and Time-constrained Communication," A CM Computing Surveys, Vol.16, No.l, 1984, pp.43-70. [51] Lazar, A. A. and Pacifici, G., "Control of Resources in Broadband Networks with Quality of Service Guarantees," IEEE Communications Magazine, October 1991, pp.66-73. [52] Lavenberg, S.S., "Computer Performance Modeling Handbook," Academic Press, Inc., 1983. [53] Lee, Y.H. and Shen, L.T., "Real-Time Communication in Multiple Token Ring Networks," Proc. of the Real Time. System Symposium, December, 1990, pp.146-154. [54] Lee, Y.H., Shen, L.T. and Nguyen, L.T., "Design and Evaluation of MultipleSlot Reservation Protocols for DQDB Networks," submitted to IEEE INFOCOM'93. [55] Liberg, 0., "Metropolitan Area Network Concepts, Architecture and Evolution towards B-ISDN," Information Network and Data Communication, HI, Elsevier Science Publishers B.V. (NorthHolland), IFIP, 1990, pp.179-192. [56] Liu, C.L. and Layland, J.W., "Scheduling Algorithms for Multiprogramming in a Hard-Real-Time Environment," Journal of the ACM, Vol.20, Jan. 1973, pp. 46-61. [57] Maxemchuk, N.F. and Zarki, M.E., "Routing and Flow Control in High-Speed Wide-Area Networks" Proceedings of the IEEE, Vol.78, No.l, January 1990, pp.204-220.

PAGE 129

123 [58] Mendelson, H. and Whang, S., "Opt imal Incentive-Compatible Priority Pricing for the M/M/l Queue," Operations Research, Vol.38, No.5, Sept-Oct. 1990, pp.870-883. [59] Minet, P., "Performance Evaluation or GAM-T-103 Real-Time Transfer Protocols," IEEE INFOCOM'89, 1989, pp.766-773. [60] Mollenauer, J.F., "Standards for Met ropolitan Area Networks," IEEE Communications Magazine, Vol.26, No. 1, April 1988, pp. 15-19. [61] Mukherjee, B., "Integrated Voice-Dat a Communication over High-Speed Fiber Optic Networks" IEEE Computer, February 1991, pp.49-58. [62] Nagle, J., "On Packet Switches with Infinite Storage," IEEE Transactions on Communications, Vol.35, 1987, pp. 135 138. [63] Newman, R.M., Budrikis, Z.L. and Hullett, J.L., "The QPSX Man," IEEE Communications Magazine, Vol.26, No. 4, April 1988, pp. 20-28. [64] Panwar, S.S., Towsley, D. and Wolf, J.K., "Optimal Scheduling Policies for a Class of Queues with Customer Deadlines to the Beginning of Service," Journal of the Association for Computing Machinery, Vol.35, No. 4, October 1988, pp.832-844. [65] Peden, J.H. and Weaver, A.C., "Performance of Priorities on an 802.5 Token Ring," ACM SIGCOM'87 Workshop, Stowe, VT, August 11-13, 1987, pp.58-66. [66] Phinney, T., et.al, "FieldBus — Real-Time Comes to OSI," IEEE International Phoenix Conference on Computers and Communications, 1991, pp. 594-600.

PAGE 130

124 [67] Pope, R., "The SIMNET Network and Protocols," BBN Systems and Technologies Corporation, Report No.7102, July, 1989. [68] Postel, J., Sunshine, C. and Dogen, D., "The ARPA Internet Protocol," Computer Networks, Vol.5, 1981. [69] Potter, P.G. and Zukerman, M., "A Discrete Shared Processor Model for DQDB," Computer Networks and ISDN Systems, 20(1990), pp.217-222. [70] Potter, P.G. and Zukerman, M., "Analysis of a Discrete Multipriority Queueing System Involving a Central Shared Processor Serving Many Local Queues," IEEE Journal on Selected Area in Communications, Vol.9, No. 2, February 1991, pp. 194-202. [71] Reiser, M., "Communication-System Models Embedded in the OSI-Reference Model, A Survey," Computer Networking and Performance Evaluation, Elsevier Science Publishers B.V. (NorthHolland), IFIP, 1986, pp.85-111. [72] Ross, F.E., "An Overview of FDDI: The Fiber Distributed Data Interface," IEEE Journal on Selected Areas in Communications, Vol.7, No. 7, September 1989, pp.1043-1051. [73] Rubin, I. and Baker, J.E., "Media Access Control for High-Speed Local Area and Metropolitan Area Communications Networks" Proceedings of the IEEE, Vol.78, No.l, January 1990, pp.168 203. [74] Saito, H., "Optimal Queueing Discipline lor Real-Time Traffic at ATM Switching Nodes," IEEE Transaction on Communication, Vol.38, No.12, December 1990, pp.2131-2136.

PAGE 131

125 [75] Sanders, R.M. and Weaver, A.C., "The Xpress Transfer Protocol (XTP) — A Tutorial," Computer Communication Review, Vol.20, No.5, Oct. 1990, pp. 6780. [76] Sauer, K. and Schodl, W., "Performance Aspects of the DQDB Protocol," Computer Networks and ISDN Systems, 20(1990), pp.253-260. [77] Schill, A. and Zieher, M., "Performance Analysis of the FDDI 100 Mbit/s Optical Token Ring," Proc. of the IF1P TC6/WG6.4 Workshop on High Speed Local Area Networks, Feb. 1987, pp. 5374. [78] Sethi, A.S. and Saydam, T., "Performance Analysis of Token Ring Local Area Networks," Computer Networks and ISDN Systems, Vol.9, No.3, March 1985, pp.191-200. [79] Sevcik, K.C. and Johnson, M.J., "Cycle Time Properties of the FDDI Token Ring Protocol," IEEE Trans. Software Eng., Vol.SE-13, No.3, March 1987, pp.376-385. [80] Shen, L.T. and Lee, Y.H., "Dynamic Priority Assignment Scheme for RealTime Communications in DQDB Networks," Proceeding of IEEE International Phoenix Conference on Computers arid Communications, April, 1992, pp. 612619. [81] Shin, K.G. and Hou, C.J., "Analysis of Three Contention Protocols in Distributed Real-Time Systems," RTSS'90, pp.136-145. [82] Sohraby, K., "Congestion Control in High Speed Networks: Current Research Activity," IEEE Communications Magazine, October 1991, pp.6.

PAGE 132

126 [83] Stankovic, J. A., "Misconceptions About Real-Time Computing: A Serious Problem for Next-Generation Systems," IEEE Computer, Vol.21, No.10, October, 1988, pp.10-19. [84] Stankovic, J. A., "Real-Time Computing Systems: The Next Generation," RealTime Systems, Feb 18, 1988, pp.14-36. [85] Stankovic, J.A. and Ramamritham. K., "The Spring Kernel: A New Paradigm for Real-Time Operating Systems," Operating System Review, July 1989, pp.54-71. [86] Strayer, W.T. and Weaver, A.C., "E valuation of Transport Protocols for RealTime Communications," University of Virginia, Report No. TR-88-18, 1988. [87] Strosnider, J.K., "Highly Responsive Real-Time Token Rings," Carnegie Mellon University, PhD Thesis, August 1988. [88] Strosnider, J.K., Marchok, T. and Lehoczky, J., "Advanced Real Time Scheduling Using the IEEE 802.5 Token Hing," Proceedings Real-time Systems Symposium, 1988, pp. 42-52. [89] Takagi, H., "Queueing Analysis of Polling Models," A CM Computing Surveys, Vol.20, No.l, March 1988, pp.5-28. [90] Takagi, Y., Hino, S., and Takahashi, T., "Priority Assignment Control of ATM Line Buffers with Multiple QOS Classes," IEEE Journal on Selected Areas in Communications, Vol.9, No.7, September 1991, pp. 1078-1092. [91] Takine, T., Takahashi, Y. and Hasegawa, T., "Average Message Delay of an Asymmetric Single-Buffer Polling System with Round-Robin Scheduling of Ser-

PAGE 133

127 vice," Modeling Techniques and Tools for Computer Performance Evaluation, Plenum Press, New York, 1989, pp.179 187. [92] Tanenbaum, A.S., "Computer Networks," Second Edition, Prentice Hall, Inc., 1988. [93] Tokuda, H., et.al, "Priority Inversions in Real-Time Communication," Proc. of the Real Time System Symposium, December, 1989, pp. 348-259. [94] Tokuda, H. and Mercer, C.W., "ARTS: A Distributed Real-Time Kernel," Operating System Review, July 1989, pp.29-53. [95] Tran-Gia, P. and Stock, T., "Approximate Performance Analysis of the DQDB Access Protocol," Computer Networks and ISDN Systems, 20(1990), pp. 231240. [96] Valenzano, A., Montuschi, P. and Ciminiera, L., "On the Behavior of Control Token Protocols with Asynchronous and Synchronous Traffic," Proc. INFOCOM'89, 1989, pp.348-355. [97] Watson, R. and Mamrak, S., "Gaining Efficiency in Transport Services by Appropriate Design and Implementation Choices," A CM Trans, on Computer System, 5(2), May 1987, pp.97-120. [98] Welzel, Th., Martini, P. and RudlofF, S., "Performance Analysis of FDDI and Multiple Token Ring Backbones in a Mixed Traffic Environment," High Speed Local Area Networks, II, Elsevier Science Publishers B.V. (North-Holland), IFIP, 1990, pp.53-68.

PAGE 134

128 [99] Yao, L. and Zhao, W., "Performance or an Extended IEEE 802.5 Protocol in Hard Real-Time Systems," IEEE INFOCOM'91, 1991, pp.469-478. [100] Zhao, W. and Ramamritham, K., "Virtual Time CSMA Protocols for Hard Real-Time Communications," IEEE Transactions on Software Engineering, Vol.SE-13, No.8, August 1987, pp.938 952. [101] Zhao, W., Stankovic, J. and Ramamritham, K., "A Window Protocol for Transmission of Time Constrained Messages," IEEE Transactions on Computers, Vol.39, No.9, September 1990, pp, 1 186-1203. [102] Znati, T., "Design and Analysis of Local Area Network Protocols for Distributed Real-Time Systems," Michigan State University, Dissertation, 1988. [103] "Military Standard (Draft): Entity Information and Entity Interaction in a Distributed Interactive Simulation," IST/UCF, January 25, 1991. [104] "Rational Document: Entity Information and Entity Interaction in a Distributed Interactive Simulation," IST/UCF, February 22, 1991. [105] "SIMNET Database Interchange Specification," BBN System and Technologies Corporation, July 31, 1989.

PAGE 135

BIOGRAPHICAL SKETCH Li-Tao Shen was born in Shanghai, Peoples Republic of China. He was awarded a Bachelor of Science degree in Computer and Information Sciences in 1984 and a Master of Science degree in 1987, both from the Fudan University in China. From 1987 to 1989, he served as a research and development member in the Computer Networking Laboratory and also as a lecturer in the Computer and Information Sciences Department at Fudan University. Since August 1989, he has been working towards the Ph.D. in the Department of Computer and Information Sciences from the University of Florida, while serving as a research assistant. His research interest includes high-speed high performance networks, real-time communications, distributed processing, and system performance evaluations. 124

PAGE 136

I certify that I have 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 dissertation for the degree of Doctor of Philosophy. Yann-Hang Lee, Chairman Associate Professor of Computer and Information Sciences I certify that I have 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 dissertation for the degree of Doctor of Philosophy. Randy Chow Professor of Computer and Information Sciences I certify that I have 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 dissertation for the degree of Doctor of Philosophy. Richard E. NewmanWolfe Assistant Professor of Computer and Information Sciences I certify that I have 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 dissertation for the degree of Doctor of Philosophy. Panos Livadas Assistant Professor of Computer and Information Sciences

PAGE 137

I certify that I have 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 dissertation for the degree of Doctor of Philosophy. Haniph A. Latchman Associate Professor of Electrical Engineering This dissertation was submitted to the Graduate Faculty of the College of Engineering and to the Graduate School and was accepted as partial fulfillment of the requirements for the degree of Doctor of Philosophy. December 1992 Winfred M. Phillips Dean, College of Engineering Madelyn M. Lockhart Dean, Graduate School