<%BANNER%>

A Unified Framework for Performance Analysis of Contention-Based Wireless MAC

Permanent Link: http://ufdc.ufl.edu/UFE0022540/00001

Material Information

Title: A Unified Framework for Performance Analysis of Contention-Based Wireless MAC Case Studies on QOS, Heterogeneity and Rate-Adaptation in 802.11
Physical Description: 1 online resource (179 p.)
Language: english
Creator: Wang, Shao-Cheng
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2008

Subjects

Subjects / Keywords: Computer and Information Science and Engineering -- Dissertations, Academic -- UF
Genre: Computer Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Performance analysis for IEEE 802.11 based protocols is essential to not only better understand the protocol, but also provide insight in developing protocol improvement schemes. In this dissertation , we propose a unified IEEE 802.11 MAC performance evaluation framework to systematically explore different facets of protocol behavior and the corresponding inherent interactions between different factors of various scenarios and system performance. Our proposed framework differs from the existing performance analysis approaches in that it categorizes the protocol operation parameters that exhibit similar behavior into the same group and uses the categorizations to interconnect the target scenarios with relevant performance affecting parameters. As a result, such weaving interconnections enable us to formulate the unified performance evaluation framework in explaining the similarity or dissimilarity of the performance impacts on different target scenarios, which can not be achieved by any existing performance evaluation models. Additionally, we aim to provide guidelines for various protocol enhancement designs by applying the insight gained from studying the protocol dynamics with the proposed performance evaluation framework. Using the proposed framework, we are able to conduct a series of systematic evaluations on the performance of various scenarios in which the performance is affected by a mixture of protocol operation parameters and environment factors such as wireless losses, collisions, and different background traffic load levels. In the first case study, we examine the mixed throughput anomaly effects in hybrid IEEE 802.11b and IEEE 802.11g networks, where various settings that are necessary to accommodate the heterogeneous environment all jointly affect the system performance. We demonstrate that our performance framework not only helps predict the aggregate system performance in hybrid 802.11b/g networks, but also clarifies and isolates the mixed effects among different contention window, different backoff stages, and different data rates and frame formats on these two versions of standard. In the second case study, we evaluate the effectiveness of various performance improvement schemes at the physical layer and MAC layer in understanding the performance limitations on throughput and total system delay of IEEE 802.11 MAC with arbitrary competing traffic. By using the proposed framework, we are able to decompose the effects of various performance improvement schemes and the corresponding interactions with the variable amount of network delay incurred by collisions, and the backoff and re-transmission procedures. We further identify a performance bottleneck beyond which the packet delay becomes infinitely high and we pinpoint the exact turning point depends on the packet arrival pattern in consideration. Furthermore, in the third case study, we first use the framework to analyze the performance rate adaptation algorithms and discover that a good rate adaptation algorithm needs to dynamically adjusting the rate selection decisions with respect to different background traffic levels, in addition to the wireless channel fluctuations that was believed to be the sole key of rate adaptation. We then apply the insight observed from the proposed framework to design a rate adaptation algorithm that accommodate the mixed effects from different background traffic levels and fluctuating wireless channel conditions. While simulation results show promising improvements over the existing rate adaptation schemes, our real-world test-bed implementations also demonstrate that our framework is suitable for being practically incorporated into real hardware to provide superb performance in real-world scenarios. We expect that the proposed framework can be further generalized to evaluate the cross-layer effects such as TCP and routing in wireless mesh networks.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Shao-Cheng Wang.
Thesis: Thesis (Ph.D.)--University of Florida, 2008.
Local: Adviser: Helmy, Ahmed H.

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2008
System ID: UFE0022540:00001

Permanent Link: http://ufdc.ufl.edu/UFE0022540/00001

Material Information

Title: A Unified Framework for Performance Analysis of Contention-Based Wireless MAC Case Studies on QOS, Heterogeneity and Rate-Adaptation in 802.11
Physical Description: 1 online resource (179 p.)
Language: english
Creator: Wang, Shao-Cheng
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2008

Subjects

Subjects / Keywords: Computer and Information Science and Engineering -- Dissertations, Academic -- UF
Genre: Computer Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Performance analysis for IEEE 802.11 based protocols is essential to not only better understand the protocol, but also provide insight in developing protocol improvement schemes. In this dissertation , we propose a unified IEEE 802.11 MAC performance evaluation framework to systematically explore different facets of protocol behavior and the corresponding inherent interactions between different factors of various scenarios and system performance. Our proposed framework differs from the existing performance analysis approaches in that it categorizes the protocol operation parameters that exhibit similar behavior into the same group and uses the categorizations to interconnect the target scenarios with relevant performance affecting parameters. As a result, such weaving interconnections enable us to formulate the unified performance evaluation framework in explaining the similarity or dissimilarity of the performance impacts on different target scenarios, which can not be achieved by any existing performance evaluation models. Additionally, we aim to provide guidelines for various protocol enhancement designs by applying the insight gained from studying the protocol dynamics with the proposed performance evaluation framework. Using the proposed framework, we are able to conduct a series of systematic evaluations on the performance of various scenarios in which the performance is affected by a mixture of protocol operation parameters and environment factors such as wireless losses, collisions, and different background traffic load levels. In the first case study, we examine the mixed throughput anomaly effects in hybrid IEEE 802.11b and IEEE 802.11g networks, where various settings that are necessary to accommodate the heterogeneous environment all jointly affect the system performance. We demonstrate that our performance framework not only helps predict the aggregate system performance in hybrid 802.11b/g networks, but also clarifies and isolates the mixed effects among different contention window, different backoff stages, and different data rates and frame formats on these two versions of standard. In the second case study, we evaluate the effectiveness of various performance improvement schemes at the physical layer and MAC layer in understanding the performance limitations on throughput and total system delay of IEEE 802.11 MAC with arbitrary competing traffic. By using the proposed framework, we are able to decompose the effects of various performance improvement schemes and the corresponding interactions with the variable amount of network delay incurred by collisions, and the backoff and re-transmission procedures. We further identify a performance bottleneck beyond which the packet delay becomes infinitely high and we pinpoint the exact turning point depends on the packet arrival pattern in consideration. Furthermore, in the third case study, we first use the framework to analyze the performance rate adaptation algorithms and discover that a good rate adaptation algorithm needs to dynamically adjusting the rate selection decisions with respect to different background traffic levels, in addition to the wireless channel fluctuations that was believed to be the sole key of rate adaptation. We then apply the insight observed from the proposed framework to design a rate adaptation algorithm that accommodate the mixed effects from different background traffic levels and fluctuating wireless channel conditions. While simulation results show promising improvements over the existing rate adaptation schemes, our real-world test-bed implementations also demonstrate that our framework is suitable for being practically incorporated into real hardware to provide superb performance in real-world scenarios. We expect that the proposed framework can be further generalized to evaluate the cross-layer effects such as TCP and routing in wireless mesh networks.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Shao-Cheng Wang.
Thesis: Thesis (Ph.D.)--University of Florida, 2008.
Local: Adviser: Helmy, Ahmed H.

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2008
System ID: UFE0022540:00001


This item has the following downloads:


Full Text

PAGE 1

1 A UNIFIED FRAMEWORK FOR PERFORMANC E ANALYSIS OF CONTENTION-BASED WIRELESS MAC: CASE STUDIES ON QOS, HETEROGENEITY AND RATEADAPTATION IN 802.11 By SHAO-CHENG WANG A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLOR IDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2008

PAGE 2

2 2008 Shao-Cheng Wang

PAGE 3

3 To my parents and my dearest wife.

PAGE 4

4 ACKNOWLEDGMENTS First and forem ost, my sincerest gratitude goes to my advisor, Professor Ahmed Helmy, for his invaluable advice, support, inspiration, and encouragement during the course of this work. Without the numerous discussions, brainstorm ing, and his acuminous insight and framework thinking on computer network research, this dissertation would never have existed. I am grateful to my dissertation committee me mbers, Dr. Shigang Chen, Dr. Sartaj Sahni, Dr. John Shea, and Dr. Ye Xia, for their valuable suggestions that finally bring the completion of this dissertation. I also thank my NOMADS group colleagues, both in University of Florida and in University of Southern California, incl uding Wei-Jen Hsu, Udayan Kumar, Sungwook Moon, Jeeyoung Kim, Yibin Wang, Sapon Tanachaiwiwat, Jabed Faruque, Ganesha Bhaskara, Shamim Begum, Shirin Ebrahimi-Taghizadeh, Fan Bai, Narayanan Sadagopan, and Yongjin Kim. I thank them for their friendship, the enjoyable resear ch environment, and many useful discussions. Last but not least, I want to thank my famil y. I want to express my deepest thanks to my wife, Ho-Chih (Grace) Chuang, for her love, support, and lovely encouragements. Her companionship and strong faith in me help me achieve this far. I am also thankful to my parents and my brothers for their full support and love du ring the course of my gr aduate studies abroad. My special thanks also go to my church family, both in Gainesville and in Los Angeles, for their spiritual support and cheer-leading.

PAGE 5

5 TABLE OF CONTENTS page ACKNOWLEDGMENTS...............................................................................................................4 LIST OF TABLES................................................................................................................. ..........8 LIST OF FIGURES.........................................................................................................................9 ABSTRACT...................................................................................................................................13 CHAP TER 1 INTRODUCTION..................................................................................................................16 1.1 Motivation .....................................................................................................................17 1.2 Contributions .................................................................................................................19 1.3 Organization of the docum ent....................................................................................... 22 2 BACKGROUND AND RELATE D WORK.......................................................................... 24 2.1 IEEE 802.11 m edium access control............................................................................ 24 2.1.1 Distributed Coordinati on Function (DCF) ........................................................ 25 2.1.2 Enhanced Distributed Channel Access (EDCA) ...............................................27 2.2 IEEE 802.11 a/b/g and protection m echanisms for hybrid 802.11b/g network............ 28 2.2.1 IEEE 802.11 a/b/g .............................................................................................28 2.2.2 Protection Mechanism for Hybrid 802.11b/g Networks...................................29 2.3 IEEE 802.11 perform ance modeling............................................................................. 30 2.3.1 IEEE 802.11 Saturation Perform ance Analysis................................................ 31 2.3.2 IEEE 802.11 Non-saturation Perform ance Analysis......................................... 36 2.3.3 Summ ary of existing IEEE 802.11 perfor mance modeling approaches and distinctions of our proposed approach.............................................................. 40 3 IEEE 802.11 PERFORMANCE ANALYSIS FRAMEWORK............................................. 43 3.1 Analysis m odel.............................................................................................................. 43 3.1.1 MAC layer service tim e.................................................................................... 44 3.1.2 Perform ance results........................................................................................... 51 3.1.3 Summ ary...........................................................................................................52 3.2 Understanding the param eters in the analysis model.................................................... 54 3.2.1 Perform ance affecting paramete rs and the target scenario............................... 54 3.2.2 Categorizing perform ance affecting parameters............................................... 59 3.3 UF-PASS: Unified Fram ework for Perfor mance AnalySiS of contention-based IEEE 802.11 MAC........................................................................................................62 3.3.1 Fram ework overview........................................................................................ 64 3.3.2 Exam ples of performance analysis using UF-PASS framework...................... 67

PAGE 6

6 4 PERFORMANCE EVALUATIONS FOR HYBRID IEEE 8 02.11B AND 802.11G WIRELESS NETWORKS......................................................................................................70 4.1 Introduction ................................................................................................................... 70 4.2 Related work .................................................................................................................73 4.3 Perform ance analysis model for hybrid 802.11b/g networks....................................... 74 4.3.1 Perform ance analysis model using UF-PASS framework................................ 74 4.3.2 Saturation perform ance model for hybrid 802.11b/g networks........................ 76 4.3.3 Discussions ........................................................................................................ 79 4.4 Perform ance evaluation using Markov-chain based saturation model......................... 83 4.4.1 Model validations .............................................................................................. 85 4.4.2 Effects of interoperability ................................................................................. 86 4.4.3 Effects of da ta packet sizes...............................................................................88 4.4.4 Effects of da ta rates...........................................................................................90 4.4.5 Discussion ......................................................................................................... 90 4.4.6 Packet delay perform ance.................................................................................93 4.5 Conclusion ....................................................................................................................94 5 PERFORMANCE LIMITS AND ANALYSIS OF CONTE NTION-BASED IEEE 802.11 MAC...........................................................................................................................96 5.1 Introduction ................................................................................................................... 96 5.2 Background and overview of analysis m odel...............................................................99 5.2.1 Background .....................................................................................................100 5.2.2 Overview of analysis m odel............................................................................100 5.3 IEEE 802.11 MAC Perform ance limit analysis.......................................................... 104 5.3.1. MAC Layer Packet Service Tim e...................................................................104 5.3.2. Theoretical Throughput Lim it......................................................................... 105 5.3.3. Theoretical Delay Lim it.................................................................................. 106 5.3.4. Model Validation with TGn Usage Scenarios ................................................109 5.3.5. Summ ary of theoretical perf ormance limitation analysis............................... 109 5.4 Discussion ...................................................................................................................110 5.4.1. Effects of Com peting Traffic Pack et Data Rates and Payload Sizes.............. 110 5.4.2. Perform ance Improvements on Frame Bursting and Block Acknowledgement..........................................................................................112 5.5 Conclusion ..................................................................................................................114 6 BEWARE: BACKGROUND TRAFFIC-AWAR E RATE ADAPT ATION FOR IEEE 802.11 MAC.........................................................................................................................116 6.1 Introduction ................................................................................................................. 117 6.2 Related W ork.............................................................................................................. 119 6.2.1 Existing Rate Adaptation Algorithm s.............................................................119 6.2.2 Loss Differentiation For Rate Adaptation ....................................................... 122 6.3 Perform ance of Rate Adaptation Al gorithms with Background Traffic..................... 123 6.3.1 IEEE 802.11 Rate Adaptation with Differe nt Level of Background Traffic .. 123 6.3.2 Perform ance of RAAs in RTS Access Mode..................................................125

PAGE 7

7 6.3.3 Perform ance of Collision-Aware Rate Adaptation (CARA).......................... 129 6.4 BEWARE Design ....................................................................................................... 131 6.4.1 Expected Packet Transmission Tim e and IEEE 802.11 Rate Adaptation....... 132 6.4.2 Rate Selection Engine .....................................................................................136 6.5 Simulation Results......................................................................................................139 6.5.1 Simulation Setup............................................................................................. 139 6.5.2 Perform ance Of Single Station With Varying Distance................................. 140 6.5.3 Perform ance Of Single Station With Dynamically Changing Background Traffic 141 6.5.4 Aggregated Perform ance w ith Different Topology........................................142 6.5.5 Aggregated Perform ance Under Va rious Channel Fading Conditions...........144 6.5.6 Perform ance With Heteroge neous RAA Deployments..................................146 6.6 Test-bed Implem entation and Evaluation................................................................... 148 6.6.1 Implementation...............................................................................................148 6.6.2 Experim ental Setup.........................................................................................150 6.6.3 Test-Bed Perform ance Evaluation Result....................................................... 154 6.6.4 Summ ary of Experimental Results..................................................................163 6.7 Conclusion ..................................................................................................................164 7 CONCLUSION AND FUTURE RESEARCH DIRECTIONS............................................ 165 7.1 Summ ary of the Dissertation....................................................................................... 165 7.2 Future Research Directions .........................................................................................167 APPENDIX A BEWARE IMPLEMENTATION......................................................................................... 169 A.1 Statistics Collection Module ....................................................................................... 169 A.2 Expected Packet Transmission Tim e Module............................................................. 170 A.3 Rate Probing Module ..................................................................................................170 A.4 Rate Selection Module ................................................................................................170 LIST OF REFERENCES.............................................................................................................172 BIOGRAPHICAL SKETCH.......................................................................................................179

PAGE 8

8 LIST OF TABLES Table page 2-1 Timing parameters of 802.11a, 802.11g, and 802.11b standard........................................ 29 2-2 Common drawbacks of existing IEE E 802.11 performance modeling approaches........... 42 3-1 Summary of parameters in IEEE 802.11 MAC analysis model........................................ 53 3-2 TGn usage models in high perform ance networks............................................................57 3-3 Categorization of perform ance affecting parameters......................................................... 61 4-1 Fram e parameters of 802.11g and 802.11b st andard used in this chapter......................... 84 4-2 Comparisons between throughput attained fr om field measurements and analysis model 85 4-3 Perform ance affecting parameters in pure 802.11g and hybrid 802.11b/g network scenarios 87 5-1 Comparison of packet delay from model and simulations............................................... 109 6-1 Perform ance in a crowded campus caf........................................................................... 163

PAGE 9

9 LIST OF FIGURES Figure page 2-1 Backoff procedure of IE EE 802.11 DCF...........................................................................25 2-2 IEEE 802.11e EDCA ......................................................................................................... 27 2-3 Classification of existing IEEE 802.11 perform ance modeling studies............................ 31 2-4 Markov-chain m odel and key equations in [4].................................................................. 33 3-1 IEEE 802.11 MAC analysis m odel....................................................................................44 3-2 Packet transm ission and collision events during IEEE 802.11 MAC backoff.................. 45 3-3 Average busy slot length, Tbusy ,with different payload si zes and transmission rates........55 3-4 In addition, we can also see other param eters such as Tfail and Tsucc also show similar effect on MAC layer service time...................................................................................... 55 3-5 Channel busyness ratio P busy and number of saturated transmission nodes in 802.11b and 802.11a standard........................................................................................... 56 3-6 Average M AC layer service time with different Pbusy (with Ploss=0, Tbusy=0.25ms, Tfail=1ms, Tslot=0.02ms), different Ploss (with Pbusy=0) and Pfail (with Pbusy=0.3)...............58 3-7 UF-PASS fr amework overview.........................................................................................63 3-8 The UF-PASS fra mework showing the interc onnections between target scenarios and performance affecting pa rameter categorizations.............................................................. 65 3-9 Effects of Pl oss on average IEEE 802.11 MA C layer service time and throughput......... 69 3-10 Effects of mixed data rate on aver age IEEE 802.11 MAC layer service time and throughput ..........................................................................................................................69 4-1 Markov chain m odel of back-off window size.................................................................. 77 4-2 Pbusy of 802.11g station and number of saturated transmission nodes in pure 802.11g and hybrid 802.11b/g networks..........................................................................................80 4-3 Field m easurement testbed configurations......................................................................... 84 4-4 Saturated T hroughput from analysis model (lines) and the simu lations (symbols)........... 85 4-5 Throughput anom aly in pure 802.11g and hybrid 802.11b/g network scenarios.............. 88 4-6 Effects of pa cket sizes in hybrid 802.11b/g networks.......................................................88

PAGE 10

10 4-7 Effects of pa cket sizes and ra tio of 802.11b/g transmission time......................................89 4-8 Effects of da ta rate and ratio of 802.11b/g transmission time...........................................90 4-9 Thought of 802.11b/g mixed networ k under different policies .........................................92 4-10 Packet delay perform ance in pure 802.11g and hybrid 802.11b/g network scenarios......94 5-1 Average MAC layer service tim e of different 802.11 specifications.............................. 104 5-2 Theoretical throughput lim it of different 802.11 specifications...................................... 106 5-3 Theoretical delay lim it of different 802.11 specifications............................................... 107 5-4 Theoretical delay lim it of 802.11e AC_VI with different arrival process....................... 108 5-5 Average pac ket delay limit with competing tr affic operates at different data bit rates and different payload sizes............................................................................................... 111 5-6 Theoretical throughput of Fra me-bursting and Block ACK schemes with different burst sizes 114 6-1 Throughput versus distance for IEEE 802.11a data......................................................... 124 6-2 Throughput versus distance for IEEE 802.11a da ta rates, with 12 background traffic stations 125 6-3 Best available data rate under different num ber of background traffic stations.............. 126 6-4 Throughput com parison for RAA-enabled station with RTS loss differentiation at 2.5m away from access point, with various num ber of background traffic stations in RTS access mode.............................................................................................................127 6-5 Throughput com parison for RAA-enabled station with RTS loss differentiation at 12.5m away from access point, with various number of background traffic stations in RTS access mode.............................................................................................................127 6-6 Normalized throughput for ARF-RTS, with various number of background traffic stations in RTS access mode............................................................................................ 128 6-7 Data ra te selection for ARF-RTS and oracle-selection strategy with various number of background traffic stations in RTS access mode......................................................... 129 6-8 Throughput com parison for ARF, CARA1, and Best ( oracle-selection strategy) with 12 background traffic stations in basic access mode....................................................... 131 6-9 Fram e error probability ( Pfail) and the MAC layer service time of 36Mbps and 24Mbps, no background traffic........................................................................................ 133

PAGE 11

11 6-10 MAC layer service tim e of 36Mbps and 24Mbps, various background traffic level......134 6-11 Throughput ratio (36Mbps over 24Mbps) unde r various background traffic level, Tbusy = 0.334ms................................................................................................................135 6-12 Throughput ratio (36Mbps over 24Mbps) unde r various background traffic level, Tbusy = 0.224ms................................................................................................................136 6-13 Structure of BEWARE design .........................................................................................137 6-14 Throughput com parison for Best ( oracle-selection strategy ), BEWARE, CARA1, and ARF with RTS/CTS, with 12 backgr ound traffic stations in RTS access mode....... 139 6-15 Data ra te selections for BEWARE and oracle-selection strategy, with various number of background traffic st ations in RTS access mode............................................140 6-16 Data rate selection for B EWARE and ARF-RTS with dynamically changing background traffic payload size....................................................................................... 140 6-17 Aggregate throughput com parison for BEWARE, CARA1, ARFRTS, and ARF in close-by topology with various number of contending stations...................................... 143 6-18 Aggregate throughput com parison for BE WARE, CARA1, ARF-RTS, and ARF in random topology with various number of contending stations........................................ 143 6-19 Aggregate throughput com parison for BEWARE, CARA1, ARF with RTS, and ARF in random topology under different Ricean Parameter K................................................ 144 6-20 Aggregate throughput com parison for BEWARE, CARA1, ARF with RTS, and ARF in random topology under different Doppler Spread fm................................................. 145 6-21 Individual and Aggregate throughput im provement of BEWARE and ARF-RTS with various number of contending statio ns in heterogeneous deployments........................... 147 6-22 Indoor experim ents layout. (floor plan pr ovided by The Facilities Planning and Construction Division of University of Florida).............................................................. 151 6-23 Outdoor experim ents layout............................................................................................. 153 6-24 Normalized throughput for BEWARE ove r ARF and BEWARE over ARF-RTS in controlled indoor environm ent with number of b ackground traffic stations................... 156 6-25 UDP perfor mance of BEWARE, ARF, and ARF-RTS with switching traffic pattern at different locations in the indoor controlled environment.............................................158 6-26 Individual and aggregate perform a nce of BEWARE and ARF-RTS under heterogeneous deployments at different locations in the indoor controlled environment.................................................................................................................... .159

PAGE 12

12 6-27 Throughput perform ance for BEWARE, BEWARE with RTS, ARF, and ARF-RTS in controlled indoor envir onment with hidden terminal.................................................. 160 6-28 UDP perfor mance of BEWARE, ARF, and ARF-RTS at different locations in the outdoor controlled environment....................................................................................... 161

PAGE 13

13 Abstract of Dissertation Pres ented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy A UNIFIED FRAMEWORK FOR PERFORMANC E ANALYSIS OF CONTENTION-BASED WIRELESS MAC: CASE STUDIES ON QOS, HETEROGENEITY AND RATEADAPTATION IN 802.11 By Shao-Cheng Wang August 2008 Chair: Ahmed Helmy Major: Computer Engineering Performance analysis for IEEE 802.11 based prot ocols is essential to not only better understand the protocol, but also provide insight in developing protocol improvement schemes. In this dissertation we propose a unified IEEE 802.11 MAC performance evaluation framework to systematically explore different facets of pr otocol behavior and the corresponding inherent interactions between different factors of various scenarios and system performance. Our proposed framework differs from the existing pe rformance analysis approaches in that it categorizes the protocol operation parameters that exhibit si milar behavior into the same group and uses the categorizations to interconnect th e target scenarios with relevant performance affecting parameters. As a result, such weaving interconnections enable us to formulate the unified performance evaluation framework in expl aining the similarity or dissimilarity of the performance impacts on different target scenar ios, which can not be achieved by any existing performance evaluation models. Additionally, we ai m to provide guidelines for various protocol enhancement designs by applying the insight gain ed from studying the protocol dynamics with the proposed performance evaluation framework. Using the proposed framework, we are able to conduct a series of sy stematic evaluations on the performance of various scenarios in whic h the performance is affected by a mixture of protocol operation parameters and environment factors such as wireless losses, collisions, and

PAGE 14

14 different background traffic load levels. In the first case study, we examine the mixed throughput anomaly effects in hybrid IEEE 802.11b and IEEE 802.11g networks, where various settings that are necessary to accommodate the heterogeneous environment all jointly affect the system performance. We demonstrate that our perf ormance framework not only helps predict the aggregate system performance in hybrid 802.11b/g ne tworks, but also clarif ies and isolates the mixed effects among different contention window, di fferent backoff stages, and different data rates and frame formats on these tw o versions of standard. In th e second case study, we evaluate the effectiveness of various performance impr ovement schemes at the physical layer and MAC layer in understanding the performance limitation s on throughput and total system delay of IEEE 802.11 MAC with arbitrary competing traffic. By using the proposed framework, we are able to decompose the effects of various performan ce improvement schemes and the corresponding interactions with the va riable amount of network delay incu rred by collisions, and the backoff and re-transmission procedures. We further identify a performance bottleneck beyond which the packet delay becomes infinitely high and we pi npoint the exact turning point depends on the packet arrival patter n in consideration. Furthermore, in the third case study, we first use the framework to an alyze the performance rate adaptation algorithms and discover that a good rate ad aptation algorithm needs to dynamically adjusting the rate selection decisions with respect to diffe rent background traffic levels, in addition to the wireless channel fluctuati ons that was believed to be the sole key of rate adaptation. We then apply the insight observed from the proposed framework to design a rate adaptation algorithm that accommodate the mixe d effects from differe nt background traffic levels and fluctuating wireless channel condi tions. While simulation results show promising improvements over the existing rate adapta tion schemes, our real-world test-bed

PAGE 15

15 implementations also demonstrate that our framework is suitable for being practically incorporated into real hardware to provide s uperb performance in real-world scenarios. We expect that the proposed framewor k can be further generalized to evaluate the cross-layer effects such as TCP and routing in wireless mesh networks.

PAGE 16

16 CHAPTER 1 INTRODUCTION W ith the large-scale deployments of wireless local area networks (WLANs) in homes, offices, and public areas, the IEEE 802.11 standa rd has become the dominant technology in providing low-cost high-bandwid th wireless connections. The IEEE 802.11 protocol is also a popular choice in various wireless mobile comm unication networks such as wireless mesh networks and mobile infostation networking paradigm [65]. Performance analysis for IEEE 802.11 based protocols has been an im portant and popular topic as it provides better understanding of the protocol. By analyzing the effects of different pr otocol parameters and impacts of various environment factors, performa nce analysis helps pinpoint the bottleneck of protocol operations and achieve better quality of service. It is a part icularly essential and challenging task to build a performance evalua tion framework that no t only provides better understanding of the protocol but also systemat ically explores different facets of protocol behavior and provides insight to guide the engineers in de veloping protocol improvement schemes. The objective of this di ssertation is to propose an insightful IEEE 802.11 MAC performance evaluation framework that systemati cally and coherently en compasses the protocol dynamics under diversified network operation scen arios. We do not want to provide yet another performance evaluation study to I EEE 802.11 MAC, which analyzes isolated case studies by special customized models without gi ving coherent viewpoints to various scenarios investigated. Our proposed framework aims to of fers one unified performance analysis model that provides thorough understandin g of the inherent interactions between different factors of various scenarios and system performance. Addi tionally, we will demonstrate that, by applying

PAGE 17

17 the insight gained from studying the protocol dynamics, the proposed performance evaluation framework provides guidelines for various protocol enhancement designs. 1.1 Motivation There have been quite a few perfor man ce modeling studies for IEEE 802.11 MAC. By assuming constant collision probability, Bianchi [4] developed a pilot model based on Discretetim e Markov Chain which provides very accurate performance predictions at asymptotical saturation conditions, i.e., each station is assume d always have packets to send. On the other hand, motivated by the need to prov ide analysis for more realistic traffic scenarios, later studies focused on performance analysis for unsaturated conditions. This type of models usually involves complex mathematical tec hniques such as Markov chains [57] [1] or Transform Theory [92] [76] to deal with the co m plicated backoff procedures of IEEE 802.11 MAC and queuing dynamics of packet delay modeling. Unsaturated performance modeling studies enable more thorough packet delay analysis to incorporate various traffic patte rns in real-world scenarios. In addition to performance evaluations and pr edictions on the effect s of basic protocol parameters specified in the st andard, numerous models have followed Bianchis approach and extended the results to cover advanced MAC mechanisms such as IEEE 802.11e EDCA [66] [50], and realistic network scen arios, such as lossy wi reless channel conditions [25]. Furthermore, som e studies focused on more complex scenario s such as heterogeneous environment where different stations operate on different data rates [27] and different backoff settings [88]. Other studies even extended the m odel to investigate system perfor mance under on-the-fly protocol improvement mechanisms such as rate adaptati on algorithms that dynamically adjust data rates for better performance [68] [17]. Despite the f act that both saturated and unsaturated perfor mance modeling studies have enabled accurate prediction and various pe rformance improvement schemes of IEEE 802.11

PAGE 18

18 MAC protocol, models from both approaches suffer from several common drawbacks. Firstly, there has been a lack of a coherent view of the effects on system performance from different factors in various scenarios. Even though there are a number of studies in both approaches built upon the same base models, the ways in which exte nd the base models are significantly different. As a result, even when the target scenarios or parameters of different studies share some same common effects on system performance, differe nt studies do not shar e or yield the same viewpoint due to significantly different mode l extension approach. It is important that performance analysis models should provide cohere nt insight to the system dynamics so that we can get a complete picture of the performance imp acts of various problem scenarios. However, it is also a very challenging task to build such a model to clearly unfold the complicated backoff dynamics and track down the traces of the significantly variable duration incurred by various environment factors such as wireless losses, co llisions, and different background traffic load levels. Secondly, as most studies are built for performance evaluation or prediction purposes, they require certain information to be known prior to the analysis, such as num ber of stations in the network or data rates used by the stations. Howeve r, in real-world scenarios, such information might not be available or is constantly changing. Therefore, the requirements of such information have restrained the models from being applie d to real-world algorith ms and applications. Furthermore, the complicated mathematical repres entations employed by most studies, especially in some studies that do not have close-form solutions but rely on numerical methods for approximations, also impose serious constraint s on applying these models to real-world implementations. Nonetheless, avoiding these design pitfalls is a nontrivial task, as we will show later in this dissertation.

PAGE 19

19 The aforementioned shortcomings motivate this dissertation. To our best knowledge, little effort has been made to come up with an insightful IEEE 802.11 MAC performance evaluation framework that provides a coherent view to the effects of differ ent factors on system performance. We aim to offer one unified perfor mance analysis model that can help to get a complete and harmonious picture about the behavior of the protocol, and not be limited by the unconnected results from most previous studies which have been independently focused on tackling different problem scenarios. In addi tion, we also attempt to go one step beyond traditional performance evaluation models. We try to design a model that can also be practically implemented into real-world algorithms in he lping the network system to perform more efficiently. We believe that, by f acilitating lightweight mathematical representations and utilizing only locally available information, our proposed model can meet this challenging goal. 1.2 Contributions Our main contribution is to present UF-PASS: A Unified Framework for Performance AnalySiS of contention-based IEEE 802.11 MAC to achieve the objectives we state above. With the detailed design presented in Ch. 3, our proposed framework offers one unified performance analysis model that provides thorough understa nding of the inherent interactions between different factors of various scenarios and system performance. By interconnecting the target scenarios with relevant protocol operation para meters, the proposed analysis model does not only provide accurate prediction to the system pe rformance, but also helps pinpoint the key performance affecting factors in different target scenarios. In Ch. 4 and Ch. 5, we further conduct a systematic evaluation to the proposed fram ework by several case st udies with detailed quantitative analysis. Such ev aluations aim to provide more thorough understanding of the protocol and a coherent vi ew of the effects of various scenarios on IEEE 802.11 MAC performance. In Ch. 6, we also utilize the insi ght obtained from the eval uation case studies to

PAGE 20

20 build algorithms that dynamically adapt the operation data rate to fluctuati ng environment factors. We further conduct real-world experimentations and implementations of the proposed rate adaptation algorithm to verify our findings in the evaluation framework. More specifically, this dissertation has made the following contributions: (1) We propose an analysis model for IEEE 802.11 MAC backoff procedure. Our approach is to decompose the MAC layer timing dynamics in to stages and identify the key performance affecting components in such stag e settings. We then incorporate the effects of these parameters and derive the expected time duration of IEEE 802.11 MAC backoff. The resultant model not only provides accurate prediction to the system pe rformance but also offers a convenient way to qualitatively estimate the impacts of certa in parameters on protocol performance. (2) Thanks to the stage decomposition approach of the analysis model, we are able to categorize the protocol operation parameters that exhibit simila r behavior into the same group and use the categorizations to interconnect the target scenarios with relevant performance affecting parameters. As a result, such weaving interconnections enable us to formulate the unified performance evaluation framework in expl aining the similarity or dissimilarity of the performance impacts on different target scenar ios, which can not be achieved by any existing performance evaluation models. (3) In the first case study (Ch. 4), we tack le the mixed throughput anomaly effects in hybrid IEEE 802.11b and IEEE 802.11g networks. The challenges are that in hybrid 802.11b/g networks, the various settings that are necessary to accommodate the heterogeneous environment, all jointly affect the system performance. We demonstrate that our pe rformance framework not only helps predict the aggregat e system performance in hybrid 802.11b/g networks, but also clarifies and isolates the mixed effects among different contention wi ndow, different backoff

PAGE 21

21 stages, and different data rates and frame fo rmats on these two versions of standard. We accurately quantify the throughput anomaly that penalizes fast 802.11g stations and privileges the slow 802.11b station. We al so discuss the effects and reasons of such anomaly under different scenarios. We learn that as the ratio of transmission durations of 802.11b and 802.11g approaches 1:2, the system throughput is more balanced by accommodating the 2:1 contention window setting. We also discuss the pros and cons of different performance improvement schemes that alleviate th e throughput anomaly effect. (4) In the second case study (Ch. 5), we evaluate the effectiveness of frame headers in high data rate scenarios. We focus on understanding the performance limitations on throughput and total system delay of IEEE 802.11 MA C with arbitrary competing traffic. The challenges to such analysis are in modeling the effects of various performance improvement schemes at the physical layer and MAC layer and the co rresponding interactions with th e variable amount of network delay incurred by collisions, and the backoff and re-transmission procedures under different background traffic load level arbi trary competing traffic. By using the proposed framework, we are able to systematically examine and provide a unified explanation to the effects of background traffic intensity, data bit rates, and payload sizes of the competing traffic on theoretical limits of IEEE 802.11 MAC throughput and delay performan ce. We identify a performance bottleneck beyond which the packet delay becomes infinitely high and we pinpoint th e exact turning point depends on the packet arriva l pattern in consideration. (5) In the third case study (Ch. 6), we use our proposed framework to evaluate the rate adaptation algorithms proposed by previous studie s in dynamically swit ching data rates to accommodate the fluctuating wireless channel cond itions. We make a painstaking investigation into the effects of wireless channel conditions and competi ng background traffic on rate

PAGE 22

22 adaptation performance. Thanks to the systema tic structure of the proposed framework, we are able to pinpoint the problems of the loss different iation technique that have been believed to be the best solution for this sett ing. We uncover that the fundame ntal challenge for a good rate adaptation algorithm is in dynamically adjusting the rate selection deci sion objectives with respect to different background traffic levels, in addition to be responsive to random and even drastic wireless channel fluctuati ons. We then propose to use our analytical model as the engine in selecting the best operating data rate influen ced jointly by wireless and collision losses. Our simulation results show that the proposed algorit hm outperforms other rate adaptation algorithms without loss differentiation by up to 250% and th at with loss differentiation by up to 25% in throughput. In addition, by implemen ting the proposed rate adapta tion algorithm into real-world drivers, we demonstrate that our framework is also suitable for be practically incorporated into real-world algorithms. Through a series of real-world test-bed experimentations, we further show that the performance of the proposed rate adaptation algorithm outpe rforms the existing algorithms under the effects of real -world scenarios such as capture effect, channel fading effect, and etc. 1.3 Organization of the document The rest of the docum ent is organized as follows: Chapter 2 provides background information of the basic operations of IEEE 802.11 MAC and parameter settings that are relevant to the analysis. In addition, we also br iefly review the existing IEEE 802.11 performance modeling studies in the literat ure. In Chapter 3, we presen t the derivation of IEEE 802.11 analytical model and the proposed performance evaluation framework. In Chapter 4, we conduct the first case study in performance of hybrid IEEE 802.11b and IEEE 802.11g network. Chapter 5 presents the second case study of performa nce limitations of IEEE 802.11 MAC. In Chapter 6, we discuss the problems of existing rate ad aptation algorithms and propose our new rate

PAGE 23

23 adaptation algorithm design by embedding our analyti cal model as the rate selection engine. We conclude this dissertation and provide future research directions in Chapter 7.

PAGE 24

24 CHAPTER 2 BACKGROUND AND RELATED WORK Before we present the proposed IEEE 802.11 perform ance analysis framework and mathematical models, we first briefly revi ew the basic operations of IEEE 802.11 MAC and parameter settings that are releva nt to the analysis. In addition, we also briefly review the existing IEEE 802.11 performance modeling studies in the literature. 2.1 IEEE 802.11 medium access control In order to m ediate the access to the shared wireless medium among hosts in the network, the 802.11 standard specifies seve ral Medium Access Control (MAC) mechanisms to maximize system efficiency. The original IEEE 802.11 MAC defines two modes of operation the mandatory Distributed Coordinati on Function (DCF), a fully distributed binary backoff medium contention mechanism, and the optional Poin t Coordination Function (PCF), a centrally controlled polling mechanism. In DCF mode, each station senses the medium and performs a binary backoff algorithm before it sends any packets to avoid collisions Since all stations perform the same backoff procedure, there is no Quality of Service (QoS) guarantee in DCF. On the other hand, in PCF mode, the point coordinator (PC) is responsible for generating a contention-free schedule such that there is only one station transmits in a specific time pe riod. Although the PCF mode was originally designed to provision QoS for real-time traffic, previous studies [79] [90] have shown its poor bandwidth efficiency com pared to DCF. In addition, PCF is specified as an optional operation in the standard and may pose difficulties to interoperate with other access points, hence is not widely adopted in most current wireless cards. Therefore, in this dissertation, we focus on the analysis of IEEE 802.11 DCF and describe its operation in more details in Sec. 2.1.1.

PAGE 25

25 Furthermore, A QoS amendment of IEEE 802.11 MAC, IEEE 802.11e [37], was later proposed to provide service differentiations for IEEE 802.11 MAC. In this specification, a Hybrid Coordination Function (HCF) is introduc ed, which employs an Enhanced Distributed Channel Access (EDCA) conten tion-based medium access and a HCF Controlled Channel Access (HCCA) contention-free medium access. In particular, EDCA enhances the contentionbased backoff mechanisms of DCF with different medium access pr iorities to different traffic types such that packets with higher priority level have higher probability to be transmitted earlier than lower priority packets. In this disserta tion, we limit our scope to EDCA (not HCCA) and describe it in more details in Sec. 2.1.2, as it is backward compatible to the popular DCF and is more widely implemented in toda ys advanced network equipments. Figure 2-1. Backoff procedure of IEEE 802.11 DCF 2.1.1 Distributed Coordination Function (DCF) The DCF o f IEEE 802.11 is a listen-before-t alk medium access scheme based on the Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA) protocol. The backoff procedure of IEEE 802.11 DCF is illustrated in Fi g. 2-1. Before initiating any packet delivery, Frame Station A CWindow Backoff Defer Defer DeferStation B Station C Station D Station EDIFS Frame CWindow Defer Frame CWindow Frame Frame CWindow R T S S I F S C T S S I F S DATA S I F S A C K DATA S I F S A C KCWindow= Contention Window = Backoff = Remaining backoff time Frame Station A CWindow Backoff Defer Defer DeferStation B Station C Station D Station EDIFS Frame CWindow Defer Frame CWindow Frame Frame CWindow R T S S I F S C T S S I F S DATA S I F S A C K DATA S I F S A C KCWindow= Contention Window = Backoff = Remaining backoff time

PAGE 26

26 the station detects the wireless medium to be idle for a minimum durati on called DCF Interframe Space (DIFS). The station randomly selects the backoff timer interval from [0, CWmin] number of slot_time, where CWmin denotes the minimu m Contention Window ( CW ), and then enters the backoff process. Note that slot_time is a parame ter that depends on the underlying physical layer (PHY), and will be described in more details in Sec. 2.2. During the count-down of backoff timer, if the station senses the medi um busy, it stops decrementing the timer and does not reactivate the paused value until the channel is sensed idle agai n for more than a DIFS. At the timer expiration, the station is free to access the medium for pack et transmission. The transmitting station has two options to initiate the transmission: the default basic access mode and optional Request-To-Send/Clear-To-Send (RTS/CTS). In ba sic access mode, the sender station transmits the data (DATA) packet directly and the recei ver station sends back an acknowledgement (ACK) frame upon successful reception of the data pa cket. On the other hand, the RTS/CTS access mode institutes a four-way handshaking techniqu e: the sender station fi rst transmits a RequestTo-Send (RTS) frame, to be acknowledged by a Clear-To-Send (CTS) frame from the receiver station, and followed by the normal DATA/ACK sequence. Note that because all frames include the information of how long it takes to transm it the rest of the transmission sequence, any neighboring nodes overhear such information will de fer their transmission accordingly. Besides, the RTS-CTS frames further prevent the hidden terminal problem [46] by decimating such inform ation so that stations close to sender and receiver understand how long they should not attempt to transmit. Upon receiving an acknowledgement frame, the transmission is considered successful; the CW is reset to minimum CWmin and the station stands-by for the next packet arrival. The transmission is considered failed if no acknowle dgement is received within a specified timeout;

PAGE 27

27 the station repeats the backoff process with CW selection range doubled up to maximum contention window, CWmax. If the transmission has been re-tried for up to RetryLimit times, the packet is discarded and CW is reset to CWmin. Figure 2-2. IEEE 802.11e EDCA 2.1.2 Enhanced Distributed Channel Access (EDCA) The EDCF i s a variant of DCF and provides pr ioritized Quality-of-Service (QoS) support among different traffic types. Each QoS-enhanced station (QSTA) maps the packets arriving at MAC layer into four different access categories (A Cs) and assigns a set of backoff parameters, namely Arbitration IFS (AIFS), CWmin, and CWmax to each AC. As illustrated in Fig. 2-2, each AC uses its own backoff parameters to conte nd for the wireless medium by the same backoff rules as legacy DCF stations in the previous paragraph. The AIFS[AC], determined by AIFS[AC] = AIFSn[AC] slot_time + aSIFSTime, replaces the fi xed DIFS in DCF. Shorter AIFS[AC] in higher priority AC provides high pr iority traffic earlier timing to unfreeze the paused timer after each busy wait period. On the other hand, smaller CW sizes probabilistically provide shorter backoff stages high priority traffic. More detail ed description of DCF and EDCF can be found in [36] and [37], respectively. Busy medium Backoff Slots DIFS/AIFS SIFS PIFS DIFS AIFS[i] AIFS[j] = AIFSn[j] *Slot_time+SIFS Select slot and decrement backoff as long Next Frame Defer Access Slot_time as medium is idle Contention window2 15 7 AC_VO 2 31 15 AC_VI 3 1023 31 AC_BE 7 1023 31 AC_BK AIFSn CWmax CWmin AC Busy medium Backoff Slots DIFS/AIFS SIFS PIFS DIFS AIFS[i] AIFS[j] = AIFSn[j] *Slot_time+SIFS Select slot and decrement backoff as long Next Frame Defer Access Slot_time as medium is idle Contention window2 15 7 AC_VO 2 31 15 AC_VI 3 1023 31 AC_BE 7 1023 31 AC_BK AIFSn CWmax CWmin AC 2 15 7 AC_VO 2 31 15 AC_VI 3 1023 31 AC_BE 7 1023 31 AC_BK AIFSn CWmax CWmin AC

PAGE 28

28 2.2 IEEE 802.11 a/b/g and protection mechanisms for hybrid 802.11b/g network The IEEE 802.11b, 802.11a [35], and 802.11g [38] are higher-speed physical layer (PHY) extensions o f the IEEE 802.11 standard. They all use the same DCF medium access mechanism described in the previous section, with somewhat different parameter settings. We describe the different settings of IEEE 802.11a/b/g in Sec. 2.2 .1, and the special protection mechanism when IEEE 802.11b/g stations interoperate in the same network in Sec. 2.2.2. As we will show later in this dissertation, we will use the proposed perf ormance analysis framework to investigate the effects of different IEEE 802.11 variants pa rameters on MAC performance and related interoperability issues among IEEE 802.11 variants. 2.2.1 IEEE 802.11 a/b/g The detailed operational param eter settings of three high-speed extension of the standard are summarized in Table I. IEEE 802.11b is the earliest high-speed PHY extension in the 2.4GHz band. It employs 8-chip Complementar y Code Keying (CCK) modulation scheme to support 11 Mbit/s and 5.5 Mbit/s data rates in add ition to 2 Mbit/s and 1 Mbit/s supported in the original IEEE 802.11 standard. IEEE 802.11a is later proposed as the high-sp eed PHY extension in the 5GHz band. IEEE 802.11a employs an advanced orthogonal frequenc y division multiplexing (OFDM) system to support further higher data rate at 6, 9, 12, 18, 24, 36, 48, and 54 Mbit/s. In addition, it also reduces some DCF operational parameters such as SlotTime to 9us and minimum contention window ( CWmin) to 15 in order to provide more efficient medium access. However, due to the different operation band, modul ation schemes, and operational parameters, IEEE 802.11a is not compatible with IEEE 802.11b. Note that, IEEE 802.11e is the MAC QoS enhancement amendment to the IEEE 802.11 standard and can be incorporated with any of the three higher-

PAGE 29

29 Table 2-1. Timing parameters of 802.11a, 802.11g, and 802.11b standard speed PHY extensions. In Ch. 5, we will use our proposed framework to systematically study the impacts of these PHY and MAC layer performance enhancement amendments. 2.2.2 Protection Mechanism for Hybrid 802.11b/g Networks Since 802.11g utilizes a different m odulation scheme when it ope rates at the data rates not supported by 802.11b, the 802.11b stations cannot dec ode these frames. As a result, the 802.11b stations may interfere with packets tran smitted by 802.11g stations and cause unfavorable performance degradation. A protection mechan ism is mandated when the access point (AP) senses that there are both types of stations associated with the ne twork. The protection mechanism bit in the beacon will be set for notifying 802.11g stations to transmit 802.11b decodable control frames before the DATA frame. As a result, the transmitting 802.11g stations can successfully reserve the medium and subsequently switch to th e higher data rates that only 802.11g stations understand during the DATA and ACK frames to maximize the throughput. There are two such protection mechanisms sp ecified in the 802.11g standard, namely RTSCTS and CTS-to-self. CTS-to-self is the mi nimum requirement by the standard, which only 802.11a 802.11g (pure/hybrid) 802.11b Band 5GHz 2.4GHz 2.4GHz SlotTime 9 s 9 s/20 s 20 s SIFS 16 s 16 s/10us 10 s DIFS 34 s 34 s/50 s 50 s Tp 16 s 16 s/72 s* 72 s* TPHY 4 s 4 s/24 s 24 s CWmin 15 15 31 Supported Data Bit Rate (Mbps) 54, 48, 36, 24, 18, 12, 9, 6 54, 48, 36, 24, 18, 12, 9, 6, 11, 5.5, 2, 1 11, 5.5, 2, 1 Tp: transmission time of physical preambles TPHY: transmission time of PHY header *short preamble mechanism

PAGE 30

30 requires the packet sender transmit a CTS packet w ith destination to itself to reserve the medium. This mechanism lives under the risk of hidden terminal problem that some other stations may not see the CTS frame. Since the protection frames have to be sent at the 802.11b decodable rate, which can be much slower than 802.11g data rates, these frames may cause a significant overhead to network performance. In this dissertation, we will study the perfor mance impacts of the interoperations among different 802.11 versions in wireless local netw orks. We aim to use our proposed performance analysis framework to provide a unified view of the performance impacts from various operating parameters in different 802.11 vers ions, including contention windows ( CW ) frame header length, and among others. In Ch. 4, we will show that the proposed performance analysis framework helps us systematically identify and quantify pot ential problems in interoperability between 802.11b and 802.11g devices. 2.3 IEEE 802.11 performance modeling As shown in Fig. 2-3, existing IEEE 802.11 perf orm ance modeling studies can be broadly classified into two categories, namely satu ration and non-saturation pe rformance studies. The early IEEE 802.11 performance modeling studies in the literature mainly focus on performance analysis in saturated-traffic condi tions, i.e., each station is assu med to always have packets to send at any given time. The saturation throughput represents the maximum throughput a system can reach as the offered load increases. Such assumption also simplifies the analysis to only concentrate on system capacity a nd ignore the effects from the tr affic arrival characteristics. Despite the fact that this assumption is unlikely to be valid in many real -world scenarios, these types of studies provide signifi cant insight in protocol operati ons and configurations. Later studies were proposed to conduct non-saturation performance anal ysis in attempt to provide

PAGE 31

31 more thorough investigations on packet delay an alysis. In this section, we broadly review different modeling approaches in both classes as detailed in Fig. 2-3. We also comment on the contributions and limitations of these analyses and the applicability of the models beyond performance analysis. By the end of this secti on, we discussion the di stinctions between our proposal and existing studies. Figure 2-3. Classification of existin g IEEE 802.11 performance modeling studies 2.3.1 IEEE 802.11 Saturation Performance Analysis There have been several approaches in analyzing the IEEE 802.11 perform ance in saturation condition. We first review the popular Ma rkov-chain based appro ach and its variants in analyzing various complicated network scenar ios, such as erroneous wireless channel, heterogeneous stations, and EDCA service differentiation. We also briefly review other approximation approaches for saturation performance analysis. Markov-chain based models One of the early studies adopted by ma ny as canonical model for modeling IEEE 802.11 DCF performance was proposed by Bianchi in [3], and further detailed in [4]. With the assum ptions of ideal channel condition and fixed number of stations in the network, Bianchis model accurately computes the system throughput in saturation condition. As shown in Fig. 2-4, IEEE 802.11 Performance Modeling Saturation Non-saturation Markovchain based Approximation based Markovchain based Transferfunction based Approximation based IEEE 802.11 Performance Modeling Saturation Non-saturation Markovchain based Approximation based Markovchain based Transferfunction based Approximation based

PAGE 32

32 the model uses a bi-dimensional Markov process { s(t),b(t) } to characterize the binary exponential backoff procedure in DCF. B(t) represents the backoff c ounter of the station, which decrements at the start of every idle backoff slot and stops when the channel is sensed busy. When b(t) decrements to zero, the station tran smits the packet. On the other hand, s(t) tracks the contention window size of the backoff stage that b(t) operates. S(t) is reset to CWmin whenever the transmission is successful, and is doubled up to CWmax whenever the transmission is failed. The transition probabilities of the Markov ch ain are governed by the events accounting for backoff counter countdown and packet tr ansmissions, and expressed in terms of p, conditional collision probability and other backoff stage parame ters. As a result, the stationary distribution is formed as bi,k= limtP{s(t)=i,b(t)=k} In addition, the union of the states with b(t) =0, i.e., i=0 mbi,0, is in fact the probability that a station will transmit in a given slot. On the other hand, in a network of n stations, p can also be expressed by p = 1(1)n-1, the probability that at least one of the other (n1) stations transmit. Therefore, assumi ng the collision probability is constant and independent of the number of retransmission and backoff stages, p and can be obtained by solving a nonlinear system of equations. It follo ws that the system throughput can be found with p, and other constant system parameters such as packet transmission time. Wu et. al. [85] propose a modification to the bi-d im ensional Markov chain to improve the accuracy of Bianchis model wh en considering the frame retry limits specified in IEEE 802.11 standard. The actual number of st ages in the model is determined by the retransmission limit count, m, i.e., 4 for data frames ( dot11ShortRetryLimit ) or 7 for RTS frames ( dot11LongRetryLimit ). They re-arrange the Markov chain su ch that the state goes back to the initial backoff stage after the packet transmission exhausts all retransmission retries. The results

PAGE 33

33 in this study significantly improve the accuracy of the model, especially when the number of stations in the ne twork is large. Figure 2-4. Markov-chain model and key equations in [4] A num ber of papers have built on Bianchis basic model and extended its capability to investigate more complicated network scenarios. In particular, much attention was focused on extending Bianchis model to relax th e ideal channel condition assumption [25][14] [78]. For exam ple, Hadzi-Velkov et. al. [25] propose a simple modification to integrate wireless channel loss pf into the error probability p considered in the model. By assuming the collision events and wireless loss events are independent, p can be expressed as p = 1-(1-p1)(1-pf) where p1 is the i,0 i,1 i,2 i,Wi-2 i,Wi-1 0,0 0,1 0,2 0,W0-2 0,W0-1 m,0 m,1 m,2 m,Wm-2 m,Wm-1 i-1,0 1 1111 1 1111 1 1111 p p p 1-p 1-p 1-p i,0 i,1 i,2 i,Wi-2 i,Wi-1 0,0 0,1 0,2 0,W0-2 0,W0-1 m,0 m,1 m,2 m,Wm-2 m,Wm-1 p i-1,0 1 1111 1 1111 1 1111 p p p p 1-p 1-p 1-p p i,0 i,1 i,2 i,Wi-2 i,Wi-1 0,0 0,1 0,2 0,W0-2 0,W0-1 m,0 m,1 m,2 m,Wm-2 m,Wm-1 i-1,0 1 1111 1 1111 1 1111 p p p 1-p 1-p 1-p i,0 i,1 i,2 i,Wi-2 i,Wi-1 0,0 0,1 0,2 0,W0-2 0,W0-1 m,0 m,1 m,2 m,Wm-2 m,Wm-1 p p i-1,0 1 1111 1 1111 1 1111 p p p p 1-p 1-p 1-p i-1,0 1 1111 1 1111 1 1111 p p p p 1-p 1-p 1-p p = = = =+ ]1,0[ /1}0,|,0{ ],1[]1,0[/}0,1|,{ ]1,0[]1,0[/)1(}0,|,0{ ],0[]2,0[ 1}1,|,{0 0 0 0WkWmkP miWkWpikiP miWkWpikP miWk kikiPi i i 1 0 0,0 1 0,)1(1 1 1 = += ==n m i m ip b p p b

PAGE 34

34 original collision probability, i.e., p1 = 1(1)n-1. On the other hand, Ergen et. al. [22] and Yang et. al. [89] modify the expression of collision du ration and successful duration (equation 6 and 7 in [4]) to accommodate the scenarios when stations use different data rates in the network. Note that, the m odifications in [22] and [89] do not consider automatic da ta r ate adaptation schemes; rather they assume the station data rates are known a prior and fixed in the analysis. Furthermore, in [66] and [50], Bianchis model is extended to handle the QoS enabled EDCA with different priority tr affic catego ries. Robinson et. al. [66] use a new contention zone concept to deal with different collision ch aracteristics resulting from AIFS service differentiations among different access categories (ACs). Because high priority transmissions only see collisions from other high priority tr ansmissions, the conditional collision probability for high priority AC is calculated with the transm ission probability and number of stations that have same AIFS setting. For low priority transm issions, the conditional collision probability is calculated with all stations that have same or higher priority AIFS setting. Kong et. al. [50] propose to augm ent one extra dimension to Bianch is Markov chain model to handle the extra backoff waiting time of different ACs in EDCA. Similarly, Kongs model also requires knowing the number of stations in differe nt AC in-prior. Inanc et. al. [39] and Tao et. al. [73] also used 3D structure to augm ent the model to support service differentiation in EDCA and show comparably accurate prediction results. Note th at, however, the use of multi-dimensional Markov chains and other complex non-linear equation systems lead to hi gher computational complexity, and thus make it impractical to obtain real-tim e solutions for these models and prevent their implementation in real-time applications. On the other hand, several papers [13] [81] attempt to derive packet delay based on Bianchis model. Their key idea in finding average packet d elay is to calc ulate the expected time

PAGE 35

35 duration a packet transmission spent on certain level of backoff procedure, multiplied by the probability a packet transmission is completed in the corresponding backoff stage. However, because the fundamental saturati on assumption of Bianchis mode l, the packet delay derived from this approach only accounts for the time peri od in backoff procedure, i.e., access delay; but it does not account for the backl og delay. As a result, this approach does not find the important overall queuing delay for real-wor ld packet delivery procedures. Approximation-based models In order to avoid the complicated Markovchain development yet provide simple and effective way to model IEEE 802.11 MAC, ther e have been other approximation-based approaches for finding saturation throughput of IEEE 802.11 DCF. For example, Cali et. al. [11] propose to estim ate the IEEE 802.11 saturation thr oughput by the ratio betw een average message length and average virtual transmission time, tv, which includes successful transmission intervals, collision intervals, and idle periods due to backoff algorithm. However, the accuracy of this model is limited by the assumption that pack et length and sampled backoff interval are geometrically distributed. Tay et. al. [75] provide a very simple close-form approximation for the m aximum system throughput in saturation condition. This model approximates the transmission episode between any two transmissions as the sum of rate of transmissions and rate of collisions. The transmission episode length can also be appr oximated by the interval from the instance of last transmission to the instance any station that picks the earliest slot and breaks the channel silence. Furthermore, Chen et. al. [15] and Lin et. al. [55] extend the mean value analysis model in [75] for AIFS and CW differentiation. Thes e approxim ation-based models provide an alternative way to estimate the saturation performance of IEEE 802.11 MAC without involving complex Markov chain and non-linea r equation system calculations.

PAGE 36

36 Summary The saturation performance analysis models have enabled accurate prediction and thorough understandings of IEEE 802.11 MAC protocol pe rformance. However, the results from saturation performance analyses are unlikely to be applicable to real-world scenarios or applications such as admission control, because in real-world networks, stations often do not operate in saturation mode. In our work, by contra st, we aim to design an analytical model that can be used for practical applications with more realistic traffic patterns. In the next subsection, we review the other thread of IEEE 802.11 perfor mance analysis studies that focus on more realistic traffic patterns, in which more thorough investigation into packet delay is enabled. We will see that as our proposed model borrows some of the concepts from these models, we further extend beyond existing work and improve our model with objectives that other existing models have not achieved. 2.3.2 IEEE 802.11 Non-saturation Performance Analysis In light of studying the impact of m ore realistic traffic pattern on IEEE 802.11 performance, there have been a number of studies that attempt to derive mathematical models that relax the restriction of saturated-traffic operation. The general idea for IEEE 802.11 nonsaturation performance analysis is to use a queui ng system point of view to model the backoff stage behaviors of IEEE 802.11 protocol as queui ng service process. The non-saturated packet arrivals from upper layer are modeled as customer arrival process, and standard queueing theory techniques are employed [45] for performance resu lts. In this section, we categorize and review three popular approaches for IEEE 802.11 non-satura tion perf ormance analysis, namely Markovchain based models, transfe r-function based models, and approximation-based models. Markov-chain based models

PAGE 37

37 Several studies attempt to model non-satu ration traffic IEEE 802.11 MAC performance by adding extra states into Bianch is saturated th roughput model to accommodate situations when stations do not have packet to send. By assuming M/G/1 queuing model, Lee et. al. [53] add one extra idle state when a node finds the buffe r e mpty after a transmission, with probability pb = /, where represents and the Poisson arrival rate at each node and represents the mean service rate. The system throughput is then obtained in a similar fashion as Bianchis model. Malone et. al. [57] and Alizadeh-Shabdiz et. al. [1] introduce a new stage of states (0,k)e to represent postbackoff situations, where a station has transm itted a packet but has none waiting. Similarly, in order to apply standard M/G/1 results, the mode l also assumes a constant probability (1-q) that the stations buffer is empty, which is a ssociated with the Poisson arrival rate In [57], the authors attempt to provide an estim ation technique to apply the proposed model to other arrival patterns, but it sacrific es the accuracy of the model and i nvolves more complex computations. Cantieni et. al. [12] extend the model in [1] to discuss performance and fairness issue when stations use different data rate. Li et. al. [54] use approach similar to [53] to provide analysis with contention window service different iation. Furtherm ore, Engelstad et. al. [21] extend the analysis to contention window and AIFS servic e differentiations. However, as m ost Markovchain based non-saturation models are limited to Po isson arrivals, it follows that using the idle states in Markov chain to model non-saturated traffic lacks of a ge neric representation to relate general non-Poisson arrival pattern to the buffer empty probability. Transfer-function based models The key idea for transfer-function based models is to analyze the general distributions of some quantities in backoff stages such as collis ion and transmission probability. Once the service process distribution is obtained, one can use stan dard second-order queueing analysis techniques

PAGE 38

38 in obtaining delay performan ce results. Tickoo et. al. [76] break down the MAC service time calcu lation as the result of the following events: i) time length between two successful transmissions, including the pr obability generating function (pgf) of number of backoff slots BO(z) and delay due to other stations transmission X(z) and ii) the time required to transmit the packet, as the distribu tion of packet length L(z) As a result, the service time distribution, B(z) is given by B(z) = BO(z)X(z)L(z) Assuming the distribution of pack et arrivals is also known, the system performance results such as total syst em delay can be obtained by standard G/G/1 discrete time queueing theory [10]. On the other hand, Zhai et. al. [92] first express the pgf of the tim e length of collision and successful transmissions as Ct(z) and St(z) respectively. They also denote H(z) as the pgf for server when it is busy but not transmitting. As a result, the generalized system transfer function Q(z) can be characterized as a series of transmission and collision events through the backoff process service syst em with corresponding time length (i.e. Ct(z) and St(z) ) and combinations of server busy and idle probability with H(z) Similarly, once Q(z) is obtained, packet delay can be found by standard queueing theory techniques. Note that the model in [92] requires an iterative algo rithm to solve the re lationship between node no-transmission probability and service time distribut ion of M/G/1/K model. In [59], the IEEE 802.11 system is considered in two separate points of view: (i) the system seen by a user (ii) the system seen by the wireless medium. The user centric view provides a wider set of arrival/channel situations by modeling each users queue as a separate G/G/1 system. They derive the MAC layer acce ss delay probability distribution by collecting different delay sources, i.e., random backoff tim e, random number of collisions, and random number of successful transmissions. On the ot her hand, the system centric view provides a simple way to derive channel access rate and to tal delay by modeling the wireless medium as a

PAGE 39

39 central resource shared among random access users in one system cycle time [58]. For hom ogeneous and uniform user arrivals, the mo del reduces to an M/G/1/PS process sharing system. Although transfer-function based IEEE 802.11 non-sa turation performance models provide accurate performance analysis to a wide range of arrival patterns, this approach usually involves complicated calculations and constructions of event probability and durations in backoff service process. In addition, as derivati on the MAC layer service process are so complex, any changes in model assumptions or operating co nditions such as erroneous channe l or heterogeneous data rate will result in major revamp of the derivation, and are thus not scalable to general purpose analysis. Other approximation models Foh et. al. [24] propose to approximate the service probabi lity distribution of the m edium access protocol into a phase-type distribution, e.g. the Erlang distribut ion, and evaluate the protocol performance using Continuous Time Markov Chain (CTMC). The results show that Erlang random variable with parameter 32 (E32) can provide fairly accu rate prediction for the service time Cumulative Distribution Functi on (CDF) of a 50-node network. However, the proposed model only works with Markovian Arrival Process type of arrival patterns such as Markov Modulated Poisson Proce ss (MMPP) or general Poiss on Process. Tantra et. al. [72] further extend the m odel to include service differentiations in EDCA. Summary Although many non-saturation performance analysis models have enabled accurate performance analysis with realistic non-saturation traffic, most models limit their focus on exponential arrival patterns. Only a few models such as [76] consider generic arrival patterns and em bed such requirement in the analysis model design. However, such models usually require

PAGE 40

40 complex mathematical augmentations to deal with the complicated G/G/1 type queuing formulations. As we will see in the following chapters, although we borrow the approaches in some of these non-saturation models such as G/G/1 formulation and backoff procedure timing analysis, we also avoid some of the design pitfalls commonly found in existing non-saturation analysis models. Furthermore, instead of just de veloping an analysis model, we aim to provide a generic performance evaluation framework that ca n coherently analyze the effects of various factors on system performance. In the next subs ection, we highlight the di stinctions between our proposed framework and existing studies with mo re detailed discussi ons on common drawbacks of existing IEEE 802.11 performance modeling approaches. 2.3.3 Summary of existing IEEE 802.11 performance modeling approaches and distinctions of our proposed approach The IEEE 802.11 perform ance analysis models, both saturation and non-saturation, have enabled accurate prediction, thorough understa nding, and various performance improvement schemes of IEEE 802.11 MAC protocol. As summariz ed in Table 2-2, we have observed some common drawbacks from the models discussed in this section. As some of the weaknesses of existing models motivate us to build a better analytical model, our proposed framework also retains some strength of existing models and build a performance evaluation framework beyond the scope of many existing studies, which is one of the main contributions of this dissertation. Firstly, we have seen models from both a pproaches suffer from some common drawbacks. For example, for those models that yield highly accurate results such as Markov-based saturation models and transfer-function based non-saturation models, th ey usually involve complex mathematical calculations and complicated linea r equation systems. Implementing these models to achieve real-time performance presents a major concern. As we will show later in this dissertation, we aim to design a model that pr ovides accurate performance predictions, while

PAGE 41

41 maintaining the simplicity to be practically impl emented into real-world algorithms to help the network perform more efficiently. In additi on, we will demonstrate the feasibility by implementing our lightweight, yet accurate framew ork with real-world experimentation results. Moreover, we also notice from Table 2-2 that most models require certain information to be known prior to the analysis, su ch as number of stations in th e network and data rates used by the stations. In real-world scenarios, previous st udies have reported that number of stations is generally hard to be accurately estimated. In add ition, station data rates are adjustable if stations turn on the rate adaptation mechanism. As a result, the requirement of such information prevents the models from being applied to cases other th an performance predictions. Our proposed model avoids such requirement and uses only informa tion that is locally available to the wireless stations. Lastly but most importantly, we can see that while many models have been built on the same Bianchis model, different studies have appr oached their specific problem scenarios in significantly different ways. In other words, there is no unified framework that can model and explain the impacts of different parameters or sc enarios in a coherent vi ewpoint. Similar problem exists among non-saturation performance modeling studies. This dissertati on aims to fill such void and take one step beyond the collective resu lts from these existing studies. In the next chapter, we will present the design of our performance evaluation framework and we will see how our framework explains the si milarity/dissimilarity of various target scenarios that do not seem to be connected by traditional modeling appro aches. We think that it is very important to provide one unified performance analysis model th at can help the readers in this field get a complete and harmonious picture about the behavior of the protocol, and not be limited by the

PAGE 42

42 unconnected results from most previous studies which have been independently conducted for different problem scenarios. Table 2-2. Common drawbacks of existing IEEE 802.11 performance modeling approaches Saturation Non-saturation Markovchain based Approximationbased Markovchain based Transferfunction based Approximationbased Unrealistic traffic assumption X X (limited) Low accuracy X X High complexity X X X Require aprior info X X X X X Drawbacks Low extensibility X X X

PAGE 43

43 CHAPTER 3 IEEE 802.11 PERFORMANCE ANALYSIS FRAMEWORK In this chapter, we present the proposed IEEE 802.11 perform ance evaluation framework. We first describe our approach to model IEEE 802.11 performance, and we present the derivation of IEEE 802.11 service time in DCF and EDCA model. Then, based on the IEEE 802.11 service time performance model, we present the unified IEEE 802.11 performance evaluation framework. 3.1 Analysis model In this section, we present th e perform ance analysis model with non-saturation traffic for IEEE 802.11 DCF and EDCA contention-based wireless medium access methods. We consider the infrastructure Basic Service Set (BSS) scenar io, which consists of a base station (or access point) connected to a wired netw ork and multiple static wireless stations randomly scattered in the network. We assume that all stations and the access point are within carrier sensing range of each other. Unlike most models in the litera ture, we do consider realistic wireless channel scenarios, which we incorporate in the overall frame error probabil ity in the model. In addition, we do not restrict the traffic among stations in the network to be homogeneous. Instead, we assume that, in steady state, the channel ac tivity observed by a give n node, irrespective of whether it is homogeneous or heteroge neous, can be aggregately represented by Tbusy, the average medium occupation time, and Pbusy, the busy probability in a given time slot. The key idea to our analysis is to model the MAC layer timing dynamics, from the instance when the packet arrives into th e sending station until the packet is received by the intended node, as a G/G/1 queuing system shown in Fig. 3-1. At each wireless station, the packets generated by the application layer are modeled as the arriva l process of the queuing system. Our model works with any type of traffic arrival as long as the mean and standard de viation of the arrival

PAGE 44

44 distribution can be expressed in closed form. On the other hand, we model the time duration during the medium access backoff process as the se rvice process of the queuing system. In other words, the MAC layer service time is defined as, from the instant that a packet becomes the head of the transmission queue and MAC layer starts cont ention backoff process, to the instant that the packet completes the backoff process by being e ither successfully received or dropped because of maximum retry limit has reached. Note that, during this backoff procedure, various factors such as background traffic, collisions, and wireless channel conditions (illustrated as different shaded areas in Fig. 3-1) make this duration hi ghly variable. Therefore, constructing an accurate expression to the MAC layer service system is the key to IEEE MAC layer performance analysis. In this section, we first derive the DCF MA C layer service time by car efully analyzing the duration and occurring probability of different ev ents that take place at backoff stages. We further extend the derivation to EDCA where a node with high prior ity traffic has early transmission opportunity over othe r nodes with low priority traffic. Then, we apply the MAC layer service time to standard queueing system formulations to obtain performance results for IEEE 802.11 MAC. Collision with other nodes Successful transmission of the tagged node Transmission by other nodes MAC Arrival packets Collision with other nodes Successful transmission of the tagged node Transmission by other nodes MAC Arrival packets MAC Arrival packets Figure 3-1. IEEE 802.11 MAC analysis model 3.1.1 MAC layer service time As we discuss in Sec. 2.1.1, any node ope rating IEEE 802.11 MAC shar es the wireless m edium with other nodes in range and defers the transmission to ot her nodes transmission during a backoff process. We illustrate this e ffect in Fig. 3.2 by an example of the channel

PAGE 45

45 activities seen by a node during th e backoff procedure. We can s ee that the overall duration of the backoff procedure is characterized by the duration and occurring probability of different events that take place during the backoff stages. As a result, we identify the event occurring parameters as follows. a) When the backoff timer decrements, the tim e slot is either sensed as idle (for Tslot, the length of one time slot) or as busy occupied by background traffic transmission (for Tbusy, the average medium occupation time used by background traffic transmissions). We define Pbusy to be the probability that, at a given time slot, the backoff timer is frozen due to busy medium in carrier sensing. It follows that the occurring proba bility of idle slot and busy slot is (1Pbusy) and Pbusy, respectively. b) When the backoff timer expires (i.e. d ecrements to zero), the attempt of packet transmission either fails (after Tfail) or succeeds (after Tsucc). We define Pfail to be the overall frame error probability. It follows that the occurring probability of packet failure and success is Pfail and (1Pfail), respectively. Note that, both collision a nd wireless loss events contribute to the overall frame error probability ( Pfail); however, we only need Pfail to model the packet transmission events, irrespective of the error cause. TbusyTbusy TbusyTfailTsuc Backoffstage j Backoff stage j+1 Collision with other nodes (Pfail, Tfail) Successful transmission of the tagged node (Tsucc) Transmission by other nodes (Pbusy, Tbusy) Tslot Backoff stage 0 TbusyTbusy TbusyTfailTsuc Backoffstage j Backoff stage j+1 Collision with other nodes (Pfail, Tfail) Successful transmission of the tagged node (Tsucc) Transmission by other nodes (Pbusy, Tbusy) Tslot Backoff stage 0 Backoff stage 0 Figure 3-2. Packet transmission and coll ision events during IEEE 802.11 MAC backoff

PAGE 46

46 We assume that Pbusy and Pfail are constant and independent at each time slot or transmission attempt. Such assumption has been evaluated by test-bed measurements in [57], and found a reasonable approxim ation for modeling IEEE 802.11 MAC behavior. Furthermore, we note that all the parameters required to model the above process except Tslot, which is specified in different version of IEEE 802.11 standard, are availa ble from monitoring channel activities. We will discuss in more detail about how to obtain these parameters in the following section. Once these parameters are collected, we can construc t a mathematical model calculating the occurring probability for combinations of all different backoff events throughout all backoff stages. MAC layer service time of DCF According to the above analysis of various events during backoff procedure, we break down the derivation of overall backoff duration into two steps, i.e., first the duration in single backoff stage, and second the duration throughout successive backoff stages. We first consider any single backoff stage j where backoff timer is selected from 0 to Wj (maximum number of backoff slots in stage j). The occurring probability, j knkF ,, that there are exactly k busy time slots and ( n-k ) idle slots in backoff stage j as, 0,)1( 1, j kn busy k busy n k j j knkWnk PPC W F bbb = (3-1) Moreover, we know that any combination of number of busy and idle slots can be a cumulative effect from successive backo ff stages. Therefore, we then define j knkS ,for probability of backoff counter being frozen ( k-j ) times and idle ( n-k ) times and transmission ends (with successful transmission when jbm-1 or with failed transmission when j=m-1 ) after backoff stage j (which implies packet transmission failed j times),

PAGE 47

47 bbb= bbbbb bbb = = = = = = = = 1 1,1 .1 ,2 1, )1( 1 0,0 ,)1(1 0 00 1 0 00 1 0 0 m i i k l kn i j iknlk j il fail j i i k l kn i j iknlk j il fail fail knkfail j knkWnkmmjfor FSP Wnkjmjfor FSPP Wnk jforFP S (3-2) Note that m in this equation is the maximum number of retries specified in the standard. For stage greater than zero (i.e. j=1,2,,m-1), j knkS includes all possible cases, from combination of previous stage(s) to the current stage, which result in ( n-k ) idle slots, ( k-j ) busy slots, and j failed transmission pe riods. In other words, the pack et transmission time when such combination happens can be characterized by, 1 ,10 *)(**)(0 ,bbbbb +++== j i i succ slot fail busy j knkWnkjmjfor TTknTjTjkT (3-3) As a result, we have the probability mass f unction (pmf) of MAC layer service time, Bt, in the sequences of time points at the multiples of the busy medium time (Tbusy) and slot time ( Tslot), failed transmission time (Tfail), and successful transmission time ( Tsucc) as []. Pr, j knk j knktTtfor SBob = = (3-4) Note that, since Pfail partially depends on the packet le ngth sent by the tagged node (i.e. Tsucc), this representation of Bt, also depends on Tsucc. For simplicity of the analysis, we assume Tsucc remains the same for the packets sent from the tagged node.

PAGE 48

48 Furthermore, the probability generating function (pgf) of Bt can be expressed with a reasonable system clock unit, e.g. in s or in system slot time, []... Pr)(2 0 11 0 1 0 0 00,2 1,1 0,1 0,0 succ busy succ slot busy succ busy succTT TTT TT T t k tzS zSzSzS zBobzB+t +t+t +t =+ + += = (3-5) MAC layer service time of EDCA The QoS-enabled IEEE 802.11e EDCA mechanis m provides prioritized medium access by assigning different AIFS and b ackoff window parameters ( CWmin and CWmax) to different traffic categories. In particular, AIFS pr ovides advance opportunity to high priority traffic to access the medium by shorten the amount of time a station defers access to the channel following a busy time slot. As a result, the time slot immediately af ter a busy time slot can only be frozen by other stations with same or higher priority traffic. Therefore, as AIFS changes the way we construct the occurring probability of busy and idle slot combinations, we need to re-define Eq. 3-1. Following a similar procedure to that used in finding DCF MAC layer service time, we first look at the event occurring probability and duration for a single backoff stage in EDCA. To simplify the derivation, we assume that only th e tagged node utilizes the short AIFS traffic category, i.e. AC_VO or AC_VI with AIFSn=2 while all other competing traffic utilize the AC_BE traffic category with AIFSn=3. Under this setting, we observe that what happens in the last backoff time slot of the tagged node decides two different scenarios: 1) When last backoff slot ( cw=1 ) was an idle slot, the transm ission is subject to collision. 2) When last backoff slot was a busy slot, the tagged node un-freezes the backo ff timer one time slot before all other traffic. As a result, the backoff timer of the tagged node expires before all other traffic un-freeze the timer, and thus the transmission is guaranteed to be successful without collision.

PAGE 49

49 Firstly, we consider the case where last backoff slot was an idle slot. We define the occurring probabilityj knkFCC that, in any single backoff stag e j, there are exactly k busy time slots and (n-k) idle slots is, 12,12/)1(0 ,)1( 112 1 j j kn busy k busy kn k j j knkWnk Wk PPC W FCC bb++bb = (3-6) Note that this formulation differs from Eq. 31 in the occurring probability of idle slots. The very first time slot and the time slots after busy slot always advance before all other low priority traffic and the countdowns are successful with probability 1. Only the slots other than these special slots and busy slots in Eq. 3-1 are classified as regular id le slots with occurring probability (1Pbusy). Secondly, we consider the case where last bac koff slot was a busy slot. We define the occurring probability j knkFNC ,that, in any single backoff stage j there are exactly k busy time slots and (n-k) idle slots is, 2,12/)1(1 ,)1( 12 1 1 j j kn busy k busy kn k j j knkWnk Wk PPC W FNC bb+bb = (3-7) Similarly, in this formulation, only (n-2k) idle slots occur with probability (1Pbusy). All the slots after busy slot always occur with probability 1. Subsequently, we use the intermediate term s SCC and SNC to define all possible cases, from combination of previous stage(s) and the current stage, which result in k busy slots and (n-k) idle slots, ( k-j ) busy slots, and j failed transmission periods, for the cases where the transmission is subject to collisions (SCC) and where th e transmission is free from collisions (SNC). Following similar derivations as Eq. 3-2, SCC and SNC are described as,

PAGE 50

50 [] [] bb++bb= bb++bbbb bb++bb= = = = = = = = = = 1 12,112/)1( ,1 1 12,112/)1( ,21 )1( .12,12/)1(0,0 ,)1(1 0 1 0 00 1 0 0 00 1 0 m i j m i j k l kn i j iknlk j il fail j i j j i j k l kn i j iknlk j il fail fail j j knk fail j knkWnk Wkjmjfor FCC SCC P Wnk Wkjmjfor FCC SCC PP Wnk Wkjfor FCCP SCC(3-8) [] [] bb+bb= bb+bbbb bb+bb= = = = = = = = = = 1 2,112/)1( ,1 1 2,112/)1( ,2 1 .2,12/)1(0,0 ,1 0 1 0 00 1 0 0 00 1 0 m i j m i j k l kn i j iknlk j il fail j i j j i j k l kn i j iknlk j il fail j j knk j knkWnk Wkjmjfor FNC SCC P Wnk Wkjmjfor FNC SCC P Wnk Wk jfor FNC SNC (3-9) Regardless of the cases where the transmissi on is subject to collisions or free from collisions, the packet transmission time for such co mbination of previous stage(s) to the current stage, which result in (n-k) idle slots, (k-j) busy slots, and j failed tr ansmission periods can again be expressed by Eq. 3-3. As a result, the probability mass functi on (pmf) of EDCA MAC layer service time,tB, in the sequences of time points j knkT is []. Pr, j knk j knk j knk tTtfor SNC SCCBob = += (3-10) Finally, the probability generating func tion (pgf) of EDCA MAC service time tB can be expressed in the same format as Eq. 3-5.

PAGE 51

51 3.1.2 Performance results Throughput From standard z-transform theory [62], we know that the aver age MAC layer service tim e can be obtained by B'(1), the first derivative of B(z) at z=1. Altern atively, the average MAC layer service time can also be expressed by .)1( ,)*(1 0 0 1 0 = == = = =m i i N k N kn m j j knk j knk avgWNwhere TS T (3-11) Hence, with the LDATA bytes long payload in I EEE 802.11 DATA frame, maximum achievable throughput can be expressed by )1(' 8 B L ThroughputDATAt = (3-12) Delay On the other hand, to derive the total packet delay, we can apply standard discrete time queuing theory [10] [67] [48] with the statistical characteristics of the arrival and service process. Here we assum e the first and second moment of the arrival di stribution are known and can be expressed in closed form. A'(1) and A"(1) repres ent the first and second de rivative of the pgf of arrival distribution, A(z), at z=1, respectively. According to [10], if the arrival process is a genera l independent (GI) arri val process, i.e., the num bers of packets entering the system during the consecutive time units are assumed to be independent and identically distributed (i.i.d.), th e mean system time, i.e. packet delay in our case, of GI/G/1 queue system can be expressed as )]1(')1('1[2 )1(')1(")1(")]1('[ )1('12 1//BA BABA BX DelayGGI + ++= (3-13)

PAGE 52

52 where B'(1)and B"(1) are the first and second de rivative of the pgf of MAC layer service time, i.e.)( zPservT, at z=1. X is the mean distance of the arriva l point from the start of the unit time slot. When unit time is small, X is negligible. However, Eq. 3-13 is not applicable for applicat ions with deterministic arrival process, e.g. Voice-over-IP (VoIP). Therefore, we refer to [67] for models of discre te-tim e D/G/1 queues. The average delay of such system is = + =1 1 1//1 1 ))1(')1('(2 )1(")1)1(')(1(' )1('N r r GDz BA B AA B Delay (3-14) where N is the inter-arrival time, in system tim e unit, of the deterministic arrival process. Zr are the roots of solving zN-B(z)=0 on or inside the unit circle. Finally, for packet arrivals th at are neither General Independ ent process nor Deterministic process, an upper bound of the system waiting time is provided in [48], [] )1(')1('2 )1(")1(" BA BA W + b (3-15) Hence the upper bound of total system delay is .)1('1//WB DelayGG+b (3-16) 3.1.3 Summary In this section, we propose a queuing system poi nt of view to analy ze the performance of 802.11 contention-based MAC, including basic DCF and priority-based EDCA. Each wireless station is modeled as a queuing system with the p acket generation process as the arrival process and the variable amount of time a packet spends on MAC layer medium contention as the packet service process. The key to the analysis is to use a two-step decompos ition approach to model IEEE 802.11 MAC layer service time duri ng the backoff process. In pa rticular, we first calculate

PAGE 53

53 Pbusy Busy medium occurring probability Tbusy Average busy medium occupation time Tslot The length of one time slot (idle medium occupation time) Pfail Packet transmission failure occurring probability Tfail Failed packet transmission time Tsucc Successful packet transmission time Wj maximum number of backoff slots in stage j m maximum number of retries Table 3-1. Summary of parameters in IEEE 802.11 MAC analysis model with standard G/G/1 representations, and thus the proposed model works with any saturated or the duration in single backoff st age by analyzing the combinations of occurring probability and duration of medium busy and idle events. We then generalize the cumulative effect from successive backoff stages to obtain the overall b ackoff duration, and the MAC layer service time is obtained by the form of probability mass f unction (pmf). With AIFS -based EDCA, we follow the similar procedure with slight changes in the advance access opportun ity during the backoff stage transitions. The performance results su ch as throughput and delay are obtained nonsaturated underlying competing traffic patterns. We note that, one of the featur es of this modeling approach is that we only use a small number of parameters, such as Pbusy/ Tbusy and Pfail/ Tfail to fully characterize the dynamic timing in IEEE 802.11 MAC layer procedure. In the next section, we focus on understanding these parameters and we will show how this unique approach enables us to construct a unified performance evaluation framework for IEEE 802.11 MAC.

PAGE 54

54 3.2 Understanding the parameters in the analysis model In the previous section, we analyzed the IEEE 802.11 MAC layer backoff procedure by a two-step decom position approach. We have s hown that, by using only a few parameters as summarized in Table 3-1, this approach enable s a simple and effective way to characterize timing dynamics and thus throughput and delay performance of IEEE 802.11 MAC. In this section, we use the proposed model to lay out a systematic performance evaluation framework for IEEE 802.11 MAC. In particular, we will first focus on understanding the how to quantitatively relate some of these parameters wi th real-world representa tions. We will also try to understand how these parameters affect th e performance. Secondly, we will categorize the event occurring parameters into according to the ro les they play in the analytical model and how they affect MAC layer service time, namely intra-stage parameters and inter-stage parameters. In the next section, we will show that this categorization is the key to a systematic framework that helps us identify similar trends with different parameters and understand the inner dynamics of the backoff procedure in a more systematic manner. 3.2.1 Performance affecting paramete rs and the target scenario In Sec. 3.1.1, we identify the parameters that determine the event o ccurring probability and duration in IEEE 802.11 MAC backoff procedure, namely, Pbusy, Tbusy, Pfail, Tfail, Tsucc, and Pslot. We summarize the definitions of these para meters in Table 3-1. We note that, except Tslot, which is specified in different version of IEEE 802.11 st andard, we can acquire all other parameters directly from either the operating parameters of the considered node or monitoring channel activities. For example, Tfail and Tsucc are directly determined by th e operating data rate of the considered wireless station. We can determine Pfail by counting the ratio of failed packet transmission attempts and total packet transmission attempts. We also obtain Pbusy and Tbusy by keeping track of the number and duration of expe rienced collisions, respectively. Note that, in

PAGE 55

55 0 1 2 3 4 5 6 7 8 9 10 0 5001000150020002500 0 0.5 1 1.5 2 2.5 3 3.5 4 00.511.522.5 T (ms)Average MAC layer service time (ms) Tbusy Tfail practice, it may be difficult to obtain some of these parameters accurately due to implementation complexity in real devices. We can consider alternative approaches [28] [32] such as using num ber of consecutive idle slots between two busy slots to estimate Pbusy and Pfail. On the other hand, we should also try to quant itatively relate some of these parameters with real-world representations, so that we can be tter understand the implications of these parameters on IEEE 802.11 MAC performance. For example, Tbusy is an indicator of how long a busy slot takes, and is a function of PHY/MAC overhead, operating data ra te, and payload size. Figure 3-3. Average busy slot length, Tbusy ,with different payload si zes and transmission rates Figure 3-4. In addition, we can also see other parameters such as Tfail and Tsucc also show similar effect on MAC layer service time.

PAGE 56

56 Figure 3-5. Channel busyness ratio Pbusy and number of saturated tr ansmission nodes in 802.11b and 802.11a standard The longer the overhead or payload is or the slower the data rate operates, the longer it takes to wait on busy slots, and consequently the MAC laye r service time is longer. Fig. 3-3 plots the length of Tbusy as a function of payload sizes, for di fferent data rates. We can see that Tbusy increases linearly with payload size and inverse proportionally with data rate. From Eq. 3-3, we see that the effect of Tbusy on average MAC layer service time is close to linear, as we illustrate in On the other hand, Pbusy is an indicator for how busy th e network is and it is usually a function of number of nodes in the network, thei r traffic patterns and corresponding traffic load. It is obvious from the model deri vation in previous subs ections that the busier the network is, the more often a packet will wait on busy slots, and consequently the MAC la yer service time is longer. Unfortunately, no existing model can be used to quantify Pbusy with arbitrary number of nodes and traffic loads, so we have difficulties in relating the amount of Pbusy with real-world scenarios and quantifiable metrics such as number of nodes or traffic loads. On the other hand, if we assume all nodes transmit in saturati on mode and operate at the same IEEE 802.11 specification, e.g. 802.11b or 802.11a, then the model in [4] and [85] can quite a ccurately quantify and related Pbusy with number of nodes in the network. By using the models in [4] and [85], Fig. 3-5 illustrates the relationship between Pbusy and the number of saturation nodes

PAGE 57

57 Application OFFERE D LOAD (MBPS) Traffic Type Number of application in TG-1, Digital Home Number of application in TG-2, Digital Office Number of application in TG-3, Public Hotspot VoIP 0.096 UDP 3 30 15 VIDEO CONFERENCING /VIDEO PHONE 0.5 TCP 1 10 0 A/V Streaming 2-4 UDP 1 0 10 STDV 4 UDP 1 1 1 HDTV 19.2-24 UDP 2 0 0 Internet File Transfer N/A TCP 1 0 10 Local File Transfer N/A TCP 0 10 2 Resulting Pbusy 0.159 0.217 0.47 Table 3-2. TGn usage models in high performance networks Additionally, we adapt the usage mode l scenarios suggested by IEEE 802.11 TGn [71] to operating at IEEE 802.11b and 802.11a scenar ios. It can be seen that Pbusy quickly approaches to 0.5 when about only 20 nodes transmit in the network. The busyness ratio ( Pbusy) tends to saturate with more nodes contend the medium. Note that, compared with IEEE 802.11b stations, IEEE 802.11a stations operate with smaller c ontention window settings and thus utilize the bandwidth more aggressively, which in turn results in higher busyness ratio under the same number of nodes in the network. further illustrate the relationship between Pbusy and real-world non-saturated traffic. As summarized in Table 3-2, we use different combinations of high-bandwidth multimedia (video and audio) and data networking applications to emulate futuristic high-performance wireless network scenarios, such as digital home, dig ital office, and public hotspots. We then obtain Pbusy of each scenario through simulations. In order to discuss the effects of Pbusy on IEEE 802.11 MAC layer service time, we recall from Eq. 3-1 that Pbusy determines the single stage busy and idle slot event occurring probability

PAGE 58

58 0 10 20 30 40 50 60 00.20.40.60.81 PAverage MAC layer service time (ms) Pbusy Ploss Pfail Figure 3-6. Average MAC layer service time with different Pbusy (with Ploss=0, Tbusy=0.25ms, Tfail=1ms, Tslot=0.02ms), different Ploss (with Pbusy=0) and Pfail (with Pbusy=0.3) in the binomial distribution expression assuming Pbusy is constant and independent at each time slot. As Pbusy increases, the expected single stage durat ion also increases. On the other hand, as we can see from Eq. 3-2, Pfail is multiplied every time the j knkS parameter advances one more stage and its effect on MAC layer service time is similar to a geometric di stribution. We note that the overall frame error probability Pfail can be attributed to two sources: collision events and wireless loss events. While we express wireless loss event probability as Ploss, we can also assume that, at the end of backoff stage, collisions happen with the same Pbusy probability. In other words, the overall frame error probability can be represented by Pfail = Pbusy + (1Pbusy)*Ploss, as frame error is caused either by collis ions or wireless loss when there is no collision. Therefore, as Pbusy increases, it also increases the number of backoff stages the backoff procedure operates. As we plot the effects of Pbusy on average MAC layer service time in Fig. 3-6, we can see that, combining the two observations of how Pbusy involves in calculation of IEEE 802.11 MAC layer service time, the overall effect of Pbusy is close to exponential with the tail slightly slower than the expone ntial distribution. Moreover, we also see the effects of Pfail and Ploss in Fig. 3-6 also show an asymptotic trend at a slower rate as a result of the fact that Pfail and Ploss only affect the backoff stage transitions in backoff procedures.

PAGE 59

59 3.2.2 Categorizing performance affecting parameters In the previous subsection, we obs erv e that some of the parameters exhibit similar trends in the way they affect the system performance, even in some case these parameters do not seem to be related in the physical world. In this subsec tion, we further discuss th e roles these parameters play in the analytical model and to categorize the parameters according to their effects on MAC layer service time. We will also show how such categorization systematically guides us to better understand the dynamics of the backoff procedure. In particular, based on the observations from Eq. 3-1 and 3-2 that some parameters only appear in the calculations of the duration of single backoff stage while other parameters involve in the calculations across multiple stages, we try to classify the parameters into two approaches, namely intra-stage parameters and inter-stage parameters. Intra-stage parameters Recall from Sec. 3.2.1, we observe that both Tbusy and Tfail exhibit similar effects on IEEE 802.11 MAC layer service time. We can see the reasons being that, from Eq. 3-3 and Fig. 3-2, Tbusy and Tfail only affect the duration of backoff stages but not the reoccurrence of successive stages. Specifically, as Tbusy represents average medium occupation time when backoff counter freezes for other stations transmission, the whol e duration of the backoff stage apparently increases linearly as Tbusy increases. On the other hand, as one count of Tfail is added to backoff stage duration every time the transmission attempt fails, it is also expected that backoff stage duration increases linearly as Tfail increases. It follows that we can define the concept of intrastage parameters since the roles of such paramete rs are limited within the backoff stages, and all parameters exhibit such mathematically sim ilar effects can be categorized accordingly.

PAGE 60

60 The concept of intra-stage parameters applie s to other parameters as well. For example, as one count of Tsucc is added to the last backoff stage duration every time the transmission attempt succeeds, the effect of Tsucc is expected to be similar to Tbusy and Tfail. Moreover, even though the length of idle slots Tslot is specified by the standard, we can expect that when Tslot decreases/increases, backoff stage duration also decreases/increases proportionally. Therefore, Tslot also fits the intra-stage parameters category. Note that, as it has been illustrated in Fig. 3-4, the degree of influence to backoff stage duration is apparently different fo r different intra-stage parameters. For example, while Tfail only affects the backoff stage duration once at the end of each backoff stage, Tbusy affects the backoff stage duration multiple times through each backoff stage depending on how often other stations transmit. As a result, the effect of Tbusy is stronger than Tfail on average MAC service time. Additionally, not only the ev ent occurring parameters, bu t some IEEE MAC operation parameters affect the MAC layer service time by the same way. From Eq. 3-1, we can see that as Wj (number of time slots at backoff stage j) increa ses, the probability of choosing longer backoff counter increases, which means the backoff st age duration also increases. As a result, Wj demonstrates similar effect to IEEE MAC layer service time, and can also be categorized as intra-stage parameters. Inter-stage parameters In Sec. 3.1, we also observe that some pa rameters are involved throughout different backoff stages in the calculation of the IEEE MAC layer service time. For example, as we can see from Eq. 3-2, Pfail is multiplied every time the j knkS parameter advances one more stage until the transmission succeeds. As a result, such eff ect on MAC layer service time is similar to a geometric distribution and effectively increases the expected number of stages (or number of

PAGE 61

61 Intra-stage parameters Tbusy, Tfail, Tsucc, Tslot, Wj Inter-stage parameters Pbusy, Pfail, m AIFS Table 3-3. Categorization of pe rformance affecting parameters Ploss is constant, Pbusy also shows the same geometric dist ribution-like effects on the backoff trials in geometric distributi on analogy) at the rate of Pfail/(1-Pfail) Thus the overall backoff duration increases as Pfail increases. Therefore, we can de fine the concept of inter-stage parameters since the effects of such parame ters are spread throughout the backoff stages. We mention in Sec. 3.1.3 that Pfail can be represented by Pfail = Pbusy + (1-Pbusy)*Ploss, as both Ploss and Pbusy contribute to overall frame error even ts probability. Therefore, assuming constant Pbusy, Ploss demonstrate the same geometric dist ribution-like effects on average MAC layer service time and thus can be categorized as inter-stage parameter. Similarly, assuming stage transitions in the backoff procedure. In the meantime, however, Pbusy also affects the length of signal-stage backoff duration according to Eq. 3-1. As a result, we can see that Pbusy demonstrates even stronger effects on average MA C layer service time as we discuss in Fig 3-6. Nevertheless, we should still categorize Pbusy as inter-stage paramete r as its effect on the backoff stage transitions is st ronger than its effect on signalstage backoff duration length. Additionally, we notice that the maximum number of retries, m also plays a role as the limit of backoff stage advancement in IEEE MAC layer service time calculation. In other words, as m increases, the backoff procedure can pursue deeper backoff stages and thus introduce longer backoff stage duration. On the other hand, in Sec. 3.1, we see AIFS changes the way we construct the occurring probability of busy/idle slot combinations and behavior of backoff stage transitions for high priority traffic in accessing the medium with advance opportunity. Hence, we can also categorize maximu m number of retries, m and AIFS as inter-stage parameters.

PAGE 62

62 As summarized in Table 3-3, we propose the categorization to group all six event occurring parameters in our mathematical model and other relevant oper ation parameters that exhibit similar effects to the av erage MAC layer service time. For intra-stage parameters, we classify those affect backoff stages individually, but not the reoccurrence of successive stages during the calculation of average MAC layer service time. In Particular, we identify that Tbusy, Tfail, Tsucc, Tslot, and Wi all fall into such category and e xhibit comparably linear impact on average MAC layer service time. On the other hand, for inter-stage parameters, we classify those affect the average MAC la yer service time across the reo ccurrence of successive stages. We identify that Pfail, Ploss, and Pbusy exhibit a geometric distributio n-like effects on the backoff stage transitions in backoff procedure, while maximum number of retries, m acts as the upper limitation point of the truncated geometric distribution. 3.3 UF-PASS: Unified Framework for Performa nce AnalySiS of contention-based IEEE 802.11 MAC In this chapter, we analyzed the IEEE 802.11 MAC layer backoff procedure by a two-step decom position approach. We have shown that, by using only a few parameters, this approach enables a simple and effective way to charac terize timing dynamics and thus throughput and delay performance of IEEE 802.11 MAC. Furthermore, in the previous section, we can see the parameters in the proposed analysis model are closely tied to the physical scenarios in investigation. We further categorized the parameters into according to the roles they play in the analytical model and how they affect MAC laye r service time. In this section, we use the proposed model and the categorized parameters to lay out a systematic performance evaluation framework for IEEE 802.11 MAC. We will show that our proposed mathematical analysis model is not only adequate for performance evaluati on and prediction purposes, but also capable of

PAGE 63

63 providing unified insight to the system dynami cs and performance impact across different parameters and various problem scenarios. In the following subsections, we first present UF-PASS: Unified Framework for Performance AnalySiS of contention-based IEEE 802.11 MAC We will show that, by linking the target scenarios with the categorized parameters a ccording to the nature of the target scenarios, the UF-PASS framework enables us to qualitativel y estimate the impacts of the target scenarios on IEEE 802.11 MAC performance at a unified pl atform. Moreover, the proposed framework helps us coherently identify the trends how th e performance of IEEE 802.11 MAC is affected by a number of problem scenarios, as opposed to th e way that most existing studies evaluate the IEEE 802.11 MAC performance with different models for different target scenarios. We then complete this chapter by showing some exam ples of using the UF-PASS framework for performance evaluations. Performance affecting parameters representation Problem scenarios Analysis: Simulations/Experimentations o/p: performance affecting Parameters (e.g. Tbusy/Pbusy) Performance evaluations o/p: performance results (e.g. throughput, delay) Protocol implementations Performance affecting parameters representation Problem scenarios Analysis: Simulations/Experimentations o/p: performance affecting Parameters (e.g. Tbusy/Pbusy) Performance evaluations o/p: performance results (e.g. throughput, delay) Protocol implementations Figure 3-7. UF-PASS framework overview

PAGE 64

64 3.3.1 Framework overview In the previous section, we id entified two categories of para m eters that exhibit similar effects on modeling the IEEE 802.11 MA C layer service time, namely intra-stage parameters and inter-stage parameters. The benefits of such categorization are manifold. 1) it provides a convenient way to quickly identif y qualitative estimate for the impacts of a particular target scenario on IEEE 802.11 MAC performance. For exam ple, if we know that the medium busy time (Tbusy) changes is the dominant effect of payload sizes of background traffic, we can roughly expect the impact of such effect to average MAC layer service time is on linear scale of Tbusy and in turn correspondingly to IEEE 802.11 MAC system performance. 2) it provides a way to compare and correlate the impacts of di fferent target scenarios on IEEE 802.11 MAC performance. For example, if we are interested in comparing two particular scenarios, in which the effect of payload sizes of background traffi c involves changes in the medium collision time ( Tbusy) and the effect of payload sizes of operatin g traffic involves changes in the medium transmission time ( Tsucc). Then we can expect the impacts of these two scenarios are roughly the same trend since they both fall into th e intra-stage parameters category. It follows that we can use Fig. 3-7 to illustrate the conceptual flow of our approach in the framework. The work starts with the target scen arios under consideration. We first identify the key performance affecting parameters and the corresponding categorizat ion of the scenarios under consideration. The resulting output of the examples in the above paragraph is Tbusy and Tsucc. The performance affecting parameters then are fed into the analys is model to produce performance results such as throughput and delay. On the one hand, for performance evaluation and prediction studies, the performance results are evaluated and feed back to conduct systematic evaluations on a series of different scenarios. The case studies we present in Ch. 4 and Ch. 5 follow this path. On the other hand, the perfor mance results can also provide insights for

PAGE 65

65 Figure 3-8. The UF-PASS framework showing the interconnections between target scenarios and performance affecting parameter categorizations protocol and application developments. In Ch. 6, we will demonstrate how we apply our framework in developing a new rate adaptation algorithm for IEEE 802.11 MAC. As illustrated in Fig. 3-8, we can further use a wide range of target scenarios that have been of interests to many previous studies as examples to demonstrate the generality of our framework. Although by no means we want to provide an exhaustive list of these scenarios, here we discuss various scenarios that have been popular in this field and their corresponding performance affecting parameters from our system atic frameworks point of view. We classify the popular target scenarios investigated by previous IEEE 802.11 MAC into three categories: 1) scenarios involving parameters intrin sic to IEEE 802.11 protocol specifications: Minimum contention window size ( CWmin): Although CW value is specified in the standard, much interest in previous studies has b een arisen to adaptively find the optimal CW value for the target scenario. The key performan ce affecting parameter corresponds to our framework is the intra-stage parameter Wj. Intrinsic parametersMAC headers Slot time Retry limits CWmin mInter-stage parameters Intra-stage parametersPfailTfail/TsuccTbusyTslotWj Target scenarios Identification of performance affecting parameters Performance Analysis & applications Pbusy AIFSEnvironment FactorsBackground traffic Wireless channel Operational parametersHeterogeneity Data rates Payload size Priority Intrinsic parametersMAC headers Slot time Retry limits CWmin mInter-stage parameters Intra-stage parametersPfailTfail/TsuccTbusyTslotWj Target scenarios Identification of performance affecting parameters Performance Analysis & applications Performance Analysis & applications Pbusy AIFSEnvironment FactorsBackground traffic Wireless channel Environment FactorsBackground traffic Wireless channel Operational parametersHeterogeneity Data rates Payload size Priority

PAGE 66

66 Maximum retry limit ( m ): The standard specifies m value as aLongRetryMax = 4 and aShortRetryMax=7 for basic and RTS access mode respectively. However, previous studies have investigated the effects of tunable m for optimal system performance. The key performance affecting parameter corresponds to our framework is the inter-stage parameter SSRC and SLRC Backoff slot time ( aSlotTime ): aSlotTime is determined by RX-to-TX turnaround time, MAC processing delay, and PHY detection time. aSlotTime has evolved from 20ms in IEEE 802.11 to 9ms in IEEE 802.11a/g as the un derlying PHY technology upgrades. This parameter corresponds to intra-stage parameter Tslot in our framework. MAC header size: MAC header size is determined by the underlying PHY layer technology and its impact to system performa nce magnifies as the supported data rates becomes higher. This parameter affects intra-stage parameters Tbusy Tfail and Tsucc in our framework Priority: The 802.11e standard provides prio ritized medium access to different traffic categories by assigning different AI FS and backoff window parameters ( CWmin and CWmax), which correspond to AIFS, Wj, and m parameters in our framework. 2) scenarios involving operational parameters that are controllable by the station: Packet payload size: The effects of pack et payload size corres pond to intra-stage parameters Tbusy Tfail and Tsucc in our framework. Data rate: The data rates selected by the opera tion stations affect the transmission duration and thus correspond to intra-stage parameters Tbusy Tfail and Tsucc. In addition, due to the different underlying modulation and coding schemes employed by different data rates, the data rate selection also aff ect the transmission failure proba bility and thus correspond to inter-stage parameters Pfail. Heterogeneity: The detailed compositions of the effects of heterogeneous network scenarios depend on the exact setting in the network. For example, hybrid IEEE 802.11b and IEEE 802.11g network, the two version of standard specify different CWmin, aSlottime MAC header and data rate. Therefore, the eff ects of such scenario correspond to a mix of CWmin, aSlottime, Tbusy Tfail and Tsucc in the proposed framework. 3) scenarios involving factors fr om the environment and are not controllable by the considered station.: Wireless channel condition: Despite various source causing packet failure in wireless error-prone conditions, the eff ect of wireless channel condition can be comprised by the inter-stage parameter Pfail. Background traffic: The effects of packet tr ansmissions from other stations can be comprised by the inter-stage parameter Pbusy and intra-stage parameter Tbusy.

PAGE 67

67 As we can see in Fig. 3-8, the key to such viab le analysis approach is the identification of the performance affecting parameter(s) and the co rresponding interconnections in our framework. Once such interconnections are constructed, we can see that our framework can effectively explain the effects and simila rity/dissimilarity of various target scenar ios IEEE 802.11 MAC performance, and provide insights to perf ormance evaluation studies and application developments. Note that these are the distinctiv e features between our proposed framework and previous approaches: we are able to study the IEEE 802.11 MAC performance in a more unified manner, as opposed to the way that most existing studies eval uate the IEEE 802.11 MAC performance with different models for different target scenarios. In the next subsection, we demonstrate the si gnificance of the proposed framework in more details by revisiting some target scenarios that pr evious studies have looked into before. We will see how our proposed framework brings some new li ghts in explaining the e ffects of those target scenarios without rigorously re-shaping the mathematical models. 3.3.2 Examples of performance anal ysis using UF-PASS framew ork In Sec. 3.1, we have discussed how different studies investigate the effects of wireless losses on IEEE 802.11 MAC performance in saturation and non-saturation condition. In particular, Hadzi-Velkov et. al. [25] propose to modify the tran sition p robability in Bianchis model in consideration of collision losses and wireless channel losses by p=1-(1-pc)(1-pe) (Eq. 2 in [25]. Zhang et al [93] adopt similar concept and place the modified transm ission failure probability p (Eq. 11 in [93]) into the complex transfer f unction to com pute pgf of MAC layer service subsystem. However, in both studies, there is no clear relationship that we can see how this modified parameter will interact with the analytical model and impact the system performance. On the other hand, using the int er-stage parameters concept in previous subsection, we can clearly see how Pfail and Ploss are involved in the tran sition of backoff stages

PAGE 68

68 and how such geometric distribution-like eff ect will impact IEEE 802.11 MAC performance. As we can see from Fig. 3-9, the average MAC laye r service time increase s in an exponential-like fashion as Ploss increases, and thus results in throughput reduction as Ploss increases. This effect is consistent with the finding reported in Figure 4-a in [25] and Figure 6 in [93]. On the other hand, we can look into one m ore example to see how our proposed framework helps not only to estimate the impact of certain parameters but also explain how different or how similar the parameters a ffect IEEE 802.11 MAC performance. Assuming ideal wireless condition, Ergen et. al. [22] propose to modify the Ts and Tc parameter in Bianchis model to accommodate the mixe d data rates in different st ations. From Eq. 6 and 7 in [22], the authors ad just Ts to weight different Ts from different data-rate stations and adjust Tc with the longest packet involved in the collision event. They di rectly plug in the modified Ts and Tc into Eq. 8 to obtain the total system throughput. Then again, by using the intra-stage parameters concept in the previous subsection, we have seen the linear effect of Tbusy and Tfail on MAC layer service time and in turn the system throughput. We illustrate such effect in Fig. 3-10 and confirm the result with Fig. 4 of [22]. Nevertheless, when we look into how Ergen et. al. and Hadzi-Velkov et. al. m odify the same Bianchi model to fit into their different objectives, we can not see how different or how similar the two target scenarios will affect the system performance. On the other hand, by using our evaluation framework, we can clearly see that the target scenario parameters in these two studies fall into the two different categories of th e performance affecting pa rameters defined in our the framework. Therefore, it is clear that these two target s cenarios will impact the system performance in very different approaches.

PAGE 69

69 Figure 3-9. Effects of Ploss on average I EEE 802.11 MAC layer service time and throughput Figure 3-10. Effects of mixe d data rate on average IEEE 802.11 MAC layer service time and throughput 0 1 2 3 4 5 6 7 0.01 0.1 1 PlossThroughput (Mbps)0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5Average MAC layer service tim e (ms) Tput Avg. MAC layer service time (ms ) 0 0.5 1 1.5 2 2.5 3 12345 #Stations with 11Mbps (The rest is with 1Mbps in 5 Station Total)Tput (Mbps)0 5 10 15 20 25 time (ms) Tput Avg. MAC layer service time Tbusy:avg. medium occupation time

PAGE 70

70 CHAPTER 4 PERFORMANCE EVALUATIONS FOR HYBR ID IEEE 8 02.11B AND 802.11G WIRELESS NETWORKS The IEEE 802.11g standard has been proposed to enhance the data rate of wireless LAN connections up to 54Mbps, while ensuring backwa rd compatibility with legacy 802.11b devices at the same time. However, in a hybrid 802.11b/ g network, the throughput of 802.11g stations is compromised because of not only the overhead to interoperate with 802.11b devices, but also the unbalanced medium contentions among devices with different versions of the standard. In this chapter, we use UF-PASS framework to analy ze this throughput reductio n effect in various 802.11b/g mixed scenarios. By decomposing effects of various factors (e .g. different contention window sizes, different backoff stages, and different data rate s) in hybrid 802.11b/g network into corresponding performance affecting parameters in the UF-PA SS framework proposed in Ch. 3, we try to offer a unified explan ation on why and how these differe nt setting changes affect the performance in hybrid IEEE 802.11b/g networks. The an alytical model is further verified with simulations and field measurements under different station numbers, data rates, and data packet sizes. In addition, a simple frame-bursting techni que is shown to balan ce the throughput between 802.11b and 802.11g stations. 4.1 Introduction In recent years, the IEE E 802.11-based [34] wireless local area networks (W LAN) have been widely deployed to provide high bandwidth wireless connec tions for various applications. For example, devices conforming the IEEE 802.11b standard [36] provide data rates up to 11Mbps at the 2.4GHz ISM band. Furtherm ore, enabled by the latest modulation techniques, the IEEE 802.11a [35] standard even pushes the wirele ss bandwidth up to 54Mbps at 5GHz band, while keep ing the same wireless medium acce ss (MAC) schemes. Such higher data rate

PAGE 71

71 extension is later implemented at 2.4GHz band as the IEEE 802.11g [38] standard. Due to the sam e operation frequency band, full backward compatibility with legacy 802.11b products becomes a major advantage of 802.11g standard; Upgr ading to a system that interoperates with the existing network protect s the investment to the existing infr astructure and faci litates a smooth, incremental upgrade without disposing of the old devices immediately. However, in order to provide backward compatibility, the 802.11g devices have to dynamically change a few communication parameter settings when operate in a mixed 802.11b/g network, which result in a compro mise to network performance. For instance, even the new physical layer technology allows the MAC layer of 802.11g stations to operate at a faster 9 s slot time, the adoption of slower 20 s slot time of 802.11b standard in hybrid environment degrades system performance. In addition, th e mandatory protection mechanism, which is designed to avoid unnecessary co llisions between packet fram es with different modulation scheme, also incurs extra overhead and causes throughput degradation. In [29], the authors report that the longer slot tim e t ogether with overhead from protection mechanism cause the throughput of 802.11g devices degrades more than 30%, compar ed with the original settings for pure 802.11g networks. Moreover, as we will show later, the conten tion and packet transmission settings also caused unbalanced throughput pe rformance in hybrid 802.11b/g network. On the one hand, by cutting the initial contention window by half, an 802.11g device is twice as likely to win the contention during the common contention peri od as an 802.11b device On the other hand, consider the case that both devi ces operate at their maximum data rates (i.e. 802.11g at 54Mbps and 802.11b at 11Mbps). Once the 802.11b device wins the medium contention, the data frame may occupy the shared medium longer than transmitting two 802.11g frames. In other words, the

PAGE 72

72 throughput of 802.11g stations is penalized by the slow data rate of 802.11b stations [9] even 802.11g devices have more chances to win the contention. Although previous studies have identified such perform ance im pacts under the scenarios of interoperability, no studies, to our best knowledg e, provide detailed mathematical models to evaluate the insights of this th roughput degradation effect. We note that, for real-world wireless LAN deployments, as long as the 802.11b devices ar e not totally ruled out by the market, they are likely to co-exist and in teroperate with new systems (e.g., 802.11g) for sometime (perhaps many years). Therefore, we argue and indeed i llustrate that a systematic analysis of the performance impacts of interoperation between different versions of 802.11 MAC protocols (in this case 802.11b/g) is in fact necessary to understand the resulting performance of real-world deployments that accommodate both technologies. Ou r goal in this paper is manifold. First, we aim to identify and quantify potential problem s in interoperability between different 802.11 versions. Second, we develop a systematic evaluation method through which we want to analyze the causes of such interoperability problems. Our evaluation method includes mathematical modeling, simulation and real e xperimentation. Third, using the insights developed in the analysis we hope to provide guidelines and solutions to the problems detected. We outline this chapter as follows. In Sec. 4.2 we briefly review previous literature for performance evaluations of 802.11-based MAC pr otocol. Section 4.3 defines detailed modeling of evaluating the throughput in a network in which both 802.11b and 802.11g devices contend the medium. Section 4.4 validates the accuracy of the model by simulation results and measurements. The effects under different scenar ios (i.e. station number combinations, data packet sizes, and data rates) and some possi ble improvements are also evaluated. Section 4.5 concludes the work.

PAGE 73

73 4.2 Related work There are co nsiderable interests in evaluati ng the performance issues of recent high data rate extensions to the 802.11 st andard, namely 802.11a and 802.11g. In [19], Doufexi et al. show that, in a typical office WLAN environm en t, 802.11g network covers about twice of the coverage in 802.11a network under the same data rate configuration. However, 802.11g suffers from a lower MAC efficiency for maintain ing backward compatibility with 802.11b. In [29], the authors provide a detailed perform ance anal ysis when 802.11b devices are present but not transmitting any traffic in a 802.11g network. In [9], the paper briefly discusses the throughput im pacts of 802.11g stations when 802.11g stations contend the medium with 802.11b stations, with and without protection m echanisms. Nevertheless, none of the above studies provide a detailed mathematical model to evaluate the insights of this throughput degradation effect. There have been several performance m odeling studies for IEEE 802.11-based WLANs. While studies such as [74][33] [16] utilize probabilistic approximations, other studies [4] [85] [94] exploit Markov chain model to quantify the throughput of the ge neric 802.11 MAC protocol. Yet, all of these studies focus on homogeneous 802.11 ne tworks, and cannot be directly applied to hybrid 802.11b and 802.11g networks. In [27], the authors analyze the performance impacts when devices operate at diffe rent data rates in an 802.11b network. Howe ver, the model in [27] is not able to address the thr oughput issues in 802.11b/g m ixed networ ks since the data rate is not the only factor that affects the performance. The related background information regard ing the operation of the Distributed Coordination Function (DCF) mode of IEEE 802.11 MAC can be found in Sec. 2.1.1 and more details in [34]. In addition, the detailed operational pa ram eter settings and the special protection feature for 802.11g devices to interoperate with 802.11b devices are detailed in Sec. 2.2.

PAGE 74

74 4.3 Performance analysis model for hybrid 802.11b/g networks In this section, we first present the perfor mance analysis model for hybrid 802.11b/g networks using UF-PASS framework. In particular by analyzing how differe nt parameters in the standard change during the transitions from pure 802.11g or pure 802.11b network to hybrid 802.11b/g network, we formulate the corresponding performance affecting parameters in UFPASS framework and derive the resultant performance results. On the other hand, we also present a Markov-chain based saturation performance model for hybrid 802.11b/g networks, which helps us better understand the performance results in real-w orld representations such as number of stations in hybrid 802.11b/g networks. 4.3.1 Performance analysis model using UF-PASS framework Recall from Ch. 3, UF-PASS uses a series of backoff event occurring probability and duration, as well as some protocol parameters to formulate the MAC layer service time and the resultant system throughput and delay. Therefore, we start our analysis by checking the different protocol setting changes in hybrid 802.11b/g network. As summarized in Table 2-1, 802.11g stations operates with short 9us SlotTime and high OFDM-based data rates when there are only 802.11g st ations exists in the network. In this case, the Tsucc parameters can be formulated by considering the frame exchanges and MAC layer timing parameters involved in a successful or collided transmission cycle with pure 802.11g settings, as 8 811 pure ACK ACK OFDM PHY OFDM p DATA DATA OFDM PHY OFDM p Pureg succDIFS R L TT SIFS R L TT T ++ +++ ++ +++=n n (4-1) Note that DIFSpure corresponds to 2 times Tslot (in this case 9us) plus one SIFS.

PAGE 75

75 In addition, we know from Sec. 2-2 that 802.11 stations start to adopt the slower 20us SlotTime as soon as the 802.11g stations are notified by the AP that some 802.11b stations just associated to the AP. In this case, Tslot changes to 20us and Tsucc slightly changes to 8 811 Hybrid ACK ACK OFDM PHY OFDM p DATA DATA OFDM PHY OFDM p hybridg succDIFS R L TT SIFS R L TT T ++ +++ ++ +++=n n (4-2) where DIFShybrid corresponds to 2 times Tslot (20us) plus one SIFS. Furthermore, when the 802.11g stations or AP overhear any data packets transmitted by 802.11b stations, 802.11g stations are required to turn on the CTS-to -self protection mechanism, which in turn changes Tsucc to 8 8 8,_11 hybrid ACK ACK OFDM PHY OFDM p DATA DATA OFDM PHY OFDM p CTS CTS CCK PHY CCK p onCTShybridg succDIFS R L TT SIFS R L TT SIFS R L TT T ++ +++ ++ +++ ++ ++=n n n (4-3) Note that, since the data rate of 802.11b stations are usually much lower than 802.11g stations, Tbusy and Tfail perceived by 802.11g station should also change accordingly. In summary, there are two scenarios fo r 802.11g stations switching from pure 802.11 networks to hybrid networks. When 802.11b stations only associate with the AP but not transmit data packets, 802.11g stations Tslot changes to 20us and Tsucc changes to Eq. 4-2. When 802.11b stations transmit data pack ets in the same network, Tslot changes to 20us and Tsucc changes to Eq. 4-3. It follows that, assuming all other parameters such as Pbusy, Pfail are obtained as we described

PAGE 76

76 in Sec. 3-2, we can formulate the average MAC layer service time as de scribed in Sec. 3.1, and further derive the system throughput and delay by Eq. 3-12 ~ Eq. 3-16. 4.3.2 Saturation performance model for hybrid 802.11b/g networks In the previous subsection, we develop a general perform a nce analysis approach for arbitrary background traffi c level (i.e. arbitrary Pbusy level) in hybrid 802.11b/g network. On the other hand, as we discuss in Sec. 3-2, it is easier to understand the implication of Pbusy and its impacts on system performance by translating Pbusy to number of stations in saturation operation scenarios. Besides, many 802.11 MAC performa nce studies use satu ration throughput as benchmarks to evaluate the accuracy of the anal ytical model. However, to our best knowledge, no existing saturation performance model can be di rectly applied to the unique situation in hybrid 802.11b/g network where the 802.11b and 802.11g devices operate with different CWmin (31 and 15, respectively), different backoff stages, and different data rates. Therefore, in this subsection, we make an effort to derive a saturation performance model for hybrid 802.11b/g networks. In particular, we fo cus on providing a mapping between Pbusy perceived by 802.11b or 802.11g stations and the number of stations in the network. We then plug in the derived Pbusy to UF-PASS framework for performance results with representations clos er to real-world scenarios. Note that, as we will show in the next subsection, the saturation performance model can be further extended to derive the throughput for saturated hybrid ne tworks. In addition, the result from this Markov-chain based satu ration model is a special case for the analysis using UF-PASS framework, and should agree with the results de rived from UF-PASS model. The accuracy of this Markov-chain based saturation throughput model has been validated [82]. In our analysis m odel, we follow the same assumptions in [4] and [85] that there are finite num bers of contending stations w ith always having a packet availa ble for transmission. Besides,

PAGE 77

77 Figure 4-1. Markov chain model of back-off window size in any fixed-length slot time the probability p that a transmitted packet collides is independent on the stage of the transmitted terminal. Let b(t) be the stochastic proces s representing the backoff window size for a give n station with state s(t) at slot time t, the bi-dimensional process { s(t), b(t) } can be modeled as a discretetime Markov chain in Figure 4-1. In this model, we define W=(CWmin+1) and 2mW = (CWmax +1). Therefore, the contention window of stage i is '', 2 2 mi mi WW WWm i i i> b = = (4-4) For example, in Direct Sequence Spread Spectrum (DSSS) PHY layer, CWmin and CWmax equal to 31 and 1023 respectively, then m is 5. The actual number of stages in the model is determined by the retransmission count, m. Considering the latest 802.11 standard, m equals 4 ( dot11ShortRetryLimit ) or 7 ( dot11LongRetryLimit ) depending on data frame or RTS frame is transmitted as the first packet i,0 i,1 i,2 i,Wi-2 i,Wi-1 0,0 0,1 0,2 0,W0-2 0,W0-1 m,0 m,1 m,2 m,Wm-2 m,Wm-1 i-1,0 1 1111 1 1111 1 1111 p p p p 1-p 1-p 1-p i,0 i,1 i,2 i,Wi-2 i,Wi-1 0,0 0,1 0,2 0,W0-2 0,W0-1 m,0 m,1 m,2 m,Wm-2 m,Wm-1 i-1,0 1 1111 1 1111 1 1111 p p p p 1-p 1-p 1-p

PAGE 78

78 frame. The state of (i,k) represents the contention wi ndow of the terminal equals k and its backoff stage is i The one-step transition probabilities are = = = =+ ]1,0[ /1}0,|,0{ ],1[]1,0[/}0,1|,{ ]1,0[]1,0[/)1(}0,|,0{ ],0[]2,0[ 1}1,|,{0 0 0 0WkWmkP miWkWpikiP miWkWpikP miWk kikiPi i i (4-5) These transition probabilities can be interprete d as (1) the decrement of back-off timer at each slot time; (2) after a successful transmission, the back-off timer of the new arriving packets starts from back-off stage 0; (3) if the transmissi on is not successful, the system step into next back-off stage; (4) at the maximum back-off stag e, the contention window will be reset no matter the transmission is successful or not. The closed-form solution for this Markov chai n model can be derived by chain regularity, b< = + = =.0 0 )1( .0,1 0, 1 0 0, ,mi bp ibbp W kW bi m m j j i i ki (4-6) We can now express the probability that a station transmits in a randomly chosen slot time as 0 0,0 1 0,1 1= + ==m i m ib p p b (4-7) where bo,o can be obtained by 1'' 1 1' 1 1 0,0' )]1)(21(2)1)(21( )1)()2(1(/[)1)(21(2 )]1)(21( )1)()2(1(/[)1)(21(2 > b ++ + = + + + + +mm mm pppWpp ppWpp pp ppWpp bmm mm m m m m (4-8) Note that 802.11b and 802.11g use different CWmin and m so we have different transmission probability, namely b and g, for 802.11b and 802.11g stations respectively.

PAGE 79

79 In the steady state, the collision probability can be expressed as at least one of the other nb+ng-1 stations transmits at the sa me slot time given the packet transmitted is from a 802.11g station or 802.11b station. ].)1(*)1[(1 ].)1(*)1[(11 1 = =g b g bn g n b g n g n b bp p (4-9) b, g, and pb, pg in equation (4-5) and (4-6) form a nonlinear system which can be solved by numerical techniques. It follows that, w ith given number of 802.11b and 802.11g stations in the network (nb and ng, respectively), pb, and pg in Eq. 4-9 are the Pbusy for 802.11b and 802.11g stations to be used in UF-PASS model. 4.3.3 Discussions Pbusy in saturated hybrid 802.11b/g network In Fig. 3-4, we have shown the mapping between Pbusy and number of stations in pure 802.11a and 802.11b networks, by using the gene ric saturation performance model for homogeneous networks proposed in [4]. In the previous subsection, we further extend such m apping for 802.11b and 802.11g stations in hybr id 802.11b/g networks, which is a new contribution to the field. In Fig. 4-2, we plot Pbusy perceived by an 802.11g station when its in pure 802.11g network as the line with diamond markers and Pbusy perceived by an 802.11g station with the same number of competing 802.11b stations as the line with square markers. We can see that Pbusy perceived by an 802.11g station drops around 50% when switching from one competing 802.11g station to one competing 802.11b st ation. In this case because the minimum contention window ( CWmin) of the competing station increase s from 15 to 31 with the transition, it is expected that the medium busy probability ( Pbusy) drops about half. On the other hand, as the number of competing stations increase s, the increase in competing stations CWmin actually

PAGE 80

80 decreases the chance of collisions be tween competing stations. Therefore, Pbusy perceived by an 802.11g station does not drop as much as the case wh en the number of competing stations is small. Figure 4-2. Pbusy of 802.11g station and number of saturated transmission nodes in pure 802.11g and hybrid 802.11b/g networks As we discussed above, the Pbusy ( pb, pg) derived by the Markov-chain saturation performance model can be plugged back into th e UF-PASS framework to obtain the performance results in hybrid 802.11b/g network. In Sec. 4-4, we will discuss the performance impacts caused by the combined effect of the Pbusy changes in this subsection and the Tsucc changes in the previous subsection in more detail. We will see that, because of the flexible structure of UFPASS framework, we are able to explain such comb ined effects in a systematic way, which is not available in any previous studies. Saturation throughput for hybrid 802.11b/g networks In the previous subsection, we use a similar approach in [4] and [85] to derive the transm ission probabilities b, g, and collision probabilities pb, pg for 802.11b and 802.11g stations in the hybrid network. We can see that, with slight modifications in accommodating the transmission opportunities between 802.11b and 802.11g sta tions, it is very easy to continue such procedure to derive the saturation throughput for hybrid 802.11b/g network.

PAGE 81

81 We let Ptr be the probability that there is at leas t one transmission in the considered slot time, .)1()1(1g bn g n b trP = (4-10) With the consideration that only eith er one 802.11b station or one 802.11g station transmits, Ps, the probability that a transmission is successful, can be expressed as )1(*)1()1(*)1(1 1 tr n g n bgg n g n bbb sP n n Pg b g b + = (4-11) Then, we can express the normalized system throughput S as the ratio, )1( )1( ][ slot time] a of E[Length ]slot time ain n informatio Payload[cs trsstr tr strTPPTPPP PEPP E S ++ = = (4-12) Here, Ts, Tc can be referred as the same symbol in [4] and [85], which means the average tim e the channel is sensed busy when there is e ither a successful transmission or a collision happened. Ts can be expressed with weighted transmission time between 802.11b and 802.11g stations as, gsg bsb sTXTXT, ,** += where Xb and Xg formulate the probability of one b or g station transmits, given that there is exactly one station transmits in the network )1()1( )1( )1(*)1()1(*)1( )1(*)1(1 1 1 bgggbb bbb n g n bgg n g n bbb n b n gbb bnn n n n n Xg b g b g b + = + = )1()1( )1( )1(*)1()1(*)1( )1(*)1(1 1 1 bgggbb ggg n g n bgg n g n bbb n g n bgg gnn n n n n Xg b g b g b + = + = (4-13)

PAGE 82

82 Furthermore, Ts,b and Ts,g depend on the access mechanism the stations choose. For example, for an 802.11b station, it can use eith er RTS-CTS mode or basic mode (DATA-ACK directly) /][,n n n n++++ +++ +++++= ACK SIFS DataRate PEH SIFSCTS SIFSRTS DIFSTb rts bs /][, n n ++++ ++= ACKSIFS DataRate PEHDIFSTb bas bs (4-14) (while n represents propagation delay and H = PHY header + MAC header superscript bas means basic mode and superscript rts means RTS-CTS mode) If the 802.11g stations use CTS-to-self mode, /][,n n n ++++ + ++++= ACK SIFS DataRate PE H SIFS CTS DIFSTg cts gs (4-15) The detailed expression of Tc is determined by the amount of time the channel is kept busy due to the longest packet involvi ng in the collision. Fo r simplicity, we assume the packet length distributions of 802.11g and 802.11b stat ions are identical. Besides, since we only focus on the performance impacts of situations when the data rate of 802.11g st ation is higher than that of 802.11b station; in such cases, the length of 802.11b DATA frame is always longer than the whole duration of 802.11gs CTS-to-self frame and DATA frame. Therefore, Tc is attributed to 802.11bs DATA frame length whenever an 802.11b pack et is involved in the collision. On the other hand, when the collision is caused only by 802.11g packets, Tc is attributed to the addition of CTS frame and 802.11gs DATA fram e. Besides, since the probability of collisions by more than two 802.11g packets is very small, we use the approximation that th e collision is caused by two 802.11g packets or otherwise to estimate the collision period.

PAGE 83

83 ). /][ (* ))1(*)1(**1( ) /][ (* )1(*)1(**2 2 2 2 2 2 b b n b n g g n g g n b n g g n base cACK SIFS DataRate PEHDIFS C ACK SIFS DataRate PEHCTS DIFS CTb g g b g g++ ++ + ++ +++ = (4-16) When 802.11b stations use RTS mode, th e collision period is simpler since Tc is only attributed to the amount of time that the medium is kept busy due to the transmission of 802.11g packet. /][ ACK SIFS DataRate PEHCTS DIFS Tg RTS c++ +++= (4-17) Lastly, the throughput is shared by 8 02.11g and 802.11b stations proportionally by of Xg and Xb. ./* /*bb b gg gnXSS nXSS = = (4-18) 4.4 Performance evaluation using Mark ov-chain based saturation model In this section, we first validate the proposed m odel by comparing the results from ns-2 [8] sim ulations and field measurements. While our m odel apply to general tr affic patterns, we only focus on the performance results in saturation scen arios due to its flexibility to interpret the results in terms of number of 802.11b and 802.11g stations. We also vary data packet size and data rate of 802.11b/g stations to evaluate the effects of throughput degradation due to the heterogeneity in such networks. Unless otherwise specified, we use UDP packet s with constant payload size of 1472 bytes as the traffic source. Other parameters used in the proposed analytical model and simulations follow the parameter settings in the standard of DSSS technology as summarized in Table 4-1. The experimental setup is illustrated as Fig. 4-3. All wireless stations are placed within 2 meters to the access point in order to mitigate the wireless interference effects.

PAGE 84

84 Table 4-1. Frame parameters of 802.11g a nd 802.11b standard used in this chapter In addition, all wireless cards ar e configured to operate at fi xed transmission rates, namely 54Mbps for 802.11g stations and 11Mbps for 802.11b stations. All wireless cards also adopt the short preamble mechanism in order to achieve the maximum throughput. A special UDP traffic generator program run in each wireless stati on continuously contends the medium to send packets to the sink which directly connects with the access point. The data rate of the wired link from access point to sink is 100Mbps, which is much higher than the available wireless bandwidth and should not be the bottlene ck of throughput performance evaluations. Figure 4-3. Field measurement testbed configurations 802.11g (with protection) 802.11b Packet payload 12000 bits 12000 bits MAC header 224bits 224bits 11b_PHY header 72bits@1Mbps +48bits@2Mbps* 72bits@1Mbps +48bits@2Mbps RTS 160bits+11b_PHY header 160bits+11b_PHY header CTS 112bits+11b_PHY header 112bits+11b_PHY header Control frame bit 11Mbps 11Mbps 11g_PHY header 136bits@6Mbps N/A ACK 112bits+11g_PHY header 112bits+11b_PHY header ACK frame bit 24Mbps 11Mbps Data frame bit 54Mbps 11Mbps Propagation Delay 1us 1us Slot Time 20us 20us SIFS 10us (16us between data and ACK) 10us DIFS 50us 50us CWmi n 15 31 CWmax 1024 1024 short preamble mechanism Access point Sink 100Mbps Wired connection Wireless clients Access point Sink 100Mbps Wired connection Wireless clients

PAGE 85

85 Table 4-2. Comparisons between throughput attain ed from field measurements and analysis model 4.4.1 Model validations Fig. 4-4 shows the throughput derived from the UF-PASS analysis model (lines) and simulations (symbols) match closely to each other. Recall from Sec. 4.3.1, we use Tslot and Tsucc changes to model the 802.11g station transitions fr om pure all-11g network with short slot time, to all-11g network with long slot time (with 802.11b stations associated to AP), to all-11g network with CTS-to-self (with overhearing of 802.11b stations data packet activities). As Tslot and Tsucc increase during the transitio n, we can see that the thr oughput of 802.11g station drops as we predicted by UF-PASS framework. Figure 4-4. Saturated Throughput fro m analysis model (lines) and the simulations (symbols). a) one 802.11g station transmits at 54Mbps, b) several 802.11b stations transmit at 11Mbps. 4.10 6.03 5.44 8.86 g Throughput (measurement) 1.69 2.85 2.31 3.76 b -8.89% -5.19% -7.80% -2.85% g Error -16.34% 0.00% -12.50% -8.07% b 42 0 2 4.50 0.217 0.174 0.045 0.094 2g-2b 2.85 6.36 0.188 0.141 0.047 0.099 2g-1b 5.90 9.12 g Throughput (analysis) 2.64 4.09 b bgbg 0.150 0.098 0.050 0.106 1g-2b 3 0.113 0.053 0.053 0.111 1g-1b 2 Collision probability Transmission probability Config. # of stations 4.10 6.03 5.44 8.86 g Throughput (measurement) 1.69 2.85 2.31 3.76 b -8.89% -5.19% -7.80% -2.85% g Error -16.34% 0.00% -12.50% -8.07% b 42 0 2 4.50 0.217 0.174 0.045 0.094 2g-2b 2.85 6.36 0.188 0.141 0.047 0.099 2g-1b 5.90 9.12 g Throughput (analysis) 2.64 4.09 b bgbg 0.150 0.098 0.050 0.106 1g-2b 3 0.113 0.053 0.053 0.111 1g-1b 2 Collision probability Transmission probability Config. # of stations 0 5 10 15 20 25 30 35 135791 1 Number of stationsThroughput of 802.11g station (Mbps) All 11g, short slot time All 11g, long slot time w/o CTS-to-self All 11g, long slot time with CTS-to-self One 11g station in the hybrid network 0 1 2 3 4 5 6 7 8 135791 1 Number of stationsThroughput of 802.11b station (Mbps) All 11b One 11g station in the hybrid network Two 11g stations in the hybrid network(a) (b)

PAGE 86

86 The results attained from field measuremen ts for hybrid 802.11b/g scenarios are presented in Table 4-2. In most cases, the differences between analysis model and measurements are no more than 15%. Note that all of the simulation and measurement results are averaged from 3 different runs. The duration of simulations are 5 minutes long and the filed measurements are conducted at least for 2 minutes long. An error-free channel is assumed in all simulations. 4.4.2 Effects of interoperability Another interesting observation we can see from Fig. 4-4 is that, com pared with the same number of stations in a networ k consisting of all 11g devices w ith protection on (CTS-to-self), the throughput of 802.11g stations drops ~15% for interoperating with the same number of 802.11b stations. In addition, in Fig. 4-4(b), th e throughput of 802.11b stations becomes even greater than operating in a pur e 802.11b network. This is a performance anomaly effect due to interoperability, which is not iden tified and quantified before. Note that, this effect is different from the anomaly of pure 802.11b networks observed by [27] that the thro ughput of higher rate stations deg rades below the level of the lowest rate, when stations with multiple data rates interoperate together. While we can not find a suitable explanati on from the Markov-chain based saturation performance model for this perf ormance anomaly effect, the UF-PASS model provides a systematic framework to decompose the mixed effects from different parameter settings in this heterogeneous scenario, and offers a unified explanation on why and how the different setting changes affect the pe rformance in hybrid IEEE 802.11b/g networks. Recall from Ch. 3, we need to first identify the key performa nce affecting parameters (e.g. Pbusy/ Tbusy, Pfail/ Tfail, Tsucc and Pslot) that correspond to the scenario s under consideration. We then further interconnect the target scenarios with thos e performance affecting pa rameters accordingly,

PAGE 87

87 Table 4-3. Performance affecting parame ters in pure 802.11g and hybrid 802.11b/g network scenarios so that we can decompose the effects of diffe rent parameters on system performance with a unified point of view. It follows that, there are two particular aspects that change with the replacement of 802.11b competing st ations into the network. Fi rst, as the decrease in transmission rates of the compe ting stations corresponds to longer packet transmission time, the UF-PASS framework models su ch effect as increase in Tbusy and Tfail, which result in linear increase in MAC layer service time, and thus decreases in throughput. Second, the increase of the competing stations contention window size (switching from 16 for 802.11g stations to 32 for 802.11b stations) can be modeled as decrease in Pbusy and Pfail, which we report as a decreasing trend similar to exponential decrease in MAC laye r service time in Ch. 3. We summarize these parameter changes in Table 4-3. As a result, in Fig. 4-5, we can further use the UF-PASS model to accurately quantify the effects of such parameter changes. We can s ee that, after the switch, while the increase of Tbusy and Tfail correspond to a lower curve in throughput, the decrease of Pbusy and Pfail results in a shift of the operating points on the new curve. The overall result of such changes is a slight decrease

PAGE 88

88 in throughput as we see in Fig. 4-5. Note that by using the traditi onal Markov-chain based saturation model, we can not isol ate the effect of data rates an d contention window sizes as the analysis above. Figure 4-5. Throughput anomaly in pure 802.11 g and hybrid 802.11b/g network scenarios. Figure 4-6. Effects of packet sizes in hybrid 802.11b/g networks 4.4.3 Effects of data packet sizes We next consider the effects of data pack et sizes. W e first decr ease the competing 802.11b stations data packet sizes from 1500 bytes to 1200 bytes, and thus result in decrease of Tbusy and Tfail from 1351 s to 1131 s. We can see from Fig. 4-6 that the resulting throughput curve rises slightly (dashed line with triangular markers) and the throughput become s comparable to the

PAGE 89

89 pure 802.11g stations setting with the same number of competing stations. We then further decrease the competing 802.11b stations data pa cket sizes to 1000 bytes, which correspond to Tbusy and Tfail as 981 s. The 802.11g stations throughput in su ch scenario then surpasses the throughput in pure 802.11g network. Figure 4-7. Effects of packet sizes and ratio of 802.11b/g transmission time In Fig. 4-7, with three compe ting 802.11b stations, we plot 802.11g stations throughput as all 802.11b stations data packet size decreases. We further plot the 802.11g stations throughput as the dashed line for reference, with the same number of competing stations transmitting fixed 1500-byte data packet size (which corresponds to Tbusy as 472us long),. We can see that the throughput anomaly effect is alleviated linearly when 802.11b stations data packet size decreases. The 802.11g stations th roughput even surpasses that in pure 802.11g network when 802.11b stations data packet size is smalle r than 1050 bytes (which corresponds to Tbusy as 1023us long). We note that, at such turning point, the ratio of 802.11b/g transmission time is around 2 ( Tbusy as 1023us and 472us, respectively). It follows that, as the contention window of 802.11b and 802.11g stations provide roughly 1:2 transmission opportunities to 802.11b and 802.11g stations respectively, the 2:1 ratio of transmission cycle of 802.11b and 802.11g actually

PAGE 90

90 balances the uneven transmission opportunity caused by different contention window size settings, and thus alleviates the throughput anomaly effect. Figure 4-8. Effects of data rate a nd ratio of 802.11b/g transmission time 4.4.4 Effects of data rates In this experim ent, we evaluate the effects of 802.11g stations data rate. In Fig. 4-8, we first plot the 802.11g station throughput when the data rate of 802.11g stations in pure 802.11g network are tuned at different levels, namely 54, 48, 36, 24, 18, 12Mbps. With the 802.11g stations data rate unchanged, we replace the comp eting stations with 802.11b stations and plot resultant 802.11g station throughpu t. There is a crossing po int where the 802.11g stations throughput remains about the same before and after the switch. It is because, as we decrease the data rate of 802.11g stations, the duration of 802.11g stations transmission cycle becomes longer and comparable to half of one transmi ssion cycle of 802.11b stations, which in turn balances the uneven transmission opportunity as we argue above. 4.4.5 Discussion From the observations made in the previ ous subsections, the throughput anomaly in an 802.11b/g mixed network can be attributed to th e unbalanced transmission opportunities and

PAGE 91

91 transmission cycle durations. Decreasing the data packet size or data rate accommodates the transmission cycle duration and mitigates the extent of throughput anomaly. However, the effective throughput in both cases is lower. Fig. 4-9 discusses the effects of different pol icies on throughput performance. By adjusting the initial contention window ( CWmin) of 802.11g stations to 8 ( CWmin of 802.11b remains unchanged) or adjusting the CWmin of 802.11b to 64 ( CWmin of 802.11g remains unchanged), the throughput of 802.11g stations is gr eatly improved (Fig. 4-9a) but conversely the throughput of 802.11b is heavily penalized (Fig. 4-9b), compared with the situations th at same number of stations in a pure 802.11g or 802.11b network respectively. On the other hand, a simple, non-proprietary fr ame bursting technique can be applied to improve the balance of contention timing w ithout compromising the system throughput. The center idea of frame bursting is to insert the next data packet at the end of every transmission cycle without initiating another run of random b ackoff. By aggregating multiple packets in one transmission cycle, the overhead of control frame and PHY layer preamble is minimized and thus the performance can be improved. Especially in the case that 802.11g stations operate at 54Mbps and 802.11b stations operate at 11Mbps, by a ggregating just one more packet in the transmission cycle of IEEE 802.11g stations makes the ratio of transmission duration approaches 1:2, which just balances the uneven transmi ssion opportunity. As seen in Fig. 4-9c, the throughput of 802.11g stations is si gnificantly improved while th e throughput of 802.11b stations is only slightly penalized (~10 %). Moreover, the total throughput of the network (Fig. 4-9c) is also improved for ~25% when burst mode is implemented in 802.11g stations. Note that other proprietary performance boos ting techniques such as channel bonding implemented by several major vendors require all stations in the network utilize the same

PAGE 92

92 technology. In addition, those techniques do not emphasize on accommodating the transmission opportunity issues and may not be applicable to the throughput anomaly discussed here. Figure 4-9. Thought of 802.11b/g mixed network under different policies. a) Effects on the throughput of 802.11g stations, b)Effects on the throughput of 802.11b stations, c) total throughput We can also use UF-PASS framework to ex plain the effects of different throughput improvement policies above. We compare frame-burst ing with the policy that changes the initial -20.00% -10.00% 0.00% 10.00% 20.00% 30.00% 40.00% 50.00% 60.00% 70.00% 24681 01 2 Number of stations Original settings 11g burst mode 11g CWMin=8 11b CWMin=64 (a) -40.00% -30.00% -20.00% -10.00% 0.00% 10.00% 20.00% 30.00% 2 4 6 8 10 12 Number of stations Original settings 11g burst mode 11g CWMin=8 11b CWMin=64 (b) 0 2 4 6 8 10 12 14 24681 01 2 Number of stationsTotal Throughput (Mbps) Original settings 11g burst mode 11g CWMin=8 11b CWMin=64 (c)

PAGE 93

93 contention window of 802.11g sta tions to 8. We know that turning on frame-bursting for 802.11g station only results in a slight increase in Tsucc for 802.11g station and correspondingly slight increases in Tbusy for 802.11b stations (especially when ther e are large number of stations in the network). As a result, the thr oughput of 802.11b stations is only slightly penalized, while the throughput of 802.11g stations is bo osted by doubling the payload in each transmission. On the other hand, if we adjust the in itial contention window of 802.11g stations to boost the throughput of 802.11g stations, 802.11b stations suffer from a significant increase in Pbusy. In addition, as we point out in Ch. 3, the increase in Pbusy will result in exponentiallike increase in MAC layer service time. Therefore, we can expect significant throughput penalty in 802.11b stations if we choose to use such policy. 4.4.6 Packet delay performance One of the key features of UF-PASS fra mewor k is the ability to analyze packet delay performance for arbitrary traffi c patterns. In Fig. 4-10, we plot the packet delay of the 802.11g station before and after switching to hybr id 802.11b/g network, assuming the 802.11g station operates a G. 711 [40] VoIP application with dete rministic 10ms inter-arrival time. We can see that, in pure 802.11g network, the 802.11g stations p acket delay remains reasonable low (<50ms) when Pbusy is less than 0.5, which correspond to the scenario where the number of competing stations is no more than 13. On th e other hand, with the same level of Pbusy (and the same number of competing stations in hybrid 802.11g networ k), the 802.11g stations packet delay becomes too high (> 150ms) for interactive convers ations as recommended by (ITU-T) G.114 [39]. Particularly, in hybrid 802.11b/g network, the maxim u m number of stations that can coexist with the 802.11g station while the 802.11g stat ion can still maintain a reas onable quality voice call is reduced to 9.

PAGE 94

94 Figure 4-10. Packet delay performance in pure 802.11g and hybrid 802.11b/g network scenarios 4.5 Conclusion In this chapter, we focus on the perfo rma nce impacts of the interoperations between 802.11b and 802.11g devices in wireless local netw orks. We use UF-PASS framework to characterize the effects of different parameter settings on network performance. Comparisons with simulation results and field measurements show that the model is able to accurately predict the system throughput. Meanwhile, we also observe a throughput anomaly that penalizes fast 802.11g stations and privileges the slow 802.11b in hybrid 802.11b/g network. The effects and reasons of such anomaly are studied under different scenarios. We learn that as the ratio of transmission durations of 802.11g and 802.11b appro aches 1:2, the system throughput is more balanced by accommodating the 2:1 contention wi ndow setting. A simple non-proprietary frame bursting technique can be applie d to improve the balance of contention timing and consequently lessen the throughput anomaly as well as increase the system total throughput. Moreover, by using the UF-PASS performan ce evaluation framework, we show that the throughput anomaly can be decomposed into the effects of data rate s and contention window sizes changes. Although the overall effect results in a sl ight throughput reduction for 802.11g

PAGE 95

95 station to switch to hybrid 802.11b/g network, we fi nd that the throughput anomaly effect can be alleviated linearly by reducing the packet size of the competing 802.11b stations. In addition, we find the UF-PASS framework to be very conveni ent in interpreting the effects of different performance improvement schemes, which can not be addressed by traditional Markov-chain based models. Finally, we use UF-PASS framew ork to evaluate the 802.11g stations packet delay performance when switching from pure 802.11g network to hybrid 802.11b/g network. In particular, we observe that, after switching to hybrid 802.11b/g network, the maximum number of stations that can coexist with the 802.11g station while the 802.11g station can still maintain a reasonable quality voice call is significantly reduced.

PAGE 96

96 CHAPTER 5 PERFORMANCE LIMITS AND ANALYSIS OF CONTE NTION-BASED IEEE 802.11 MAC Recent advance in IEEE 802.11 based standard has pushed the wireless bandwidth up to 600Mbps [22] while keeping the same wireless medium access control (MAC) schemes for full backward compatibility. However, it has been show n that the inefficient protocol overhead casts a theoretical throughput upper lim it and delay lower limit for the IEEE 802.11 based protocols, even the wireless data rate gr ows infinitely high. Such limits are important to understand the bottleneck of the current technology and develop insight for protocol performance improvements. This chapter uses a queuing system appro ach to extend the discussions of IEEE 802.11 protocol throughput and delay limits to the situation that arbitrary non-saturated background traffic is present in the network. We use the UF-PASS performance evaluation framework proposed in Ch. 3 to quantify the limits for Di stributed Coordination F unction (DCF) of legacy 802.11a/b/g and Enhanced Distributed Coordi nation Access (EDCA) of IEEE 802.11e. We find such limits are function of the underlying MAC layer backoff parameters and algorithms, and are highly dependent on the load that background tr affic injects into th e network. Surprisingly, depending on the rate of arrival traffic, the packet delay limit may become unbounded such that no delay sensitive services can be operated under such condition. Moreover, we also discuss the effects of different frame aggregation schemes on the performance limits when the data rate is infinite. The developed model and analysis pr ovide a comprehensive understanding of the performance limitations for IEEE 802.11 MAC, and ar e useful in gauging the expected QoS for the purposes such as admission control. 5.1 Introduction In recent years, the IEE E 802.11-based wireless local area networks (WLANs) [34], nam ely IEEE 802.11b [36], 802.11g [38], and 802.11a [35], have been incr easingly popular in

PAGE 97

97 providing low-cost high-bandwid th (up to 54Mbps) wireless c onnections. With the growing demands of higher bandwidth for applications su ch as high-definition video streaming, network storage, and online gaming, the industry has be en working to seek higher data rate (HDR) extensions [42] [77] [31] to the family of IEEE 802.11 specifications. In early 2006, IEEE Working Group m eeting approved the first proposal of IEEE 802.11n [22], in which the data rate is expected to be as high as 600Mbps. Mo reover, the 802.11n specification adopts the sam e medium access control (MAC) schemes to ensure backward compatibility with existing IEEE 802.11 specifications. The industry also seeks advancement in providi ng better Quality-of-Service (QoS) at the MAC layer. A QoS amendment of IEEE 802.11 MAC, IEEE 802.11e [37], aims to provide service differentiations to different traffic types. In particular, the Enhanced Distributed Channel Access (EDCA) contention-based m edium acce ss improves the legacy IEEE 802.11 Distributed Coordination Function (DCF) by providing differentiated medium contention opportunities to high priority traffic. Despite the efforts on advancing data rate and QoS of IEEE 802.11, an analysis of theoretical throughput and delay limit was first discussed in [86] by Xiao and Rosdahl. The paper em phasized on the 802.11 MAC overhead effectivene ss and proved the existe nce of theoretical throughput and delay limits for IEEE 802.11 DCF pr otocol. The authors c oncluded that, given that the PHY data rate has advanc ed to infinitely high and only one station transmits in the ideal channel condition, the minimum time required for completing one packet transmission task is bounded by PHY and MAC headers as well as MAC layer backoff waiting time, and consequently bounds the maximum achievable th roughput and minimum achievable packet delay.

PAGE 98

98 In [87], the authors extended the de rivation of packet transm ission time to consider collisions and backoff freeze in the case that multiple stations transmit in saturation mode. However, the results in [86] and [87] only represent the thro ughput and delay lim it in aforementioned special cases but are unsuitable to real-world scenarios, which typically consist of multiple wireless stations operating in non-saturation mode w ith various traffic patterns. Besides, the delay an alysis presented in [87] only considers the med ium access delay and fails to address the queuing delay for the waiting tim e packets spent when backlogged. On the other hand, as the models used in [86] and [87] are only applicable to legacy IE EE 802.11 DCF, it is also important to expand the explorations of theoretical limits to the QoS enhanced IEEE 802.11e specification. In particular, it is essent ial to answer the following questions: will the similar performance boundaries exist in the ED CA MAC protocol? If so, how do we quantify such boundaries in different prioritized traffic ca tegories and what are the subsequent impacts in fulfilling the QoS requirements promised by IEEE 802.11e EDCA? Therefore, this chapter aims to provide a comprehensive understanding of the performance limitations on throughput and total system delay of both IEEE 802.11 DCF and EDCA MAC protocols with arbitrary amount of non-saturated competing traffic. Such analysis is critical in pinpointing the performance bot tleneck of state-of-the-art IEEE 802.11 technologies and in developing insight for future protocol performance improvements. We use UF-PASS framework to provide a queuing system point of view to directly analyze the access dynamics of 802.11 contention-based MAC. The proposed model wo rks with any satura ted or non-saturated underlying competing traffic patterns. Each wirele ss station is modeled as a queuing system with the packet generation process as the arrival pr ocess and the variable amount of time a packet spends on MAC layer medium contention as the p acket service process The packet throughput

PAGE 99

99 and delay bound is then derived wi th infinitely high operating data rate. The results are validated through extensive simulations under various network loading and operation conditions. The challenges to such analysis are in modeling the dynamic interactions between the arrival pattern of the considered node and the significantly va riable amount of network delay incurred by the backoff, collision, and re-transmission procedures under different background traffic load level. In this chapter, we make the following contributions: We construct a lightweight mathematical model for characterizing the throughput and delay limits and performance of conten tion-based IEEE 802.11 MAC. The proposed model enables us to systematically explore the e ffects of backoff settings, arrival processes, competing traffic characteristic s, and frame aggregation sche mes on theoretical throughput and delay limit of differen t versions of IEEE 802.11 MAC protocols, including legacy DCF and QoS enhanced EDCA. We discover a performance bottleneck of the 802.11 DCF and EDCA under the presence of background traffic: there is a turning point when packets a rrive faster than the packet service rate, packet delay becomes unbounde dly high beyond such network condition. The rest of the chapter is outlined as fo llows. Section 5.2 provides related background information and an overview of the analysis model. Using UF-PASS performance evaluation framework, we present the evaluations and simu lation comparisons for packet throughput and delay of IEEE 802.11 MAC in Section 5.3. In Section 5.4, we further discuss the effects of competing traffic packet data rates and payload sizes and the effects of some performance improvement schemes on the performance limits of IEEE 802.11 MAC. Section 5.5 concludes and provides future work directions. 5.2 Background and overview of analysis model In this sec tion, we first brie fly highlight the different bac koff settings, which affect the derivation of the proposed model. Then, we provide an overview of how do we apply the UFPASS performance evaluation framework to th is case study in analyzing the theoretical

PAGE 100

100 throughput and delay limit with non-saturation background tr affic for IEEE 802.11 DCF and EDCA contention-based wireless medium access methods. 5.2.1 Background As this chapter analyzes the perfor m ance limits for IEEE 802.11 DCF and EDCA contention-based wireless medium access met hods, interested readers can find the related background information regarding the operation of the DCF and EDCA in Sec. 2.1 and more details in [34] and [37], respectively. On the other ha nd, as w e will apply our performance limitation model to various higher-speed physic al layer (PHY) extensions of the IEEE 802.11 standard including the IEEE 802.11b, 802.11a, and 802.11g, the readers should also familiar themselves with the different backoff and head er settings discussed in Sec. 2.2. Note that, although these three versions of PHY extensions use different backoff and header parameter settings, they all can be incorporated with DCF and EDCA QoS enhancement amendment for medium access control. 5.2.2 Overview of analysis model In this subsection, we describe how we apply the UF-PASS perform ance evaluation framework to derive the theoretical throughput and delay limit with non-saturation background traffic for IEEE 802.11 DCF and EDCA. We consider the infrastructure Basic Service Set (BSS) scenario, which consists of multiple wireless nodes and a base station connected with wired networks. Following the bestcase scenario philosophy in [86] and [87], we make the f ollowing assumptions: 1) The wireless channel is ideal without errors. 2) All nodes are within carrier sensing range of each other. 3) All nodes use the basic access operation (no RTS/CTS) for shorter transmission cycles. The key idea to our analysis is to model the MAC layer timing dynamics, from packet arriving into the sending station until the packet received by the intended n ode, as a G/G/1

PAGE 101

101 queuing system. The theoretical throughput and delay limit are thus derived with infinite data rate. We will show that, even with infinitely high data rate, the overhead of background packets causing non-negligible amount of time in the bac koff stages is the dominant factor that bounds the MAC layer throughput and delay limit. Packet Arrivals Depending on the application layer, the traffi c arriving at each wireless station can be characterized with different probabilistic models In our proposed model, we treat the packet arrivals as the general arrival process of G/G/1 queue. For spec ial case arrival process such as Voice over IP (VoIP) with deterministic arrival rate, it can be treated as D/G/1 queue in our model. MAC Layer Service Time In Ch. 3, we develop a comprehensive MAC la yer service time anal ysis model with nonsaturation traffic for IEEE 802.11 DCF and EDCA by carefully characterizing the variable amount of time spent on busy and silent slots and the corresponding occurring probabilities throughout the backoff stages. In Eq. 3-5 and 3-10, we formulate the MAC layer service time of DCF and EDCA with the duration and occurrin g probability of different events including Pbusy, Tbusy, Pfail, Tfail, Tsucc, and Pslot. We note that, as we assume that Pbusy (and Pfail) is constant in steady-state and independent of the backoff st ages of the node under c onsideration (i.e. the tagged node). 1 we can obtain Pbusy (and Pfail) by monitoring the channe l activity and gathering the long-term statistics of the ratio th at medium is busy over all time slots [5]. Likewise, Tbusy and Tfail can also be obtained by channel activity monitoring. 1 Previous work has shown that this assumption has very meager effects on model accuracy [5]

PAGE 102

102 On the other hand, consider the case in deriving the theoretical throughput and delay limits with infinitely high data rate, Tbusy, Tfail, and Tsucc can be formulated by considering the frame exchanges and MAC layer timing parameters involved in a successful or collided transmission cycle. For example, Tsucc can be expressed by the durati on of DATA and ACK frame for pure 802.11a/b/g traffic, or by the duration of CT S, DATA, and ACK frame when the tagged node operates at hybrid 802.11g environment and has CTS-to-self protection turned on: 8 8 DIFS R L TT SIFS R L TTTACK ACK PHYP DATA DATA PHYp pure succ++ +++ ++ ++=n n (5-1a) 8 8 8_11DIFS R L TT SIFS R L TT SIFS R L TT TACK ACK OFDM PHY OFDM p DATA DATA OFDM PHY OFDM p CTS CTS CCK PHY CCK p hybridg succ++ +++ ++ +++ ++ ++=n n n (5-1b) where LCTS, LDATA, and LACK is the size (in bytes) of CTS, DATA and ACK frame, respectively. RCTS, RDATA, and RACK is the data rate (in bps) of CTS, DATA, and ACK, respectively. SIFS is the mandatory Short IFS inserted between frames. is the propagation delay. Other parameters such as Tp and TPHY of different versions of standard in can be found Table 2.1. In the case when the attempt of packet transmission fails, consid ering the ACK timeout effect, Tfail is expressed with the longest data frame i nvolved in the collision. In other words, the LDATA in Equation 5-1a and 5-1b is the size (in bytes) of the longe st data frame involved in the collision.

PAGE 103

103 The duration of busy slot, Tbusy, can be expressed either by Tsucc when the busy slot is occupied by successful transmission of the background traffic, or by Tfail when the busy slot is occupied by packet collisions. In the case when the wireless nodes operate at finite data rates, Tbusy can be collected by the long-term st atistics of channel activity monitoring. Finally, limits with infinitely high data rate, th e time duration to carry the payload of CTS, DATA, and ACK frames become infinitesimal. As a result, depending on the network operates in pure 802.11a/b/g, or in hybrid 802.11b/g environment, Tsucc, Tfail, and Tbusy can be expressed by 222 DIFS SIFS TT TTTPHY p pure busy pure fail pure succ++++= ==n (5-2a) and 23 2 2_11 _11 1111 _11 _11 _11DIFSSIFS T TTT T T Tpureg PHY pureg p b PHY b p hybridg busy hybridg fail hybridg succ+++ + ++= = =n (5-2b) Throughput and Delay Once these parameters are obtained, we can use Eq. 3-5 and Eq. 3-10 to formulate the probability distribution of DCF and EDCA, resp ectively. Furthermore, we use Eq. 3-12 to express the maximum achievable throughput for given LDATA bytes long payload in IEEE 802.11 DATA frame. On the other hand, by applying standa rd discrete time queuing theory, we use Eq 3-13 and Eq. 3-14 to derive the to tal packet delay depending on the statistical characteristics of the arrival process. Finally, for packet arrivals that can not be expressed by neither General Independent process nor Deterministic process, Eq. 3-15 and Eq. 3-16 provides an upper bound of the system waiting time and total system delay, respectively.

PAGE 104

104 5.3 IEEE 802.11 MAC Performance limit analysis In this section, we use the queuing system based packet throughput and delay model to quantify and explore the theoretical throughput and delay limits of different IEEE 802.11 MAC specifications. 5.3.1. MAC Layer Packet Service Time To study the theoretical packet throughput and delay lim it of IEEE 802.11 MAC, we first examine the MAC layer packet service time. Note that even with infinitely high data rate, this is the minimum required time that the packets need to wait during MAC backoff due to finite protocol overhead of background traffic in busy slots. Figure 5-1. Average MAC layer service time of different 802.11 specifications Fig. 5-1 plots the average MAC layer service time of different versions of 802.11 standard in the presence of non-saturated b ackground traffic. We can see that packet service time of all 802.11 specifications increases with Pbusy. In a network with business ratio Pbusy as low as 0.5-0.6, the MAC layer packet service time can be in the order of tens of millisec onds. As we will see in later subsections, this significant amount of medium access time limits the achievable packet throughput and delay of I EEE 802.11-based protocols. 0 10 20 30 40 50 0 0.2 0.4 0.6 0.8 1 Busyness ratioAverage packet service time (ms) 0 10 20 30 40 50 0 0.2 0.4 0.6 0.8 1 Busyness ratioAverage packet service time (ms)

PAGE 105

105 Besides, we can refer to the MAC para meters and overhead of different IEEE 802.11 specifications in Table 2-1 and see such pa rameters do affect the packet service time significantly. For example, because the minimum contention window ( CWmin) of 802.11g is only half of that of 802.11b, it takes r oughly half of the time for a packet from 802.11g station to be served, compared to a packet fr om 802.11b station. Due to shorter Tbusy, packet service time of pure 802.11a/g network is further decreased, compared to 8 02.11g station in hybrid 802.11 b/g networks. Moreover, through the medium access advantage of 802.11e AIFS differentiation, the packet service time in 802.11e networks reduces by an order of magnitude, compare to legacy 802.11 protocols. On the other hand, by compari ng pure 802.11g in long slot and short slot setting, we find it interes ting that shorter slot time does not he lp in reducing the packet service time too much. It is because shorter slot time only saves the time spent in idle slots by couple of microseconds, which are relativel y insignificant compared to th e duration of busy slots in hundreds of microseconds. From the observations we made above, we find that, through the progression of recent PHY and MAC amendments on 802.11 standard, the minimum medium access time has been reduced significantly. In the following subsections, we will examine such effects on packet throughput and delay. 5.3.2. Theoretical Throughput Limit In this subsection, we exam ine the theo retical maximum throughput of different 802.11 specifications with infinite data rate. From previous subsection we know that, even with infinitely high data rate, there is a minimum me dium access time that the packets need to wait during MAC backoff. As a result, the maximum am ount of data delivered in a given period of time is bounded by this minimum medium access time and thus a theoretical throughput limit of 802.11 MAC exits.

PAGE 106

106 Figure 5-2. Theoretical throughput limit of different 802.11 specifications Fig. 5-2 plots the theoretical th roughput limits with arbitrary background traffic at different level of busyness ratio. When ther e is no background traffic, i.e. Pbusy =0, our result is consistent with the theoretical thro ughput limit presented in [86]. As Pbusy increases, we can see the maximum throughput decreases expo nentially. In a network with Pbusy about 0.5, the throughput limit has decreased for more than an order of magnitude. Moreover, as the latest IEEE 802.11n proposal aims to provide 100Mbps effective thr oughput at the MAC layer, our result indicates that theoretically, even with in finite data rate, such goal can only be achieved at the condition that the busyness ratio of the network is le ss than 0.3 (or 0.2) with 802.11e AC_VO (or AC_VI) being employed. Such QoS performance boundary is not identified before. 5.3.3. Theoretical Delay Limit From the results we present above, the existenc e of finite MAC service time, in the case of infinitely high data rate, bounds the minimum packet delay that can be achieved by IEEE 802.11-based MAC. Moreover, recall from Sec. 5-2, depending on the type of packet arrival process, additional backlog waiting time will be a dded to the total packet delay. In this section, by using the queuing system based MAC layer packet delay model, we present the results of MTU=2346 bytes 0 100 200 300 400 500 600 700 00.20.40.60.81 Busyness ratioTheoreticalThroughput (Mbps) 802.11e, AC_VO 802.11e, AC_VI 802.11a/g, short slot 802.11g, long slot 802.11b

PAGE 107

107 theoretical delay limit of different 802.11 MAC specifications in the presence of arbitrary background traffic. Figure 5-3. Theoretical delay lim it of different 802.11 specifications Fig. 5-3 shows the theoretical packet delay limit of different 802.11 specifications with infinite operation data rate. Here we use the deterministic arrival process of a G. 711 [40] VoIP application with 10m s inter-arri val time as a case study. We can see the delay limits increase exponentially as Pbusy increases. It is because the MAC layer service time increases as the network get busier, and thus increases the b acklog waiting time significantly. An observation that worth special attention in this figure is that there is a point where the packet delay becomes unbounded. Recall from the queuing model in UF-PASS framework, this is because the fact that whenever the queue service time approaches or even exceeds the packet arrival time (i.e. 10ms in this case), the queue becomes unable to handl e packets in a timely manner, and eventually packets become indefinitely backlogged. The exis tence of such turning point demonstrates an important performance limitation of 802.11-based MA C that is not discovered in previous D/G/1 inter-arrival=10ms0 30 60 90 120 150 00.20.40.60.81 Busyness ratioTheoretical packet delay (ms)

PAGE 108

108 literature: with the presence of background traffi c, the packet delay of 802.11-based MAC can be boundlessly high even with infinite operating data rate. Our model identifies a network condition boundary, such as Pbusy > 0.45 for 10ms-frame G. 711 VoIP in 802.11b networks, that any delaysensitive application may never be able to meet the delay requirement beyond such boundary. Figure 5-4. Theoretical delay limit of 802.11e AC_VI with different arrival process Furthermore, from previous section we know the MAC layer service time reduces as the MAC parameters, such as CWmin, protocol overhead, and AIFS, improve from 802.11b, to 802.11g, to 802.11a, and then to 802.11e. Therefore, we see the turning point of boundless delay moves toward higher busyness ratio in the same order. On the other hand we also explore the effects of arrival process on packet delay limits particularly for the turning point of boundless packet delay. As the packet delay limit of 802.11e appears to be finite in all level of busyness ratio in Fig. 5-3, we plot Fig. 5-4 for the packet delay li mit of prioritized 802.11e MAC with decreasing arrival time. We can see a turning point of unbounded delay eventually emerges as the arrival time decreases less than 2.5ms. It is because when arrival time decreases, it poses stricter delay constraints to the queue system, c onsequently results in packet delay limit curve 802.11e AC_VI, D/G/1 arrival0 30 60 90 120 150 00.20.40.60.81 Busyness ratioAverage packet delay (ms)

PAGE 109

109 moving toward less busy environments. Out result indicates the performance bottleneck also exists for the latest QoS-enabled IEEE 802.11e EDCA. Table 5-1 Comparison of packet delay from model and simulations 5.3.4. Model Validation with TGn Usage Scenarios We further validate the accuracy of our m odel with ns-2 simulations of non-saturated TGn scenarios listed in Table 3-2. With the traffi c characteristics specified in Table 3-2, each TGn scenario corresponds with a particular Pbusy value. We collect simulation results of packet delay for G.711 10ms inter-arrival VoIP traffic and comp are to the delay derived from the analytical model under the same Pbusy value. Table 5-1 shows a close match between our model and simulations (error < 10%). 5.3.5. Summary of theoretical perfor ma nce limitati on analysis In this section, we use the MAC layer service tim e analysis model proposed in UF-PASS framework to investigate the throughput and delay performance limits of different IEEE 802.11MAC specifications. We first show that, even with infinitely high da ta rate, there exists the minimum required MAC layer service time that the packets need to wait during MAC PBusy Delay (Model) Delay (Simulation) Error (%) TGn-1 0.159 2.000ms 1.945ms 2.75% TGn-2 0.217 3.239ms 3.120ms 3.67% 802.11b TGn-3 0.47 656.669ms N/A TGn-1 0.159 0.996ms 0.957ms 3.92% TGn-2 0.217 1.471ms 1.450ms 1.43% 802.11g TGn-3 0.47 15.020ms 13.926ms 7.28% TGn-1 0.159 0.533ms 0.502ms 5.82% TGn-2 0.217 0.609ms 0.664ms 9.03% 802.11e, AC_VI TGn-3 0.47 0.949ms 0.912ms 3.90% TGn-1 0.159 0.899ms 0.810ms 9.90% TGn-2 0.217 1.060ms 1.075ms 1.42% 802.11e, AC_VI TGn-3 0.47 1.798ms 1.773ms 1.39%

PAGE 110

110 backoff due to finite protocol overhead of bac kground traffic in busy slots. We find that, as the MAC parameters and overhead reduce from 802.11b to 802.11a/g, they change Tbusy, Tsucc, and Tfail in the model, and thus affect the packet se rvice time accordingly. On the other hand, we also observe that other backoff settings such as the change of minimum contention window ( CWmin) between 802.11b and 802.11g, and AIFS differentiation of 802.11e affect the packet service time even more significantly as we modeled in UF-PASS framework. Furthermore, we find the maximum amount of da ta delivered in a given period of time is bounded by the minimum medium access time and thus a theoretical throughput limit of 802.11 MAC exits. Such boundary can be quantified by our proposed MAC layer service time model. Besides, as the latest IEEE 802.11n proposal aims to provide 1 00Mbps effective throughput at the MAC layer, our result indicates that theoretically, even with infinite data rate, such goal can only be achieved at the condition that the busyness ratio ( Pbusy) of the network is less than 0.3 (or 0.2) with 802.11e AC_VO (or AC_VI) being employed. Finally, we show that, in the presence of b ackground traffic, the theo retical packet delay limit of the IEEE 802.11 MAC becomes unbounded as the MAC layer service time approaches or even exceeds the packet arrival time. We find that such unbounded delay exists for all variants of IEEE 802.11 MAC, including QoS enabled 802.11e EDCA. 5.4 Discussion 5.4.1. Effects of Competing Traffic Packet Data Rates and Payload Sizes In real-world IEEE 802.11-based wireless networ k deploym ents, the operating data rate is not only finite, it is changing dynamically. A wire less node usually degrade the data bit rate (to incorporate a more resilient modulation scheme) due to increased distance or obstructions such as walls between the access point and the wirele ss node, or due to repeated unsuccessful frame

PAGE 111

111 Figure 5-5. Average packet delay limit with compe ting traffic operates at di fferent data bit rates and different payload sizes transmissions. Besides, the payload sizes of the co mpeting traffic also vary from time to time. It is apparent from the equations in Sec. 5-2, that it takes more time for a wireless station to wait on busy slots when other wireless nodes operate at sl ower data rate or longe r payload size. As a result, increased busy slot time increases the serv ice time and delay and results in significant performance degradation. This problem becomes more important in hybrid IEEE 802.11 networks. As specified in the latest IEEE 802.11n proposal, the fastest data rate is expected as high as 600Mbps. Meanwhile the network is backward compatib le to legacy IEEE 802.11a/b/g devices, which support data rate as low as 1Mbps. In a network that the operating da ta rates vary in such a wide range, the resulting packet service time, and consequently network throughput and delay, may also fluctuate greatly. Th erefore, it is essential to investig ate the performance under the scenarios that nodes operate at different data bit rates and different payload sizes. 802.11e AC_VI, D/G/1 inter-arrival=10ms 0 10 20 30 40 50 60 70 80 90 100 00.10.20.30.40.50.60.70.80.91 Busyness RatioAverage packet delay (ms) Tbusy=0.206ms (8Mbps) Tbusy=0.227ms (600Mbps, 1000bytes) Tbusy=0.351ms (54Mbps, 1000bytes) Tbusy=0.544ms (24Mbps, 1000bytes) Tbusy=0.939s (11Mbps, 1000bytes) Tbusy=1.738ms (5.5Mbps, 1000bytes) Tbusy=2.34ms (2Mbps, 500bytes) Tbusy=4.34ms (2Mbps, 1000bytes) Tbusy=6.34ms (2Mbps, 1500bytes)

PAGE 112

112 As we use Tbusy in UF-PASS framework to model such effect, Fig. 5-5 plots the average packet delay of prioritized 802.11e AC_VI access category with competing traffic operates at different data bit rates and differe nt payload sizes. We vary the da ta rate from infinite to 2Mbps, in which cases the average busy slot length increa ses an order of magnitude. As a result, we can see delay performance varies sign ificantly. For the cases the operati ng data rate is higher than 11Mbps (up to infinite data rate ), the average packet delay main tains below 10ms for all medium busyness ratio. When the operating data rate dr ops to 5.5Mbps, the turning point for boundless delay emerges. In the case when all other node s operate at 2Mbps, the average packet delay might have reached to an unacceptable level (> 100ms) with busyness rati o as little as 0.4. On the other hand, when we fix the data rate at 2M bps and increase the payload sizes of background traffic, we can see the delay turns to be infinitely high at even less busyness ratio ( Pbusy ~0.3). We should keep in mind that data bit rates and payload sizes of competing traffic may change from time to time and are not controlled by any other node in the network. In other words, even the considered node operates at the highest possible data rate and uses the highest priority access category, the performance of the considered node can be constraint by the operating characteristics of other nodes in the network. Such performance limitation is not properly identified and quantified in previous literature. 5.4.2. Performance Improvements on Frame Bu rsting and Block Acknow ledgement From the observations we made in previ ous subsections, we know that the major contributor to the delay in 802.11 based networks is the delay in troduced during backoff stages. It is especially inefficient in terms of channe l utilization that such backoff medium contention has to be repeated for every arriving packet. On the other hand, the amendments of IEEE 802.11 standard specify two frame aggregation schemes to improve channel utilizations by aggregating

PAGE 113

113 multiple transmissions in one medium conten tion, namely Frame-Bursting (FB) and Block Acknowledgment (BTA). Frame-Bursting inserts a burst (say, m number) of DATA frames and corresponding ACK frames back-to-back without initiating anothe r round of random backo ff. In addition, FrameBursting does not require any explicit signa ling between the source and receiver nodes, and hence can be implemented in any IEEE 802.11 based networks. On the other hand, Block ACK mitigates the inefficiency of protocol overhea d by placing a burst of DATA frames separated by a SIFS period without being acknowledged. At th e end of the burst, th e sender initiates an explicit Block ACK Request (BAR) to enquire th e number of frames successfully received by the receiver. The receiver then responds with a Block ACK (BA) frame. The number of frames in a BTA burst (say n) is broadcasted by the access point or pre-negotiated between the sender and the receiver. It is obvious that, with th e same number of data frames in a burst ( m = n), BTA transmits fewer frames and thus saves more overhead compared to FB. Assuming the considered node always has da ta packets to send, we can express the theoretical throughput of FB and BTA aggregation schemes by sli ght modifications in Equation 8 2) ()1(' 8 mSIFSTTB mL TputPHYp DATA FBt+++ tt = (5-3) )2() ()1(' 8 +t+++ tt = nSIFSTTB nL TputPHYp DATA BTA (5-4) Fig. 5-6 shows the theoretical throughput of Frame-bursting and Block ACK schemes with different burst sizes in the presen ces of non-saturated competing tra ffic. In particular, we can see that, with the same burst size m = n=16, the performance improvement s from BTA is greater than that from FB in low busyness environments. The difference in throughput improvements

PAGE 114

114 between BTA and FB, however, becomes insignifi cant as the wireless medium becomes busier. It is because the waiting time in backoff conten tion becomes the dominant factor that limits the theoretical throughput under such network conditions. Figure 5-6. Theoretical throughput of Frame-bursting and Block ACK schemes with different burst sizes 5.5 Conclusion In this chapter, we investigate the th eoretical lim its of I EEE 802.11 MAC throughput and delay performance in the presence of non-satura ted background traffic. A queuing system based analytical model is proposed to evaluate the throughput and dela y bounded by PHY and MAC overhead and backoff waiting time even the operati ng data rate is infinitely high. We present a detailed analysis for theoretical throughput and delay limits in different IEEE 802.11 specifications. We identify a performance bot tleneck beyond which the packet delay becomes infinitely high. Such bottleneck exists for all IEEE 802.11 contention-based DCF and EDCA MAC protocol, although the exact turning point depends on the packet arrival pattern in consideration. By using the UF-PASS performance evaluation framework, we show that such theoretical limits are functions of the MAC layer paramete rs the nodes operate on and the busyness of the MTU=2346 bytes 0 20 40 60 80 100 120 140 160 180 00.20.40.60.81Busyness ratioThroughput (Mbps) BTA, n=64 BTA, n=16 FB, m=16 FB, m=2 legacy 802.11g

PAGE 115

115 wireless medium caused by competing traffic. We study the effects of factors like backoff contention windows, protocol overhead, system slot time, and interframe space time on throughput and delay limits. One of the key obser vations is that the a dvanced medium access opportunity enabled by AIFSn in high priority EDCA voice and video access categories is the primary contributor which significantly improve s the QoS in terms of maximum achievable throughput and minimum achievable delay. The effects of Frame-bursting and Block ACK frame aggregation schemes on packet performance are also discussed.

PAGE 116

116 CHAPTER 6 BEWARE: BACKGROUND TRAFFIC-AWARE RATE ADAPT ATION FOR IEEE 802.11 MAC IEEE 802.11-based devices employ rate adaptation algorithms to dynamically switch data rates to accommodate the fluctuating wireless channel conditions. Many studies observed that, when there are other stations transmitting in the network, existing rate adaptation performance degrades significantly due to the inability of differentiating losses between wireless noise and contention collisions. Previous studies proposed to exploit optional RTS frames to isolate the wireless losses from collision losses, and thus improve rate adaptation performance. In this chapter, we conduct a systematic evaluation on the effectiveness of various existing rate adaptation algorithms and related proposals for loss differentiations, with multiple stations transmitting background traffic in the network. Our main contributions are two-fold. Firstly, we observe that existing RTS-ba sed loss differentiation scheme s do not perform well in all background traffic scenarios. In addition, our study reveals that RTS-based loss differentiation schemes can mislead the rate adaptation algorithms to persis t on using similar data rate combinations regardless of background traffic level, thus result in performance penalty in certain scenarios. The fundamental challenge is that a good rate adaptation algo rithm must dynamically adjust the rate selection decisi on objectives with respect to di fferent background traffic levels. Secondly, we design a new Background traffic awar e rate adaptation algo rithm (BEWARE) that addresses the above challenge. BEWARE uses a ma thematical model to calculate on-the-fly the expected packet transmission time based on current wireless channel and background traffic conditions. Our simulation and te st-bed experiment results s how that BEWARE outperforms other rate adaptation algorithms without RTS loss differentiation by up to 250% and with RTS by up to 25% in throughput.

PAGE 117

117 6.1 Introduction W ith the large-scale deployments of wireless local area networks (WLANs) in homes, offices, and public areas, the IEEE 802.11 standa rd has become the dominant technology in providing low-cost high-bandwidth wireless connections. A large part of the success of IEEE 802.11 protocol can be attributed to the implementation of several simple yet fully distributed algorithms in dealing with the fundamental challenges for wireless communications. For example, the IEEE 802.11 standard employs multiple data rates with different levels of complexity and redundancy in signal modulation and coding schemes to combat the volatile nature of wireless channel. IEEE 802.11-based stat ions then implement rate adaptation algorithm (RAA) to dynamically select the best transmission rate that yields the highest performance in the given wireless channel conditions. The key challenges are that RAA must not only accurately estimate the channel condition in order to infer the most suitable data rate, but also be very responsive to the rapidly fluctuating wireless channel dynamics. A number of approaches have been proposed [43] [7][30] [61] [26][52] [6] [84] to use various metrics such as received signal strength, local Acks, and packet s tatistics to design a RAA in addres sing the above challenges. The effectiveness of RAAs has been extensively evaluated under various wireless channel conditions, when there is only one station in the network. On the other hand, in multiple-user environment, several studies [44] [60] reported that the perfor m ance of some types of RAAs e.g. Automatic Rate Fallback (ARF) [43], degrades drastically b ecause the RAA m istakenly lowers its data rate when the consecutive frame losses are caused by coll ision losses not by wireless losses. There have been a few studies attempting to aid rate adaptati on algorithms in dealing with the collision effects in multi-user environment. Th eir key idea is to provide RAAs the ability to differentiate between wireless losses and col lision losses. For example, by assuming the only

PAGE 118

118 cause for the data frame transmission failure after a successful Reque st-To-Send/Clear-To-Send (RTS/CTS) exchange is due to channe l error not collisi on, the authors in [44] [60] use RTS/CTS to f ilter out collision losses from rate decisi on process. On the other hand, the authors in [60] [91] suggest to add extra fram es and fields to e xplicitly notify the sendi ng station whether the transmission failure is due to collision or channel errors. While these proposals provide significant improvements compared to RAAs without loss differentiation capability, it is un clear whether loss differentiation is good enough to deal with all kinds of mixed wireless and collision loss scenarios. The fundamental problem is, as we will show later in this paper, that background traffic from other c ontending stations changes the throughput ranking of the operating data rates In other words, under the same wireless condition, the data rate yielding the highest throughput in no bac kground traffic scenarios is not necessarily the best one when background traffic exists. This is particularly problematic for existing loss differentiation schemes as they filter out all collision losses for RAA, the RAAs become insensitive to the throughput ranking cha nges caused jointly by wireless losses and collision losses, thus resulti ng in performance degradation. In this chapter, we design a new Bac kground traffic aware Rate Adaptation Algorithm (BEWARE) that explicitly addresses the mixed effects from wireless and collision losses. Our contributions of this paper are: 1) we systema tically evaluate the perf ormance of RTS-based loss differentiation in different mixed wireless and col lision losses scenarios. We identify when and why RTS-based loss differentiation does not work in certain scenarios, 2) we use the insight in these systematic evaluations to identify a novel metric the expected packet transmission time to explicitly address the mixed effects from wireless and collision losses on all available data rates. We propose an online algorithm to estimate th is parameter for all data rates and embed this

PAGE 119

119 information into the RAA design as the key rate decision maker, and 3) by using both simulations and test-bed experi ments, we compare the performance of BEWARE with other RAAs with and without loss differentiation, and observe up to 250% and 25% performance improvement, respectively. The rest of the paper is organized as follo ws. Section 6.2 reviews the existing RAAs and related loss differentiation appro aches. Section 6.3 evaluates the performance of existing RAAs and loss differentiation schemes. Section 6.4 pres ents the design of our background traffic aware rate adaptation algorithm, and Section 6.5 evaluates its perf ormance under various background traffic scenarios via simulations. Section 6.6 pr ovides the real-world driver implementation details and test-bed experiment results Section 6.7 concl udes this chapter. 6.2 Related Work In this sec tion, we briefly review the ex isting rate adaptation algorithms (RAAs) and related loss differentiation schemes that help RAAs deal with collisions in multiple-user environment. We discuss the pros and cons of each approach. 6.2.1 Existing Rate Adaptation Algorithms As the 802.11 standard intentionally leaves the rate adaptation algorithm s open to vendors implementation, there have been quite a few RAAs proposed by academia and industry. They can be broadly classified into three categories based on the information they collect for rate selection decisions: 1) statistics based RAAs, 2) received signal strength (RSS) based RAAs, and 3) hybrid RAAs. 1) Statistics-based rate adaptation algorithms Statistics-based RAAs collects frame transmi ssion statistics such as number of retries, number of frame success and failures. These stat istics are further processed and compared for different rates or pre-set thresholds to infer for current wireless channe l conditions. Based on the

PAGE 120

120 statistics the RAA uses for rate decisions, we can further categorize this class of RAAs into three different approaches. i) Retry-based rate adaptation : This approach [52] [7] uses number of transm ission successes/losses as the indi cator of good/bad wireless condition, and increase/decrease data rate accordingly. For example, ARF [52] decreases the data rate upon two consecutive transm ission losses and increases data rate after ten consecutive transmission successes. However, despite its easy design, previous study [84] has shown that, due to random ness of the wireless loss behavior, ther e is very weak correlation between past consecutive transmission successes/losses and fu ture channel condition, and consequently this approach tends to yield pessi mistic rate estimations. ii) Frame-Error-Rate(FER)-based rate adaptation : this approach [7] [84] calculates FER by the ratio of the num ber of received ACK frames to the number of transmitted frames. The RAA decreases and increases the operation data rate if FER exceeds some pre-dete rmined thresholds. The major drawback is the pre-determined FER thresholds. As wireless channels are so vulnerable to many factors such as multipath, channel fading, and obstructions, it is difficult for one set of pre-determined FER thresholds to fit in all circumstances. iii) Throughput-based rate adaptation : This approach [6] calculates each data-rates throughput based on the packet length, bit-rate, and th e num ber of retries collected during a predefined decision window (~1 sec). Th e major drawback of this approach is the excessive length of the decision window. As th e decision window has to be large enough to collect meaningful statistics, it causes the rate adaptation algo rithm to be less responsive to sudden wireless condition changes. 2) Signal-strength-based ra te adaptation algorithms This class of RAAs [30] [61] relies on wireless signal strength inform ation, such as Received Signal Strength Indicator (RSSI) or Signal-to-Noise Ratio (SNR), to make the rate

PAGE 121

121 adjustment decisions. They assume a strong corr elation between received signal information and the delivery probability of a data rate. The RAAs pick the data rate based on a pre-determined mapping between the received signal streng th and throughput. Meanwhile, there are two approaches to overcome the communication issue of piggybacking the signal strength measurement taken at the receiver side to sende r so that sender can adjust the data rate accordingly. One has to either use explicit signaling [30], which is incomp atible to the IEEE 802.11 standard, or assum e the channel is symmetry [61], which is clearly not the case in realworld scena rios, and thus of little practical value. In addition, this class of RAAs suffers from other drawbacks. Firstly, the rate adjustment mechanism requires a priori channel model to map the received signal information to corresponding data rate throughput In reality, such mapping is highly variable and a model established before-hand may not be applicable to any environments later. Secondly, it is not trivial to obtain reliable signal strength estimation from the radio interfaces. 3) Hybrid rate adaptation algorithms. The hybrid RAA [26] collects both frame transmissi on statistics and received signal streng th, and use statistics-based controller as the core rate adaptation engine. The rate decision can be overridden by signal strength based contro ller if it detects sudde n changes in received wireless signal strength. As hybrid RAA design still assumes sy mmetric wireless channel and pre-established RSSI-to-rate thresholds, this approach is not immune from the drawbacks we discussed in signal-strength-based RAAs section. In summary, all types of RAAs strive to obtain accurate channel estimations from different kinds of loss characteristics and decide when to decrease and when to increase the rate. However, in multiple-user environment, packet collisions in cur new sources of frame losses. None of these

PAGE 122

122 RAAs explicitly address this issue. In the next section, we review several proposals that try to aid RAAs in dealing with collision effects. 6.2.2 Loss Differentiation For Rate Adaptation Previous studies reported that, because AR F treats collision lo sses no different than wireless losses, ARF excessively decreases its rate upon contention collisions, even when wireless channel is close to perfect. As this effect causes severe perf ormance degradation for ARF when background traffic exists, we refer this effect as rate poisoning. There have been two approaches to aid rate adaptation algorithms in differentiating wireless losses from collision losses. i) Loss differentiation by RTS/CTS: With RTS/CTS exchanges preceding data transmissions, [44] and [60] assume that the only cause for the data frame transmission failure af ter a successful RTS/CTS exchange is due to channel error not collision. Therefore, filtered by RTS/CTS, RAA rate decision proc ess reacts only on wireless losse s and is no longer affected by the collision effect. Kim et al. [44] further propose Collision-Aware Rate Adaptation (CARA) to reduce the extra RTS/CTS overhead by selectiv ely turning on RTS/CTS after data frame transmissions fail at least once without RTS/CTS. The data rate is increas ed as the consecutive success count reaches 10, similar to ARF. ii) Loss differentiation by explicit notification: [60] and [91] propose to add extra frames and fields to explic itly notify th e sending station of the source of losses. However, both proposals re quire incompatible changes to the IEEE 802.11 standard and thus are not favorab le for real-world deployments. In summary, loss differentiation is the dominating approach for RAAs in dealing with collision effects when there are other stations transmitting traffic in the network. However, it is not clear whether loss differentia tion is sufficient to guide R AAs to perform well in various multiple-user environments with mixed wire-le ss and contention conditions. As we will show

PAGE 123

123 later in the paper, while RTS-based loss different iation works in certain circumstances, we also find other scenarios that RTS-based loss differentiation performs poorly. 6.3 Performance of Rate Adaptation Algorithms with Background Traffic In this section, we first explain briefl y how does IEEE 802.11 rate adaptation work. In particular, we analyze how rate selection objective vari es with the level of background traffic. Furtherm ore, we systematically evaluate th e performance of various RAAs with RTS loss differentiation schemes under differe nt scenarios. We explore th e interactions between RAAs and mixed wireless and collision effects by varyi ng the number of stations in the network and the distance between stations and access point. Although there have been a few studies [44][60] evaluated ARFs perform ance in multiple-user e nvironment, to our best knowledge, there is no comprehensive study on how other popular RAAs perform when there is background traffic present in the network. As we will show in this section, it is critical to examine how and why these RAAs do not perform well w ith background traffic. By su ch investigation, we not only better understand the necessity fo r a RAA that does take backgr ound traffic into consideration, but also gain insight into how to design such a RAA. 6.3.1 IEEE 802.11 Rate Adaptation with Diff erent Level of Background Traffic To visualize the th roughput-distance tradeo ff among multiple data rates employed by IEEE 802.11 standard, in Fig. 6-1, we use ns-2 [8] to simulate an 802.11a stations maximum throughput as it m oves away from the access poi nt (AP) in Ricean fading environment [63]2. As seen in Fig. 6-1, among the 8 data rates availa ble in IEEE 802.11a, higher data rates can achieve higher throughput, but their transm ission ranges are shorter. The crossing points of two adjacent data rates indicate that, at a gi ven location, the error rate of th e high data rate is becoming too 2 Refer to Sec. 6.5.1 for detailed simulation parameters

PAGE 124

124 high that its actually more favorab le to use the next lower data rate to benefit from the lower error rate. Clearly, rate adaptation mechanism should try to follow su ch transitions as close as possible to select the best da ta rate according to the current wireless channel condition experienced by the link. Ideally, if a rate adap tation mechanism has pe rfect knowledge of the current network condition, its data rate selections follow closely w ith the outer envelope (plotted as thick solid line) of Fig. 6-1. In this wa y, the throughput yielded by the rate adaptation mechanism is always maximized given a particular channel condition. We w ill refer to this outer envelope concept as the oracle-selection strategy and its performance as maximum throughput throughout the paper. Figure 6-1. Throughput versus distance for IEEE 802.11a data On the other hand, Fig. 6-2 plots the performance of the same data rate set under the same wireless channel condition, but with 12 other stations continuously transmitting background traffic in the network. We can see that, not onl y the shape of staircas e like throughput-distance curves changes, but also the pe rformance ranking of data rates for a given location. As a result, the rates selected by the oracle-selection strategy also change for the same location. Fig. 6-3 further plots the rate selections by the oracle-selection strategy when operating with different number of saturated background traffic stations or unsaturated resident ial traffic benchmark 0 5 10 15 20 25 30 051015202530354045Distance (m)Throughput (Mbps) FIX-54M FIX-48M FIX-36M FIX-24M FIX-18M FIX-12M FIX-9M FIX-6M OracleFIX-6M FIX-9M FIX-12M FIX-18M FIX-24M FIX-36M FIX-54M FIX-48M Oracle-selection strategy 0 5 10 15 20 25 30 051015202530354045Distance (m)Throughput (Mbps) FIX-54M FIX-48M FIX-36M FIX-24M FIX-18M FIX-12M FIX-9M FIX-6M OracleFIX-6M FIX-9M FIX-12M FIX-18M FIX-24M FIX-36M FIX-54M FIX-48M Oracle-selection strategy

PAGE 125

125 scenario as specified in [71]. As shown in Fig 6-3, the rate selected by the oracle-selection strateg y varies widely with background traffic inte nsity. In other words, the rate adaptation strategy that works well in one background traffi c scenario may not work in other background traffic scenarios, hence the rate adaptation mechanism needs to explicitly address this phenomenon. We note that, as we defer the discu ssion of the cause of such phenomenon to Sec. 6.4.1, it is not surprise to see the combined effect from both wireless losses and collision losses changes the performance ranking of data rates for a given location. Nonetheless, it is very critical for rate adaptation designs to be aware of such changes and adjust its rate selection strategy to accommodate such changes; otherwise, as we w ill show in the next su bsection, it will suffer from serious performance degradation. Figure 6-2. Throughput versus distance for IEEE 802.11a data rates, with 12 background traffic stations 6.3.2 Performance of RAAs in RTS Access Mode Previous s tudies identified that the lack of ability in differentiating between wireless losses and collision losses is the main problem for ARF to suffer fr om rate poisoning in background traffic scenarios. They reported the superior performance of ARF with RTS on over that with RTS off. However, those studies did not provide systematic inves tigation into whether RAA with 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 051015202530354045Distance (m)Throughput (Mbps) FIX-54M FIX-48M FIX-36M FIX-24M FIX-18M FIX-12M FIX-9M FIX-6M OracleFIX-36M FIX-24M FIX-48M FIX-54M FIX-9MOracle-selection strategy FIX-18M FIX-6M FIX-12M 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 051015202530354045Distance (m)Throughput (Mbps) FIX-54M FIX-48M FIX-36M FIX-24M FIX-18M FIX-12M FIX-9M FIX-6M OracleFIX-36M FIX-24M FIX-48M FIX-54M FIX-9MOracle-selection strategy FIX-18M FIX-6M FIX-12M

PAGE 126

126 RTS really achieves the optimal throughput a nd why it does or does not Besides, we convey detailed comparisons among different RAAs with RTS on, which are also not offered by previous studies. We include representable RAAs from all three classes of statistics-based RAAs, i.e., ARF, ONOE [56], Sample-Rate (SMPL) [6], and RRAA-basic [84], in addition to signal strength based RBAR [30]. Figure 6-3. Best available data rate under different number of background traffic stations With the same simulation settings in the previo us section, we first place all stations at 2.5m away from the AP and turn on RTS for all st ations. We isolate the effects of RTS loss differentiation in performance comparisons by enab ling only one station w ith RAA on, and other background traffic stations with fi xed data rate. When there is little wireless loss for the RAAenabled station, we observe from Fig. 6-4 that all RAAs perform almost the same as the oracleselection strategy regardless of how many stations tr ansmitting background traffic in the network. We then move the RAA-enabled station to 12.5m away from the access point, we can see from Fig. 6-5 that RAAs start to lose track from the best rates and even drop their throughput lower than that is offered by the lowest data rate. In order to inves tigate such performance degradation in more details, in Fig. 6-6, we pl ot the throughput of AR F with RTS (ARF-RTS) 0 12 24 36 48 60 051015202530354045Distance (m)Rate selection by best available strategy (Mbps) # of BK STA=0 # of BK STA=2 # of BK STA=8 # of BK STA=12 TGn residential tfc

PAGE 127

127 (normalized against the maximu m throughput) as it moves away from the AP, with different number of background traffic sta tions. As the RAA -enabled station moves away from the access point, we can see that ARF-RTS starts to lose track of the best available rate. Figure 6-4. Throughput comparison for RAA-enabled station with RTS loss differentiation at 2.5m away from access point, with various num ber of background traffic stations in RTS access mode Figure 6-5. Throughput comparison for RAA-enabled station with RTS loss differentiation at 12.5m away from access point, with various number of background traffic stations in RTS access mode 0 5 10 15 20 25 N=0N=2N=5N=8N=12 Number of background traffic stationsThroughput (Mbps) Best (54Mbps) FIX-6M ARF ONOE RRAA SMPL RBAR 0 2 4 6 8 10 12 14 16 18 N=0N=2N=5N=8N=12 Number of background traffic stationsThroughput (Mbps) Best FIX-6M ARF ONOE RRAA SMPL RBAR

PAGE 128

128 In addition, with more background traffic sta tions transmitting in the network, ARF-RTS stations performance deviates further from the maximum throughput. Finally, ARF-RTS stations performance approaches back to ~90% of maximum throughput when the station is reaching the transmission edge (>35m) of the AP. Figure 6-6. Normalized throughput for ARF-RTS, with various number of background traffic stations in RTS access mode To further explain such scenari o, we plot Fig. 6-7 to illustrate rate selection breakdowns of ARF-RTS along with that of the oracle-selection strategy as distance to access point increases. We can see that the rate selections of ARF-RTS remain almost the same regardless the number of background traffic stations in the networ k, as opposed to the rate selections of oracle-selection strategy that vary widely with background traffic le vel as we discussed above. This is because RTS frames isolate the wireless losses from collision losses and make the rate decisions solely on wireless losses. As a result, RAAs become insensitive to the throughput ranking changes caused jointly by wireless losses and collision losses, and persist in using the rate selections that is only suitable in no bac kground traffic scenarios. We note that, when the distance from station to AP is close-by to intermediate (5m~25m), the optimal rate selections (as selected by of oracle-selection strategy ) for background scenarios

PAGE 129

129 deviate significantly from that for no background tr affic scenarios. Therefore, the rate selection strategy of ARF-RTS becomes part icular problematic and results in significant performance degradation. On the other hand, when the RAA-enabled station is far away from AP, the only operable data rate is the lowest rate. In this ca se, ARF-RTSs rate selection coincides with the Figure 6-7. Data rate selection for ARF-RTS and oracle-selection strategy with various number of background traffic stations in RTS access mode oracle-selection strategy and thus performs close to 100% of maximum throughput. We further examine the rate selections of other statis tics-based RAAs (ONOE, Sample-Rate, and RRAAbasic) with RTS-on in background traffic scenar ios, and find the same phenomenon exists. It follows that, as it is never identified in previous literature, turning on RTS misleads RAAs into using rates only suitable for no-background-tra ffic in scenarios with background traffic, where these rates are not always suitable. As a result, RTS loss differentiation only works well when the rate selections are similar for all other background traffic scenarios. 6.3.3 Performance of Collision-Aware Rate Adaptation (CARA) In this sec tion, we evaluate the performance of another RTS-based loss differentiation scheme, CARA [44], in multi-user environment. In [44], the authors show that CARA outperform s RTS-based loss differentiation by reducing the extra with adaptive RTS-CTS 0 12 24 36 48 60 2.51017.52532.540Distance (m)Rate selection (Mbps) ARF-RTS, # of BK STA=0 ARF-RTS, # of BK STA=2 ARF-RTS, # of BK STA=5 ARF-RTS, # of BK STA=12 Best, # of BK STA=0 Best, # of BK STA=2 Best, # of BK STA=5 Best, # of BK STA=12

PAGE 130

130 exchanges. The mechanism, as called CARA-13 in [44], works in the following manner. By default, the data fram es are transmitted without RTS. When the consecutive failure count reaches probe activation threshold (Pth), the RTS-CTS exchange is activated. If the consecutive failure count further reaches consecu tive failure threshold (Nth), the transmission data rate is decreased. The default values of Pth and Nth are set as 1 and 2 in [44], respectively. The data rate is increased as the consecutive succe ss count reaches 10, sim ilar to ARF. In this subsection, we compare the performan ce of CARA with ARF in basic access mode and the oracle-selection strategy As shown in Fig. 6-8, we vary the distance between the RAAenabled station and access point from 2.5m~45m, and s how the RAA station throughput when operating with 12 other stations transmitting back ground traffic at fixed 54Mbps data rate. We can see that, deviate from what reported in [44], CARA does not always offer superior perform ance over rate-poisoned ARF. In particul ar, when the station is far away from AP (>25m), the most suitable rates turn to be lower rates. In these cases, CARAs adaptive RTS/CTS mechanism only adds overhead to pack et transmissions, and no longer functions as loss differentiator for underlying RAA. On th e other hand, when CARA does outperform ARF, its performance is 15%~25% less than the oracle-selection strategy. It follows that, while RTSbased loss differentiation schemes help RAAs dis tinguish between wireless and collision losses in some scenarios, they do not perform well in some other scenarios. In summary, by systematic evaluations on how di fferent RAAs perform with different operating modes in mixed wireless and collision environm ent, we made the following observations and 3 In this study, we only consider the adap tive RTS/CTS mechanism (called CARA-1 in [44]), as the optional Channel Collision Assessment (CCA) detection, wh ich is called CARA-2 in [44], only provides marginal performance gain over CARA-1

PAGE 131

131 conclusions: i) The oracle-selection strategy varies significantly with th e level of background traffic. We argue that any rate adaptation mechanism should be aware of such change at the presence of background traffic, or it will suffer from serious performance degradation. ii) We show that none of the existi ng RAAs we investigated perform we ll in every background traffic scenario. iii) We see that, even with RTS loss diff erentiation or CARA, there are also situations 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 051015202530354045 Distance (m)Throughput (Mbps) FIX-54M FIX-48M FIX-36M FIX-24M FIX-18M FIX-12M FIX-9M FIX-6M Best ARF CARA-1Best Selection Strategy CARA-1 ARF 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 051015202530354045 Distance (m)Throughput (Mbps) FIX-54M FIX-48M FIX-36M FIX-24M FIX-18M FIX-12M FIX-9M FIX-6M Best ARF CARA-1Best Selection Strategy CARA-1 ARF Figure 6-8. Throughput comparison for ARF, CARA1, and Best ( oracle-selection strategy) with 12 background traffic stations in basic access mode where these mechanisms perform poorly. In fact, in those cases, RTS loss differentiation or CARA hurt the performance. With these valuable observati ons, we present a new background traffic aware RAA design in the next section. 6.4 BEWARE Design From the lessons we learn from the previous section, we know that the key for RAA algorithm to perform well in bac kground traffic scenarios is to incorporate not only wireless channel statistics but also background tra ffic condition as indicators in accessing the effectiveness of each available data rate. Therefore, in Sec. 6.4.1, we first conduct detailed investigations on how to select the best operating data rate which dynamically changes with wireless channel errors and backgr ound traffic level as we observed in previous section. With the

PAGE 132

132 help from UF-PASS framework, we herein identify a metric the expected packet transmission time to explicitly address the mixed effects from wireless and collision losses on all available data rates. Then, in Sec. 6.4.2., we presen t the design of BEWARE, a Background traffic aWaAre RatE adaptation algorithm for IEEE 802.11based MAC. The center part to this design is to use a mathematical model to calculate the expected packet transmission time of each data rate that attributes the combined costs of wireless ch annel errors and background traffic contentions. The rate selection engi ne then uses this metric to find the data rate that yields the highest throughput in the given wireless channel and background traffic condition. The goals to design such rate se lection strategy are two-fold: it has to be robust against any degree of background traffic; m eanwhile, it is also responsive to random and even drastic wireless channel changes. Although using th e expected-packet-trans mission-time as rate selection metric may seem at first as straight forward, it became clear only after our thorough and systematic investigations (in the previous sec tion) on how and why various existing RAAs do not perform well with background traffic. In addition, this concept is novel as no existing studies, to the best of our knowledge, have used such a rigorous metric in RAA design. 6.4.1 Expected Packet Transmission Ti me and IEEE 802.11 Rate Adaptation Recall from Ch. 3, UF-PASS framework uses a se ries of parameters to model the impacts of various factors includ ing background traffic (by Pbusy & Tbusy), wireless losses (by Pfail), and operation data rates (by Tfail & Tsucc) on IEEE 802.11 MAC performance. Therefore, we find it very suitable to explain the performance rankin g changes of data rates with different background traffic levels we observe in previous section. Particularly, as we recall from Eq. 3-12, the throughput of an 802.11 station is inversely proportional to the exp ected time duration the packet spent in MAC backoff procedures. It follows that we can use the MAC layer service time model,

PAGE 133

133 i.e., the expected packet transmission time we de rive in Ch. 3, to explain the performance of different data rates. Figure 6-9. Frame error probability ( Pfail) and the MAC layer service time of 36Mbps and 24Mbps, no background traffic We begin the investigation by considering the case when there is no background traffic. Using 36Mbps and 24Mbps, two adjacent data rate s available in IEEE 802.11a, as example, we plot the average MAC layer service time and corresponding frame error probability ( Pfail) as the distance between station and AP increases in Fig. 6-9. When the station is close to AP (<10m), both data rates show very low Pfail, and the expected packet transmission time corresponds to their packet transmission time ( Tsucc: 0.446ms for 36Mbps packets and 0.625ms for 24Mbps) plus average idle slot times ( Tslot: 0.009ms) spent in the first ba ckoff stage (roughly 15/2 slots). As a result, the performance for 36Mbps packets is better than 24Mbps when the station is closeby. As the station moves away from AP, Pfail of 24Mbps increases slower than that of 36Mbps due to more robust modulation schemes. Therefore, the expected packet transmission time of 24Mbps also increases slower than that of 36Mbps. As shown in Fig. 6-9, when Pfail of 36Mbps (~35%) is significantly higher than that of 24Mbps (~15%), there is a crossing point where the

PAGE 134

134 expected packet transmission time of 36Mbps b ecomes higher than that of 24Mbps. It follows that beyond such point, the performance for 24Mbps packets is better than 36Mbps. Figure 6-10. MAC layer service time of 36Mbps and 24Mbps, va rious background traffic level On the other hand, in Fig. 6-10, we plot the expected packet transm ission time of 36Mbps and 24Mbps with different background traffic levels We first can see that, when the station is close to AP (<10m), the expected packet tran smission time of 24Mbps and 36Mbps both increase significantly. It is because, recall from Ch. 3, th e effects of background traffic contribute to longer ( Tbusy) and more frequent ( Pbusy) wait time spent in backoff stages, and thus longer MAC layer service time. More importantly, we can furt her observe that the cros sing point at which the performance for 24Mbps becomes better than 36Mbps has moved closer to AP (where Tfail is smaller) as more background traffic stations jo in the network. We know from the UF-PASS model discussed in Ch. 3 that higher Pfail increases the chances for the packets to undergo more number of backoff stages. Therefore, as the wa iting time in all backoff stages becomes longer with background traffic, with ev en just little increase in Pfail, the packets suffer from excessive waiting time by undergoing more number of backoff stages, and thus a significant increase in

PAGE 135

135 MAC layer service time. As a re sult, the performance of 36Mbps starts to become worse than 24Mbps when Pfail of 36Mbps (~5%) is only slightly high er than that of 24Mbps (~2%). Furthermore, in Fig. 6-11, we plot the rati o for throughput of 36Mbps over that of 24Mbps with different number of background traffic stat ions. We can see the turning point (where throughput ratio = 1) moves closer to AP wh en the number of background traffic stations increases. On the other hand, we plot the same throughput ratio with smaller background traffic medium occupation time ( Tbusy) in Fig. 6-12. We can see the tu rning point changes are not as wide as that in Fig. 6-11 when Tbusy is smaller. Figure 6-11. Throughput ratio (36Mbps over 24M bps) under various background traffic level, Tbusy = 0.334ms In summary, as we use the parameters in UF -PASS framework to evaluate the performance of IEEE 802.11 MAC data rates in background traffic scenarios, we find that the turning point of adjacent data rates changes dynamically with bac kground traffic. We also find that the expected packet transmission time is a good indicator to estimate the effec tiveness of different available data rates, as it accommodates th e effects of background traffic by a series of parameters, i.e. Pbusy Pfail, and Tbusy, as we proposed in UF-PASS framework. In the presence of background

PAGE 136

136 traffic, all these parameters should all be cons idered when making rate selection decisions, as opposed to concentrating on just wireless losses or the differentiation of wireless losses and collision losses as it is believed in previous studies. In addition, instead of what previous study [64] identified that the core of RAAs is on when to decrease and when to increase the transmission rate, here, we argue that the rate selection objectives in terms of where (or which level) to decrease/increase to which change with background tr affic intensity, are essential to RAA design in multi-user environment. As we show in Sec. 6.3, RAAs suffer from serious performance degradation if thei r rate adaptation designs are no t aware of such changes and do not adjust their rate selection stra tegy to accommodate such changes. Figure 6-12. Throughput ratio (36Mbps over 24M bps) under various background traffic level, Tbusy = 0.224ms 6.4.2 Rate Selection Engine In the previous subsection, we can see that the expected pa cket transm ission time is a good indicator to gauge the mixed effects from wirele ss channel condition and co llisions for the data rates. In this section, we de scribe how BEWARE makes rate selection decisions by using the UF-PASS framework to estimate the expected packet transmission time as the rate selection metric. Moreover, BEWARE adopts more caref ul measures in probing data rates, and

PAGE 137

137 implements various schemes in dealing with mo re dynamic packet transmission statistics from the mixed wireless channel and background traffic environment. Figure 6-13. Structure of BEWARE design As shown in Fig. 6-13, the BEWARE design can be broken down into the following tasks: 1) Statistics collection/processing: As we will use the models in UF-PASS framework to estimate the data rate performance, we need to ga ther the necessary parameters we described in Ch. 3. To be more specific, after the packet transmission completes, we collect transmission environment statistics, including Tbusy, Pbusy, and Pfail, and process them by exponentially weighted moving average (EWMA) to smooth out th e biases to the sudden changes in current wireless channel and collision conditions. We can determine Pfail by counting the ratio of failed packet transmission attempts and total pack et transmission attempts. We also obtain Pbusy/ Tbusy by keeping track of the number/duration of experi enced busy medium slots, respectively. On the other hand, Tfail and Tsucc are directly determined by the operating data rate and Tslot is specified in different version of IEEE 802.11 standard. In addition, BEWARE keeps track other statistics such as number of successful/failed packets of different data rates. 2) Expected packet transmission time calculation: With the environmental parameters collected in the above module, this module uses the mathematic al model described in Sec. 3.1.1

PAGE 138

138 to calculate the expected packet transmission time. Particularly, we use Eq. 3-2 and Eq. 3-3 to find j knkS and j knkT ,, respectively. We then use the followi ng equation to obtain the expected packet transmission time as .)1( )*(1 0 0 1 0 = == = = =m i i N k N kn m j j knk j knk avgWN where TS T (6-1) Tavg is further updated with recent history values by EWMA and fed into rate selection module for processing. 3) Rate probing : Periodically, BEWARE sends packets at a data rate other than the current one to update the expected transmission time of other data rates. In order to avoid the common rate-probing pitfalls reported in [84], BEWARE adopts various m easures to ensure probing other data rates is not done very often and the cost is not too high. BEWARE limits the frequency of packet probing to a fraction (~5 %) of the total transmission time. BEWARE also limits the number of retries allowed for probing packets to 2 to save costly waiting time for unsuccessful probing. In addition, BE WARE does not probe data rate s that suffer from excessive failures for most recent packet attempts (4 rece nt successive packets have been unacknowledged), and those whose expected transmission time w ith no background traffic already exceed the expected transmission time of current operating data rate. 4) Rate selection decisions: The rate selection module cons tantly compares the expected packet transmission time of current data rate and that of others and decides to change operating data rate whenever it finds a data rate yields the shorter transmission time (and thus highest throughput) beyond a certain thresh old. BEWARE also implements a short-term frame loss reaction mechanism in case wireless channel cond itions change too rapidly. The rate selection

PAGE 139

139 module forces data rate to decrease one level wh en the packets exhaust all retries for three times consecutively. 6.5 Simulation Results In this section, we use ns-2 [8] to evaluate the performance of BEWARE and other RTSbased loss differentiation RAAs, including ARF w ith RTS/CTS (as referred to ARF-RTS) and CARA-1 under various m ixed wireless and background traffic scenarios. 6.5.1 Simulation Setup We enhance the ns-2 simulator to support 802.1 1a Physical layer (PHY) and port various RAAs from previous studies or the real-world driver implementations [56]. We simulate scenarios in an infrastructure-b ased networ k, which contains one Access Point (AP) and a number of static wireless stations spreading in the network. We cons ider realistic wireless channel conditions by using and Ricean fadi ng model with parameter K=6, and environment maximum velocity v=10m/s. The traffic sources are UDP flows unless stated otherwise. Figure 6-14. Throughput comparison for Best ( oracle-selection strategy ), BEWARE, CARA1, and ARF with RTS/CTS, with 12 backgr ound traffic stations in RTS access mode 0 0.5 1 1.5 2 051015202530354045 Distance (m)Throughput (Mbps) Best BEWARE CARA1 ARF-RTS

PAGE 140

140 Figure 6-15. Data rate selections for BEWARE and oracle-selection strategy, with various number of background traffic st ations in RTS access mode Figure 6-16. Data rate selection for BE WARE and ARF-RTS with dynamically changing background traffic payload size 6.5.2 Performance Of Single Stat ion With Varying Distance We first focus on RAAs performance with varying distance under background traffic scenarios. We place 2~12 stations on a circle around the AP within 2.5 meter radius, and all stations transmit UDP background traffic with RTS access mode. The transmission data rate of background traffic stations is locked at 54Mbps b ecause of their proximity to AP. We then add one RAA-enable station in the network and measure the RAAs performance by varying the distance between RAA-enable station and AP. We show results with 12 stations transmitting 0 12 24 36 48 60 2.51017.52532.540Distance (m)Rate selection (Mbps) BEWARE, # of BK STA=0 BEWARE, # of BK STA=2 BEWARE, # of BK STA=5 BEWARE, # of BK STA=12 Best, # of BK STA=0 Best, # of BK STA=2 Best, # of BK STA=5 Best, # of BK STA=12

PAGE 141

141 background traffic as an exampl e in Fig. 6-14, while results w ith other number of background traffic stations show similar tr end. In all cases, the performance of BEWARE follows closely to what is offered by the oracle-selection strategy by only 10% less in throughput, and the performance of CARA-1 trails behind BEWARE by another 10%-15%. On the other hand, similar to what we discuss in Sec. 6.3.2., the pe rformance of ARF-RTS significantly derails from the oracle-selection strategy when the distance from station to AP is close by to moderate (5m~25m). From Fig. 6-15, we can further see that the rate sel ections by BEWARE do change according to different background traffic level an d stay close to the rate selections by the oracleselection strategy 6.5.3 Performance Of Single Station With Dy namically Changing Background Traffic In this subsection, we further investigate how different RAAs adapt with dynam ically changing background traffic levels. We place 12 b ackground traffic stations randomly scattered in the network and always use th e lowest transmission rate (6M bps) to guarantee high packet delivery rate. We synchronize the traffic patterns of the background traffic stations so that, for every 3~5 seconds, they all change packet payl oad size around the same time. We then add one RAA-enabled station in the network a nd measure the RAAs performance. We can see from Fig. 6-16, as the average pa cket size of background traffic changes, the average data rate used by ARF-RTS does not sh ow notable changes. On the other hand, we can see that BEWARE tries to adapt its rate selec tions as background traffic packet size changes. Recall from the discussions in Sec. 6.4, as the high er data rates are more vulnerable to have more backoff stages due to high wireless loss rates ( Pfail), the longer backoff stages caused by increased background traffic payload sizes make the higher data rates less favorable to operate in situations with large background traffic payload si ze. Therefore, we can see that BEWARE tries to adapt to the lower data rates when it sens es such changes. On the other hand, when the

PAGE 142

142 background traffic payload sizes ar e small, BEWARE tries to use the highest possible data rate for optimal performance. We observe that BEWARE outperforms ARF-RTS for ~20% in throughput in this dynamically changing background tr affic environment. We further investigate the effects of other background traffic changing pa tterns, such as increas ing the frequency of background traffic payload size fluctuations and changing th e number of simultaneously transmitting stations, and we observe that BE WARE consistently outperforms ARF-RTS for 25%-50% in various dynamic changi ng background traffic scenarios. 6.5.4 Aggregated Performance with Different Topology We now evaluate aggreg ate performance when all stations turn on RAA and operate with the same RAA homogeneously. We first simulate a topology with minimum wireless losses, in which various numbers of stations are uniformly placed at 2.5m away from AP and each station transmits fixed size 1500-byte long UDP traffi c. As shown in Fig. 6-17, ARFs aggregate performance degrades severely due to the rate poisoning effect we discussed in Sec. 6.3. On the other hand, with the help from RTS loss differentiation, ARF-RTS performs well for any number of contending stations. Furthermore, BEWARE and CARA-1 pe rform closely and both outperform ARF-RTS in most cases, thanks to the overhead reduction design in CARA-1 and accurate background traffic effect estimation in BEWARE. Secondly, we simulate a random topology with various numbers of stations randomly scattered in the network with maximum distan ce 45m away from AP to guarantee no hidden terminals. Each station transmits UDP traffic with random size. As shown in Fig. 6-18, the performance ranking differs from what we observe in Fig. 6-17. While ARF still suffers from rate poisoning and performs the worst, CARA -1 no longer outperforms ARF-RTS and ranks second from the worst. It is because, as nodes spreading at different distance to AP, both wireless loss and contention losses affect the first data frame transmissions (without-RTS by default)

PAGE 143

143 specified in CARA-1 algorithm, which cause CA RA-1 stations de-crease data rate over aggressively. On the other hand, BEWARE still performs the best in random topology. On average, BEWARE outperforms ARF by 200% to 250% and ARF-RTS, the best proposed by previous studies, by 20% to 25% in aggregate performance. Figure 6-17. Aggregate thr oughput comparison for BEWARE, CARA1, ARFRTS, and ARF in close-by topology with various number of contending stations Figure 6-18. Aggregate thr oughput comparison for BEWARE, CARA1, ARF-RTS, and ARF in random topology with various number of contending stations 0 5 10 15 20 25 30 051015 Number of contending stationsAggregated throughput (Mbps ) BEWARE CARA-1 ARF-RTS ARF 0 5 10 15 0 5 10 15 Number of contending stationsAggregated throughput (Mbps) BEWARE CARA-1 ARF-RTS ARF

PAGE 144

144 Figure 6-19. Aggregate thr oughput comparison for BEWARE, CARA1, ARF with RTS, and ARF in random topology under different Ricean Parameter K. a) K=0, b) K=6, c) k=12. 6.5.5 Aggregated Performance Under Various Channel Fading Conditions We now compare the perform ance of diffe rent RAAs under various channel fading conditions. We vary the Ricean pa rameter K and Doppler spread fm. Note that, as K increases, the line-of-sight component is st ronger and the overall channel SNR increases. On the other hand, as fm increases, the channel condi tion changes more rapidly. Fi g. 6-19 plots the aggregate performance of different RAAs under different K in a random topology similar to what we used 0 5 10 15 0 5 10 15 Number of contending stationsAggregated throughput (Mbps) BEWARE CARA-1 ARF-RTS ARF(a) 0 5 10 15 0 5 10 15 Number of contending stationsAggregated throughput (Mbps) BEWARE CARA-1 ARF-RTS ARF(b) 0 5 10 15 0 5 10 15 Number of contending stationsAggregated throughput (Mbps) BEWARE CARA-1 ARF-RTS ARF (c)

PAGE 145

145 in previous sub-section. We can see that, as K increases, the overal l throughput all RAAs increases as expected. However, the ranki ng of RAA performance remains unchanged. BEWARE outperforms ARF-RTS, CARA-1, a nd ARF under all different K parameters. Figure 6-20. Aggregate thr oughput comparison for BEWARE, CARA1, ARF with RTS, and ARF in random topology under different Doppler Spread fm. a) fm=3.5HZ (v= 0.2m/s), b) fm=10HZ (v= 0.6m/s), c) fm=17HZ (v= 1m/s) We then plot Fig. 6-20 with the aggregate pe rformance of different RAAs under different Doppler spread. We can see that, as fm decreases, BEWARE still out performs ARF-RTS in most cases, but the performance gap between BEWARE a nd ARF-RTS closes. To be more specific, as BEWARE outperforms ARF-RTS by 25% when fm =17Hz, this advantage decreases to 5% when 0 5 10 15 051015 Number of contending stationsAggregated throughput (Mbps) BEWARE CARA-1 ARF-RTS ARF(a) 0 5 10 15 051 01 5 Number of contending stationsAggregated throughput (Mbps) BEWARE CARA-1 ARF-RTS ARF(b) 0 5 10 15 051 01 5 Number of contending stationsAggregated throughput (Mbps) BEWARE CARA-1 ARF-RTS ARF(c)

PAGE 146

146 fm decreases to 3.5Hz. Previous studies [91] [64] reported that, as ARF is designed to increase its rate af ter several consecutiv e packet successes, ARF-base d RAA tends to yield higher throughput by taking advantage of the slower ch anging channel environment. However, the performance of ARF degrades when the wireless channel condition changes rapidly. On the other hand, we can see that, as BEWARE yields comparable performance in different fm environments, BEWARE is robust to both fast-changing and slow-changing wireless channel conditions. 6.5.6 Performance With Heterogeneous RAA Deployments As rate adaptation is an option that is left open for wireless card vendors to im plement, it is not uncommon that there are stations equipped with di fferent RAAs in real world scenarios. Therefore, it is essential to evaluate the pe rformance of different RAAs in heterogeneous scenarios. In this experiment, we evaluate how different RAAs improve the individual and aggregate performance with a gradual upgrade deployment. We consider a network with 12 stations randomly placed within the transmission range of the AP, and transmit UDP traffic with random size. We start with the baseline scenario where all stations ope rate with ARF without RTS/CTS, in which all stations operate at the lowest data rate due to rate poisoning problem. We then gradually upgrade a number of stations with BEWARE or ARF-RTS, and evaluate the aggregate performance improvement over base line scenario and individual performance improvement of the same station after upgrade. We can see from Fig. 6-21 that, as the aggr egate performance of ARF-RTS improves when upgraded stations added to the network, the individual performance of ARF-RTS actually decreases when less than half of the stations in the network ar e upgraded. When there are just a few stations upgraded with ARFRTS, individual performance of upgraded stations decrease due to excessive use of higher data rates as we discuss in S ec. 6.3.2. Meanwhile, aggregate performance increases as other stations ta ke advantage of the excess loss transmission

PAGE 147

147 opportunities incurred by upgraded stations. On th e other hand, when there are more and more stations upgraded with ARF-RTS, ARF-RTS stati ons mutually take advantage of other upgraded stations loss transmission oppor tunities, and collectively resu lt in higher aggregate throughput even the rate selections made by these sta tions are not the most suitable ones for the corresponding scenario. By contrast, both individual and aggregate perf ormance of BEWARE start to improve when just 1 station is upgrad ed. In addition, as the stations upgraded with BEWARE start to use data rates that is appr opriate for the given wireless and collision conditions, other stations benefit from the extra free transmission time spared by BEWARE stations, and thus yields higher throughput even they are not upgr aded with BEWARE. Note that this is an essential feature that, when incorpor ating any new algorithm to interoperate with other existing algorithms, the new algorithm should not hurt the performan ce of other existing algorithms. -50% 0% 50% 100% 150% 200% 250% 024681012 Number of upgraded stationsThroughput improvement (%) ARF-RTS, individial ARF-RTS, aggregate BEWARE, individual BEWARE, aggregate Figure 6-21. Individual and Aggregate throughput improvement of BEWARE and ARF-RTS with various number of contending st ations in heterogeneous deployments In summary, with the homogeneous and he terogeneous background traffic scenarios we evaluate in this section, we observe that while the effectiven ess of RTS-based loss

PAGE 148

148 differentiation RAAs differ in different scenario s, BEWARE always yields the best performance for most cases. In addition, even with only one station equipped with BEWARE in the network, both individual performance of BEWARE and aggregate networ k performance improve over the rate-poisoned all-ARF network. 6.6 Test-bed Implementation and Evaluation In the previous section, we have seen BEWA REs superior perform ance over other RAAs under various simulated background tr affic and wireless scenarios. In order to demonstrate that our ideas can be practically implemented in real hardware and still exhibits desirable performance, in this section, we implement our BEWARE algorithm in Atheros-based Linux device drivers. In addition, we use the UF-PASS framework to design a series of experiments to systematically evaluate and compare the perf ormance of BEWARE and other RAAs under the real-world scenarios. In Sec. 6.6.1, we first describe our implementation efforts including the necessary modifications and relevant trade-offs when we deal with the real hardware. In Sec. 6.6.2, we present the test-bed evaluation results. 6.6.1 Implementation We i mplement the BEWARE algorithm on open source MADWIFI [56] driver based on Atheros chipsets. W e also port the ARF rate ad aptation algorithm for comparison purposes. Our implementation follows the system design struct ure we described in Sec. 6.4.2, in which we estimate the expected packet transmission time with the backoff procedure parameters (i.e. Tbusy, Pbusy, Pfail, Tsucc, and Tfail). In the following, we describe the different challenges that we face in implementing the system modules in real hardware and the necessary modifications to accommodate such challenges.

PAGE 149

149 1) Statistics collection/processing: The first challenge we face is in obtaining some of the parameters needed for the algorithm; particularly the parameters depend on individual backoff stages, i.e. Tbusy and Pbusy. While MADWIFI may be the mo st accessible open source WLAN driver available in the commun ity that implements many packet transmission details in the software, such as packet encapsulations, Qo S settings, and transmission buffers, MADWIFI leaves the control and feedback of backoff procedur e details in the firmware. In other words, it is not possible for us to control or even know exactly how many backoff counters used in a particular backoff stage. As a result, we are unable to keep track the number/duration of busy medium slots to obtain Tbusy and Pbusy, as we described in the simu lations. While this information may ultimately be available from the 802.11 chipsets if we have the access to the firmware, we develop an alternative estimation-based approach to this problem as follows. Recall from the UF-PASS framework presented in Ch. 3, we know that Tbusy and Pbusy (as well as Tslot) determine the length of individual backoff stages. As a result, instead of collecting Tbusy and Pbusy and use them to characterize length of individual backoff stages, an obvious alternative is to keep track the length of individual backoff stages directly. Particularly, we can obtain the of 1st backoff stage by logging the length of all non-retransmitting successful transmissions, and subtract it by th e actual packet transmission time (Tsucc). We note that logging the length of transmissions that involve re-tra nsmissions is not a good choice since they involve different backoff stages and potentially different packet transmission time ( Tfail and Tsucc). It is important to note that k eeping track individual backoff stage length ta kes longer time to provide the up-to-date channel information for the rate adaptation decisions, compared with our original approach by using time-slot level statistics (i.e. Pbusy and Tbusy). We will show later in the experiment results that this estimation-ba sed approach does not seem to affect the overall

PAGE 150

150 performance of BEWARE algorithm in real-world scenarios. In a ddition, there is an interesting trade-off that collecting time-slot level statistics (i.e. Pbusy and Tbusy) might prevent the stations from going into sleep mode, which is critical fo r energy savings. In other words, using the new estimation-based approach might be advantag eous from the power consumption standpoint. However, since power consumption issues are not the focus of this dissertation, we leave the issues in exploring the tradeoff between energy sa vings and collecting most up-to-date statistics as one of the topics in our future work. 2) Expected packet transmission time calculation: Since Pbusy and Tbusy are no longer available in our driver implement ation, we need to modify the model for calculating the expected packet transmission time. Recall from the model in Ch. 3, we construct the derivation of overall backoff duration by the cumulative effects from th e successive backoff stages. Therefore, as we get the 1st backoff stage length, T1st-stage, in our new approach, we can calculate the overall backoff stage duration as, = ++ =m n fail n fail succ fail stagest n avgP PTTnT T1 )1( 1 )1()].1(**))1(*2[((6-2) Note that, compared with the model in Ch.3, this model also simp lifies the computation complexity and makes it more suitable to be implem ented in real-world driv er. On the other hand, since we have made several estimations to th e environment dynamics in this new design, it is important, as we will show in the next subsection, to design a series of experiments to fully expose the new implementation approach to a rich set of real-world dynamics. 6.6.2 Experimental Setup Our experim ental setup consists of one Cisco AP-1230 access point, which supports all 802.11a/b/g channels and data rates, and la ptops equipped with Proxim Orinoco Gold

PAGE 151

151 802.11a/b/g combo PCMCIA cards. The laptops run Ubuntu Linux with kernel version 2.6.24.5 and modified MADWIFI driver based on version 0.9.4. We conduct both indoor and outdoor experiments in the University of Florida campus. The indoor experiments are carried out in the 4th floor of Computer Scie nce/Engineering building (CSE). The outdoor experiments are carried out at the open ar ea between Computer Science/Engineering building and Marston Scienc e Library building (MSL). We also carry out some uncontrolled field trials in the Reitz Student Union (REI) food court. Figure 6-22. Indoor experiments la yout. (floor plan provided by The Facilities Planning and Construction Division of University of Florida) 20m : Background traffic stations LOC#1 LOC#2 LOC#3 : Hidden Terminal

PAGE 152

152 In order to provide fair co mparisons among different RAAs, we perform most experiments in a controlled manner so that there is minimum impact from external factors, such as people walking around, operation of microwave ovens, and tr affic from existing AP or client devices. Specifically, we choose channel 40 of 802.11a to conduct the indoor and outdoor experiments, and we confirm with WLAN sniffer that there is no other AP or background traffic exist in that channel. The indoor experiments are done during late evenings when the offices are mostly empty and no one is walking in the nearby ha llways. The outdoor experiments are done during a holiday long weekend afternoon when there is only very occasional pedestrian and vehicle traffic in the area. On the other hand, in order to de monstrate the usability of the proposed BEWARE algorithm, we conduct a series of uncontrolled fi eld trials in a busy campus food court, deployed with legacy 802.11g APs that are setup by the Un iversity Information Technology Office. There are a number of students using their laptop wire less connection to watch video, download music, and collaborate projects while we conduct the field trials. Fig. 6-22 shows the layout of the floor we r un indoor experiments. The environment is an office/lab setting with concrete walls separati ng the rooms and many meta l cubical partitions within the lab. We place the AP and 3 backgr ound traffic laptops in E401 lab. The background traffic laptops are situated within 1m to the AP and the operation data rates and packet sizes are specified in different experiment s. We choose three different loca tions to test RAA performance. Location #1 is in lab E401 and within 1m to the AP so that we can examine the RAAs performance when the wireless cond ition is almost perfect. Location #2 is at the corridor outside E401, between room E467 and E468A, which is obs tructed by 2 concrete walls from the AP location and the direct distance is about 12m. At this location, the average SINR is 26 to 24 db. Location #3 is further down the hall to room E466 with direct distance about 20m. The average

PAGE 153

153 SINR at this location is 16 to 18 db. We further choose an extra location at the other side of the corridor, the corner between E404 and E403, to place a hidden terminal background traffic laptop so that we can evaluate the RAAs perf ormance under hidden terminal situations. More details of hidden terminal setup is outlined in the next subsection. CSE Building Marston Science Library LOS-1LOS-2 NLOS : Background tr affic stations : Measurement points for RAA-enabled stations 28m N Figure 6-23. Outdoor experiments layout. Fig. 6-23shows the layout of the area we r un outdoor experiments. We place the AP in front of the main entrance to computer lab E211 We choose two Line-of-Sight locations (LOS-1

PAGE 154

154 and LOS-2) in the open area outs ide E211 with direct distance about 28m and 38m, respectively. We also choose one non-LOS (NLOS) location at the side that is blocked by two building poles and the direct distance is about 35m away from the AP. We conduct each experiment with multiple runs, and present the results that are averaged over all runs. Each run last for 60~90 seconds an d the performance result s are normalized to per second basis. We use the Linux kernel-level traffic generator, pktgen, to generate continuous UDP traffic to the sink that is connected with the access point via 100Mbps Ethernet connection. We also use TCP traffic in the field trials to te st BEWAREs performance in real-world settings. We vary the data rates, packet payload sizes, a nd traffic on-off patterns in different experiments. The detailed settings are described along with the experiment results presented in the next subsection. 6.6.3 Test-Bed Performance Evaluation Result In this subsection, we conduct a series of syst em atic experiments to evaluate and compare BEWAREs performance in real-world scenarios. The objective of the experiments is to not only help us understand BEWAREs performance in different scenarios, but also expose BEWARE in the dynamics of real-world situations where simulations ma y not be able to capture. We compare the performance of BEWARE with ARF and ARF-RTS. From the analysis in the previous sections, we know that ARF is one of the pione er and popular rate adaptation algorithms and is employed in many wireless devices. However, ARF suffers from ratepoisoning problem when there is some background traffic in the network. ARF-RTS is the solution proposed by later studies that has been widely accepted by the community for its ability in helping RAAs deal with background traffic. However, we have shown in previous sections that, using RTS to differentiate the losses betw een wireless and collisions can sometimes be misleading and resulting in performance degrad ation. We believe that comparing BEWAREs

PAGE 155

155 performance with these algorithms provides a good overall picture for understanding how different rate adaptation algorithms perform in real-world scenar ios with different wireless loss and background traffic environments. We design a series of systematic experiments th at vary a number of dimensions in order to investigate how different factors, and their corresponding parameters as described in UF-PASS framework in Ch.3, impact the RAAs performance. We also note that, as we use an estimationbased approach for the driver implementations ot her than the detailed pa rameter-based approach in the simulations, it is important to design the experiments with mix and match these different factors such that we will expos e the new design approach to a ri ch set of real-world dynamics and we can understand how these factors or the mix of different factors affect the performance of the new design. These test-bed experimental sensitivity analyses help us carefully study the performance impacts of a wide range of factors on the new de sign, so that we can understand whether the new design can capture the mixe d effects from the environment and make appropriate rate selection deci sions without overestimating or underestimating the environment dynamics. Traffic pattern: There are a number of different tr affic parameters that we vary to test the RAAs performance in different traffic scen arios. We vary the number of background traffic stations to evaluate the effect of Pbusy as we described in Sec. 6.3.1. We then will set the stations in non-saturation m ode with different level of Pbusy, and different payload sizes that corresponds to the effect of Tbusy in UF-PASS framework. Wireless loss: As we described above, we put RAA-enabled stations in different spots with different wireless signal levels to test the effect of Pfail in UF-PASS framework. We also test the performance in different wireless fading environments, in cluding indoor office building setting and outdoor open-area setti ng with Line-of-Sight (LOS) and Non-LOS locations. We even create a hidden-terminal s cenario to evaluate the RAAs performance in such situation. RTS & Basic mode: We test the effect of RTS loss differentiation as we discussed in Sec. 6.3.2. We will turn on various rate adaptation algorithms and evaluate their performance with and without RTS.

PAGE 156

156 Individual and aggregate RAA performance: We evaluate the individual and aggregate performance of RAAs as we replace different number of RAA-enabled stations with either the RTS-based loss different iation RAAs or BEWARE. In the following, we present the test-bed expe riment results in different scenarios we investigate. Figure 6-24. Normalized throughput for BE WARE over ARF and BEWARE over ARF-RTS in controlled indoor environment with number of background traffic stations 1) Indoor performance at different locati ons with different number of background traffic stations: We place 3 background traffic stations ne xt to the AP to create 4 different background traffic levels, i.e. w ith 0, 1, 2, 3 stations transmit ting background traffic. Each background traffic station uses th e lowest data rate to ensure that the background traffic is detectable at the farthest range of the AP. Each background tra ffic station is configured to transmit continuous UDP packets with payloa d size 500 bytes long. We then place one RAAenabled station, transmitting 1500 bytes long packet s, in the three diffe rent indoor locations referred in Fig. 6-22 to investigate the RAAs effectiveness under mixed wireless loss and contention conditions. In Fig. 6-24, we plot the performance of BEWARE normalized by either ARF-RTS or ARF, and are shown with square markers and tr iangular markers, respectively. The two thin

PAGE 157

157 solid lines show that, at locati on #1 where RAA-enabled station is just next to the AP, BEWARE does not provide significant performance improve ment against ARF-RTS & ARF. In some cases when there are two background traffic stations, BEWAREs throughput is even slightly lower (< 2%) than ARF-RTS. It is because, even the num ber of background traffic stations are fixed, the background traffic level indicator, i.e. Pbusy, fluctuates s over time. As BEWARE constantly adjusts rate selections to the changing backgroun d traffic levels, it sometimes mis-uses the lower data rates while the highest data rate should always be the best c hoice for this particular location. On the other hand, when we move the RAA-enab led station to locati on #2 (dotted lines) and location #3 (thick solid lines) we can see from Fig. 6-24 that BEWARE consistently outperforms ARF-RTS, ARF, in all background traffic scenarios. At location #3, BEWAREs performance improvements over ARF-RTS are more significant, when compared with the performance at location. #2. Recall from Sec. 6.5, when the RAA-enabled station is farther away from the AP, ARF-RTS is more vulnerable to mislead the rate adapta tion algorithm to choose wrong data rates. In addition, BEWAREs perf ormance improvement increases with more background traffic in the network. 2) Indoor performance with switching traffic: In this experiment, we set one of background traffic stations to have an on-off sw itching traffic pattern with 10-sec duty cycle. Other background stations are set to have stable UDP traffic. We again place one RAA-enabled station in the three different i ndoor locations to investigate the RAAs effectiven ess under this switching traffic situation. It is important for us to test the RAAs ability, especially for BEWARE, in dealing with vary ing background traffic levels. In Fig. 6-25, we plot the th roughput performance of BEWARE, ARF, and ARF-RTS at the three locations (Fig. 6-25a, Fig. 6-25b, and Fig. 6-25c), when there is just one switching traffic

PAGE 158

158 station (blue bars, marked as BK STA=1) or combined one switching traffic and two stable background traffic stations (purple bars, marked as BK STA=3). As we can see from Fig. 6-25a, at location #1, BEWARE is slightly outperf ormed (2%~10%) by ARF & ARF-RTS when there are 1 switching traffic station plus 2 stable background tra ffic stations. This is an effect similar to what we observe in the previous experiment. When the BEWARE station is very close to the AP, BEWARE sometimes overly chases the fluctua ting background traffic levels, such that it occasionally misuses the lower data rates, while th e highest data rate should always be the right selection at this location. Fortunately, we can see that this problem only causes marginal performance degradation for BEWARE, while we see much higher performance improvements at other locations. Figure 6-25. UDP performance of BEWARE, ARF, and ARF-RTS with switching traffic pattern at different locations in the indoor controlled environment. a) at Location #1, b) Location #2, c) Location #3 (a) (b) ( c )

PAGE 159

159 As we can see in Fig. 6-25c, BEWARE outperforms ARF and ARF-RTS up to 2.5 times in throughput, under both background traffic scenarios. Besides, as ARF-RTS suffers from very high packet error rate (up to 35%) at location #3, we can see that BEWARE always keeps the packet error rate very low (<5%) in all scenarios, which proves that BEWARE always tries to find the best data rate that is most suitable for the given wireless loss and contention conditions. 3) Indoor performance with heterogeneous deployment: In this experiment, we use three RAA-enabled stations, which all transm it 1500-byte UDP traffic continuously. We keep one in location #1 and move the other two together to either loca tion #2 or location #3. Initially we intentionally set all stations operate ARF as the rate adaptation algorithm, which suffers from the rate poisoning problem. We then gradually upgrade the stations to either ARF-RTS or BEWARE to combat the ARFs rate poisoning pr oblem. This experiment is essential not only to understand how different upgraded RAA stations deal with the rate poisoning problem, but also understand how the RAA-enabled st ations interact with each other. Figure 6-26. Individual and aggregate pe rformance of BEWARE and ARF-RTS under heterogeneous deployments at different locations in the indoor controlled environment. a) at Location #2, b) Location #3 As we can see in Fig. 6-26a, when there are two RAA-enabled stations at location #2, both upgrading to ARF-RTS and BEWARE can improve the aggregate network throughput and the ( a ) ( b )

PAGE 160

160 individual throughput of the upgrad ed stations. In addition, there is no significant differences to upgrade to ARF-RTS or BEWARE. However, in Fig. 6-26b, when the two RAA-enabled stations move to location #3, upgrading the ARF station to ARF-RTS can result in severe performance degradation. Only by upgrading th e ARF stations to BEWARE stations, both aggregate performance and upgrad ed stations individual performance consis tently increase as the upgrade gets deployed. In a ddition, we can see that the up graded BEWARE station can interoperate well with other stations which either uses ARF or BEWARE. 4) Indoor performance with hidden terminal: The last experiment for indoor office environment is to investigate the effect of hidden terminal. We keep one background traffic station next to the AP and place another back ground traffic station to the hidden terminal location we marked in Fig. 6-22, and set it to continuously transmit 500-byte UDP packets at 36Mbps data rate. At location #2, we use a wireless sniffer to make sure that stations at this location can only decode the packets from the b ackground traffic station next to the AP, but not the hidden terminal station. We then place one R AA-enabled station at location #2 to test its ability to deal with hidden terminal (and mixed with other decoda ble background traffic). Figure 6-27. Throughput perfor mance for BEWARE, BEWARE with RTS, ARF, and ARFRTS in controlled indoor envir onment with hidden terminal

PAGE 161

161 As we can see from the light blue bars in Fi g. 6-27, when the hidden terminal is the only competing traffic, BEWARE (with or without RT S) performs comparably with ARF-RTS. ARF suffers from performance degradation in this hidden terminal scenario, even the amount of background traffic is relatively light. On the other hand, when we turn on the background station next to the AP, we can see from the purple ba rs in Fig. 6-27, while BEWARE (without RTS) only outperform ARF-RTS for 15%, BEWARE with RTS outperfor ms ARF-RTS for about 60%. As we know that RTS is an important in comb ating hidden terminal problem, this experiment demonstrates that BEWARE can also be incorporated with RTS to maximize the performance in hidden terminal scenarios. Figure 6-28. UDP performance of BEWARE, ARF, and ARF-RTS at different locations in the outdoor controlled environment. a) at Location LOS-1, b) Location LOS-2, c) Location NLOS (a) (b) (c)

PAGE 162

162 5) Outdoor performance at different locati ons with different num ber of background traffic stations: We place 2 background traffic stations next to the AP and one RAA-enabled station in the three different locations as shown in Fig. 6-23. We compare the performance of BEWARE, ARF, and ARF-RTS at these three lo cations and when both the background traffic stations are either turned on or turned off in the experiment. As we can see from Fig. 6-28, BEWARE consis tently outperforms ARF and ARF-RTS, in all 3 locations and in both bac kground traffic levels. BEWAREs performance advantage is more significant when there are more background traf fic in the network. In addition, BEWAREs packet loss rate is always < 2% in all scenar ios evaluated. On the other hand, in this outdoor experiment, both ARF and ARF-RTS suffer from substantial packet loss rate, up to 18% in no background traffic scenario and up to 35% in 2 background traffic station scenario. 6) Indoor performance in a crowded campus caf: In this experiment, we take one RAA-enabled laptop to a crowded caf in Univ ersity of Florida campus, during a busy lunch period. We use the 802.11g access points installed in the caf to test the RAAs performance. The WLAN sniffer detects more than 40 unique devices operating at the time we conduct the measurement. However, since the over-the-air traffic in this scenario is not stable at all times, we do not intend to compare the performance of different RAAs, but we want to show that BEWARE does work in the legacy system with uncontrolled environment. We use VisualWare online VoIP Connection Quality and Bandwidth Speed test suite [80] to test the bandwidth, delay, and jitter of the connection between the RAA-enabled laptop and an outside server. As we can see from the the pe rformance test result in Table 6-1, BEWARE and ARF-RTS show comparable result s and both outperform ARF in most categories. We further use

PAGE 163

163 BEWARE to make a Skype [69] 3-way voice conference call to our lab in CSE building, all calling parties report very good ca ll quality throughout the call. Table 6-1 Performance in a crowded campus caf 6.6.4 Summary of Experimental Results In the previous subsection, we conduct a series of system atic test-bed experiments to evaluate the performance of BEWARE, ARF, and ARF-RTS under the mixed effects from various factors, including traffic pattern, wireless environment, a nd protocol settings. The results show that BEWARE consistently outperforms other algorithms in most scenarios. The performance advantage of BEWARE becomes more significant when there are more background traffic in the network and the test station is in termediate to far away from the AP, which is consistent to our simulation findings in S ec. 6.5. We also observe a minor weakness of BEWARE, in which BEWARE overly adjusts the rate selections when the station is actually very close to the AP. However, this probl em only accounts for less than 10% performance degradation, while the performance gain of BEWARE in other mixed wireless and contention scenario far outweigh its weakne ss. Furthermore, we conduct performance field trials in an uncontrolled environment with legacy 802.11g systems, and observe superb performance in using BEWARE. Download bandwidth (Mbps) Delay (ms) Jitter (ms) Mean Opinion Score (MOS) BEWARE 16.8 0.1 0.1 4.1 ARF 14.5 1.2 6.2 4 ARF-RTS 16.2 0.1 0.1 4.1

PAGE 164

164 6.7 Conclusion In this chapter, we first identify that data rate selection strategies of 802.11-based stations should acco mmodate the different rate selectio n criterions in different background traffic scenarios. This observation further helps us ex plain why RTS-based loss differentiation schemes, which are proposed by previous studies to aid rate adaptation algorithms in dealing with collision effects, do not perform well in certain scenario s. In particular, RTS-based loss differentiation hurts the performance by persisten tly using the same rate sele ctions regardless of background traffic level. Therefore, these observations motivat e us to design a rate adaptation algorithm that explicitly addresses wireless and contention factors in its design. We propose a novel background traffic-aware ra te adaptation, BEWARE, that uses an accurate mathematical model to estimate the effec tiveness of the data rates in given wireless and contention conditions. We show th at the rate selections of BE WARE are close to what are selected by the oracle-selection strategy that has global knowledge of network conditions. We also show that, compared to other RTS-based loss differentiation scheme s, BEWARE yields the best performance in scenarios we investigated in the paper. In additi on, we implement BEWARE into the real 802.11a wireless card driver, and co nduct a series of systematic experiments to evaluate BEWAREs performance in real-wor ld scenarios. We observe that BEWARE outperforms other RAAs under various wireless loss and contention conditions. Meanwhile, we plan to investigate the interac tions between rate adaptation algorithms and upper-layer protocols such as TCP. We believe that, as the design of BEWARE fully addresses the wireless and contention factor s in MAC layer, it should render the best performance when integrated with upper-layer protocols.

PAGE 165

165 CHAPTER 7 CONCLUSION AND FUTURE RESEARCH DIRECTIONS 7.1 Summary of the Dissertation In this dissertation, we propose UF-PASS fra mework, a Unified Framework for Performance AnalySiS of contention-based wireless MAC. The key idea of UF-PASS framework is that, by categorizing the protocol operation parameters that exhibit similar behavior into the same group and interconnecting the target scenarios with relevant performance affecting parameters, we can formulate a system atic framework that provides unified viewpoints to the performance impacts on different target sc enarios. Through a series of case studies, we demonstrate that we can reuse the same framework to explain the similarity or dissimilarity of the performance impacts that are affected by a mixture of protocol operation parameters and environment factors such as wireless losses, co llisions, and different background traffic load levels. We also show that, by using UF-PASS framework, we gain thorough understandings of the inherent interactions betw een different protocol dynamics, such as different Quality of Service (QoS) settings and differe nt protocol parameters from he terogeneous protocol standard variants (e.g. 802.11a/b/g). Particularly, we have applie d our framework to the following case studies and made the following key findings: i) In hybrid 802.11b/g netw orks, we use the framework to provide a unified explanation to the mixed effects among different contention windo w, different backoff stages, and different data rates and frame format s on these two versions of standard. We identify and quantify a throughput anomaly that penalizes fast 802.11g stations and privileges the slow 802.11b station, and provide insight to different performance improvement schemes that alleviate such throughput anomaly effect. ii) We use the framework to study the performance limitations on throughput and delay of IEEE 802.11 MAC, under the effects of background

PAGE 166

166 traffic intensity, data bit rates, and payload sizes of the competing traffic. We identify a performance bottleneck beyond which the packet delay becomes infinitely high and we pinpoint the exact turning point depends on the packet arri val pattern in consideration. iii) We use the proposed framework to evaluate the effects of wireless channel conditions and competing background traffic in multi-rate environments. We are able to pinpoint the problems of the existing rate adaptation techni ques and uncover the fundamental challenge in designing a rate adaptation algorithm is to address explicitly the mixed effects of wireless channel conditions and competing background traffic levels. We then propos e to use our analytical model as the engine in selecting the best operating data rate influe nced jointly by wireless and collision losses. Through a series of systematic simulation and test -bed experiments, we show that our proposed rate adaptation algorithm exhibits up to 250% performance improvement over existing rate adaptation algorithms in various scenarios. In summary, UF-PASS framework is not only an analytical model that provides accurate performance predictions to a wide range of s cenarios in which the system performance is affected by a mixture of protocol operation para meters and environment factors. It is also a performance evaluation framework that systemati cally and coherently en compasses the protocol dynamics and provides thorough understanding of th e inherent interactions between different factors of various scenarios a nd system performance. Furthe rmore, by applying the insight gained from studying the protoc ol dynamics, the UF-PASS framework provides guidelines for various protocol enhancement designs and a flexible model that can be inte grated into real-world hardware implementations that signi ficantly improves network efficiency.

PAGE 167

167 7.2 Future Research Directions As we have seen the UF-PASS model b een very effective in modeling IE EE 802.11 performance in a wide range of aspects in the MAC layer, we expect to extend and generalize this framework in various dimensions. Firstly, we can generalize the framework to evaluate the performance issues in multi-hop environments. In particular, we can see that, by concatenating the perlink MAC layer service time calculated by UF-PASS framework, we can generalize the UF-PASS framework to provide detailed end-to-end performance evaluation. We expect that the UF-PASS framework can provide insightful and coherent viewpoints to multi-hop network performance in complicated operating scenarios, such as heterogeneous data rates and arbitrary b ackground traffic, which have not been evaluated before. In addition, as there have been a number of literature [2][18] [20] focusing on expectedm edium-time based routing metrics in multi-hop wireless networks, we can also use the expected-packet-transmission-time concept in UF -PASS framework to offer comparative studies in multi-hop routing. We expect that, by usi ng UF-PASS based expected-packet-transmissiontime as routing metric, the link capacity is more accurately estimated with the considerations of effects of background traffic contention, interference. The multi-hop wireless mesh networks case stud ies can be carried out with the following different aspects. We first apply the UF-P ASS framework to evaluate the end-to-end performance by concatenating the performance hop-by-hop. Furthermore, with the reuse of routing protocols built by the existing studies, we evaluate the effectiveness of replacing simple retransmission rate and link ba ndwidth as routing metrics [18] [20] by UF-PASS based routing m etric. We can further generalize the framew ork to evaluate the wireless multi-hop routing performance with rate adaptation mechanism deploye d at the wireless links. We expect that, as

PAGE 168

168 the design of BEWARE fully addresses the wi reless and contention factors in MAC layer, BEWARE would also perform very well when incorporated with UF-PASS based expectedpacket-transmission-time routing metrics. Secondly, we can use the UF-PASS framework to investigate the intera ctions between rate adaptation algorithms and upper-layer protocols such as TCP. Such ripple effect studies are essential for end-to-end performance in wirele ss and wired-cum-wireless networks. Previous study [51] has reported a correlation between TCP window size and MAC data rate (with and without au to rate adaptation) in wireless multihop routing performance. We can use UF-PASS framework to further investigate such effects with various rate adaptation algorithms, including BEWARE, under various background traffic levels. We then follo w the design rationale of UFPASS framework to further construct a model that incorporate the effects of upper-layer protocols with MAC protocols.

PAGE 169

169 APPENDIX A BEWARE IMPLEMENTATION In this appendix we provide the high-level descriptions of the BEWARE im plementation that we adopted from MADWIFI driver. As the implementation follows the design we presented in Fig. 6-13, we describe the implementation of the four modules in the following four subsections. A.1 Statistics Collection Module As we described in Sec. 6.6, the BEWARE i mplementation collects a number of transmission statistics including T1st-stage, Pfail, Tfail, and Tsucc. While Tfail and Tsucc are known to the transmission, we need to modi fy the driver to provide us T1st-stage and Pfail. We follow the technique described in [57], which determines the receiving time of the ACK fra mes by requesting an inte rrupt after each succe ssful transmission. We also record the time that the packet was added to the hardware queue. As a result, we can determine the MAC layer processing time by choosing the lesser of the timestamps between these two events and the timestamps between the current ACK frame recei ving time and previous ACK frame receiving time. We implement this MAC layer processing tim e recording subroutine in if_ath.c file, under the main madwifi code segments (located in ath directory) On the other hand, the MAC layer processing tim e is further sent to the tx_complete() function in the rate control code segments (locat ed in ath_rate directory). We note that, in the rate control modules, the tran smit descriptors also report th e number of retry attempts ( ds_txstat.ts_shortretry and ds_txstat.ts_longretry ) for each packet. Therefore, we then obtain T1st-stage by subtracting the MAC layer processing time of the packet transmissions that do not experience any MAC layer retransmissions ( ds_txstat.ts_shortretry = ds_txstat.ts_longretry = 0) with the corresponding Psucc. Furthermore, we can also determine the packet failure probability,

PAGE 170

170 Pfail = R /(P + R ), where R is the total number of retries and P is the number of successful packet transmissions. A.2 Expected Packet Transmission Time Module In this m odule, we implement exactly Eq. 6-2 in tx_complete() functi on with the statistics we collect from the above process, namely T1st-stage, Pfail, Tfail, and Tsucc. The expected packet transmission time is recorded according to the data rate used and the packet payload size bins it belongs to (currently the size bins are set as 0~250, 251~1500, >1500 bytes). A.3 Rate Probing Module Periodically, BEWARE s ends packets at a data ra te other than the current one to probe and update the expected transmissi on time of other data rates. We implement such module in findrate() function in the rate control code segments. The find rate() function is executed every time the packet enters the MAC layer buffer. Partic ularly, about every 20 packets, we try to pick a data rate that satisfies the following rules. If th ere is no data rate that satisfies the rules at the moment, we do not send any probe packets. We do not probe the currently used rate. We do not probe the rate that has failed the la st 4 packet transmissions (regardless they are probed packets or regular packets). We do not probe the data rate that is 2 or more rate levels higher than the current data rate, unless such data rates Pfail is smaller than current data rates Pfail plus 0.25. We never probe the 9Mbps data rate in 802.11a/g standard. A.4 Rate Selection Module In findrate() function, we also im plement the fo llowing rules to select the data rate based on the shortest expected packet transmission time. Whenever the expected packet transmission tim e of the current data rate is longer than twice the shortest expected packet transmissi on time. The data rate of next packet is switched to the data rate with the shor test expected packet transmission time.

PAGE 171

171 When 2 seconds past and no data rate switch ha ppened, we switch to the data rate with the shortest expected packet transmission time if the expected packet transmission time of the current data rate is longer than 1.05 times the shortest exp ected packet transmission time and the data rate is smaller than the current da ta rate but larger than the current data rate minus three rate levels. When 2 seconds past and no data rate switch ha ppened, we switch to the data rate with the shortest expected packet transmission time if the expected packet transmission time of the current data rate is longer than 1.05 times the shortest exp ected packet transmission time and the data rate is larger than the current da ta rate but not smaller than the current data rate plus three rate levels. When 2 seconds past and no data rate switch ha ppened, we switch to the data rate with the shortest expected packet transmission time if the expected packet transmission time of the current data rate is longer than 1.1 times the shortest exp ected packet transmission time and the data rate is larger than the current data rate plus three rate levels.

PAGE 172

172 LIST OF REFERENCES [1] F. Alizadeh-Shabdiz and S. Subramaniam, "A nalytical Models for Single-Hop and MultiHop Ad Hoc Networks," in Proc. ACM Broadnets 2004. [2] B. Awerbuch, D. Holmer, and H. Rubens, H igh throughput route sel ection in multi-rate ad hoc wireless networks, in Wireless On-Demand Network Systems (WONS), January 2004. [3] G. Bianchi, "IEEE 802.11-Saturation throughput analysis," IEEE Commun. Lett., vol. 2, pp. 318-320, Dec. 1998. [4] G. Bianchi, "Performance analysis of the IEEE 802.11 distributed c oordination function," IEEE Journal on Selected Areas in Communications 2000. [5] G. Bianchi and I. Tinnirello, Kalman Filte r Estimation of the Number of Competing Terminals in an IEEE 802.11 network, INFOCOM 2003. [6] J. Bicket, Bit-rate Selection in Wire less Networks, MIT Masters Thesis, 2005. [7] B.E. Braswell, J.C. McEachen, Modeling Data Rate Agility in the IEEE 802.11a WLAN Protocol, OPNETWORK [8] L. Breslau, D. Estrin, K. Fall, S. Floyd, J. Heidemann, A. Helmy, P. Huang, S. McCanne, K. Varadhan, Y. Xu, and H. Yu. Advances in Network Simulation. IEEE Computer, 33 (5 ), pp. 59-67, May, 2000 [9] Broadcom Inc., Xpress Technology: Maxi mizing performance in 802.11 wireless LANs, Broadcom BRCM-WP2-802.11-030617. [10] H. Bruneel and B. G. Kim, Discrete-tim e models for communicat ion systems including ATM, Kluwer Academic Publisher, 1993. [11] F. Cali, M. Conti, and E. Gregori, "I EEE 802.11 Wireless LAN: Capacity Analysis and Protocol Enhancement," in Proc. IEEE Infocom 1998. [12] G. R. Cantieni, Q. Ni, C. Barakat, and T. Turletti, "Performance Analysis under Finite Load and Improvements for Multira te 802.11," Comp. Commun., pp. 1095-1109, June 2005. [13] P. Chatzimisios, A.C. Boucouvalas, V. Vitsas, "IEEE 802.11 Packet Delay-A Finite Retry Limit Analysis", in Proc. IEEE Globecom 2003. [14] P. Chatzimisios, V. Vitsas, and A. C. Boucouvalas, "On improving performance for IEEE 802.11 wireless lans under congested and error-prone environments," IASTED international conference on Internet and multimedia systems and applications 2006.

PAGE 173

173 [15] Y. Chen, Q.-A. Zeng, and D. P. Agrawal, "Performance Analysis of IEEE 802.11e Enhanced Distributed Coordinati on Function," in Proc.IEEE ICON 2003. [16] H.S. Chhaya and S. Gupta, Performance m odeling f asynchronous data transfer methods of IEEE 802.11 MAC protocol, Wireless Networks, Vol 3, pp.217-234, 1997 [17] J. Choi, K. Park, and C. Kim, "Cross-Laye r Analysis of Rate Adaptation, DCF and TCP in Multi-Rate WLANs," in Proc. IEEE INFOCOM 2007 [18] D. S. J. D. Couto, D. Aguayo, J. Bicket, a nd R. Morris, A high-th roughput path metric for multi-hop wireless networks, in 9th annual international conference on mobile computing and networking (MobiCom 03), September 2003. [19] A. Doufexi, S. Armour, B.-S. Lee, A Nix a nd D Bull, An Evaluation of the Performance of IEEE 802.11a and 802.11g Wireless Local Ar ea Networks in a Corporate Office Environment, ICC 2003. [20] R. Draves, J. Padhye, and B. Zill, Routing in multi-radio, multi-hop wireless mesh networks, in ACM MobiCom, September 2004. [21] P. E. Engelstad and O. N. Osterbo, "Analysis of the Total Delay of IEEE 802.11e EDCA and 802.11 DCF," in Proc. IEEE ICC June 2006. [22] Enhanced Wireless Consortium, Medium HT MAC Specification, Interoperability MAC Specification v1.24, Enhanced Wirele ss Consortium publication, Jan. 2006. [23] M. Ergen and P. Varaiya, "Throughput Formulation and WLAN Optimization in Mixed Data Rates for IEEE 802.11 DCF Mode," in Proc. IEEE Globecom 2004. [24] C. H. Foh and M. Zukerman, "A New Tec hnique for Performance Evaluation of Random Access Protocols," in Proc European Wireless 2002. [25] Hadzi-Velkov and B. Spasenovski, "Satura tion Throughput Delay Analysis of IEEE 802.11 DCF in Fading Channel," in Proc. IEEE ICC 03. [26] I. Hatatcherev, K. Langendoen, R. Lagendijk, H. Sips, Hybrid rate control for IEEE 802.11, ACM MobiWac [27] M. Heusse, F. Rousseau, G. Berger-Sabbatel and A. Duda, Performance Anomaly of 802.11b, INFOCOM 2003. [28] M. Heusse, F. Rousseau, R. Guillier, and A. Duda, Idle Sense: An Optimal Access Method for High Throughput and Fairness in Rate Diverse Wireless LANs, ACM SIGCOMM [29] M.-J. Ho, J. Wang, K. Shelby and H. Haisch, IEEE 802.11g OFDM WLAN throughput performance VTC 2003-Fall.

PAGE 174

174 [30] G. Holland, N. Vaidya, and P. Bahl, A Rate-Adaptive MAC Protocol for Multi-Hop Wireless Networks, ACM MobiCom [31] S. Hori, Y. Inoue, T. Sakata, and M. Mo rikura, "System capacity and cell radius comparison with several high data ra te WLANs," IEEE 802.11-02-159r1, Mar. 2002. [32] C. Hu and J. Hou, A Novel Approach to Contention Control in IEEE 802.11e-Operated WLANs, IEEE INFOCOM [33] K.C. Huang and K.C. Chen, Interference an alysis of nonpersistent CSMA with hidden terminals in multi-cell wireless data networks, IEEE PIMRC 1995. [34] IEEE Standard for Wireless LAN Medium A ccess Control (MAC) and Physical Layer (PHY) Specifications, Sep. 1999 [35] IEEE Standard for Wireless LAN Medium A ccess Control (MAC) and Physical Layer (PHY) Specifications: High-Speed Physical Layer in the 5 GHz band (802.11a), Sep. 1999. [36] IEEE Standard for Wireless LAN Medium A ccess Control (MAC) and Physical Layer (PHY) Specifications: Higher Data Rate Extension in the 2.4 GHz Band (802.11b), Sep. 1999. [37] IEEE Standard for Wireless LAN Medium A ccess Control (MAC) and Physical Layer (PHY) Specifications, Amendment 8: Medium Access Control Quality of Service (QoS) Enhancements (802.11e), Jan. 2005. [38] IEEE Standard for Wireless LAN Medium A ccess Control (MAC) and Physical Layer (PHY) Specifications, Amendment 4: Further Higher Data Rate Extension in the 2.4 GHz Band (802.11g), June. 2003 [39] ITU-T Recommendations, "G.114 : One-way transmission time," International Telecommunication Union, 2003. [40] ITU-T Recommendations, "G.711 : Pulse code modulation (PCM) of voice frequencies," International Telecommunication Union, 1988. [41] I. Inan, F. Keceli, and E. Ayanoglu, "Saturation Throughput Analysis of the 802.11e Enhanced Distributed Channel Acce ss Function," in Proc. IEEE ICC 2007. [42] V.K. Jones, R. DeVegt, and J. Terry, "I nterest for HDR extension to 802.11a," IEEE 802.11-02-081r0, Jan. 2002. [43] A. Kamerman, L. Monteban, WaveLAN-II: A High-Performance Wireless LAN for the Unlicensed Band, Bell Labs Techni cal Journal, pp. 118, Summer [44] J.S. Kim, S.K. Kim, S.H. Choi, D. Qiao, CARA: collision-aware rate adaptation for IEEE 802.11 WLANs, IEEE INFOCOM

PAGE 175

175 [45] L. Kleinrock, Queueing Syst ems. John Wiley and Sons, 1975. [46] L. Kleinrock and F. Tobagi, Packet switc hing in radio channels, PartIIThe hidden terminal problem in carrier sense multiple access and the busy tone solution, IEEE Trans. Comm., vol. COM-23, no. 12, Dec. 1975 [47] L. Kleinrock, Queueing Systems, Volume II: Computer Applications, Wiley Interscience, New York, 1976. [48] L. Kleinrock, Queueing Systems, Volume II: Computer Applications, Wiley Interscience, New York, 1976. [49] L. Kleinrock and F. Tobagi, Packet switc hing in radio channels, PartIIThe hidden terminal problem in carrier sense multiple access andthe busy tone solution, IEEE Trans. Comm., vol. COM-23, no. 12, Dec. 1975 [50] Z. Kong, D. H. K. Tsang, B. Bensaou, and D. Gao, "Performance Analysis of the IEEE 802.11e Contention-Based Channel Access," IEEE J. Select. Areas Commun., pp. 20952106, December 2004. [51] D. Koutsonikolas, J. Dyaberi, P. Garime lla, S. Fahmy, and Y. Hu, On TCP throughput and window size in a multihop wire less network testbed, Wintech07 [52] M. Lacage, M. H. Manshaei, T. Turlet ti, IEEE 802.11 rate adap tation: a practical approach, ACM MSWiM [53] W. Lee, C. Wang, and K. Sohraby, "On Us e of Traditional M/G/1 Model for IEEE 802.11 DCF in Unsaturated Traffic C onditions," in Proc. IEEE WCNC 2006. [54] B. Li and R. Battiti, "Analysis of th e IEEE 802.11 DCF with Service Differentiation Support in Non-Saturation Conditions," in Proc. QoFIS 2004. [55] Y. Lin and V. W. Wong, "Saturation Throughput of IEEE 802.11e EDCA Based on Mean Value Analysis," in Proc. IEEE WCNC 2006. [56] MADWIFI, http://sourcefor ge.net/projects/madwifi/ [57] D. Malone, K. Duffy, and D. J. Leith, "Modeling the 802.11 Distri buted Coordination Function in Non-Saturated Conditions," IEEE Commun. Lett., pp. 715-717, August 2005. [58] K. Medepalli and F. A. Tobagi, "Thro ughput analysis of IEEE 802.11 wireless LANs using an average cycle time approach," in Proc. IEEE Globecom 2005. [59] K. Medepalli and F. A. Tobagi, "System Centric and User Centric Queueing Models for IEEE 802.11 based Wireless LANs," in Proc. IEEE Broadnets 2005. [60] Q. Pang, V.C.M. Leung, and S.C. Liew, A Rate Adaptation Algorithm for IEEE 802.11 WLANs Based on MAC-Layer Loss Differentiation, IEEE BRAODNETS

PAGE 176

176 [61] J. P. Pavon and S. Choi, Link Ad aptation Strategy for IEEE 802.11 WLAN via Received Signal Strength Measurement, IEEE ICC. [62] A. D. Poularikas and S. Seely, Signals a nd Systems, Second Edition, Krieger Publishing Co., Malabar, Florida, 1994. [63] R. J. Punnoose, P. V. Nikitin, and D. D. St ancil, Efficient Simulation of Ricean Fading within a Packet Simulator, IEEE VTC-Fall. [64] D. Qiao and S. Choi, Fast-Responsive Link Adaptation for IEEE 802.11 WLANs, in Proc. IEEE ICC [65] G. Rajappan, J. Acharya, H. Liu, N. Manda yam, I. Seskar, and R. Yates, "Mobile infostation network technology," Proceed ings of SPIE -Wireless Sensing and Processing 2006 [66] J. W. Robinson and T. S. Randhawa, "Saturation Throughput Anal ysis of IEEE 802.11e Enhanced Distributed Coordination Function," IEEE J. Select. Areas Commun., pp. 917928, June 2004. [67] L.D. Servi, D/G/1 queue with vacation, Oper. Res. (1986). [68] A. Singh and D. Starobinski, "A Semi Ma rkov-based Analysis of Rate Adaptation Algorithms in Wireless LANs," in Proc. IEEE SECON 2007 [69] Skype, http://www.skype.com [70] A. P. Stephens et al., IEEE P802.11 Wireless LANs Usage Models, IEEE 802.1103/802r23, May 2004 [71] A. P. Stephens et al., "802.11 TGn Functional Requirements," IEEE 802.11-03/802r23, May 2004. [72] J. W. Tantra, C. H. Foh, I. Tinnirello, a nd G. Bianchi, "Analysis of the IEEE 802.11e EDCA Under Statistical Tra ffic," in Proc. IEEE ICC 2006. [73] Z. Tao and S. Panwar, "An Analytical Model for the IEEE 802.11e Enhanced Distributed Coordination Function," in Proc. IEEE ICC 2004. [74] Y.C. Tay and K.C. Chua, A capacity an alysis for the IEEE 802.11 MAC protocol, Wireless Networks, Vol. 7, 1997. [75] J. C. Tay and K. C. Chua, "A Capacity Analysis for the IEEE 802.11 MAC Protocol," Wireless Netw., pp. 159-171, July 2001. [76] O. Tickoo and B. Sikdar, "Queueing An alysis and Delay Mitigation in IEEE 802.11 Random Access MAC based Wireless Netw orks," in Proc. IEEE Infocom 2004.

PAGE 177

177 [77] M. Tzannes, T. Cooklev, and D. Lee, "Extended Data Rate 802.11a," IEEE 802.11-02232r0, Mar. 2002. [78] V.Vishnevsky and A. Lyakhov "802.11 LANs: Sa turation Throughput in the Presence of Noise," in Proc. IFIP Networking 2002. [79] M. A. Visser and M. E. Zarki, "Voice a nd data transmission over an 802.11 wireless network," in Proc. IEEE PIMRC 95. [80] VisaulWare, http://myspeed.visualware.com/voip/ [81] I.N. Vukovic and N Smavatkul, "Delay analysis of different backoff algorithms in IEEE 802.11," in Proc. IEEE VTC-Fall 2004. [82] S.C. Wang, Y.M. Chen, T.H. Lee, A. Helm y Performance Evaluations for Hybrid IEEE 802.11b and 802.11g Networks, in Proc. IEEE IPCCC 05. [83] S.C. Wang and A. Helmy, Performance Limits and Analysis of Contention-based IEEE 802.11 MAC, IEEE LCN [84] S. Wong, H. Yang, S. Lu, V. Bharghavan, R obust Rate Adap-tation in 802.11 Wireless Networks, ACM MOBICOM [85] H. Wu;Y Peng; K Long; S Cheng; J Ma, P erformance of Reliable Transport Protocol over IEEE 802.11 Wireless LAN: Analys is and Enhancement, INFOCOM 2002. [86] Y. Xiao and J. Rosdahl, Throughput and delay limits of IEEE 802.11, IEEE Communications Letters, vol. 6, Aug. 2002, pp. 355 357. [87] Y. Xiao and J. Rosdahl, Performance an alysis and enhancement for the current and future IEEE 802.11 MAC protocols, ACM Mobile Computing and Communications Review, vol. 7, April 2003, pp. 6-19. [88] Y. Xiao, "Performance analysis of IEEE 802.11e EDCF under saturation condition," in Proc. IEEE ICC 2004 [89] D.Y. Yang, T.J. Lee, K. Jang, J.B. Chang, S Choi, "Performance Enhancement of MultiRate IEEE 802.11 WLANs with Geographically-Sca ttered Stations," IEEE Transactions on Mobile Computing, 2006. [90] J. Y. Yeh and C. Chen, "Support of multimedia services with the IEEE 802-11 MAC protocol," in Proc. IEEE ICC 02. [91] J.H. Yun and S.W. Seo, Novel collision det ection scheme and its applications for IEEE 802.11 wireless LANs, Computer Communications Volume 30, Issue 6, March

PAGE 178

178 [92] H. Zhai, X. Chen, and Y. Fang, "How We ll Can the IEEE 802.11 Wireless LAN Support Quality of Service?," IEEE Transactions on Wireless Communications, vol. 4, no.6, pp. 3084-3094, Nov. 2005. [93] Y. Zheng, K. Lu, D. Wu, and Y. Fang, "Performance Analysis of IEEE 802.11 DCF in Imperfect Channels," IEEE Transactions on Vehicular Technology, vol. 55, no. 5, pp. 1648--1656, Sept. 2006. [94] E. Ziouva and T. Antonakopoulos, CSMA/CA performance under high traffic conditions: throughput and delay analysis, Computer Communications, Vol. 25, pp.313321, 2002. [95] J. Zyren, E. Enders, and T. Edmonds on, .11g Starts Answering WLAN Range Questions, CMP Media LLC (w ww.CommsDesign.com), Jan 2003.

PAGE 179

BIOGRAPHICAL SKETCH Shao-Cheng W ang received his B.S. and M.S. degrees in electrical engineering from National Taiwan University, Taip ei, Taiwan, in 1997 and 1999 respectively. He enrolled in the Ph.D. program at University of Southern Calif ornia and graduated with Engineer degree in electrical engineering in 2006, by satisfying all Ph.D. course requirements and the Engineer degree qualifying exam. He then continued his Ph .D. study in the Departme nt of Computer and Information Science and Engineering, University of Florida. His rese arch interests include wireless network protocol development, protocol performance an alysis, medium access control, and quality of service. He is a student member of IEEE.