Performance Improvement of IEEE802.11Wireless LAN Networks

IEEE802.11無線 LANネットワークの性能向上に関する研究

Shojiro TAKEUCHI

Major in Computer Science,Graduate School of Science and Engineering,

at

Waseda University

Directed by Prof. Yasuhiko YASUDA

Feb. 2006

Acknowledgment

I may never have undertaken without the enthusiastic encouragement from the person whobecame my advisor, Prof. Yasuhiko Yasuda. He often steered me in the right direction when Iwasn’t certain of which path to investigate. I acknowledge his significant contribution to my PhDstudies.

I would like to thank Prof. Fumio Takahata, Prof. Shigeki Goto and Prof. Jiro Katto for theirvery significant support and guidance. Their valuable comments on this thesis based on variousoutlooks could lead me to the right direction.

I also would like to express my gratitude to Associate Prof. Kaoru Sezaki in the Universityof Tokyo. He deserves special mention, not only for providing the valuable comments on myresearch but also for providing the opportunity for me to work on this thesis in the Sezaki Lab. inthe University of Tokyo.

I am also grateful to Dr. Yoshikatsu Nakagawa who is head of NRC Tokyo (Nokia ResearchCenter Japan) and Mr. Marko Teittinen who is a Competence Area manager. They gave me anopportunity to work in NRC Tokyo as a research engineer and it has been very stimulating and re-warding to work there. And also, Mr. Toshiaki Jozawa and Mika Kasslin were often acknowledgedfor their significant supports and understandings on my Ph.D. studies. My working experience to-gether with them was able to enhance this Ph.D. thesis. Other colleagues in NRC Tokyo and NRCHelsinki deserves special mention for giving me the opportunity to have valuable discussions withthem.

I would like to thank all of other members in Yasuda Lab. in Waseda University and in SezakiLab. in the University of Tokyo.

Finally, I would like to appreciate my parents encouraging my interest in computer and networkengineering and many sacrifices they have made for my studies.

1

Contents

Acknowledgment 1

Abbreviations 9

1 Introduction 111.1 Research Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121.2 Research Challenges in IEEE802.11 WLAN . . . . . . . . . . . . . . . . . . . . . 121.3 Thesis Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131.4 The Structure of This Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2 IEEE802.11 Wireless LAN 172.1 Overview of This Chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182.2 Overview of IEEE802.11 Standard . . . . . . . . . . . . . . . . . . . . . . . . . . 182.3 IEEE802.11 Standard Family . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192.4 MAC Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

2.4.1 Channel Access Mechanism in IEEE802.11 MAC . . . . . . . . . . . . . 212.4.2 Management functions provided by IEEE802.11 MAC . . . . . . . . . . . 26

2.5 Physical Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272.5.1 DSSS (Direct Sequence Spread Spectrum) . . . . . . . . . . . . . . . . . . 272.5.2 OFDM (Orthogonal Frequency Division Multiplexing) . . . . . . . . . . . 28

2.6 Wireless Networks consisting of IEEE802.11 WLAN . . . . . . . . . . . . . . . . 292.7 Conclusion of this Chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3 Performance Improvement of Ad Hoc Networks 323.1 Overview of This Chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333.2 Power-saving in Ad Hoc Networks . . . . . . . . . . . . . . . . . . . . . . . . . . 333.3 Improved Power Saving Mechanism for MAC protocol in Ad Hoc Networks . . . . 34

3.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343.3.2 IEEE802.11 PSM in DCF . . . . . . . . . . . . . . . . . . . . . . . . . . 353.3.3 Prior Arts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363.3.4 Improved Power Saving Mechanism (IPSM) . . . . . . . . . . . . . . . . 393.3.5 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . 453.3.6 Conclusion of IPSM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

2

3.4 Battery Cost Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523.4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523.4.2 Prior Arts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523.4.3 Battery Cost Routing in Consideration of Transmission Power . . . . . . . 553.4.4 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 603.4.5 Conclusion of Battery Cost Routing in Consideration of Transmission Power 64

3.5 Conclusion of this Chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

4 Performance Improvement of IEEE802.11e WLAN Networks 674.1 Overview of This Chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 684.2 IEEE802.11e WLAN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

4.2.1 Channel Access Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . 694.2.2 Admission Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 694.2.3 Continuous TXOP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

4.3 Protection of Real-time Traffic under IEEE802.11e EDCA via Dynamic Adapta-tion of EDCA Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 724.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 724.3.2 Related Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 734.3.3 Dynamic Adaptation of CW Sizes . . . . . . . . . . . . . . . . . . . . . . 744.3.4 Appropriate TXOP allocation . . . . . . . . . . . . . . . . . . . . . . . . 804.3.5 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . 824.3.6 Conclusion of Real-time Traffic under IEEE802.11e EDCA via EDCA Pa-

rameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 884.4 Quick Data-retrieving for U-APSD in IEEE802.11e WLAN Networks . . . . . . . 92

4.4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 924.4.2 IEEE802.11e U-APSD . . . . . . . . . . . . . . . . . . . . . . . . . . . . 924.4.3 Periodic U-SP for quick data retrieving . . . . . . . . . . . . . . . . . . . 954.4.4 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . 1004.4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

4.5 Conclusion of this Chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108

5 Performance Improvement of Large IEEE802.11e WLAN Networks Consisting ofMultiple APs 1095.1 Overview of This Chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1105.2 Access Point Selection Strategy in IEEE802.11e WLAN networks toward Load

Balancing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1105.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1105.2.2 Prior Arts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1115.2.3 HRFA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1125.2.4 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . 1155.2.5 Conclusion of Access Point Selection Strategy in IEEE802.11e WLAN

networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1205.3 Conclusion of This Chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120

3

6 Summary 1246.1 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1256.2 Future Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

Publication 128

4

List of Figures

2.1 Open Systems Interconnection (OSI) reference model . . . . . . . . . . . . . . . . 182.2 IEEE802.11 Standard Family . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202.3 MAC Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212.4 DCF mechanism (DATA/ACK exchange) . . . . . . . . . . . . . . . . . . . . . . 222.5 Exponential Backoff . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232.6 Basic Access . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232.7 Hidden Node Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242.8 DCF mechanism (RTS/CTS/DATA/ACK exchange) . . . . . . . . . . . . . . . . . 252.9 Frame Transfers in PCF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252.10 IEEE802.11 Standard Procedures to Connect an AP . . . . . . . . . . . . . . . . . 262.11 Transmitter configurations in DSSS (Direct Sequence Spread Spectrum) . . . . . . 282.12 Receiver configurations in DSSS (Direct Sequence Spread Spectrum) . . . . . . . 282.13 Transmitter configurations in OFDM (Orthogonal Frequency Division Multiplexing) 292.14 Receiver configurations in OFDM (Orthogonal Frequency Division Multiplexing) . 292.15 BSS (Basic Service Set) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302.16 ESS (Extended Service Set) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302.17 IBSS (Independent BSS) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3.1 Power Saving Mechanism for DCF . . . . . . . . . . . . . . . . . . . . . . . . . . 353.2 Long end-to-end delay in multi-hop wireless networks with PSM . . . . . . . . . . 363.3 NPSM for DCF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383.4 Transition to awake state . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413.5 Transition to awake or doze state . . . . . . . . . . . . . . . . . . . . . . . . . . . 423.6 Flowchart of awake period transition . . . . . . . . . . . . . . . . . . . . . . . . . 423.7 Flowchart of transmission mechanism in IPSM . . . . . . . . . . . . . . . . . . . 433.8 Standard frame control field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 433.9 Frame control field in IPSM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 433.10 Announcement of awake period . . . . . . . . . . . . . . . . . . . . . . . . . . . 443.11 Packet delivery ratio, 5 long-lived CBR connections . . . . . . . . . . . . . . . . . 473.12 Latency, 5 long-lived CBR connections . . . . . . . . . . . . . . . . . . . . . . . 483.13 Energy goodput, 5 long-lived CBR connections . . . . . . . . . . . . . . . . . . . 483.14 Comparison of packet delivery ratio, 5 on-off CBR connections . . . . . . . . . . . 493.15 Comparison of end-to-end delay, 5 on-off CBR connections . . . . . . . . . . . . . 503.16 Comparison of energy goodput, 5 on-off CBR connections . . . . . . . . . . . . . 50

5

3.17 Relay region with path loss model of (a)p ∼ 1/d2 (b) p ∼ 1/d4 . . . . . . . . . . . . . 533.18 Overusing problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 543.19 Our proposed protocol operations . . . . . . . . . . . . . . . . . . . . . . . . . . 573.20 Calculation of battery cost . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 573.21 Comparison of power efficiency in different power-aware routing protocols . . . . 583.22 The way of deciding the communication range . . . . . . . . . . . . . . . . . . . . 603.23 Comparison of packet delivery ratio for different waiting time in MBCR . . . . . . 623.24 Comparison of network lifetime for different waiting time in MBCR . . . . . . . . 623.25 Comparison of packet delivery ratio for differentαs in our proposal . . . . . . . . 633.26 Comparison of network lifetime for differentαs in our proposal . . . . . . . . . . 643.27 Comparison of packet delivery ratio in different power-aware routing protocols . . 653.28 Comparison of network lifetime in different power-aware routing protocols . . . . 65

4.1 Implementation model for EDCA . . . . . . . . . . . . . . . . . . . . . . . . . . 704.2 Example of channel access timing in EDCA . . . . . . . . . . . . . . . . . . . . . 704.3 Continuous TXOP in EDCA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 724.4 Adaptation policy in case of increasing CW sizes . . . . . . . . . . . . . . . . . . 774.5 Adaptation policy in case of decreasing CW sizes . . . . . . . . . . . . . . . . . . 774.6 First Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 794.7 Second Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 804.8 Transmission Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 814.9 Algorithm for TXOP Limit selection . . . . . . . . . . . . . . . . . . . . . . . . . 834.10 Comparison of throughput for voice flows in Scenario1 . . . . . . . . . . . . . . . 844.11 Comparison of delay for voice flows in Scenairo1 . . . . . . . . . . . . . . . . . . 854.12 Comparison of throughput for voice flows in Scenario2 . . . . . . . . . . . . . . . 854.13 Comparison of throughput for video flows in Scenario2 . . . . . . . . . . . . . . . 864.14 Comparison of delay for voice flows in Scenario2 . . . . . . . . . . . . . . . . . . 864.15 Comparison of delay for video flows in Scenario2 . . . . . . . . . . . . . . . . . . 874.16 Comparison of throughput for voice flows in Scenario3 . . . . . . . . . . . . . . . 884.17 Comparison of throughput for video flows in Scenario3 . . . . . . . . . . . . . . . 894.18 Comparison of throughput for data flows in Scenario3 . . . . . . . . . . . . . . . . 894.19 Comparison of delay for voice flows in Scenario3 . . . . . . . . . . . . . . . . . . 904.20 Comparison of delay for video flows in Scenario3 . . . . . . . . . . . . . . . . . . 904.21 Comparison of delay for data flows in Scenario3 . . . . . . . . . . . . . . . . . . . 914.22 U-APSD operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 944.23 STA operations on the receipt of EOSP=1 and MoreData=1 . . . . . . . . . . . . . 944.24 Large buffering delay in the use of ON-OFF traffic . . . . . . . . . . . . . . . . . 954.25 UPTT operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 974.26 Timer cancellation in UPTT when sending an uplink data frame . . . . . . . . . . 984.27 UTWR operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1014.28 Timer cancellation in UTWR when sending an uplink data frame . . . . . . . . . . 1014.29 Dropping probability of voice traffic . . . . . . . . . . . . . . . . . . . . . . . . . 1034.30 Average delay of downlink voice traffic . . . . . . . . . . . . . . . . . . . . . . . 104

6

4.31 Comparison of energy goodput . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1044.32 Dropping probability of voice traffic in Scenario2 . . . . . . . . . . . . . . . . . . 1064.33 Average delay of downlink voice traffic in Scenario2 . . . . . . . . . . . . . . . . 1064.34 Comparison of energy goodput in Scenario2 . . . . . . . . . . . . . . . . . . . . . 1074.35 Average throughput of Data traffic in Scenario2 . . . . . . . . . . . . . . . . . . . 107

5.1 Congested AP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1115.2 QBSS Load element . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1135.3 Simulation Area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1165.4 Average throughput of voice traffic in Case1 . . . . . . . . . . . . . . . . . . . . . 1175.5 Average throughput of video traffic in Case1 . . . . . . . . . . . . . . . . . . . . . 1175.6 Average throughput of data traffic in Case1 . . . . . . . . . . . . . . . . . . . . . 1185.7 Average delay of voice traffic in Case1 . . . . . . . . . . . . . . . . . . . . . . . . 1185.8 Average delay of video traffic in Case1 . . . . . . . . . . . . . . . . . . . . . . . . 1195.9 Average throughput of voice traffic in Case2 . . . . . . . . . . . . . . . . . . . . . 1205.10 Average throughput of video traffic in Case2 . . . . . . . . . . . . . . . . . . . . . 1215.11 Average throughput of data traffic in Case2 . . . . . . . . . . . . . . . . . . . . . 1215.12 Average delay of voice traffic in Case2 . . . . . . . . . . . . . . . . . . . . . . . . 1225.13 Average delay of video traffic in Case2 . . . . . . . . . . . . . . . . . . . . . . . . 122

7

List of Tables

3.1 Awake Period field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443.2 Energy Consumption Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453.3 Simulation Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

4.1 Default Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

8

Abbreviations

WLAN Wireless LANWM Wireless MediumQoS Quality of ServiceMAC Medium Access ControlPHY PhysicalPSM Power Saving MechanismAM Active ModeIPSM Improved PSMU-APSD Unscheduled Automatic Power Save DeliveryUPTT U-APSD with Periodic Transmission of Trigger frameUTWR U-APSD with Timer-based Wake-up RuleRRM Radio Resource ManagementOSI Open System InterconnectionMSDU MAC Service Data UnitLLC Logical Link ControlSTA StationAP Access PointOFDM Orthogonal frequency Division MultiplexingDSSS Direct Sequence Spread SpectrumTKIP Temporal Key Integrity ProtocolAES Advanced Encryption StandardEAP Extensible Authentication ProtocolMIMO Multiple-Input Multiple-OutputDCF Distributed Coordination FunctionPCF Point Coordination FunctionCSMA/CA Carrier-sense Multiple Access with Collision AvoidanceNAV Network Allocation VectorSIFS Short Inter-Frame SpaceRTS Ready to SendCTS Clear to SendPIFS PCF IFSDIFS DCF IFSEIFS Extended IFS

9

CFP Contention Free PeriodWEP Wired Equivalent PrivacyFHSS Frequency Hopping Spread SpectrumIR InfraredCCK Complementary Code KeyingDBPSK Differential Binary Phase Shift KeyingDQPSK Differential Quadrature PSKRF Radio FrequencyQAM Quadrature Amplitude ModulationIFFT Inverse Fast Fourier TransformFFT Fast Fourier TransformBSS Basic Service SetIBSS Independent BSSESS Extended Service SetATIM Announcement Traffic Indication MessageID IdentificationCBR Constant Bit RateUDP User Datagram ProtocolMBCR Minimum Battery Cost RoutingMTPR Minimum total Transmission Power RoutingRREQ Route REQuestRREP Route REPlyVoIP Voice over IPHCF Hybrid Coordination FunctionEDCA Enhanced Distributed Channel AccessHCCA HCF Controlled Channel AccessU-APSD Unscheduled Automatic Power Save DeliveryTXOP Transmission OpportunityUP User PriorityAC Access CategoryAIFS Arbitration Inter-Frame SpaceCW Contention WindowADDTS Add Traffic StreamTSPEC Traffic SpecificationU-SP Unscheduled Service PeriodEOSP End of Service PeriodRSSI Received Signal Strength IndicatorBE Best EffortHRFA High Rate First AssociationLA Link AdaptationHO Hand-Over

10

Chapter 1

Introduction

11

1.1 Research Background

WLAN (Wireless LAN) technology is rapidly a critical component of computer networks andis growing even now. This growth comes from the fact that IEEE802.11 standardized its WLANspecification. Thus IEEE802.11 WLAN has become very popular as an open solution for providingmobility as well as essential network services [1]. Furthermore, because IEEE802.11 definedseveral amendments to enhance WLAN performances with respect to data rate, QoS (Quality ofService) and etc., it can now provide a firm basis for high performance WLANs.

Compared to wired LANs, WLAN can provide the following benefits: mobility, ease of LANinstallation. Mobility allows users to move physically while they use a handheld PC or otherdevices to access , for example, a data server, a printer, intranet or internet. Therefore, even if theychange a workplace to another, they can access a data server or a network if they can receive strongradio power enough to correctly decode received frames.

Installation of a wired LAN is very time-consuming task and if the LAN changes its physi-cal configuration, network engineers have to spend much time on re-setting up the LAN again,installing new wires. On the other hand, the use of WLAN enables users to get rid of time toinstall cables and thus can reduce installation time. Furthermore, even in case where users want toconfigure a LAN in two places, which are physically separated by some obstacles, they can easilyform the LAN with the use of WLAN equipments.

Moreover, most of WLAN products are now IEEE802.11-compatible and thus WLAN users canhave wireless network applications based on open systems provided by IEEE802.11 standard. Be-sides, most WLAN suppliers follow IEEE802.11 standard, and therefore the price of WLAN de-vices can be lowered and they can be made by multivendors. In addition to advantage in cost(devices themselves and installation cost), IEEE802.11 WLAN can have an advantage in band-width. Thus WLAN devices are being embedded into cellular phones as one of complementarywireless accesses because cellular phones provide narrower bandwidth than WLAN.

The fact mentioned above results in a situation where IEEE802.11 WLAN is currently used notonly in home but also in office or university campus and is becoming more and more popular. Thisthesis is motivated by them and is concerned with promising IEEE802.11 WLAN technologies.

1.2 Research Challenges in IEEE802.11 WLAN

While in IEEE802.11 WLAN there are attractive aspects presented in the previous section, thereare some research challenges which have to be overcome as shown in the followings.

In order for users to utilize the capability of mobility provided by IEEE802.11 WLAN, WLANdevices have become small and been embedded into mobile handheld devices. Considering thefact that most of handheld devices are battery powered, power conservation is a crucial issue forIEEE802.11 WLAN. In fact, several cellular phone products now include a WLAN device usedfor a wireless access method which complements lower data rate in cellular networks, and when

12

WLAN devices are embedded into cellular phones, large amount of power consumption is consid-ered as one of the weakest points for WLAN. Therefore,power consumption in a IEEE802.11WLAN device has to be minimized to prolong battery life and operating time of WLAN ter-minals.

IEEE802.11 standard provides a specification about MAC (Medium Access Control) and PHY(Physical) layers. Channel access scheme in MAC layer is based on CSMA/CA (Carrier SenseMultiple Access with Collision Avoidance), and thus packets transmitted by different STAs con-flict with each other. Even though a retransmission method is defined in IEEE802.11, retransmittedpackets are suffered from delay and in the worst case they are dropped. In IEEE802.11 [1], trafficcollides with other traffic, regardless of traffic types (real-time and non real-time traffic). There-fore, although real-time traffic in general has delay sensitivity, it may be caused a large delayin IEEE802.11. It is in fact difficult to manipulate real-time traffic in IEEE802.11 WLAN eventhough there are large demands on the use of real-time application over WLAN.

When real-time traffic contends for the channel access with non real-time traffic it has to begiven opportunity to access the wireless medium earlier than non real-time traffic. This is dueto the fact that real-time traffic in general requires specific features such as delay sensitivity andbandwidth requirement. In this sense, A task group (TGe) was organized and defined a new MACprotocol to provide differentiated channel access [2]. However, although the differential channelaccess gives real-time traffic many opportunities to access the wireless channel, it does not guar-antee the distinct QoS and the performance of real-time traffic must be degraded due to collisionswith other traffic.Therefore, it is desirable to provide a method to enhance the performanceof real-time traffic in IEEE802.11 WLAN networks.

Moreover, as mentioned in the previous, since IEEE802.11 defines only MAC and PHY layerspecifications, it does not intend to have detailed specifications for higher layers above MAC layer.However, IEEE802.11 WLAN specification allows large WLAN networks consisting of multi-ple APs to be formed. In fact, large WLAN networks recently have been installed in enterpriseand public spaces such as an airport and a hotel. In these environments, users who are accessingWLAN APs can move from a place to another place. But, there will be a situation where specificAPs have much traffic even though traffic is sparse in others. In this case, users in crowded areamust be suffered from degradation of WLAN performance such as low throughput and large la-tency.Hence, in the environment of large WLAN networks, radio resource management andload balancing strategy are needed. Furthermore, STAs joining the networks have to takecare about these things to efficiently share the limited wireless bandwidth among STAs inIEEE802.11.

1.3 Thesis Contributions

This section presents contributions in this thesis. They consist of several contributions for 2 typesof WLAN networks such as ad hoc networks and infrastructure networks. Each of the contributionstries to overcome one of research challenges, which are presented in previous section, either in adhoc networks or infrastructure networks, and is shortly and briefly presented in the followings.

13

The first contribution (Chapter 3) is about performance improvement of ad hoc networks withregard to power conservation and consists of two researches.

Due to the fact that most of terminals are battery-powered in ad hoc networks, power con-sumption is one of the most critical issues. Energy consumed by WLAN devices is classifiedinto two parts: power consumption in actual and non actual communications. Energy is con-sumed by receiving and transmitting packets in actual communication and by listening to thewireless medium in non actual communication. In fact, IEEE802.11 WLAN supports Power Sav-ing Mechanism (PSM) to reduce power consumption in non communication part and ad hoc net-works with IEEE802.11 PSM can largely decrease power consumption. However they are sufferedfrom throughput degradation. This is due to the fact that a sender and receiver operating withIEEE802.11 PSM exchange information about whether the sender buffer packets destined for thereceiver before they start data exchange. As a result, the sender needs to buffer data packets be-fore it transmits them to the receiver. Therefore, we propose an Improved PSM (IPSM) to avoidthroughput degradation. Compared to IEEE802.11 PSM, nodes operating with IPSM exchangethe length of period in which they continue to be awake, so called awake period. Therefore asender can immediately relay data packets to a receiver without buffering them if the sender knowsthe receiver is awake. Simulation results show IPSM can overcome throughput degradation inIEEE802.11 PSM.

Routing protocols in ad hoc networks influence power consumption by actual communicationbecause packets are conveyed by wireless multi-hop communications from a source to a destinationnode. First they have to minimize power consumption needed for packet delivery from a sourceto a destination. Second, energy has to be fairly utilized among nodes in an ad hoc network sincethey have to relay packets as a router. In this research, we propose a novel battery cost routing toachieve these two requirements. Consequently it can achieve to prolong lifetime of nodes and alsoof ad hoc networks.

The second contribution (Chapter 4) is about performance improvement of infrastructure networksand consists of two researches.

The first one focuses on performance improvement of real-time traffic in IEEE802.11e WLAN[2]. IEEE802.11e provides a MAC protocol to support real-time applications in WLAN. Howeverit provides only differential channel access to give real-time traffic many opportunities to accessthe wireless channel and does not completely ensure performance of real-time traffic. In fact,real-time traffic is suffered from throughput degradation because of contentions between real-timeflows and between real-time and non real-time flows. This phenomenon stems from the fact thatchannel access parameters defined in [2] do not adapt to WLAN network conditions which aredynamically changed by the number of flows, types of flows and etc. The first research in thesecond contribution therefore proposes a method to control channel access parameters consideringthe number of real-time flows. Thus, simulation results show that dynamically changing themcan reduce collisions between real-time flows and between real-time and non real-time flows andtherefore an AP can accommodate more real-time traffic.

The second one tries to reduce delay of downlink traffic which uses U-APSD (UnscheduledAutomatic Power Save Delivery) [2]. U-APSD is defined in IEEE802.11e and an efficient PSM forreal-time traffic. Even though it efficiently works when it is used with real-time traffic periodically

14

generated both in uplink and downlink, downlink real-time traffic is, however, suffered from largedelay when it is used for real-time traffic with silence period such as ON-OFF voice traffic. Thisis because a STA operating with U-APSD enters doze mode and cannot receive downlink trafficduring time period when uplink traffic is not generated. Therefore, in order to periodically receivedownlink traffic, we propose two methods called UPTT (U-APSD with Periodic Transmission ofTrigger frame) and UTWR (U-APSD with Timer-based Wake-up Rule), respectively. These twoproposals enable downlink traffic to be periodically transmitted to STAs. Simulation results showthat using them enable a STA to periodically enter into awake mode and to receive downlink trafficand thus delay of downlink traffic can be reduced.

The last contribution (Chapter 5) is about performance improvement of large WLAN networksconsisting of multiple APs.

IEEE802.11 allows multiple APs to construct a large WLAN network such as cellular net-works. Thus large WLAN networks are being installed in enterprise or public spaces such as airport, hotel and etc. However, in the environment of large WLAN networks, there is a situationwhere many STAs is associating with a particular AP and other APs have little STAs. In this case,throughput in the AP having many STAs will be degraded due to traffic congestion. Therefore,RRM (Radio Resource Management) has to be taken into account and load balancing among APshas to be performed in large WLAN networks. Considering the fact that in IEEE802.11 WLANSTAs have the right to select an AP with which they communicate, an AP selection strategy im-plemented in STAs becomes an important role for load balancing in IEEE802.11 WLAN. Hencewe propose an AP selection strategy toward load balancing. Since prior arts do not take into ac-count a STA’s transmission rate, it may select an AP which is located far from it and choose lowtransmission speed. Thus they may inefficiently use the radio resource. Compared to them, ourproposal can enable WLAN networks to efficiently utilize the wireless resource because it selectsan AP considering both a STA’s transmission rate and channel load in an AP. Simulation resultsshow that WLAN networks implemented in our proposal can enhance utilization of the limitedwireless resource and have higher throughput.

1.4 The Structure of This Thesis

This thesis consists of an introductory chapter (Chapter 2) which presents an overview of IEEE802.11WLAN technologies, three research chapters (Chapter 3∼5), and a concluding chapter (Chapter6).

Chapter 3 consists of two researches focusing on power saving in ad hoc networks. The firstone is about performance improvement of IEEE802.11 Power Saving Mechanism (PSM). The nextone focuses on energy efficient routing protocols in ad hoc networks

Chapter 4 includes two researches on performance improvement of WLAN infrastructure net-works configured by IEEE802.11e. The first one is on performance improvement of real-timetraffic in IEEE802.11e WLAN networks. The second one focuses on reducing delay of downlinktraffic operating with U-APSD defined in IEEE802.11e.

Chapter 5 works on load balancing in large WLAN networks consisting of multiple APs and

15

proposes an AP selection mechanism.In Chapter 6, we conclude this thesis and present future works.

16

Chapter 2

IEEE802.11 Wireless LAN

17

IEEE802.2Logical Link Control (LLC)

IEEE802.3Carrier Sense

IEEE802.4Token Bus

IEEE802.5Token Ring

IEEE802.11Wireless

MAC

PHY

OSILayer2

OSILayer1

Figure 2.1: Open Systems Interconnection (OSI) reference model

2.1 Overview of This Chapter

This chapter introduces IEEE802.11 Standard which is the most well known WLAN standardspecification [1]. The following sections briefly provide an overview of IEEE802.11 Standard,IEEE802.11 Standard Family, IEEE802.11 MAC (Medium Access Control) Layer, IEEE802.11PHY (Physical) Layer and possible two types of IEEE802.11 WLAN network architectures fordetail on the IEEE802.11 standard.

2.2 Overview of IEEE802.11 Standard

IEEE802.11 WLAN standard is included as one of IEEE802 LAN standards which standardizedLocal and Metropolitan Area Network Standards. As depicted in Fig.2.1, IEEE802 includes afamily of standards and falls within the scope of layer 1 and 2 of the OSI (Open Systems Inter-connection) reference model. The IEEE802.11 standard, which is officially called IEEE Standardfor Wireless LAN MAC (Medium Access) and PHY (Physical Layer) Specifications, defines over-the-air protocols necessary to support networking in a local area. As with other IEEE802-basestandards (such as 802.3 and 802.5), the primary service of the 802.11 standard is to deliver MS-DUs (MAC Service Data Units) between peer LLCs (Logical Link Controls). Typically, a radiocard and access point support functionalities of the IEEE802.11 standard.

The IEEE802.11 standard provides MAC and PHY functionalities for wireless connectivity offixed, portable, and moving stations (STAs) moving at pedestrian and vehicular speeds within alocal area. IEEE802.11 standard includes the following specific features:

• Support of asynchronous and time-bounded delivery service.

• Continuity of service within extended areas via a distribution system, such as ethernet.

• Accommodation of transmission rates of 1Mbps and 2Mbps (802.11b, 802.11a and 802.11gextensions offer higher data rates than the base standard).

• Support of most market applications.

18

• Multicast (including broadcast) services.

• Network management services.

• Authentication services.

Furthermore, IEEE802.11 targets the following environments:

• Inside buildings, such as offices, banks, malls, hospitals, manufacturing plants.

• Outside areas, such as campuses, building complexes, and outdoor plants.

Compared to the wired standard, IEEE802.11 WLAN standard considers wireless features andprovides the following additional functionalities:

Power Management Since most WLAN Network Interface Cards (NICs) are available in PCM-CIA TypeII format, they can be used as portable devices. However, they rely on batteriesto power the electronics within them. Furthermore, when they are used in mobile handhelddevices, power consumption is one of the most critical problems to overcome. IEEE802.11provides power management functions to conserve battery power. They enable IEEE802.11WLAN devices to switch to lower-power mode periodically when not transmitting or receiv-ing data. In fact, MAC layer defines these functions by putting the radio to be off when notransmission and reception activity occur for some specific or user-definable time period.

Security Wireless signals transmitted by a WLAN device spread in a large area and thus otherusers in the area can intercept them. From privacy point of view, however, they must beprotected. IEEE802.11 provides security mechanisms to encrypt frames. Besides, it providesauthentication services which are used when a STA joins a WLAN network.

2.3 IEEE802.11 Standard Family

The IEEE802.11 WLAN standard, called Legacy standard, was approved and the correspondingstandard specification was published in 1997. After the finalization of Legacy standard, severalstandard amendments were released and even currently other amendments have been discussed inworking groups in IEEE802.11. Figure 2.2 shows standards family consisting of Legacy standardand enhancements, which are shortly explained in the followings:

802.11a The 802.11a [3] defines operation at up to 54Mbps using OFDM (Orthogonal frequencyDivision Multiplexing) modulation in 5GHz frequency band.

802.11b The 802.11b [4] is a data rate extension of Legacy standard DSSS (Direct SequenceSpread Spectrum), providing operation at up to 11.0Mbps in 2.4GHz frequency band.

802.11g The 802.11g [5] is an amendment of Legacy standard, 802.11a and 802.11b, and definesoperation at up to 54Mbps using OFDM in 2.4GHz frequency band.

19

2.4GHzDS-SS

2.4GHzFH-SS IR

802.11a5GHz,54Mbps

802.11b2.4GHz,11Mbps

802.11g2.4GHz,54Mbps

802.11iSecurity

enhancementOriginal

IEEE802.11Standard

PHY Enhancement

MAC Enhancement

MAC802.11e

QoSenhancement

802.11k ( TGk)RRM

enhancement

802.11n (TGn)

High Throughputenhancement

Figure 2.2: IEEE802.11 Standard Family

802.11e The 802.11e [2] defines a MAC protocol to support real-time applications, which havedelay sensitivity or bandwidth requirement, over WLAN. The standard specification waspublished in Nov. 2005.

802.11i The 802.11i [6] defines management functions and procedures to provide confidentialityof user information over the wireless medium and authentication of IEEE802.11 compatibledevices. It supports 802.1x security functions such as TKIP (Temporal Key Integrity Proto-col), AES (Advanced Encryption Standard), and EAP (Extensible Authentication Protocol).The standard specification was published in 2004.

802.11k The 802.11k [7], which is currently being standardized, provides functions related to ra-dio resource management and defines radio resource measurement functions such as channelload, Noise, hidden terminals and procedures to report radio performance such as histogramreports and statistical reports. If this function is implemented in WLAN networks, a STAcan adapt to dynamically changing wireless conditions.

802.11n The 802.11n [8], which is under discussion and has not issued its draft document at thismoment, defines functionalities in PHY and MAC layers to support transmission speed atleast over 100Mbps. It will probably include 802.11e in MAC layer. In PHY layer, Multiple-Input Multiple-Output (MIMO) will be adopted to increase transmission speed.

2.4 MAC Layer

This section introduces IEEE802.11 MAC layer [1] which consists of two functionalities: ChannelAccess Mechanism and Management parts. Although, for both of them, IEEE802.11 provides

20

DistributedCoordination Function

(DCF)

PointCoordination

Function(PCF)

MACExtent

Figure 2.3: MAC Architecture

amendment functions such as IEEE802.11e, IEEE802.11k and etc., this section focuses on MAClayer in Legacy standard [1]. The successive subsections describe those two functions.

2.4.1 Channel Access Mechanism in IEEE802.11 MAC

IEEE802.11 MAC provides two channel access controls, DCF (Distributed Coordination Function)and PCF (Point Coordination Function), as shown in Fig.2.3 [1]. DCF works based on CSMA/CA(Carrier-sense Multiple Access with Collision Avoidance). PCF provides contention-free channelaccess but is supported only in infrastructure networks.

DCF

The DCF works based on CSMA/CA (Carrier-sense Multiple Access with Collision Avoidance)and is the primary access protocol. In Fig2.4 [1], a sender’s, receiver’s and others behaviors in DCFare presented when the sender transmits a DATA frame to the receiver. The sender can send a DATAframe to the receiver when it determines that the medium is idle for greater than or equal to DIFSperiod. On the receipt of the DATA frame, the receiver replies with an ACK (Acknowledgement)frame after SIFS period expires. If other STAs receiving the DATA frame, they set NAV (NetworkAllocation Vector) value using duration field in the header in the DATA frame and cannot accessthe medium during that time. Using NAV to prohibit others from accessing the medium is calledVirtual Carrier Sense.

If a STA has transmitted a DATA frame and does not receive the corresponding ACK frame,it can retransmit the DATA frame until retransmission limit is reached. Before retransmission, itperforms backoff which is randomly selected as depicted in Eq 2.1.

Backoff T ime = Random(0, CW ) (2.1)

21

Source

Receiver

Other

Data

ACK

DIFS

DIFS

SIFS

Contention Window

Defer Acess Backoff After Defer

NAV

Figure 2.4: DCF mechanism (DATA/ACK exchange)

,where CW shows the current contention window. As shown in Fig.2.5 [1], contention window sizeis increased every time when transmissions fail. This is because the STA wants to avoid collisionswith frames transmitted by others and if it uses increased contention window size the collisionprobability will decrease.

In addition to the use of contention window, the DCF provides different inter-frame spacingsuch as DIFS, SIFS, PIFS. The SIFS (Short Inter-frame Space) is the shortest of inter-frame spaces,providing the highest priority level by allowing some frames to access the medium before others.Following the SIFS, ACK, CTS (Clear to Send) frames are transmitted. The PIFS (PCF IFS) is theinterval that STAs working under PCF use to gain access to the medium. The PIFS gives priority toframes transmitted by the PCF. The DIFS (DCF IFS) is the interval used in STAs operating underDCF and used for transmission of DATA and management frames. As presented in Fig.2.6 [1],the interval of SIFS is the smallest among these three IFSs. Besides them, EIFS (Extended IFS) isdefined as the longest IFS. STAs use it when a frame is received but cannot be correctly decodeddue to bad channel condition or other reasons.

As shown in Fig.2.4 [1], basic channel access scheme in DCF is accomplished by exchangingDATA and ACK frames. However, it is not strong enough to successfully transmit DATA frames.This is due to hidden node problem as shown in Fig.2.7. Hidden node problem takes place in asituation where if STA1 communicating with an AP transmits a DATA frame the DATA frame con-flicts with frames transmitted by STA2 which is outside of STA1’s communication range. In casewhere the length of DATA frame transmitted by a STA is too long, unsuccessful time due to colli-sions is also very long and thus the STA will have long time until it notices the transmission fails.To reduce unsuccessful time due to collisions with transmissions by hidden nodes, IEEE802.11prepares control frames which are called RTS (Ready To Send) and CTS (Clear To Send) and ex-changed between a sender and a receiver. Length of these control frames is generally shorter thanDATA frame and so even if exchange of RTS and CTS frames fails, unsuccessful time required

22

CW max

CW min 715

31

63

127

255

Initial attempt

First retransmission

Figure 2.5: Exponential Backoff

DIFS

Busy Medium

SIFS

PIFS

DIFS

Backoff

Slot Time

Contention window

Figure 2.6: Basic Access

23

AP

STA1

STA2

Communication Range of STA1

Communication Range of STA2

Collisions Happen!!

Figure 2.7: Hidden Node Problem

for exchanging these frame is shorter, compared to exchanging DATA and ACK frames. Figure2.8 shows an example of DATA frame transmission using RTS and CTS frames. However, in casewhere network traffic load is low and collisions with other transmissions rarely occur, exchangingof RTS and CTS frames becomes overhead.

PCF

Frame transfers between STAs and an AP under PCF are controlled by Point Coordinator (PC) im-plemented in the AP and available during a particular period called CFP (Contention Free Period).Figure 2.9 shows frame transfers between STAs and an AP under PCF. After a beacon transmis-sion, the AP transmits a CF-Poll or DATA+CF-Poll frame if it wants to start CFP. On receivingthe polling frame, STAs polled by the AP transmit Data+CF-ACK frame for data transmission andacknowledgement of the CF-Poll frame. When it further wants to poll STAs, it transmits anotherpolling frame the period of SIFS after the data exchange. Since in PCF the AP can control frameexchanges sending polling frames prior to other transmissions of frames , STAs receiving them settheir NAV values and cannot independently access the wireless medium. To end a CFP, the APsends a CF-end frame and STAs clear their NAV values. The schedule of polling is not defined in[1] and is dependent on vendors implementation.

In PCF, collisions among frames are not expected during CFP. Therefore, when an AP is in-stalled in a place where there is no adjacent AP, the PCF can work well. However, if adjacentAPs interfere frame transfers in an AP, PCF cannot work well because it assumes that there is nocollision and that it can control all frame transfers.

24

Source

Destination

Other

Data

ACK

DIFS

DIFS

SIFS

Contention Window

Defer Acess Backoff After Defer

NAV (RTS)

RTS

SIFS CTS SIFS

NAV (CTS)

Figure 2.8: DCF mechanism (RTS/CTS/DATA/ACK exchange)

Beacon

PIFS

SIFS

D1+poll

SIFS

U1+ack

SIFS

D2+poll

SIFS

U2+ack

SIFS

D3+poll

U3+ack

SIFS

SIFS

CF-End

ContentionPeriod

Contention-free Period

NAV

ResetNAV

Figure 2.9: Frame Transfers in PCF

25

STA AP1

Scanning Process(Beacon Reception)

Authentication Request to AP1

Authentication Response from AP1

Association Request to AP1

Association Response from AP1

Standard Processes defined in IEEE802.11

AP2Get some information from surrounding APs

Figure 2.10: IEEE802.11 Standard Procedures to Connect an AP

2.4.2 Management functions provided by IEEE802.11 MAC

Management functions in IEEE802.11 MAC enable a STA and AP to exchange some informationwith each other. Basic management functions are used for setting up a reliable connection betweenan AP and STA or between STAs. In fact, since new amendment standards such as IEEE802.11e[2] and IEEE802.11i [6] were approved and added into IEEE802.11 standard family, a lot of man-agement functions have been accordingly included. However, only basic management functionsare presented here. These management functions consist of Scan for searching surrounding APsor STAs, Authentication for secure communication and Association for confirmation of joining anetwork. Therefore, when an STA wants to join a network managed by an AP, it has to proceedwith these actions as shown in Fig.2.10. Each of them are presented in the following.

First, the STA does Scan to know the AP’s operating channel and AP’s capability (e.g., whatkind of options are supported in the AP) and Scan is divided into Active Scan and Passive Scan. Inthe Active Scan, it sends a Probe request frame which includes SSID (Service Set Identifier) andinformation of data rates supported in the STA. SSID is an unique ID allocated to the AP. Aftersending the Probe request, the STA waits for Probe response transmitted by an surrounding APduring a certain period. On the receipt of Probe response frame, it can obtain the correspondingAP’s capability information. It continues to send a Probe request frame in different channels until itends Active Scan. On the other hand, an STA proceeding with Passive Scan changes its operatingchannel and just waits for beacon frames transmitted from surrounding APs. Beacon frame isone of management frames and periodically transmitted from APs. The interval of consecutivebeacon transmissions is called beacon interval which is in general 100ms or 200ms. On receivingbeacon frames, the STA stores information contained in beacon frames and changes its operationalchannels until it ends Passive Scan. Both in case of Active Scan and Passive Scan, it decides anAP with which it will associate after Scanning process.

26

As shown in Fig.2.10, the STA next has to do Authentication process sending an authentica-tion frame destined for the target AP. Authentication process is categorized into Open system andShared key authentication. Open system authentication involves a two-step authentication and isknown as null-authentication algorithm. This is because even though the AP can refuse a requestissued from the STA, data frames are not encrypted after authentication process. Shared key au-thentication uses WEP (Wired Equivalent Privacy) and consists of a 4-step authentication. In thisauthentication, 4 frames are exchanged between the STA and AP, and the third and fourth framesare encrypted by WEP key. Besides, data frames are also encrypted when the STA associates andcommunicates with the AP.

If authentication succeeds, the STA proceeds with Association process. In Association process,in order for the STA to inform the target AP of STA’s capability, the STA first transmits a associa-tion request frame to the AP and waits for association response transmitted by the AP. Therefore,the association request includes STA’s capability information such as supported transmission rate,power saving mechanism capability and etc. If association result shows the association processsucceeds, it can start data transmission or reception to or from the AP.

As explained above, MAC layer provides management functions for STAs or APs to exchangetheir capabilities with their peers. If they dynamically change their capabilities, they informs thepeers of the change with the use of management frames. Moreover, when STAs take handover andjoin an AP, they have to perform the processes presented in Fig.2.10 since they have to inform theircapabilities of the AP.

2.5 Physical Layer

Legacy IEEE802.11 standard defines DSSS (Direct Sequence Spread Spectrum), FHSS (FrequencyHopping Spread Spectrum), IR (Infrared) as PHY layer protocols and further in [1] OFDM (Or-thogonal Frequency Division Multiplexing) is defined. Besides, as shown in Fig.2.2 in section 2.3,IEEE802.11a [3], IEEE802.11b [4] and IEEE802.11g [5] defined detailed PHY layer specifica-tions according to their operating frequency bands. Since DSSS and OFDM are mostly supportedin the current products, this section presents only DSSS and OFDM as PHY layer protocols.

2.5.1 DSSS (Direct Sequence Spread Spectrum)

DSSS is involved in Legacy IEEE802.11 as a PHY layer protocol. In Legacy standard, DSSSprovides 1 and 2Mbps and in IEEE802.11b [4] transmission speed is increased to 5.5 and 11Mbpsusing CCK (Complementary Code Keying) which is a extension of DSSS.

Figure 2.11 shows basic transmitter configurations in DSSS. First, Scrambling is performedfor input data, using multiplication of pseudo random values. This process is done to avoid thesituation where the same bits are continued and power is gathered in a certain frequency. Thenext process is modulation of input data coming to modulator. DBPSK (Differential Binary PhaseShift Keying) is used for 1Mbps and DQPSK (Differential Quadrature PSK) is used for 2Mbps.After modulation, multiplication of spreading code and modulated data is performed in modulationblock. As spreading code, Barker Code is used and its length is 11bits. Finally, D/A (Digital to

27

Scrambling Modulation SpreadingTx Data D/ARF

Processing

Figure 2.11: Transmitter configurations in DSSS (Direct Sequence Spread Spectrum)

De-Scrambling

De-Modulation

De-Spreading

DecodedData A/D

RFProcessing

Figure 2.12: Receiver configurations in DSSS (Direct Sequence Spread Spectrum)

Analog) conversion is performed and RF (Radio Frequency) processing follows.Figure 2.12 presents basic receiver configurations in DSSS. Receiver processings are the re-

verse of transmitter ones. That is, de-spreading is first performed using spreading code which isthe same one used in the transmitter, de-modulation follows and finally de-scrambling is processed.

Using spreading process, bandwidth of transmitted signal is spread. Using de-spreading pro-cess desired signal is obtained and is less affected by interference signal.

2.5.2 OFDM (Orthogonal Frequency Division Multiplexing)

OFDM is added in IEEE802.11 standard [1] and further re-defined in IEEE802.11a [3] and IEEE802.11g[5] in 5GHz and 2.4GHz bands, respectively.

OFDM is known as a wireless communication system which is robust in the environment ofmulti-path and keeps high transmission speed. This is because transmission data is separately con-veyed using multiple low rate data sequences mapped to multiple subcarriers. Moreover, sinceeach subcarrier overlaps with neighbor subcarriers, bandwidth is efficiently utilized. Each sub-carrier can be detached from neighbor ones because subcarriers are orthogonally overlapped withneighbors.

Figure 2.13 shows basic transmitter configurations in OFDM. First, input data is coded usingconvolutional codes for the purpose of error recovery. Next, Interleaver block performs interleav-ing and thus bit order in coded data is permuted because convolutional coding is weak to consec-utive bit errors. Third, data mapped to each subcarrier is modulated using either BPSK, QPSK,16QAM (Quadrature Amplitude Modulation) or 64 QAM. Then, using IFFT (Inverse Fast Fourier

28

Convolutional Coding

Interleave SubcarrierModulation IFFT Guard

Interval D/ATx Data RFProcessing

Figure 2.13: Transmitter configurations in OFDM (Orthogonal Frequency Division Multiplexing)

Viterbi De-Coding

De-Interleave

SubcarrierDe-Modulation

FFTRemoveGuardInterval

A/DDecoded

DataRF

Processing

Figure 2.14: Receiver configurations in OFDM (Orthogonal Frequency Division Multiplexing)

Transform) multicarrier signals are generated. After guard interval is inserted into each subcarrier,D/A conversion is performed. Guard interval is used to increase robustness against performancedegradation due to multipath and is a copy of the last part of symbol.

Figure 2.14 presents basic receiver configurations in OFDM. Receiver processings are the re-verse ones performed in transmitter. After A/D conversion, first guard interval is removed and thenext processing is FFT to detach each subcarrier. Thirdly, de-modulation is performed in everysubcarrier. After de-interleaving process and viterbi decoding are performed, decoded data areobtained.

2.6 Wireless Networks consisting of IEEE802.11 WLAN

IEEE802.11 defines two types of WLAN networks, called BSS (Basic Service Set) and IBSS(Independent BSS), respectively. BSS consists of AP and STA as shown in Fig.2.15 and is knownas Infrastructure Networks. STAs in an BSS communicate with an AP connecting to a backbonenetwork and can get access to the backbone network and further internet. The AP is responsiblefor management of the BSS and configures the BSS’s settings. Moreover, IEEE802.11 providesdefines ESS (Extended Service Set) consisting of several BSSs as shown in Fig.2.16, and thereforecan configure large WLAN networks. Each STA can select an BSS with which it communicatesand if it moves from an BSS to another BSS, it performs handover. When an ESS is established,WLAN users can seamlessly access a backbone network performing handover process even if theymove their places where they work.

29

AP (Access Point)

STA(Station)

STA(Station)

BSS(Basic Service Set)

Ethernet (Backbone Network)

Wireless Cell

Figure 2.15: BSS (Basic Service Set)

Ethernet (Backbone Network)

AP (Access Point)

STA(Station)

STA(Station)

BSS(Basic Service Set)

AP(Access Point)

STA(Station)

STA(Station)

BSS(Basic Service Set)

ESS(Extended Service Set)

Wireless Cell

Wireless Cell

Figure 2.16: ESS (Extended Service Set)

As presented in Fig.2.17, an IBSS consists of only STAs and is known as Ad Hoc Networks.STAs participating in the IBSS communicate with each other and manage the IBSS in distributedmanner. Furthermore, even though IEEE802.11 does not specify higher layer protocols, if wirelessterminals with IEEE802.11 WLAN have functionalities of IP router they can communicate with anSTA, which stays far from them, using wireless multihop communications. Routing functionalitiesimplemented in an ad hoc network decide a role to convey packets from a source to a destinationnode. Since IEEE802.11 provides functionalities of MAC and PHY layers, wireless terminals caneasily construct an ad hoc network if IEEE802.11 WLAN is installed in them. Ad hoc networks areconfigured independently from backbone networks, thus they can be useful in case where backbonenetworks are down due to disaster and other problems.

30

IBSS(Independent BSS)

Wireless Multi-hop Communication

STA(Station)

Figure 2.17: IBSS (Independent BSS)

2.7 Conclusion of this Chapter

This chapter briefly explained an overview of IEEE802.11 standard and functionalities defined inIEEE802.11.

IEEE802.11 defines MAC and PHY specifications. In MAC layer, channel access mechanismand management functions are provided and in PHY layer several modulation schemes are pro-vided. Each of them were presented in this chapter.

Moreover, this chapter explained WLAN networks which are defined in IEEE802.11. Theyare divided into BSS (known as infrastructure networks) and IBSS (known as ad hoc networks).IEEE802.11 allows BSS to be large WLAN networks called ESS and IBSS can be wireless multi-hop networks by means of routing functionalities.

The following chapters present our researches which work both on infrastructure and ad hocnetworks.

31

Chapter 3

Performance Improvement of Ad HocNetworks

32

3.1 Overview of This Chapter

Recently portable and small wireless devices has created enormous opportunities for the field ofwireless computing, and it is now possible to deploy many devices with low computation, com-munication and battery power. Wireless ad hoc networks can provide data communication amongthese devices irrespective of their physical locations. For this reason, ad hoc networks have be-come a hot research topic. They are autonomous systems consisting of mobile nodes that do notrely on the presence of any fixed network infrastructures. In order to configure ad hoc networksand communicate with other nodes, a node has to have wireless communication port (supportingPHY, MAC and network layer protocols). Since IEEE802.11 supports ad hoc mode and providesMAC and PHY layer protocols, if a node uses a IEEE802.11 WLAN interface, it can easily config-ure an ad hoc network. Furthermore, because IEEE802.11 WLAN is a widely accepted standard,compatibility between protocols supported by nodes joining an ad hoc network can be maintained.Therefore, IEEE802.11 WLAN has received attention from many researches and testbeds in adhoc networks as PHY and MAC protocols. This chapter focuses on ad hoc networks configured byIEEE802.11 WLAN ad hoc mode. In ad hoc networks each node communicates with other nodesin wireless multi-hop manner, and thus many constraints are imposed on them by the environment.Furthermore, since ad hoc networks typically consist of energy constraint nodes and their power isprovided by batteries, power conservation is one of the most crucial issues.

Hence this chapter is concerned with power saving issues in ad hoc networks. The followingsections includes two researches aim to power saving in ad hoc networks. The first one works forPSM (Power Saving Mechanism) supported in IEEE802.11[1] and we propose IPSM (ImprovedPSM) to overcome throughput degradation happens in ad hoc networks with IEEE802.11 PSM.The second one works on power efficient routing protocols in ad hoc networks and proposes batterycost routing in consideration of transmission power. Before going into them in detail, we firstdescribe in the next section what issues have to be carefully considered in terms of power savingin ad hoc networks.

3.2 Power-saving in Ad Hoc Networks

Since in ad hoc networks nodes exchanges data packets with others, energy is consumed by com-munication. Each node transmitting or receiving packets consumes energy. Furthermore severalexperimental results, [10],[11],[13], show that energy in ad hoc networks is not always consumedby actual communication. This means wireless network interfaces in the idle state waste a signif-icant amount of energy. Energy dissipation in the idle state cannot be ignored because networkinterfaces often stay in the idle state for a long time. Thus to conserve this energy, it is gener-ally desirable to turn the radio off when they are not in use. IEEE802.11 PSM has a capability ofswitching off the radio for power saving. In later section, first we show our research on IEEE802.11PSM and propose IPSM.

Moreover, based on the fact that each node plays a role of router in ad hoc networks, datapackets are conveyed from a source to a destination node in wireless multi-hop manner. In thissense, routing is a key to maximize battery life of each node and of ad hoc networks. In fact, if a

33

routing protocol implemented in an ad hoc network selects a route which consumes much energy,the ad hoc network will quickly exhaust its energy and its lifetime will be short. In this sense,a route from source to destination node has to be chosen to prolong the lifetime of each nodeand ad hoc networks. As the second part of the following sections, we show our study on powerefficient routing protocols in ad hoc networks and propose battery cost routing in consideration oftransmission power.

3.3 Improved Power Saving Mechanism for MAC protocol inAd Hoc Networks

3.3.1 Introduction

Since most nodes are battery operated in ad hoc networks, power conservation is one of the mostimportant issues. To conserve energy of nodes staying in the idle state, they have to turn theirradio off when they do not use their radio. This observation has led to several energy conservationprotocols, [13], [14] being proposed to reduce energy consumption in dense ad hoc networks byturning off devices that are not necessary for global connectivity. However, in these protocols,geographical or topological information decides which set of nodes have their radios turned on,thus those nodes still consume their energy supply even when there is no actual traffic load on thenetworks. As a result, nodes waste their energy to maintain global connectivity [15].

Some MAC protocols have also been proposed to conserve energy [15],[16], which were basedon IEEE 802.11 PSM. The general idea of PSM is for all nodes with it to keep their radios offwhen they do not have to send or receive packets, and they turn on their radios at the same time,maintaining the awake state for a specified period. For that period, a sender announces bufferedpackets to a receiver via an ATIM (Announcement Traffic Indication Message) frame. A node thatreceives such an announcement frame recognizes that the sender wants to transmit packets to it,and stays awake until the packet is delivered. Of course the nodes must be synchronized to awakenat the same time. Both MAC protocols in IEEE802.11, DCF and PCF, support PSM. Nodes inPSM must be synchronized by periodic beacon transmissions. In this research we focus on PSMin DCF.

MAC protocols have the ability to sense medium and decide when packets can be transmittedor received. Therefore they can recognize when to transmit or receive packets. They are thussuitable for playing a role in turning off the radio when it does not have to be used. From theaforementioned reason, our proposal uses MAC layer information to switch off wireless networkinterfaces. As with [15] and [16], it is also based on IEEE 802.11 PSM because IEEE 802.11provides ad-hoc mode and has received attention from many researches and testbeds in ad hocnetworks.

The remainder of this research is organized as follows. The next section 3.3.2 reviews PSM inDCF. Section 3.3.3 shows related works. Section 3.3.4 presents our proposed protocol. In section3.3.5, we describe our simulation model and discuss simulation results. Finally, section 3.3.6concludes this research.

34

beacon interval

ATIM window

beacon

ATIM

ATIM ACK

DATA

ACK B

C Doze State

beacon interval

(receiver node)

(idle node)

(sender node)A

doze state

awake state

Figure 3.1: Power Saving Mechanism for DCF

3.3.2 IEEE802.11 PSM in DCF

In this section, we present PSM in DCF in detail in Fig.3.1. Figure 3.1 shows that time is dividedinto beacon intervals, and at the beginning of each beacon interval, there exists a specific interval,called ATIM window. During an ATIM window, every node is awake, and for nodes to wake upat the same time, they need to be synchronized by beacon transmissions. Because of the absenceof a centralized timer in ad hoc networks, each node is responsible for generating a beacon. Afterthe beacon interval, all nodes compete for transmission of the beacon using a standard backoffalgorithm. The first station wins the competition and all others have to cancel their beacon trans-missions and adjust their local timers to the time stamp of the winning beacon. Packets for a nodein the doze state have to be buffered by the sender until the end of the beacon interval. When anode, like node A in Fig.3.1, has a pending packet to transmit, it transmits an ATIM frame to a re-ceiver node, like node B in Fig.3.1, during an ATIM window. When the receiver node receives theATIM frame, it replies with an ATIM-ACK. After the ATIM and ATIM-ACK handshakes, both thesender and receiver stay awake for the remaining beacon interval to perform the data transmission.A node that has not performed the ATIM and ATIM-ACK handshakes during the ATIM windowfalls back into the doze state after the ATIM window.

The performance of PSM is affected by the size of the ATIM window. It was shown in [17]that PSM performed well when the length of the ATIM window was approximately 25% of thebeacon interval. Furthermore, during an ATIM window, only ATIM and ATIM-ACK frames canbe transmitted. Overheads in energy consumption occur when transmitting or receiving additionalATIM and ATIM-ACK frames. There is a time overhead in time due to an ATIM window, sincedata can be transmitted after the ATIM window. From these perspectives, PSM using DATA win-dow was proposed in [16], and was callednew PSM(NPSM). NPSM exhibited better performancewith respect to aggregate throughput and energy conservation. The next section explains NPSM in

35

. . . . .source node

intermediate node 1

destination node

buffer

. . . . .destination node

buffer

ATIM frame exchanges

ATIM frame exchanges

(a) Source node has to inform intermediate node 1 of pending packets by an ATIM frame in every beacon interval before data transmissions.

(b) Intermediate node 1 also has to inform intermediate node 2 of pending packets by an ATIM frame in every beacon interval before data transmissions. This means that it is difficult for intermediate node 1 to send data packets to intermediate node 2 immediately after the intermediate node 1 receives them.

intermediate node 2

intermediate node 3

intermediate node 1

intermediate node 2

intermediate node 3

source node

Figure 3.2: Long end-to-end delay in multi-hop wireless networks with PSM

detail.

PSM still suffers some problems with end-to-end delay and throughput, since in ad hoc net-works source nodes use multi-hop wireless communications to deliver packets to destination nodes.For example, Fig.3.2 (a) shows that the source node informs the intermediate node 1 of bufferedpackets by an ATIM frame during an ATIM window when it has pending packets. After exchang-ing ATIM and ATIM-ACK frames, the intermediate node 1 receives data packets from the sourcenode. It cannot, however, transmit them to the intermediate node 2 immediately after receivingthem from the source node, because it did not exchange ATIM frames with the intermediate node2. At the next beacon interval, the intermediate node 1 transmits an ATIM frame to the interme-diate node 2, as indicated in Fig.3.2 (b). After the successful transmission of ATIM frames, theintermediate node 1 can transmit data packets to the intermediate node 2. ATIM frame exchangeshave a role in announcing buffered packets to receiver nodes and data packets can be transmit-ted only after successful ATIM frame transmissions. However, since sender nodes have to bufferpackets to transmit the ATIM frame to receiver nodes, announcing buffered packets generates longend-to-end delay in multi-hop wireless networks.

Furthermore, if a network has large traffic load, then a sender in PSM cannot inform a receiverof its pending packets by ATIM frames for a specific period like an ATIM window, so throughputdeclines. Therefore, ad hoc networks in PSM have long end-to-end delay and degrade throughput.

3.3.3 Prior Arts

We show related works in this section.

36

On-demand Power Management

This subsection presents a MAC protocol that uses on-demand power management [15]. It is across-layer protocol designed by network layer information and cooperates with on-demand rout-ing protocols. Nodes usually operate in PSM. If nodes receive routing control packets or datapackets, they switch from PSM to the active mode (AM). In AM, a node is in the awake stateand transmits or receives data at any time. Transitions from AM to PSM are determined by asoft-state timer, and timer values depend on the type of packets received. When the timer expires,a node reverts from AM to PSM. The soft-state timer is refreshed by communication events thattrigger a transition to the active mode. As for the trade-off between delay, throughput and energyconsumption, communication events are useful for balancing those.

When nodes in PSM have packets to send, they have to transmit the ATIM frame to receivernodes during an ATIM window. However, if neighbor nodes of a sender node that has to send ATIMframe to a receiver node are in AM, data packets transmitted by them may disturb the ATIM frameexchanges. Furthermore, this protocol can cooperate only with on-demand routing protocols, andif new on-demand routing protocols are developed, the PSM proposed in [15] will not be able toadapt to them. Therefore, we believe that it is desirable for PSM to be developed separately fromnetwork layer.

New PSM (NPSM)

This subsection describes the NPSM proposed in [16]. NPSM does not use ATIM and ATIM-ACKframes, since they waste bandwidth and consume energy. Figure 3.3 illustrates NPSM in DCF. InFig.3.3, nodes A and B are a sender and a receiver, respectively. Node C does not send any packets,but can overhear packets traveling from node A to node B.

Time is divided into beacon intervals in NPSM as well as in PSM. At the beginning of eachbeacon interval, every node in NPSM enters the awake state for a specific duration, called a DATAwindow; all nodes in NPSM stay awake during the DATA window. They do not send ATIM andATIM-ACK frames, but can transmit data packets during a DATA window. The ATIM window inPSM plays a role in announcing pending packets to receiver nodes, though NPSM has a differentway to achieve the same function. In NPSM, each nodeX has the following counters to indicatethe number of packets to transmit or receive.:

• T (i): the number of packets pending at nodeX for nodei.

• R(i): the number of packets destined for nodeX. NodeX knows they are pending at nodei.

• Rtotal(X): sum ofR(i) over all neighbors of nodeX.

• Up(i): the number of packets that the neighbor nodei needs to transmit and receive.

DATA, RTS, CTS, and ACK packets have some of these counters. When nodei transmits a DATApacket to nodej, it includesT (j) andRtotal(i). When nodej receives the DATA packet from nodei, it updatesR(i). An RTS from nodei to nodej includesT (j) + Rtotal(i). Furthermore, a CTS

37

beacon interval

DATA window

beacon

ACK

A

B

CDoze State

DATA

5ms

(sender node)

(receiver node)

(idle node)

DATA window

DATA

ACK

Figure 3.3: NPSM for DCF

and ACK from nodej to i includesT (i) + Rtotal(j). When each node receives or overhears anRTS CTS, DATA or ACK packet from nodei to nodej, it updatesUp(i). Here,Up(i) shows datatransmissions that the nodei will receive or send while staying awake. Therefore if nodek hasUp(i) greater than zero, it can recognize nodei stays awake.Up(i) is reset to zero at the beginningof each beacon interval.

In NPSM, every node stays awake during a DATA window. When a DATA window expires,nodes extend the DATA window or go into the doze state. When the following conditions aresatisfied, they extend the DATA window. IfRtotal(k) at any nodesk is greater than zero, theyextend the DATA window to receive packets. As shown in Fig.3.3, sender node A has packetsto transmit to receiver node B on the expiration of the DATA window. In this condition, node Ainfers the state of node B by means ofUp(B). If Up(B) is not maintained by node A orUp(B) iszero, node A cannot transmit packets to node B. It will transmit them in the next beacon intervaland go to the doze state except whenRtotal(A) is greater than zero. In [16], DATA window size isincreased in increments of 5ms. When the extended DATA window expires, the same process isrepeated as when the initial DATA window expires.

Since NPSM removes the ATIM window, it conserves energy for ATIM and ATIM handshakes;bandwidth for data transmission is used more efficiently. However, a sender node in NPSM in-forms a receiver node of its pending packets as well as in PSM , so that end-to-end delay is stilllong through multi-hop connections. The reason for this phenomenon was shown in the previoussection. Furthermore, when network traffic load is high, a sender node cannot transmit packets to areceiver node during a DATA window due to traffic congestion. Since the sender infers the state ofthe receiveri byUp(i) that is reset to zero at the beginning of each beacon interval, it cannot knowthe receiver’s state if it does not successfully transmit packets during each initial DATA window.Consequently, throughput deteriorates and long delays occur in NPSM. This condition similarlyhappens in PSM. Therefore, we need to develop a novel PSM that can adapt to a high network

38

traffic load and reduce delay. In the next section, we present our proposed IPSM. It improvesthroughput, delay, and energy efficiency. It can work only with information from the MAC layer.

3.3.4 Improved Power Saving Mechanism (IPSM)

We now present our proposed power saving mechanism, referred to as IPSM. The goal of IPSMis to achieve performance almost the same as in normal IEEE 802.11 with respect to end-to-enddelay and throughput, and furthermore, not to degrade the performance of energy conservation.IPSM can operate, using only MAC layer information.

Overview of IPSM

Before we express our proposed algorithm of IPSM, we summarize conditions required for eachperformance measure, energy consumption, throughput, and end-to-end delay.

• energy consumption: To reduce energy consumption, nodes must remain in the doze statefor as long as possible.

• throughput: To adapt to a high network traffic load, nodes in routes have to be awake duringdata transmissions.

• end-to-end delay: For PSM to achieve the same performance as normal IEEE 802.11, nodesin routes have to stay in the awake state. It generates long delays for a sender to inform areceiver of its pending packets in PSM and NPSM.

Considering the above requirements, we find that the relation among energy consumption, through-put and delay is a trade-off, and that the performance of each factor depends on traffic patterns.Therefore we need to balance the trade-off and consider traffic to decide when a node goes into thedoze state. However, since it is very difficult to predict packet arrivals and all network conditions,we cannot completely optimize the awake period.

The key idea of our proposed IPSM is that nodes in routes stay awake and others continueto doze. Furthermore, regarding energy efficiency and bandwidth utilization, IPSM uses a DATAwindow instead of an ATIM window. However, nodes in IPSM do not maintain counters of pendingpackets, since announcing buffered packets results in long delays. Instead, each node in IPSMpossesses a neighbor table that holds neighbor node IDs and the awake period of neighbor nodes.Consequently, when a node receives packets, it can immediately relay them if it knows from itsneighbor table that the next hop is awake. To balance the trade-off between throughput, delay andenergy consumption, the awake period can be varied in IPSM.

In the following subsections, we show how to decide the awake period and how to announcethe awake period of each node to neighbor nodes.

Awake Period

Figures 3.4 and 3.5 illustrate how nodes in IPSM change their state. Figure 3.4 (a) shows thatif node A receives a data packet, it stays awake for the rest of the current and the next beacon

39

interval to wait for packet arrivals. If node B receives a data packet, as shown in Fig.3.4 (b), it alsocontinues to stay awake for the rest of the current and the next beacon interval. Moreover, when anode is awake and receives a data packet after a DATA window, it stays awake during the rest ofthe current and the next beacon interval.

Figure 3.5 presents how node A reduces its awake period. In Fig.3.5, node A does not receiveany data packets during a beacon interval, thus it reduces its awake period in the following beaconinterval to T

2, which includes a DATA window. Here,T denotes beacon interval andTd represents

a DATA window. If node A does not receive any data packets duringT2, its awake period,T

2, is

further reduced by half, i.e.,T4. As far as energy consumption is concerned, a shorter awake period

results in more efficient energy conservation. However, receiver nodes need to remain in the awakestate and wait for packet arrivals for a while, since the interval of packet arrivals fluctuates and thetrade-off previously mentioned should be balanced. Therefore, if data packets do not arrive, theawake period is set to half of the previous awake period. If an awake period is under a DATAwindow, it will be set to DATA window,Td. Figure 3.6 shows the flowchart to explain how eachnode changes its awake period.

Announcing Awake Period

Since every node is awake during a DATA window, sender nodes can transmit packets to receivernodes during that period. However, to transmit packets after a DATA window, a sender node has torecognize whether a receiver node is awake. IPSM enable nodes to inform neighbor nodes of theirawake period in the following way: Each frame, RTS, CTS, DATA, and ACK, contains the lengthof each originator node’s awake period in frame control field in MAC header. The minimum lengthof the awake period is a DATA window,Td. Figure 3.8 shows the standard frame control field andFig.3.9 presents a structure of frame control field in IPSM. Comparing Fig.3.8 to Fig.3.9, IPSMneeds additional 8 bits to include 3 bits of sender’s awake period, 1 bit of next interval flag, and4 bits of reserved field. That is, MAC header in IPSM is 1 octet larger than the standard header.The reserved field in frame control field in IPSM is not used for any purposes at the moment. Thevalue of Awake Period ranges from 0 to 7. The field of Awake Period is interpreted as shown inTable 1. If the field of next interval flag is set to 1, it shows the sender node remains awake in thewhole of the next beacon interval. When it is set to 0, it presents the sender reduces the awakeperiod by half in the next beacon interval. Thus if we implement IPSM into IEEE 802.11 standardprotocol, we need to change header format and additionally have 8 bits to inform neighbor nodesof senders’ awake period. Even though we have additional 8 bits in frame control field in IPSM, itcan improve the performance of delay and throughput. Since nodes with ISPM does not announcebuffered packets but awake period, they can smoothly relay packets.

Each node has a neighbor table that contains neighbor node ID and the length of its awakeperiod. If a node receives or overhears each frame, it will update the length of the awake periodin a corresponding entry, as shown in Fig.3.10. Each entry in a neighbor table is updated at thebeginning of each beacon interval; that is, the length of the awake period in each entry is reducedby half. However, if a packet with next interval flag set to 1 in frame control field is receivedfrom a node, the awake period in the corresponding node’s entry is not reduced and it remains thewhole of beacon interval at the beginning of the next beacon interval. When a sender node wants

40

(a) Transition to awake state at station A: When station A receives a data packet, it continues to be awake during the rest of and the next beacon interval

beacon

A

B

awake

DATA ACKbeacon

A

B

awake

awake

(b) Transition to awake satate at station B: When station B receives a data packet, it continues to awake during the rest of and the next beacon interval

beacon intervalDATA window

T =Td =T

Td

TTd

awake

awake

awake

DATA ACK

DATA ACK

Figure 3.4: Transition to awake state

to transmit packets and does not hold the entry of a receiver node, it decides that the length of theawake period of the receiver node is the DATA window. In this case, it recognizes the receiveris in the doze state after the DATA window and will transmit the packet in the following beaconinterval. Figure 3.7 shows the flowchart to explain how each node makes a decision of whether itcan transmit a packet to a receiver.

Sender nodes in NPSM have to announce buffered packets to receiver nodes. When sendernodes in NPSM have no buffered packets, they enter the doze state except for when their counterfor receiving packets is greater than zero. In other words, they have to buffer packets to stayawake. Wireless multi-hop networks with NPSM have long delays due to buffering at intermediatenodes. Nodes in NPSM increase their awake period in increments of 5 ms when they have packetsto transmit or receive, and in particular, sender nodes decide to extend their DATA window bypending packets and theUp(i) of receiver nodei. Up(i) is reset to zero at the beginning of eachbeacon interval. When traffic congestion occurs and sender nodes cannot receive any packets fromreceiver nodes, they will not know the counter value of receiver nodei, thus they will not be ableto transmit packets after the DATA window.

On the other hand, since nodes in IPSM inform their neighbors of only their awake period,sender nodes do not have to buffer packets to remain in the awake state. In IPSM, intermediatenodes can relay packets immediately after receiving them. Consequently, networks in IPSM haveshorter delay than other PSMs. Furthermore, since each node stays awake during the rest of thecurrent and the next beacon interval when it receives only one packet, nodes in IPSM do not haveto try to send or receive packets to stay awake in every beacon interval. As a result, IPSM canadapt to high traffic load and show higher throughput than other PSMs.

41

beacon intervalDATA window

A

B

awake

DATA ACK

awake

awake

awake

T =Td =

TD T T

T/2

Transition to awake or doze state at each station: When station B receives a data frame and station A does not receive a data frame, Station B stays awake state during the whole of next beacon interval and station A awakes during T/2.

beacon

Figure 3.5: Transition to awake or doze state

Receive a DATApacket duringAwake Period

Wake-up or continue to awake at the beginning of DATA window

Keep awake in the remaining and the next beacon interval

Set the next awake period if Ta / 2 <= Td Ta = Tdelse Ta = Ta/2

No Yes

T=beacon intervalTa = awake periodmaximum value of Ta is TTd= DATA window

Figure 3.6: Flowchart of awake period transition

42

Buffer has a packet to transmit

DATA window

I awake ? &A receiver awakes?

Send the packet

Send the packet tothe receiver inthe next beacon interval

Send the packetto the receiver

Yes No

No Yes

Figure 3.7: Flowchart of transmission mechanism in IPSM

ProtocolVersion

Type Subtype ToDS

FromDS

MoreFrag

Retry

PwrMgt

MoreData

WEP

Order

2 bits 2 bits 4 bits 1 bit1 bit 1 bit1 bit1 bit 1 bit1 bit 1 bit

Figure 3.8: Standard frame control field

ProtocolVersion

Type Subtype ToDS

FromDS

MoreFrag

Retry

PwrMgt

MoreData

WEP

Order

2 bits 2 bits 4 bits 1 bit1 bit 1 bit1 bit1 bit 1 bit1 bit 1 bit

AwakePeriod

3 bits

nextintervalflag

1 bit

Reserved

4 bits

Figure 3.9: Frame control field in IPSM

43

Sender Receiver

RTS

CTS

DATA

ACK

neighbor of sender

neighbor of Receiver

overhearing RTS

overhearing DATA

overhearing CTS

overhearing ACK

insert or updatesender node’s ID andawake period in the neighbor table

insert or updatereceiver node’s ID andawake period in the neighbor table

insert or updatereceiver node’s ID andawake period in the neighbor table

insert or updatesender node’s ID andawake period in the neighbor table

insert or updatesender node’s ID andawake period in the neighbor table

insert or updatereceiver node’s id andawake period in the neighbor table

insert or updatereceiver node’s id andawake period in the neighbor table

insert or updatesender node’s ID andawake period in the neighbor table

Figure 3.10: Announcement of awake period

Awake Period ValueDATA window 0

T/64 1T/32 2T/16 3T/8 4T/4 5T/2 6

Beacon Interval(T ) 7

Table 3.1: Awake Period field

44

Transmit Receive Idle Sleep1400mW 1000mW 830mW 130mW

Table 3.2: Energy Consumption Model

3.3.5 Performance Evaluation

To better understand the performance of IPSM, we evaluated it via simulations. We implementedIPSM using NS-2 [18], and performed several simulations using different patterns of traffic, long-lived CBR traffic and on-off traffic. We consider four different types of MAC protocols: normalIEEE802.11, PSM, NPSM, and IPSM we propose. For each simulation, every node is equippedwith IEEE 802.11 based WaveLAN wireless radios with a bandwidth of 2 Mbps and nominal trans-mission range of 250 m. DSR is used as the routing protocol. The simulation area is 1500 m×500 m and contains 30 stationary nodes. The energy model is shown in Table 2 as the same modelas in [8]. The length of each data packet is 512 bytes. The beacon interval is 100 ms, and eachATIM and DATA window is set to 25 ms [17]. For NPSM, the incremental awake time is 5 msas was used in [16]. In our simulation, we prepared 10 network topology files. Furthermore, weused 10 CBR and 10 on-off traffic pattern files in CBR and on-off traffic simulations, respectively.We used the network topology files for setting up network topology and thus in each simulation 30nodes were randomly deployed in the simulation area. We used the traffic pattern files for gener-ating simulation traffic, and in each simulation 5 source-destination pairs were randomly selectedand simulation traffic was generated. We had 10 simulations with those topology and connectionpattern files in both CBR and on-off traffic simulations, and each data point in simulation resultswas the average of 10 simulation runs. In on-off traffic simulation, both busy and idle intervalsfollow exponential distribution with mean 312 ms and 325 ms, respectively [10]. We performedboth CBR and on-off traffic simulations at different transmission rates. These simulation condi-tions are presented in Table 3. The efficacy of power consumption can often be evaluated by howlong networks remain operational. However, the lifetime of a network is closely related to howit is used. If a protocol can promote high throughput, the lifetime of the network on which it isimplemented may be short. Therefore, we useenergy goodputdescribed by Eq.(3.1) to evaluatepower efficiency [15]:

energy goodput =total bits transmitted

total energy consumed(3.1)

where the total bits transmitted are calculated for application layer data packets only. The unit ofenergy goodput is bits/J, which in essence captures the energy utilization of the network with allcontrol overheads considered. We evaluate the efficiency of data delivery by the end-to-end delayand the packet delivery ratio, which is defined as the total number of packets received divided bythe total number of packets transmitted.

45

Simulation area 1500 m×500 mNumber of nodes 30

MAC&Physical settings IEEE802.11 Wave LAN[18]Bandwidth 2 Mbps

Beacon interval 100 msATIM and DATA window 25 ms

Transmission range 250 mAd hoc Routing Protocol DSR

Packet length 512 bytesBusy period 312 msIdle period 325 ms

Table 3.3: Simulation Conditions

Simulations for long-lived CBR traffic

The simulation network has 30 stationary nodes, and 5 source-destination pairs are randomly cho-sen during the simulation time, 300s. Figure 3.11 shows packet delivery ratio as traffic loadchanges, clearly indicating that, compared with normal IEEE 802.11, the packet delivery ratioin IPSM declines, but IPSM has a higher packet delivery ratio than other power saving mecha-nisms. This is because nodes with IPSM do not have to announce information on pending packetsand can stay awake to wait for arrivals of data packets, even if the network has a high traffic load.If a receiver node receives only one packet, it can remain in the awake state during the remainderof the current and the next beacon interval. However, in other power saving mechanisms, sendernodes have to announce information about buffered packets to receiver nodes during each ATIM orDATA window to stay in the awake state after that period. Therefore, if traffic congestion occurs,sender nodes cannot inform receiver nodes of pending packets to stay in the awake state after theATIM or DATA window. For example, it is difficult for sender nodes with NPSM to know thecounter values of receiver nodes during every DATA window when the traffic load is high. As aresult, the packet delivery ratio declines. This condition often happens when networks have heavytraffic loads.

Figure 3.12 shows a comparison of end-to-end delay in different power saving mechanisms.Since nodes with IPSM do not need to announce pending packets, they can relay packets imme-diately after they receive the packets. In other power saving mechanisms, announcing pendingpackets causes long delays in wireless multi-hop networks because sender nodes have to bufferpackets before informing receiver nodes of them. For these reasons, IPSM gives shorter delaysthan other power saving mechanisms; consequently, there is little difference between IPSM andnormal IEEE 802.11.

As shown in Fig.3.13, our proposed IPSM achieves the highest energy goodput at high trafficloads. It is clear that IPSM can realize energy efficient data delivery. At low traffic loads, NPSMshows better energy goodput than IPSM because nodes in IPSM stay awake to wait for packetarrivals longer than in NPSM. However, at higher traffic loads, sender nodes in NPSM cannot

46

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

10 20 30 40 50 60 70 80

Packet D

eliv

ery

Ratio

traffic load (kbps)

802.11PSM

NPSMIPSM

Figure 3.11: Packet delivery ratio, 5 long-lived CBR connections

inform receivers of pending packets during a DATA window, and then they cannot continue to stayawake to transmit packets. Moreover, since nodes with NPSM or PSM have to announce bufferedpackets to stay awake and to transmit them, data packets received by intermediate nodes cannot besmoothly relayed.

Upon analyzing these simulation results, we can see that IPSM can balance the trade-off be-tween throughput, delay and energy consumption and exhibit performance superior to other powersaving mechanisms.

Simulations for on-off traffic

The simulation time is 600 s. The simulation network has 30 stationary nodes, and 5 source-destination pairs, which is randomly selected. We evaluate the packet delivery ratio, end-to-enddelay and energy goodput vs. traffic load for each scheme. Figure 3.14 presents a comparisonof packet delivery ratio. Since, in simulations of on-off traffic, networks are less congested thanthose with long-lived traffic, NPSM and PSM have higher packet delivery ratio in on-off trafficsimulations than in long-lived traffic ones. However, as shown in the previous section, becauseNPSM and PSM cannot adapt to high traffic loads, the packet delivery ratio in NPSM and PSM islower than in IPSM. On the other hand, in IPSM, receiver nodes can stay in the awake state in theremainder of the current and the next beacon interval and wait for packet arrivals if they receiveonly one packet. Therefore, in IPSM the packet delivery ratio is higher.

A comparison of end-to-end delay is shown in Fig.3.15. IPSM shows shorter delay than otherpower saving mechanisms and almost the same result as the normal IEEE802.11 because inter-mediate nodes with IPSM can relay packets immediately after receiving them. In power savingmechanisms other than IPSM, buffered packets generate long delays as explained in the previoussection. Therefore they must suffer long delays when the traffic load is high. On the other hand, in

47

0

0.1

0.2

0.3

0.4

0.5

0.6

10 20 30 40 50 60 70 80

late

ncy (

sec)

traffic load per flow(kbps)

802.11PSM

NPSMIPSM

Figure 3.12: Latency, 5 long-lived CBR connections

0

5000

10000

15000

20000

25000

10 20 30 40 50 60 70 80

Energ

y g

oodput (b

its/J

)

traffic load per flow(kbps)

802.11PSM

NPSMIPSM

Figure 3.13: Energy goodput, 5 long-lived CBR connections

48

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

10 20 30 40 50 60 70 80

Packet D

eliv

ery

Ratio

traffic load per flow (kbps)

802.11PSM

NPSMIPSM

Figure 3.14: Comparison of packet delivery ratio, 5 on-off CBR connections

IPSM, receiver nodes can stay awake in the remainder of the current and the next beacon intervaland wait for packet arrivals, if they receive only one packet. This means that in IPSM, the packetdelivery ratio is higher and delay is shorter than in other PSMs.

In Fig.3.16, we find that IPSM shows the highest energy goodput when the traffic load is high.At low rates of traffic, however, energy goodput for NPSM is higher than for IPSM because nodesin IPSM have to wait for packet arrivals, thus staying awake longer than NPSM, and NPSM caninform receiver nodes of pending packets at low rates of traffic. However, of course, nodes inNPSM cannot adapt to high traffic load.

These simulation results indicate that our proposed IPSM strongly outperforms other protocolsand balance the trade-off previously mentioned. Furthermore, it can achieve almost the sameperformance with respect to delay and throughput as the normal IEEE802.11 and superior energygoodput to other power saving mechanisms. Even though, in this on-off traffic simulation, wechose 312 ms for an average busy interval and 325 ms for an average idle interval, we could obtainalmost the same results when we selected longer or shorter on and off intervals. That is, IPSMconsumes more energy at low traffic load than other power saving mechanisms, however, it canadapt to high traffic load and further balance the trade-off.

3.3.6 Conclusion of IPSM

This research discussed several power saving mechanisms for MAC protocol in ad hoc networks.Ad hoc networks with power saving mechanisms developed in previous researches caused longdelays and could not adapt to high traffic loads. To solve this problem, in this research we pre-sented an IPSM (improved power saving mechanism) which can operate only with MAC layerinformation. IPSM achieved shorter delay, higher throughput and high energy efficiency. Fur-thermore, it balanced the trade-off among these performance measures. Moreover, sender nodes

49

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

10 20 30 40 50 60 70 80

Late

ncy (

sec)

traffic load per flow (kbps)

802.11PSM

NPSMIPSM

Figure 3.15: Comparison of end-to-end delay, 5 on-off CBR connections

0

2000

4000

6000

8000

10000

12000

14000

10 20 30 40 50 60 70 80

Energ

y g

oodput (b

its/J

)

traffic load per flow (kbps)

802.11PSM

NPSMIPSM

Figure 3.16: Comparison of energy goodput, 5 on-off CBR connections

50

in IPSM could inform receiver nodes of their awake period instead of pending packets. Conse-quently, receiver nodes can wait for arrivals of packets, staying awake and relay them immediatelyafter receiving them.

Simulation results showed that our proposed scheme strongly outperformed other PSMs withrespect to throughput, delay and energy goodput.

Nodes with power saving mechanisms, including IPSM, must be synchronized by beacon trans-missions. Since synchronization between nodes influences performances of ad hoc networks, weplan to address this problem in future research.

51

3.4 Battery Cost Routing

3.4.1 Introduction

Since ad hoc networks are multi-hop wireless networks where all nodes cooperatively maintainnetwork connectivity, routing protocols greatly influence power consumption in ad hoc networks.If some nodes become responsible for routing packets between many source-destination pairs,they will deplete their energy earlier than other mobile nodes. Therefore energy efficient routingprotocols is of primary importance in mobile ad hoc networks. In this research, we focus ondesigning a routing protocol that increases the lifetime of mobile nodes and networks.

To prolong the lifetime of mobile ad hoc networks, some power-aware routing protocols wereproposed. Minimizing the energy consumed for each packets is an obvious solution that optimizesthe power consumption locally. This is because more power is consumed during the transmissionof packets than the reception of packets. Based on this observation, optimizing the transmissionpower for each connection yields prolonging the lifetime of mobile nodes and of entire ad hocnetwork[20], [21]. But if a specific mobile node is selected as a relaying node, it consumes its bat-tery earlier than other nodes. Consequently, it might lead to the early partitioning of the network.To solve this problem, other useful routing metrics were proposed for extending the lifetime ofmobile ad hoc networks. They include minimizing the variance in the power level of each nodeand minimizing the ratio of cost/packet. Routing protocols which use these metric can evenly dis-tribute packet-relaying loads to each node and prevent a node from being overused [22],[23]. In[22], it used the remaining battery capacity of each host as a routing metric, andMBCR(MinimumBattery Cost Routing) was proposed.MBCRselects the route with the maximum remaining bat-tery capacity. Therefore, it can distribute traffic loads for each node which takes part in a ad hocnetwork. However, if routing protocols consider only distributing energy consumption among adhoc networks, a path which needs large amount of energy to route packets will be selected. So, toprolong the lifetime of mobile ad hoc networks, mobile ad hoc networks require a routing algo-rithm that can minimize the overall transmission power for each connection as much as possibleand evenly distribute the power consumption rate of each mobile node. These two constraints mustbe simultaneously satisfied. In our approach, we propose the routing protocol that satisfies them.

The remainder of this research is organized as follows. In the next section 3.4.2 we presentprior arts about power-aware ad hoc routing protocols. In section 3.4.3 we show our proposedrouting protocol. Section 3.4.4 presents the results of our simulations. Finally, we conclude thisresearch and outlines our future research.

3.4.2 Prior Arts

Minimum Total Transmission Power Routing (MTPR)

In wireless communications, radio propagation can be modeled effectively with a1/dn (n ≥ 2),transmit power roll-off (usually,n = 2 for short distance andn = 4 for long distance). And lowerpower transmission by changing transmission power can reduce power consumption and channelinterference. In multi-hop wireless communications, given the path loss model , in some case,

52

sr

Relay region

sr

Relay region(a) n=2 (b) n=4

h h

Figure 3.17:Relay region with path loss model of (a)p ∼ 1/d 2 (b) p ∼ 1/d4

relaying packets using additional nodes may result in lower power consumption than communi-cating directly. In Fig.3.17, a region described by oblique lines illustratesrelay region. It meansthat if the senders sends packets to the destinationh, using the noder as the next hop is morepower-efficient than directly communicating froms toh. Therelay regionis presented by Eq.(3.2).

bdnsr + bdn

rh ≤ bdnsh (3.2)

whereb denotes predetection threshold (in mW) at each receiver, in other words, minimum powerwhich a transmitter must radiate in order to allow detection at distancedmeters away isbdn, wheren is the exponent in the path loss model. To maximize the lifetime of all nodes, a routing protocolthat minimizes transmission power is proposed in [20]. In [20] the transmission powerP (n i, nj)between hostsni andnj was used as a metric. The total transmission power for routel, Pl, can bederived from

Pl =

D−1∑

i=0

P (ni, ni+1) for all node ni ∈ route (3.3)

wheren0 andnD are the source and destination node, respectively, other nodes are relaying nodes,andD is the number of nodes in the route. The desired routek can be obtained from

Pk = minl∈A

Pl (3.4)

whereA is the set containing all possible routes. However, since this algorithm considers onlytransmission power, it may select a route with more hops than other routing algorithms. To over-come this problem, transceiver power as well as transmission power are considered as a metric[22]. At nodenj, it computes

Ci,j = Ptransmit(ni, nj) + Ptransceiver(nj)

+Cost(nj) (3.5)

53

1 2

34

5

node5 is overused.

Figure 3.18: Overusing problem

whereni is a neighboring node ofnj, Ptransceiver(nj) is the transceiver power at nodenj, andCost(nj) is the total power cost from the source node to nodenj . This value is sent to nodeni. Atnodeni it computes its power cost as follows:

Cost(ni) = minj∈NH(i)

Ci,j (3.6)

NH(i) ={ j ;nj is a neighbor node ofni }The path with minimum cost from the source node to nodeni is selected. This procedure isrepeated until the destination node is reached.Ptransceiver(nj) helps the algorithm find routes withfewer hops.

This algorithm can reduce the total power consumption of overall network, but has one seriousdrawback, as shown in Fig.3.18 [23]. Here, node 5 will be selected as the route for packets goingfrom 1−3 and 2−4. As a result, node 5 will consume its battery at much faster rate than the othernodes in the network. Therefore, routing protocols also have to consider other metrics for gettingover the problem in Fig.3.18. To incorporate the power level of each node into routing metrics,the remaining battery capacity of each node were considered as a metric to describe the lifetime ofeach node [22].

Minimum Battery Cost Routing (MBCR)

MBCR selects a route that has the maximum sum of remaining battery capacity in all possibleroutes [22]. Letcti be the battery capacity of a nodeni at timet. In [23], fi(c

ti) is defined as the

following battery cost function of a nodeni.

fi(cti) =

1

cti(3.7)

54

Equation (3.7) shows whether nodeni has much enough battery to transmit packets or not. Thatis, as the remaining battery capacity decreases, the value of cost function for nodeni will increase.The battery costRj for routej, containing ofD nodes, is,

Rj =D−1∑

i=0

fi(cti) (3.8)

Therefore, to find a route with the maximum remaining battery capacity, the route that has theminimum battery cost is selected by using the following Eq. (3.9).

Ri = min{Rj | j ∈ A} (3.9)

whereA is the set containing all possible routes.This algorithm prevents nodes from being overused, because battery capacity is directly in-

corporated into the routing protocol. But this protocol does not consider transmission power. Apath selected byMBCRmay need more transmission power to deliver packets and consume moreenergy.

Discussion of prior arts

Among the conventional power-aware routing algorithms, the first one,MTPR, can minimize theoverall transmission power for each connection from a source node to a destination node. Butit does not consider distributing the power consumption rate of each mobile node. Therefore aspecific node can be selected as a relaying node, and it consumes its battery earlier than othernodes. The second one,MBCR, can achieve evenly distributing the power consumption rate ofeach mobile node. However, since it does not take account of transmission power, a selected routemay need larger amount of energy to forward packets from a source node to a destination nodethan other routes.

To prolong the lifetime of mobile ad hoc networks, mobile ad hoc networks require a routingalgorithm that can minimize the overall transmission power for each connection as much as possi-ble and evenly distribute the power consumption rate of each mobile node. These two constraintsmust be simultaneously satisfied. In our approach, we propose a routing protocol that satisfiesthem simultaneously. The next section presents our proposal.

3.4.3 Battery Cost Routing in Consideration of Transmission Power

The Idea of Our Proposal

Since our proposed algorithm is based on source routing like DSR [25], route requests (RREQs) arepropagated to a destination node and a route reply (RREP) is delivered along the selected sourceroute. We suppose that each node can control its transmission power based on a distance betweena transmitter and a receiver and a source node knows the amount of data to send to a destinationnode.

Figure 3.19 (a) and (b) show the way of selecting a path from nodenj to nodenk in MBCRand our proposal, respectively. In Fig.3.19,ctni1

andctni2respectively present remaining battery of

55

nodeni1 and nodeni2, anddji is the distance between nodenj andni. Now suppose thatctni1is

larger thanctni2

, dji2 is shorter thandji1, anddi2k is shorter thandi1k. Furthermore,Rl denotesthe sum of battery cost of routel. Rl is calculated by using battery cost of each node.MBCRuses remaining battery as a metric. So, comparing the sum of battery costR1 with R2, we findR1 is smaller thanR2. ThenRoute1is selected from nodenj to nodenk (Fig.3.19 (a)). However,when using path loss model1/dn,(n ≥ 2), transmission power from nodenj to nodeni1 is largerthan to nodeni2. Transmission power from nodeni1 to nodenk is also larger than from nodeni2. If there is little difference betweenct

ni1and ctni2

, selectingRoute1which consumes muchlarger amount of energy is not realistic. Therefore, our proposed routing protocol simultaneouslyconsiders transmission power and remaining battery of each node, and selects a route which canminimize overall transmission power from a source node to a destination node and evenly distributepower consumption of each node.

In our routing algorithm, each node computes the remaining capacity of battery, consideringbattery needed to transmit data. That is, we substitute(remaining battery)−(battery needed to send data) into ct

i in Eq.(3.7) as opposed to only remaining battery.cti which

we define can achieve evenly distributing power consumption to each node and minimizing trans-mission power as much as possible. Sum of battery cost is also calculated with Eq.(3.8) in ourproposal, and then the path selection is done by using Eq.(3.9). In Fig.3.19 (b), we calculate bat-tery cost of nodeni1 and nodeni2, considering transmission power, and sum of battery cost of eachroute. IfR1 is larger thanR2, we will selectRoute2as the path from nodenj to nodenk.

The details of our proposal are presented in Fig.3.20 and the followings,

1. Source nodeswrites in a RREQ the amount of data to send, the position of its own geograph-ical position and its remaining battery capacity, and broadcasts RREQ with the maximumcommunication range.

2. A node receiving RREQ calculates the distance between the source node and itself and de-cides transmission power of the source node. Using the amount of data indicated in RREQand remaining battery capacity of the source node, it calculates the battery cost of the sourcenode after transmitting data. Thusct

s is calculated.

3. If a node receiving RREQ is the destination node, it selects the route among available mul-tiple routes, using Eq.(3.9). Otherwise, that is to say, a node is a intermediate node, it writesits own geographical position and battery capacity in RREQ, and broadcasts RREQ with themaximum communication range.

4. These procedures are repeated until the destination node is reached

5. Finally the destination node forwards RREP to the source node along the selected sourceroute.

Performance Comparison of Different Routing Protocols

In order to better understand different routing protocol in terms of power efficiency, here we per-form a simple simulation. This simulation is not done at the packet level. It is assumed that a node

56

njni2

ni1

dji2

dji1

cni2

cni1

(a) the way of selecting the next hop in MBCR

(b) the way of selecting next hop in our proposal

nk

di2k

di1k

njni2

ni1

dji2

dji1

cni2

cni1

nk

di2k

di1k

Route1

Route2

Route1

Route2

t

t

t

t

Figure 3.19: Our proposed protocol operations

nj ni

RREQsend

The RREQ includes nj’s position and nj’s battery cost.

ni calculates the distance between nj and ni and decidesnj’s battery cost.

Figure 3.20: Calculation of battery cost

57

7.60 10^4

7.80 10^4

8.00 10^4

8.20 10^4

8.40 10^4

8.60 10^4

8.80 10^4

9.00 10^4

0 5 10 15 20 25

Exprira

tion T

ime (

s)

Expiration Sequences

ProposalMBCRMTPR

Figure 3.21: Comparison of power efficiency in different power-aware routing protocols

which receives packets and is not selected in a route does not consume its energy. The simula-tion result shows a relation between the lifetime of nodes and the expiration sequence. We alsoassume when a route request occurs according to a Poisson process, two nodes are randomly se-lected as a source and a destination. If a route is constructed, nodes in the route consume the sameamount of battery. The request arrival rate is proportional to the number of nodes that power up.In this simulation, three different route selection schemes,MTPR, MBCR and our proposal, areconsidered.

30 mobile nodes are randomly distributed in a confined space of 100m×100m. Each nodemoves at a speed of 1m/s according to therandom waypoint model[25]. A sender can reach a node25 meters away at its maximum transmission power. We use the path loss model as1/dn, n = 2,whered is the distance between nodes. Power consumption is 1.4 W for transmission and 1.0 Wfor receiving [13].

From Fig.3.21, we findMTPR do not evenly distribute the power consumption rate of eachnode. So there is a big difference between the time the first node expires its battery and the thetime the last node expires its battery. Other two ways can consume battery evenly.

ComparingMBCRwith our proposal, it is clear that our proposal has longer lifetime of nodesand satisfies two factors simultaneously, distributing the power consumption rate of each node andminimizing the overall transmission power for each connection as much as possible.

Transmission of Packets in Our Proposal

In previous subsections, we do not talk about our proposal at the packet level and transmissionpower is simply set to the distance between nodes. When evaluating our proposed algorithm at thepacket level, we have to decide the details of how to set transmission power between nodes. Inmobile ad hoc networks, nodes move even during transmitting data. If transmission power is not

58

enough strong to transmit packets, packets cannot be successfully received. Here, we present theway of deciding transmission power.

Figure 3.22 presents the communication from nodenj to nodeni. Decision of transmissionpower is done by using the maximum speed of each node. In Fig.3.22, we define the average speedof nodeni andnj asvni andvnj respectively. The range where nodeni andnj move for t[s] isdotted circles with radiusdi = vni × t anddj = vnj × t respectively in Fig.3.22. And in Fig.3.22d is the distance between nodenj andni calculated when RREQ is transmitted. Therefore, if thecommunication time is t[s], nodeni decides nodenj ’s communication range asd + dj + di andtransmission power is calculated. The time t[s] is decided by the amount of data a source nodeknows and the transmission rate. This time is called the communication time. But in fact, we haveto consider extra time added to the communication time. This is the latency time between RREQis sent and RREP is received. So we redefine the communication time, t[s], as T[s],T = t(1 + α).α is expressed as the proportion of t[s]. For example, ifα is 20% of t[s], T [s] = t(1 + 0.2). Ifd + dj + di is over the communication range, the communication range is set to maximum one.The details of our proposal are presented in the followings.

1. A source node writes in RREQ the amount of data to send, its own geographical position,its remaining capacity of battery, time needed to complete communication and its maximumspeed and broadcasts RREQ with the maximum communication range.

2. A node received RREQ calculates the distance between the source node and itself and de-cides transmission power of the source node. Using the amount of data indicated in RREQand the remaining capacity of battery of the source node, it calculates the battery cost of thesource node based on battery after transmitting data.

3. If a node received RREQ is the destination node it waits for several RREQs during a waitingperiod. When the waiting timer is expired, it selects the route among available multiple path,using Eq.(3.9). This waiting timer is defined in the next section. If it acts as an intermediatenode, it writes its own geographical position, its remaining battery capacity and its maximumspeed in RREQ, and broadcasts RREQ with the maximum communication range.

4. These procedure are repeated until the destination node is reached.

5. Finally the destination node forwards RREP to the source node along the selected sourceroute.

A destination node and intermediate nodes transmit RREP with the maximum communicationrange.

Waiting Period

When a destination node selects an energy efficient path, it has to wait for arrival of several RREQs.This issue is very important, but not much related with how to set routing metrics to routingprotocols. Previous works ,[22] and [23] did not define this waiting period. Even though the issueis also out of the scope of this research, we simply define this parameter in order to take computersimulations.

59

nid

dj dinj

Figure 3.22: The way of deciding the communication range

We define how long a destination node waits for RREQs as Eq.(3.10) and Eq.(3.11).

Tave =Trecv − Tgen

h(3.10)

,whereTave denotes the average time required for RREQs to be forwarded by one hop,Trecv de-notes the time when the first RREQ arrives at a destination node,Tgen indicates the time when theRREQ is generated at a source node, andh presents the number of hops that the first arrival RREQneeds to be forwarded. The waiting period,Twait, is expressed in Eq.(3.11).

Twait = Tave ∗HP (3.11)

,whereHP indicates how many hops a destination node additionally needs to wait for RREQs.The waiting period will influence the performance of power-aware routing protocols. However,generally, if a destination node waits for RREQs for a long period, a RREQ that reaches theretoward the end of the period cannot be selected, since a route with many hops generally showslarge amount of battery cost in Eq.(3.9). Therefore destination nodes do not have to wait for a verylong time. Additionally, it is very difficult to optimize the waiting period. As the simple method,we consider the case whereHP is constant. We can, however, adaptively change according to thecircumstances.

3.4.4 Evaluation

Simulation Conditions

We implemented our proposed routing protocol using NS-2 [18]. In our simulation, four differentrouting protocols are considered. They areMTPR, MBCR, Shortest Path Routing(Shortest) andour proposal.

In our simulation model, a network is placed in a confined space of 1000m×1000m and nodeshave 802.11 network interface card. A sender can reach a node 250 meters away at its maximumtransmission power. We assume that a radio consumes 1.4W for transmitting and 1.0W for receiv-ing [13]. Path loss model isTwo Ray Ground Model. Each node moves according to therandomwaypoint model[25], with a maximum velocity of 3 m/s and with a pause time of 0 seconds.

60

In all scenarios, we use 20 traffic nodes acting as sources and sinks, and not acting as relayingnodes. They have full of battery and it is not expired during simulation running. When we examinenode density, we set the number of relaying nodes to 5, 10, 20, 25, and 30. That is, the number ofall nodes in the simulation area is varied from 25 to 50.

Simulation traffic is CBR. The packet length is 512 bytes. A source and destination are se-lected according to a Poisson process. Then 100 packets are transmitted from a source node to adestination node. The packet rate was set to 10 pkts/s.

Evaluation of waiting period at destination nodes

In order to decide waiting period, we evaluatepacket delivery ratioand network lifetime in MBCRfor differentHPs. We definepacket delivery ratioas (number of successfully deliverd packets)

(number of transmitted packets). When

evaluatingpacket delivery ratio, all nodes in the network do not expire their battery. Networklifetime is evaluated by the relation between the lifetime of nodes and the expiration sequence.Figure 3.23 showspacket delivery ratio. We can find there is little difference among differentHPs. Figure 3.24 shows network lifetime for each value ofHP . We can see there is a littledifference among eachHP . However, we find that, whenHP is 4, the lifetime of nodes is thelongest. This is because, whenHP is set to 2, waiting period for RREQs is too short to selectenergy efficient path, and, whenHP is set to 6, a destination node can select a path which needsmore hops. At the packet level simulation, a node which is not included in a route can overhearpackets and consumes its battery. Therefore, if longer path is selected, the network lifetime can beshorter. Since, in our simulation, the network size is set to a square region of 1000 meters on eachside and maximum transmission range is 250 meters, source nodes need at most 4 hops to deliverpackets to destination nodes.

From these simulation results, we can see that we have to consider path length to select energyefficient routes. MBCR does not care about it and transmission power to deliver a packet. Ifpower-aware routing protocols consider total transmission power for each connection, they canselect shorter path thanMBCR. Therefore it is necessary to consider not only remaining battery ofeach node, but also transmission power. We select waiting period asHP is 4, and use that value insubsequent simulations.

Performance Comparison in Our Proposal

In our proposed algorithm, we useT = t(1 + α) to set the communication time, considering thelatency time between the time when RREQ is sent and the time when RREP is received. Therefore,in this subsection, we evaluate our proposal for different values ofα.

We choose0.2, 0.4, and0.6 asα and evaluatepacket delivery success ratioin Fig.3.25 and thenetwork lifetime in Fig.3.26, respectively. We varied the number of nodes for evaluation ofpacketdelivery success ratioand set 50 nodes in the simulation area for the evaluation of the networklifetime.

From Fig.3.25, we can find that there is not a big difference among values ofalpha, especiallybetween whenα is 0.4 and whenα is 0.6. However, when we selectα of 0.6, packet deliverysuccess ratiois the highest. This is because transmission power is larger than otherαs and is

61

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

25 30 35 40 45

Packet D

eliv

ery

Success R

atio

number of nodes

MBCR HP = 2MBCR HP = 4MBCR HP = 6

Figure 3.23: Comparison of packet delivery ratio for different waiting time in MBCR

1760

1780

1800

1820

1840

1860

1880

0 5 10 15 20 25 30

Exprira

tion T

ime (

s)

Expiration Sequences

MBCR HP = 2MBCR HP = 4MBCR HP = 6

Figure 3.24: Comparison of network lifetime for different waiting time in MBCR

62

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

25 30 35 40 45

Packet D

eliv

ery

Success R

atio

number of nodes

alpha = 0.2alpha = 0.4alpha = 0.6

Figure 3.25: Comparison of packet delivery ratio for differentαs in our proposal

set to enough strong. From Fig.3.26, we see that the network lifetime is the shortest, whenα is0.6. Larger transmission power results in shorter network lifetime. As the other reason, when wechoose small transmission power,packet delivery success ratiois lower than when larger one isselected, and then nodes consume less energy for transmitting packets. As a result, we find therelation betweenpacket delivery success ratioand the network lifetime is trade-off. Therefore, wehave to select the value ofα, considering that relation.

We choose0.4 asα in the following simulations. The reason is that the difference forpacketdelivery success ratiobetween whenα is 0.4 and whenα is 0.6 is little.

Performance Comparison in Different Power-aware Routing Protocols

In Fig.3.27, we show a simulation result, comparingpacket delivery ratioin different power-awarerouting protocols and a shortest path routing protocol. They areMBCR, MTPR, Shortest, and ourproposal. From this result, we see that, in MBCR andShortest, packet delivery ratiois higherthan other routing algorithms, but the difference betweenpacket delivery ratios is very little. Ourproposal can deliver packets as well as other conventional protocols.

Figure 3.28 shows a comparison of network lifetime in different power-aware routing protocolsand a shortest path routing protocol. We find that the network lifetime in our proposal is the longestamong them. Our proposed algorithm simultaneously considers minimizing total transmissionpower for each connection as much as possible and evenly distributing power consumption rate ofeach mobile node. Even though MBCR achieves longer lifetime thanShortest, it needs more hopsto deliver packets from a source to a destination and consequently more energy than our proposalsince MBCR considers only remaining battery of nodes.

In wireless networks, a node not acting as a transmitter or a receiver consumes its energy whenit overhears transmission of neighbors. Even if energy efficient path is selected, nodes wastes their

63

1820

1840

1860

1880

1900

1920

1940

1960

1980

2000

0 5 10 15 20 25

Exprira

tion T

ime (

s)

Expiration Sequences

alpha = 0.2alpha = 0.4alpha = 0.6

Figure 3.26: Comparison of network lifetime for differentαs in our proposal

energy by overhearing packets. However, in order to avoid overhearing, MAC protocol, such asIEEE802.11 [1], supports power saving mechanism. Nodes using this protocol can turn their radiooff if they do not act as a sender or a receiver. Supposing this types of MAC protocol, nodes doesnot have to overhear packets.

3.4.5 Conclusion of Battery Cost Routing in Consideration of TransmissionPower

In this research, we discussed power-aware routing protocols in mobile ad hoc networks and pre-sented our proposed routing protocol. It simultaneously satisfies two factors needed to prolong tothe lifetime of ad hoc networks. The simulation results showed our proposal performed efficientpower consumption and prolonged the network lifetime.

There are a number of works to be done as further researches. Since our proposed routingprotocol is based on source routing, it has to extend to other routing protocols. In our proposedprotocol, now, transmission power is controlled based on distance between two nodes. Therefore,we will develop transmission power suitable for ad hoc networks, considering dynamic networktopology.

3.5 Conclusion of this Chapter

This chapter focused on ad hoc networks consisting of IEEE802.11 WLAN. It contained tworesearches, mainly working on power saving in ad hoc networks.

The first research was on enhancing throughput in ad hoc networks operating with IEEE802.11PSM. Since, in ad hoc networks with IEEE802.11 PSM, sender nodes have to transmit ATIM

64

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

25 30 35 40 45

Packet D

eliv

ery

Success R

atio

number of nodes

ProposalMBCRMTPR

Shortest

Figure 3.27: Comparison of packet delivery ratio in different power-aware routing protocols

1700

1750

1800

1850

1900

1950

2000

0 5 10 15 20 25

Expiration T

ime (

s)

Expiration Sequences

ProposalMBCRMTPR

Shortest

Figure 3.28: Comparison of network lifetime in different power-aware routing protocols

65

frames to receiver nodes in order to continue to be awake and to send DATA frames to the re-ceiver nodes, they have to buffer DATA frames before they transmit DATA frames. As a result,throughput declines due to buffering delay at transmitter nodes. We, therefore, proposed IPSMto improve the performance of throughput in ad hoc networks operating with IEEE802.11 PSM.Nodes operating with IPSM does not have to exchange ATIM frames and instead exchange theperiod of time when they continue to be awake. Consequently, they do not have to buffer DATAframes to exchange ATIM frames and can transmit them to receiver nodes immediately after theyreceive them if they know receivers are awake. Simulation results showed IPSM could overcomethroughput degradation in ad hoc networks with IEEE802.11 PSM.

The second research worked on power efficient routing protocols in ad hoc networks. Sincenodes in ad hoc networks convey packets by means of multihop wireless communications, routingprotocols largely influence power consumption in ad hoc networks. For power conservation, it isimportant for source nodes to select a route which minimizes total transmission power from themto destination nodes. Therefore, routing protocols in ad hoc networks have to consider transmissionpower as a routing metric. However, if routing protocols consider only transmission power, theyselect a route which selects specific nodes, those particular nodes consume their battery earlier thanother nodes and networks may be partitioned. Therefore, routing protocols have to also encouragenodes to fairly use battery among them. Our proposed routing protocol considered these tworequirements and achieved selecting a route which minimizes transmission power from a sourceto a destination node and evenly utilizes battery power among nodes taking part in an ad hocnetwork. Simulation result showed that our proposal achieved power efficient packet delivery, fairpower consumption among nodes, and consequently increased the lifetime of ad hoc networks.

66

Chapter 4

Performance Improvement of IEEE802.11eWLAN Networks

67

4.1 Overview of This Chapter

As explained in section 2.6, IEEE802.11 WLAN can configure infrastructure networks. In infras-tructure networks, each STA communicates with an AP and accesses to backhaul networks suchas wired LANs and internet. Nowadays, WLAN is mainly used for Internet access, but real-timeapplication like VoIP (Voice over IP) and video conference are identified as promising applicationsfor WLAN. Since these applications require distinct specific features, such as delay sensitivity orbandwidth requirement, it is desired to support differential services in IEEE802.11 standard [1].MAC protocol in [1] employs a contention-based channel access, called DCF. The DCF operateswith CSMA/CA. However the DCF does not work well with real-time applications due to the factthat a STA, having real-time traffic, may wait for long time to access the wireless medium re-gardless of its requirement. When real-time traffic contends with best effort traffic, both of traffichas the same opportunity to access the wireless medium. Therefore, real-time traffic, which hasdelay sensitivity, does not meet its requirement under DCF. To overcome the problem presentedabove, IEEE802.11e working group standardized new 802.11 MAC protocol, which provides QoS[2]. The 802.11e HCF (hybrid coordination function) can support QoS in IEEE802.11 WLAN net-works. The HCF provides both a contention-based channel access, called EDCA, and a controlledchannel access, referred to as HCCA. In order for WLAN to be more widely used and to supportQoS in WLAN networks, the functionalities of IEEE802.11e have to be provided in WLAN de-vices. This chapter therefore focuses on infrastructure networks consisting of IEEE802.11e WLANand EDCA as a channel access scheme. In the following sections, we propose several methods toenhance the performance of IEEE802.11e WLAN networks.

Although IEEE802.11e can provides QoS more than DCF and real-time traffic is given moreopportunities to access the wireless medium than non real-time traffic, it cannot completely guaran-tee QoS for real-time traffic. Real-time traffic is sometime suffered from performance degradationdue to contentions between real-time flows and between real-time and non real-time flows. Toovercome such performance degradation, first we research on and propose a method called Dy-namic Adaptation of EDCA Parameters in the following section.

Power consumption is one of the most critical issue when IEEE802.11 WLAN is used for ahandheld device. IEEE802.11e defines U-APSD (Unscheduled Automatic Power Save Delivery)for energy-efficient data delivery. However, STAs operating with U-APSD are sometime sufferedfrom large delay of downlink packets. Second, we therefore research on and propose a method todecrease delay of downlink packets in U-APSD.

Before each research part is presented, we introduce several features defined in IEEE802.11eWLAN [2] in the next section.

4.2 IEEE802.11e WLAN

IEEE802.11e provides some functionalities to support QoS in WLAN networks. This sectionintroduces these IEEE802.11e features, such as channel access method called EDCA, admissioncontrol under EDCA and continuous TXOP.

68

4.2.1 Channel Access Mechanism

EDCA provides differentiated and distributed channel access to the wireless medium based on 8different UPs (User Priorities). Each frame arriving at the MAC from the higher layers carries anUP value. Each higher layer UP is mapped into an AC (Access Category) as presented in Fig.4.1.The mapping rule from UPs to ACs is defined in [2]. As shown in Fig.4.1, four transmissionqueues are implemented in a STA and each queue supports one AC. Each AC behaves as a singleDCF entity [1], contends for the channel access, and independently starts its backoff procedure.Differentiated channel access in EDCA is achieved through setting in each AC different channelaccess parameters such as AIFS (arbitration inter frame space) and the size of CW (ContentionWindow). AIFS is the amount of time a station would sense the channel to be idle and the lengthof CW is used for the value of backoff counter. To express EDCA parameters in ACk (0 ≤ k ≤ 3),the minimum CW size isCWmin[k], the maximum CW size isCWmax[k], and the arbitration interframe space isAIFS[k]. Further the arbitration inter frame space number isAIFSN [k]. Therelation betweenAIFS[k] andAIFSN [k] is as follows.

AIFS[k] = AIFSN [k]× slotT ime+ SIFS

slotT ime is the time unit defined in each PHY layer setting such as IEEE802.11a [3], andSIFSis the nominal time that MAC and PHY layer require to receive the last symbol of a frame at theair interface, process the frame, and respond with the first symbol on the air interface [1], [3]. Thebackoff counter in ACk is selected from[0, CW [k]], whereCW [k] is the current CW size in ACk. At first transmitting attempt,CW [k] is assigned to the value ofCWmin[k]. After consecutivetransmissions (due to collisions), the value ofCW [k] is increased up to the maximum value ofCWmax[k] in binary exponential manner. Figure 4.2 shows an example of the timing relationbetween AC0 and AC3. In Fig.4.2, CW[3] and CW[0] present the current CW sizes in AC3 andAC0, respectively. If AC3 has smaller values of AIFS and CW size than AC0, the AC3 needs ashorter time to sense the channel to to be idle before starting backoff and its backoff time is alsoshorter than AC0. Therefore it can have more chance to access the wireless medium earlier. InEDCA, an AC for real-time traffic has smaller values of AIFS and CW size than other ACs of nonreal-time traffic. The AP announcesCWmin[k], CWmax[k] andAIFSN [k] as a part of EDCAparameter-set in beacon frames. Default EDCA parameters are shown in [2]. When more than oneAC within a STA has their backoff timers expire at the same time, the collision among them istreated in a virtual manner by virtual collision handler shown in Fig.4.1. Then the highest priorityframe is chosen and transmitted, and the other ACs increase values of CW and start their backoff.

However, even though, in a STA, an AC for real-time traffic can have more opportunities toaccess the wireless medium than non real-time traffic, collisions often happen when there aremany transmissions of not only real-time traffic but also non real-time traffic from other STAs. Asa result, real-time traffic may not be able to meet its requirement. Therefore, it is necessary toprotect real-time traffic from other real-time traffic and non real-time traffic

4.2.2 Admission Control

This subsection shows admission control method under EDCA which is defined in [2] and is wellknown astoken bucket.

69

AIFS[0]CW[0]

AIFS[1]CW[1]

AIFS[2]CW[2]

AIFS[3]CW[3]

Mapping User Priority to Access Category

Virtual Collision Handler

AC0 AC1 AC2 AC3

Transmission Attempt

Each DCF entity handles backoff

transmission queue

Figure 4.1: Implementation model for EDCA

AIFS[3]

AIFS[0]

Channel is busy

slotTime

Backoff windowselected from [0, CW[3]]

Backoff windowselected from [0, CW[0]]

Frame transmission

Figure 4.2: Example of channel access timing in EDCA

70

If admission control is needed for an AC, a STA has to send an ADDTS (add traffic stream)request frame to the AP. The ADDTS request contains TSPEC (traffic specification), such as meandata rate, nominal MSDU (MAC service data unit) size, delay bound and etc.. When the APreceives an ADDTS request, it makes a determination as whether to accept or deny the request. Ifit accepts the request, it calculates from information conveyed in the request the amount of timefor requested traffic to access the wireless medium per one second, which is calledmediumtime.Even though any algorithms can be used for derivingmediumtime, a recommended procedure ispresented in [2]. After calculating it, the AP sends to the STA an ADDTS response frame, whichcontains derivedmediumtime. On receipt of the response from the AP, the STA addsmediumtimeto a local variable,admittedtime, if the request is admitted. It also has another local variable, calledusedtime. Theusedtime presents how long the STA has accessed the wireless medium. Usingtheadmittedtimeand theusedtime, the STA controls the channel access to the wireless medium.Theusedtimeare updated after each successful or unsuccessful (re) transmission attempt of dataframe as follows,

used time = used time+DataExchangeT ime (4.1)

Further, at seconds interval,

used time = max((used time− admitted time), 0) (4.2)

DataExchangeTime is the duration needed to transmit a data frame, considering the duration ofSIFS or ACK transmission [2]. Ifusedtimereaches or exceeds theadmittedtime, the correspond-ing AC cannot transmit any frames using its EDCA parameter-set until theusedtime is reset asshown in Eq.(4.2). Since admission control prevents large number of real-time traffic from access-ing the wireless medium, it is very important for protecting the existing multimedia traffic fromother real-time traffic. However, even if admission control is used for multimedia traffic, it cannotensure to maintain requirements of real-time traffic under EDCA. Thus the traffic can be disturbedby other real-time and non real-time traffic.

4.2.3 Continuous TXOP

In addition to the differentiated channel access, the EDCA supports an useful function. It is referredto as a continuation of TXOP. As shown in Fig.4.3, a STA using the continuation of TXOP canconsecutively access the wireless medium a SIFS period after the successful transmission of aDATA frame [2]. That is, the STA does not have to contend for channel access every time andit can continuously transmit several frames once it has the right to access the wireless medium.How long the STA can consecutively transmit frames is specified by the TXOP Limit in the EDCAparameter-set transmitted in beacon frames. As presented in Fig.4.3, the whole transmission timeof several data frames belonging to ACk and the corresponding ACK frames is less than TXOPLimit whose value is specified for ACk. And in each AC the value of TXOP Limit is specified.Any algorithms for the determination of the TXOP Limit are not specified in [2], and they dependon vendors’ implementation. If an appropriate TXOP Limit is given to real-time flows, they canbe protected from other flows or best effort traffic under EDCA. In order to maintain desiredthroughput of real-time traffic and efficiently utilize the wireless bandwidth, TXOP Limit should

71

AIFS[k]

slotTime

Backoff windowselected from [0, CW[k]]

Data ACK Data ACK

SIFS SIFS SIFS

TXOP Limit [k]

Time gap

Post Backoff

Figure 4.3: Continuous TXOP in EDCA

be determined, considering channel load or interference from other STAs. In section 4.3.4, wepropose an algorithm for selecting appropriate TXOP Limit based on desired throughput of real-time flows.

4.3 Protection of Real-time Traffic under IEEE802.11e EDCAvia Dynamic Adaptation of EDCA Parameters

4.3.1 Introduction

IEEE 802.11e working group defined new 802.11 MAC protocol, which provides QoS [2]. The802.11e HCF (hybrid coordination function) can support QoS in 802.11 networks. The HCF pro-vides both a contention-based channel access, called EDCA, and a controlled channel access,referred to as HCCA. In this research we focus on the EDCA. The EDCA ensures that a STA withhigh priority traffic (i.e. traffic with real-time requirement) can have more opportunities to accessthe wireless medium than low priority traffic transmitted from other STAs or itself. The EDCAachieves the service differentiation through setting different CW sizes used for backoff conter andinter-frame spaces. Comparing with the DCF, the EDCA can guarantee the service differentiation.But it does not completely meet requirements of real-time traffic. If an AP accepts a lot of flows,the network will become saturated and then they are suffered from performance degradation. Toavoid excess accesses, the 802.11e supports an admission control scheme. The next section ex-plains admission control under the EDCA.

However, even though the EDCA provides both the service differentiation and the admissioncontrol, it does not fully protect real-time traffic. Since the EDCA provides contention-basedchannel access, contentions between real-time flows or between real-time and non real-time flowsdegrade performance of throughput and delay of real-time traffic. Hence this research proposestwo methods to protect real-time traffic under EDCA. First, we focus on reducing collisions amongreal-time flows. Basically CW sizes for real-time traffic are smaller than non real-time traffic sothat real-time traffic gets more chance to access the wireless medium. But, especially if thereis a large number of real-time traffic, contentions between these flows often occur and degradethroughput of real-time traffic. This phenomenon happens due to the fact that the default CW sizes

72

defined in [2] for real-time traffic are very small. Short backoff time generated by small CW sizesresults in many collisions between real-time traffic and the network cannot accommodate manyreal-time flows. One solution to overcome this problem is to set larger CW sizes as default values.But larger value of CW size makes larger backoff time (i.e. channel access delay) and reducesefficiency of bandwidth utilization. For example, in case where the network has little real-timetraffic, non real-time traffic has to wait for long time to access the wireless medium. Therefore, itis desirable for CW sizes to be adaptively changed. In this research, we propose a method for anAP to dynamically control CW sizes.

Next, we focus on a method to protect real-time traffic from non real-time traffic. In fact, nonreal-time traffic, such as best effort traffic, disturbs multimedia traffic under EDCA even if ad-mission control is used. One solution to avoid this problem is to use admission control for nonreal-time traffic. However, since best effort traffic generally does not have flows or delay con-straint, it’s difficult to use admission control for non real-time traffic. The authors in [26] proposeda protection mechanism of existing real-time traffic from non real-time traffic. Although their pro-posed method achieved their goal, this mechanism inefficiently utilized the wireless bandwidth. Toprotect real-time traffic from non real-time traffic, we focus on continuation of TXOP. The contin-uous TXOP is introduced in [2] as one of new features in the 802.11e. TXOP is the interval of timewhen a STA can initiate frame exchange onto the wireless medium. Duration of TXOP is specifiedin the TXOP Limit, which is included into the EDCA parameter-set in beacon frames transmittedfrom the AP. The TXOP Limit is calculated at the AP, but how to calculate it is dependent onvendors’ implementation [2]. In this research, we propose a method for selecting an appropriatevalue of TXOP Limit referred to asAdaptiveTXOP. It monitors throughput of admitted flows andadaptively calculates a value of TXOP Limit. If an appropriate TXOP Limit is given to multimediatraffic, the traffic is protected from other traffic. Furthermore, since TXOP Limit does not createlong idle time for STAs to access the wireless medium, the bandwidth is efficiently used.

The remainder of this research is organized as follows. The next section introduces relatedworks. In Section 4.3.3, we show our proposed algorithm for Dynamic adaptation of CW sizes.Section 4.3.4 presents an algorithm for AdaptiveTXOP. Performance evaluations of our proposalsare carried in the section 4.3.5. Finally in section 4.3.6, we present the conclusion of this research.

4.3.2 Related Works

Regarding admission control mechanism under EDCA, there were several proposals [26], [27].These proposals monitor available or utilized bandwidth. When a new request arrives at the AP,their proposed admission controllers implemented in the AP check whether the requested band-width is acceptable, and then decide to accept or deny it. Furthermore, a protection mechanism ofreal-time traffic from non real-time traffic was proposed in [26]. It uses two procedures for datatraffic class (AC0). When a transmission of a frame belonging to AC0 fails and the correspondingCW size increases, [26] uses window-increasing factorσi (i = 1, · · · , Lretry, whereLretry is theretry limit), and the CW size quickly becomes larger backoff window size than the original binaryexponential backoff (referred to as fast-backoff). Another procedure is called dynamically adjust-ing parameters. When a transmission of data traffic reaches the retry limit and fails,CWmin[0] and

73

AIFS[0] is increased, as shown in Eq.(4.3) and (4.4), until a limit is reached.

CWmin[0] = θ × CWmin[0](θ > 1) (4.3)

AIFS[0] = ψ × AIFS[0](ψ > 1) (4.4)

When a STA successfully transmitsm consecutive frames,CWmin[0] andAIFS[0] are adjustedas shown in Eq.(4.5) and (4.6) until the low limit is reached.

CWmin[0] = CWmin[0]/θ(θ > 1) (4.5)

AIFS[0] = AIFS[0]/ψ(ψ > 1) (4.6)

These procedures are locally implemented in each STA in [26]. Consequently, higher priority traf-fic can more often access the wireless medium than lower priority traffic. However, even whenthere is little high priority traffic in a WLAN network, low priority traffic takes long time to ac-cess the wireless medium due to fast-backoff and dynamically adjusting parameters, and then idleperiod caused by long backoff time wastes the wireless bandwidth. Therefore, channel access pa-rameters, such as CW size, have to be decided considering network conditions. The AP had bettergive appropriate channel access parameters to STAs because they do not know what kind of trafficis on the fly in their associated WLAN networks.

4.3.3 Dynamic Adaptation of CW Sizes

Small CW sizes set as default values in 802.11e [2] for real-time traffic result in small backofftime. As a result, when there are many real-time flows, collisions between them often occur andthey cannot meet their requirements (e.g. throughput or delay). Hence we propose a method todynamically adjust CW sizes in each AC to reduce collisions between real-time traffic. Dynamicadaption of CW sizes is done for protecting traffic in AC3 and AC2 (i.e. ACs for real-time traffic).Therefore, our proposed algorithm focuses on whether CW sizes in AC3 and AC2 are increased ordecreased. However, for example, when CW sizes in AC3 are updated, changing only them willresult in the collapse of service differentiation defined in IEEE802.11e. That is, in case where CWsizes in AC3 or AC2 are updated, whether those in other ACs are accordingly updated has to beconsidered. For example, when CW sizes in AC3 are updated, those in AC2, AC1, and AC0 mayhave to be updated. In the following, we first show an adaptation policy of CW sizes in all ACswhen CW sizes in AC3 or AC2 are updated and, next, describe our proposed adaptation algorithmin detail. In these subsections, we assume that AP takes admission control presented in previoussection for real-time traffic, i.e. traffic belonging to AC3 and AC2.

Adaptation Policy

Table 4.1 shows default EDCA parameters defined in IEEE802.11e. In Table 4.1,wmin andwmax are specified in each physical setting such as IEEE802.11a [3]. These EDCA parametersare sent from AP to STAs through beacon transmissions. The service differentiation defined inIEEE802.11e is that an AC for real-time traffic can have more opportunities to access the wireless

74

AC CWmin CWmax AIFSN0 wmin wmax 71 wmin wmax 32 wmin+1

2− 1 wmin 2

3 wmin+14

− 1 wmin+12

− 1 2

Table 4.1: Default Parameters

medium earlier than other ACs for non real-time traffic. Our policy completely follows the policyused in IEEE802.11e even when CW sizes are updated. That is, minimum and maximum sizes ofCW in an AC do not become equal to or smaller than those in ACs for real-time traffic. AP runsproposed algorithm presented later to decide whether CW sizes in AC3 and AC2 are increased ordecreased before every beacon transmission. The algorithm, first, considers if CW sizes in AC3 areincreased or decreased. If CW sizes in AC2 are also updated in addition to ones in AC3, whetherCW sizes in AC2 are increased or decreased is not next considered. Otherwise whether those inAC2 are updated is deliberated next. When CW sizes in each AC are increased, bothCWmin

andCWmax in the AC are gained by double. If they are decreased, their values are reduced byhalf. However, they do not become less than default values presented in Table 4.1, and both ofCWmax[0] andCWmax[1] are not increased and also not decreased since default values of them areset towmax. The below shows how to increase or decrease CW sizes in each AC when increasingor decreasing CW sizes in AC3 or AC2.

Adaptation policies for increasing and decreasing CW sizes in AC3 or AC2 are presented inFig.4.4 and 4.5, respectively. In case where CW sizes in AC3 are increased, those in all otherACs may have to be increased to maintain the service differentiation presented in Table 4.1. Butthey might have been increased when CW sizes in AC2 were increased. Therefore, CW sizes inAC3 is compared to ones in AC2. Since bothCWmin andCWmax in an AC are updated whenthey are increased or decreased, all we have to do is just to compareCWmin[3] with CWmin[2]. Ifhalf ofCWmin[2] is larger thanCWmin[3], onlyCWmax[3] andCWmin[3] are increased by double.This is because even ifCWmax[3] andCWmin[3] are increased by double, updatedCWmax[3]andCWmin[3] are not beyondCWmax[2] andCWmin[2], respectively. On the other hand, if halfof CWmin[2] is equal toCWmin[3], thenCWmax[3], CWmin[3], CWmax[2] andCWmin[2] areincreased by double to maintain the service differentiation presented in Table 4.1, and accordinglyCWmin[1] andCWmin[0] are also increased by double. For example, suppose that the current CWsizes in all ACs are those presented in Table 4.1. If CW sizes in AC3 are increased by double,CWmax[3] andCWmin[3] becomewmin and (wmin + 1)/2 − 1, respectively. In this case, sincethese values are the same as ones in AC2, CW sizes in AC2 have to be increased by double tomaintain service differentiations defined in Table 4.1. ThusCWmin[1] andCWmin[0] have to beaccordingly increased by double. As a result,CWmax[2] andCWmin[2] become2× (wmin +1)−1andwmin, respectively, and thenCWmin[1] andCWmin[0] become2 × (wmin + 1) − 1. In casewhere CW sizes in AC3 are decreased, whether CW sizes in other ACs are decreased is alsoconsidered. If there are admitted flows in AC2, onlyCWmin[3] andCWmax[3] are decreased byhalf as shown in Figure 4.5. This is due to the fact that if CW sizes in AC2 are decreased inaddition to those in AC3, the admitted real-time flows in AC2 may be suffered from performance

75

degradation. Therefore, when there are admitted flows in AC2, it is better to separately considerwhether CW sizes in AC2 are reduced. Besides, ifCWmax[2] andCWmin[2] are not reduced inaddition to CW sizes in AC3,CWmin[1] andCWmin[0] cannot be accordingly decreased since theservice differentiation between AC2 and AC1 or AC0 has to be maintained. On the contrary, whenthere is no admitted flow in AC2,CWmax[2], CWmin[2], CWmin[1] andCWmin[0] are decreasedin addition to CW sizes in AC3. This is because if there is no admitted traffic in AC2, no trafficin AC2 is suffered from performance degradation even though CW sizes in the AC is reducedin addition to ones in AC3. For example, suppose that the current CW sizes in all ACs are thefollowings:CWmax[3] = wmin, CWmin[3] = (wmin + 1)/2− 1, CWmax[2] = 2× (wmin +1)− 1,CWmin[2] = wmin, CWmin[1] = 2× (wmin + 1)− 1, andCWmin[0] = 2× (wmin + 1)− 1. Whenthere is no admitted flow in AC2 and CW sizes in AC3 are decreased, CW sizes in all ACs becomethose presented in Table 4.1. If there are admitted flows in AC2,CWmax[3] andCWmin[3] becomewmin+1

2− 1 andwmin+1

4− 1, respectively, and others are not decreased.

When CW sizes in AC2 is not updated in addition to ones in AC3, the proposed algorithm nextconsiders whether they are increased or decreased. If CW sizes in AC2 are increased as shownin Fig.4.4, the service differentiation between AC2 and AC1 or AC0 must be maintained. Hence,when they are increased,CWmin[1] andCWmin[0] are also increased in addition toCWmax[2],andCWmin[2], and then relation presented in Table 4.1 amongCWmin[1], CWmin[0], CWmax[2],andCWmin[2] is maintained even if CW sizes in AC2 are updated. For example, suppose that thecurrent CW sizes in all ACs are those presented in Table 4.1. If CW sizes in AC2 are increasedby double,CWmax[2] andCWmin[2] become2 × (wmin + 1)− 1 andwmin, respectively. Besides,CWmin[1] andCWmin[0] become2 × (wmin + 1) − 1. In case where CW sizes in AC2 are de-creased, relation betweenCWmin[2] andCWmin[3] is considered. If half ofCWmin[2] is larger thanCWmin[3], thenCWmax[2],CWmax[2], CWmin[1] andCWmin[0] are reduced by half. On the otherhand, if half ofCWmin[2] is equal toCWmin[3], CW sizes in any ACs are not decreased since CWsizes in AC2 cannot become equal to or smaller than those in AC3. For example, suppose that thecurrent CW sizes in all ACs are the followings:CWmax[3] = wmin+1

2−1, CWmin[3] = wmin+1

4−1,

CWmax[2] = 2 × (wmin + 1) − 1, CWmin[2] = wmin, CWmin[1] = 2 × (wmin + 1) − 1, andCWmin[0] = 2 × (wmin + 1) − 1. In this case, if CW sizes in AC2 are decreased,CWmax[2],CWmin[2], CWmin[1] andCWmin[0] becomewmin, wmin+1

2− 1, wmin andwmin, respectively.

Algorithm for Dynamic Adaptation of CW sizes

The proposed algorithm is implemented in AP and adaptively controls CW sizes. When CW sizeset in an AC for real-time traffic is a small value and there are a huge number of real-time flows,they can more often access the wireless medium but their transmissions often conflict with others.Hence it is necessary to adaptively control CW sizes, taking into account the number of real-timetraffic. If we assume that a network is saturated, all we have to do is to care about the number ofadmitted real-time traffic in order to control CW sizes. However, in fact, whether a STA is readyto transmit a frame in a given time depends on packet arrival rate from upper layer. Even when thenumber of admitted flows is large, small CW sizes are acceptable for them if their rates are low.Therefore the algorithm considers whether a STA transmitting a real-time flow has frames in itstransmission queue in order to calculate the number of real-time traffic in saturated condition. In

76

(CWmin[2]+1)/2 > (CWmin[3]+1)

CWmin[3] and CWmax[3]are increased by double

CWmin[3],CWmax[3],CWmin[2],CWmax[2],CWmin[1]and CWmin[0]are increased by double

CWmin[2],CWmax[2],CWmin[1]and CWmin[0]are increased by double

YesNo AC2

CW sizes in AC3 are increased CW sizes in AC2 are increased

Figure 4.4: Adaptation policy in case of increasing CW sizes

CWmin[3] and CWmax[3]are decreased by half

CWmin[2],CWmax[2],CWmin[1]and CWmin[0]are decreased by half

admitted flows belonging to AC2exist ?

CWmin[3],CWmax[3],CWmin[2],CWmax[2],CWmin[1]and CWmin[0]are decreased by half

(CWmin[2]+1)/2 > (CWmin[3]+1)

Any CW sizes are not decreased

Yes No

Yes No

CW sizes in AC3 are decreased CW sizes in AC2 are decreased

Figure 4.5: Adaptation policy in case of decreasing CW sizes

77

IEEE802.11e, the AP can know queue size of an admitted real-time flow from QoS control field inMAC header [2]. When a STA transmits a data frame, it sets its queue size in QoS control field inMAC header. The proposed algorithm assumes that the AP records the queue size upon receivingdata frames from admitted flows. Before beacon transmission, the AP counts the number of STAswhose queue size is larger than zero, and runs the proposed algorithm to calculate optimum CWsizes. Decided CW sizes are contained in a beacon frame and then are transmitted.

The algorithm has two phases. Flow charts in 1st and 2nd phases are shown in Fig.4.6 and 4.7,respectively. It assumes in Fig.4.6 and 4.7 that the number of flows whose queue size is larger thanzero isnk for AC2 and AC3 (k = 2 or 3). Before beacon transmissions, the algorithm presented in1st phase is first considered. The first phase deals with whether CW sizes in AC3 are increased ordecreased. When they are increased or decreased, CW sizes in other ACs may be gained or reducedto maintain service differentiation defined in Table 4.1 as presented in the adaptation policy. Onlywhen CW sizes in AC2 are not changed in the first phase, the algorithm presented in the secondphase is next considered.

Here, the algorithm in the first phase is presented. Collision probability is very related tothe number of STAs, which are ready to transmit frames, and CW sizes [28]. Therefore, takinginto account these two parameters, the algorithm makes a decision of whether CW sizes in AC3are updated. Since CW sizes are adapted in order to reduce collisions between real-time traffic,it takes care about the number of STAs which are ready to transmit real-time frames, i.e. thesum ofn3 andn2, as a key to update CW sizes in AC3. Besides it also focuses on the valueof CWmin[3] because the value is definitely used in the initial backoff and if the value is small,collisions between real-time traffic in the AC often occur. Moreover, the average value of backoffcounter ofCWmin[3] is basically treated as the half ofCWmin[3]. Hence the algorithm comparesthe half ofCWmin[3] with the sum ofn3 andn2. If it is smaller than the sum ofn3 andn2, CW sizesin AC3 are increased to reduce collisions between real-time traffic. When they are increased, CWsizes in other ACs may have to be increased in order to maintain the service differentiation definedin Table 4.1. In this case, as explained in the adaptation policy, the half ofCWmin[2] is comparedwith CWmin[3] since CW sizes in AC3 must not be equal to or beyond CW sizes in AC2. If thehalf ofCWmin[2] is larger thanCWmin[3], onlyCWmin[3] andCWmax[3] are increased by double.OtherwiseCWmin[2], CWmax[2], CWmin[1] andCWmin[0] are increased by double in addition toCWmin[3] andCWmax[3]. On the other hand, when comparing the half ofCWmin[3] with the sumof n3 andn2, if the half ofCWmin[3] is larger than the sum ofn3 andn2, CW sizes in AC3 willbe possibly decreased because smaller CW sizes in AC3 may be acceptable for admitted real-timeflows. To determine whether they can be reduced,(CWmin[3] + 1)/(2 ∗ 2) is compared with thesum ofn3 andn2 since the half of the currentCWmin[3] will be the value ofCWmin[3] if CW sizesin AC3 can be decreased. In case where(CWmin[3]+1)/(2∗2) is larger than the sum ofn3 andn2,CW sizes in AC3 can be reduced. When reducing CW sizes in AC3 by half, those in other ACs willbe able to be decreased. As mentioned in the previous subsection, the number of admitted flowsin AC2 is a key to decide it. If no admitted flow exists in AC2,CWmin[2], CWmax[2], CWmin[1]andCWmin[0] are decreased by half in addition toCWmin[3] andCWmax[3]. Otherwise, onlyCWmin[3] andCWmax[3] are decreased by half.

If CW sizes in AC2 are not updated in 1st phase, the algorithm presented in 2nd phase is nextprocessed. As similar to the algorithm presented in the 1st phase, the 2nd phase algorithm con-

78

(CWmin[3]+1)/2 < (n3+n2)

(CWmin[2]+1)/2 > (CWmin[3]+1)

CWmin[3] and CWmax[3]are increasedby double

CWmin[3],CWmax[3],CWmin[2],CWmax[2],CWmin[1]and CWmin[0]are increased by double

Yes

Yes

No

(CWmin[3]+1)/(2*2) < (n3+n2)

n2 = 0

CWmin[3],CWmax[3],CWmin[2],CWmax[2],CWmin[1] and CWmin[0]are decreased by half

CWmin[3] and CWmax[3]are decreasedby half

Any CW sizes are not changed

Yes

No

No

No Yes

Go to the SecondPhase

Go to the SecondPhase

Figure 4.6: First Phase

siders whether CW sizes in AC2 are increased or decreased, taking into account the number ofSTAs which are ready to transmit real-time traffic and the half ofCWmin[2]. Hence, the half ofCWmin[2] is first compared with the sum ofn3 andn2. If it is smaller than the sum ofn3 andn2, CW sizes in AC2 has to be increased to reduce collisions between real-time flows, and as ex-plained in the previous subsection,CWmin[1] andCWmin[0] are accordingly increased to maintainthe service differentiation defined in Table 4.1. Otherwise they will be possibly decreased. Todecide whether they are reduced, the algorithm considers whether the half of the currentCWmin[2]is acceptable for real-time flows because it will be the value ofCWmin[2] if CW sizes in AC2can be decreased. Therefore(CWmin[2] + 1)/(2 ∗ 2) is first compared with the sum ofn3 andn2. If (CWmin[2] + 1)/(2 ∗ 2) is larger than the sum ofn3 andn2, CW sizes in AC2 will bepossibly decreased. And then, as explained in the adaptation policy, relation betweenCWmin[2]andCWmin[3] is next considered. If the half ofCWmin[2] is equal toCWmin[3], CWmin[2] cannotbe decreased because CW sizes in AC2 must not be equal to and smaller than those in AC3. Incase where the half ofCWmin[2] is larger thanCWmin[3], CWmax[2], CWmax[2], CWmin[1] andCWmin[0] are decreased by half.

Since the proposed algorithm does not take into account the number of best effort flows, ifthere are small number of high priority flows and much best-effort traffic, the high priority trafficwill be disturbed from the best-effort traffic. However, as mentioned earlier, since best effort trafficgenerally does not have flows, it is difficult for an AP to estimate the number of best effort trafficin real-time. Instead, we focus on the use of continuation of TXOP to protect high priority trafficfrom low priority traffic. The next section presents an algorithm to determine value of TXOPLimit.

79

Yes

(CWmin[2]+1)/(2*2) < (n3+n2)

(CWmin[2]+1)/2 > (CWmin[3]+1)Any CW sizes

are not changed

Yes

No

No

Yes No

CWmin[2],CWmax[2],CWmin[1]and CWmin[0]are increased by double

CWmin[2],CWmax[2],CWmin[1] and CWmin[0]are decreased by half

Any CW sizes are not changed

(CWmin[2]+1)/2 < (n3+n2)

Figure 4.7: Second Phase

4.3.4 Appropriate TXOP allocation

This section shows an algorithm for selecting an appropriate TXOP Limit to protect high prioritytraffic from low priority traffic. First we show a simple throughput model under EDCA. This modelis based on [28] and [27], and also considers a continuation of TXOP. Then, using the model, weshow our proposed algorithm for selecting a value of TXOP Limit.

Throughput Model

We assume that a STA hasK traffic classes with distinct QoS requirement. Using renewal theory[28], throughput of a flow in queuek (0 ≤ k ≤ K−1) for one transmission cycle can be expressedas:

ρ[k] =Lk

E[Tk](4.7)

Lk is the average length of a payload in a classk and is expressed in a slot time.E[Tk] denotesthe average transmission cycle, which consists of the average length of idle periods resulting frombackoff, some unsuccessful periods, (due to collision and error), and a successful period as shownin Fig.4.8. It is expressed as follows:

E[Tk] = (E[CNk] + 1)E[Idk] + E[CNk](τ +

AIFS[k] + TRTS + TSIFS) + E[Sk] (4.8)

In this equation, we assume that a collision does not happen if RTS and CTS are successfullyexchanged between a sender and receiver. Further,E[Idk] is the average slot time of idle time of

80

unsuccessidle idle

AIFS[k]

unsuccess …… idle success

Transmission Cycle

Figure 4.8: Transmission Cycle

a flow in a classk, E[CNk] denotes the average number of collisions occurred in a classk, TRTS

presents time needed to transmit RTS,τ is propagation delay,TSIFS shows the length of SIFSperiod, andE[Sk] is expressed as shown in Eq.(4.9)

E[Sk] = TRTS + TCTS + TACK + 4TSIFS + 4τ

+Lk + AIFS[k] (4.9)

In Eq.(4.9),TCTS andTACK are transmission time of CTS and ACK, respectively.We consider TXOP Limit into Eq.(4.7) and then present throughput model. When TXOP limit,

m (seconds), is given from AP, the number of frames,C, that the STA can additionally transmit inone TXOP duration is expressed in Eq.(4.10).

�m− E[Sk]

B� = C

B = 2TSIFS + Lk + TACK (4.10)

Since a STA, which is given TXOP, can transmit an additional frame SIFS period after successfultransmission of the first frame as presented in Fig.4.3, transmission time of an additional frame isexpressed asB. If C is larger than 1, the STA can transmit at least one frame SIFS period afterreceiving the first ACK.

Considering TXOP Limit into Eq.(4.7) and (4.9), they are further re-expressed as the followingequations,

ρ[k] =Lk(1 + C)

E[Tk](4.11)

E[Sk] = TRTS + TCTS + TACK + 4TSIFS

+4τ + Lk(1 + C) +AIFS[k]

+2C(TSIFS + TACK + τ ) (4.12)

When a flow in a classk in a STA has the right to access the wireless medium, it can continu-ously transmitC frames. Other STAs wait for channel access during time set by NAV (networkallocation vector). Note that in Eq.(4.12) we assume that transmissions of additional frames arenot disturbed by transmissions from other STAs. If AP gives appropriate TXOP Limit to admit-ted real-time flows, they can be protected from other best-effort traffic and maintain their desired

81

throughput Furthermore, STAs using continuation of TXOP have short idle time before additionalframe transmissions since they do not have to start backoff. The TXOP Limit is given from APwith beacon frames. In next subsection, we explain our proposed algorithm to select appropriatevalues of TXOP Limit.

Algorithm for appropriate TXOP allocation

This section presents an algorithm for selecting TXOP Limit. It runs on the AP. When admissioncontrol is needed for an ACk and a STA has flowi in the AC k, the STA sends to the AP anADDTS request, which contains mean data rate,MRi,k, and nominal MSDU size,Si,k. After theAP decides whether the request can be accepted, the AP records these traffic characteristics into alist, called TSPEC List. The TSPEC List contains several entries. For example, an entry for flowiconsists of mean data rate,MRi,k, and nominal MSDU size,Si,k, estimated throughput of flow i,αi,k. The AP measures the throughput of flowi andαi,k is calculated as averaged throughput usingmoving average with window sizew. In order to calculate throughput, other radio performancesare also estimated by using throughput model presented in the previous subsection. Our proposedalgorithm monitors each flow and decides a value of TXOP Limit to maintain desired throughput,as shown in Fig.4.9. Every time before beacon transmissions, TXOP Limitmk is calculated. Sinceit should be allocated to real-time traffic, calculation of TXOP Limit is performed for an AC whichneeds admission control (i.e. AC for real-time traffic). To decide TXOP Limitmk in AC k, first,mk is set to a default value defined in [2]. Then, for all admitted flowsi (0 ≤ i < hk) in ACk, TXOP Limit mi,k is calculated by using Eq.(4.10), (4.11) and (4.12) (hk shows the number ofadmitted flows in ACk). In Eq.(4.11), mean data rate of each admitted flow,MRi,k, is used forρ[k] andSi,k is used forLk. The AP selects the largest value ofmi,k as a value of TXOP Limitin AC k. If the largest value ofmi,k is smaller than the default value of TXOP Limit in ACk, thedefault value is selected as TXOP Limit in the AC. After calculating values of TXOP Limit in allACs, the AP delivers a beacon frame containing calculated TXOP Limit. Since use of appropriateTXOP Limit can allocate adequate bandwidth to admitted real-time flows, it protects them fromother real-time traffic or best effort traffic.

4.3.5 Performance Evaluation

We implemented our proposed algorithms in NS-2 [18] and evaluated the performance of mul-timedia flows with different channel loads. First, we evaluate our proposed algorithm (referredto as CWAdap) for dynamic adaptation of CW sizes comparing with 802.11e EDCA (referredto as EDCA). Next, the CWAdap using AdaptiveTXOP is compared to EDCA. All of them useadmission control presented in section 4.2.2 [2]. Throughput and delay are used as performancemeasures. We assume that AP and STAs operate with IEEE 802.11a [3]. Physical data rate andbasic rate are set to 12 Mbps and 6 Mbps, respectively. Channel error is not considered becausewe want to evaluate link-level performance. The beacon interval is set to 100 ms. As defaultvalues of TXOP Limit for each AC, we used values of TXOP Limit specified in [2]. Transmis-sion queue size for each AC is set to 50. For each AC, we have the following parameters [2]:CWmax[0] = 1023, CWmin[0] = 15; CWmax[1] = 1023, CWmin[1] = 15; CWmax[2] = 15,

82

mk is set to default value

AC k needs admission control?

mi,k is calculated based on Eq. (10),

(11), and (12)

mi,k > mk

mk = mi,k

i += 1

i < hk

i = 0

k+= 1

k < 4

k = 0

End of Processall mk are delivered in a beacon frame

Yes

No

Yes

No

Yes

Yes

No

No

Every time before beacon transmissions

Figure 4.9: Algorithm for TXOP Limit selection

CWmin[2] = 7; CWmax[3] = 7, CWmin[3] = 3; AIFSN [0] = 7, AIFSN [1] = 3, AIFSN [2] =2, AIFSN [3] = 2. In order to calculate averaged radio performances using moving average, thewindow sizew is set to 30. Each voice flow is 83.2 Kbps, which is generated by a constant inter-val, 20 ms and has a fixed payload size of 208 bytes. This flow corresponds to G.711-coded VoIP[26]. Each video flow is 256 Kbps, which is generated by a constant interval, 20 ms and has afixed payload size of 640 bytes. As background traffic, UDP traffic is generated by an exponentialinterval with mean inter-arrival time, 12.5 ms and has a fixed payload size of 1024 bytes.

In each simulation run, simulation runs for 60 seconds and all flows start their transmissionsduring the first 30 seconds. In order to calculate throughput and delay, we use simulation outputsin the interval from 30 to 60 seconds in each simulation run. Even though an admission controlleris implemented in the AP, it does not deny all admission requests because we do not intend toevaluate admission control mechanism in this research.

Performance evaluation of CWAdap

We first performed simulations using only voice flows (Scenario1). A STA has one voice flow.Simulation results show relations between throughput or delay and the number of voice flows, aspresented in Fig.4.10 and 4.11 respectively. In EDCA, throughput decreases and delay increaseswhen the AP has many voice flows. Collisions among voice flows often happen due to short back-off time generated by small CW sizes. As a result, the AP cannot accommodate many voice flowseven if wireless bandwidth is not fully utilized. On the other hand, in CWAdap, the AP can control

83

20

30

40

50

60

70

80

90

15 20 25 30 35 40

Th

rou

gh

pu

t (K

bp

s)

The number of voice flows

EDCACWAdap

Figure 4.10: Comparison of throughput for voice flows in Scenario1

CW sizes considering whether STAs have frames in their transmission queue. If the current CWsizes are inadequate, the AP updates CW sizes and STAs with voice flows can use longer back-off time. Consequently, CWAdap can reduce collisions among voice flows and can accommodatemore voice flows than EDCA. Next, we performed simulations using voice and video flows (Sce-nario2). Each STA has one voice and one video flows in Scenario2. In order to increase simulationtraffic, we add new STA in the simulation network. Therefore, if we add a STA in the simulationnetwork, both a voice and a video flows are added. As similar to the Scenario1, relations betweenthroughput and delay and the number of voice and video flows are evaluated. Figure 4.12 and 4.13show comparisons of throughput for voice and video flows, respectively. When small number ofvoice and video flows associate with the AP, default CW sizes are acceptable both in EDCA andCWAdap. However, if large number of voice and video flows are transmitted, collisions amongthem often occur due to short backoff time generated by small CW sizes and throughput is de-graded in EDCA. On the contrary, since in CWAdap CW sizes are updated to reduce collisions,both throughput of voice and video flows are maintained. Comparisons of delay for voice andvideo flows are shown in Fig.4.14 and 4.15. As similar to the simulation results of throughput, de-lay for both voice and video flows in EDCA increases when the AP accommodates large numberof flows. On the other hand, low delay for both voice and video flows is maintained in CWAdapsince CWAdap can reduce collisions between those flows.

Performance evaluation of AdaptiveTXOP

Finally we performed simulation with voice, video and data traffic (Scenario3) for three schemes,EDCA, CWAdap and CWAdap+AdaptiveTXOP (referred to just as AdaptiveTXOP). We evaluateeach scheme through varying the number of data traffic. In the simulation network, 10 STAs haveboth voice and video flows and other STAs have only data traffic. In order to vary the number of

84

0

0.5

1

1.5

2

2.5

15 20 25 30 35 40

De

lay (

s)

The number of video flows

EDCACWAdap

Figure 4.11: Comparison of delay for voice flows in Scenairo1

30

35

40

45

50

55

60

65

70

75

80

85

2 4 6 8 10 12 14

Th

rou

gh

pu

t (K

bp

s)

The number of voice flows

EDCACWAdap

Figure 4.12: Comparison of throughput for voice flows in Scenario2

85

140

160

180

200

220

240

260

2 4 6 8 10 12 14

Th

rou

gh

pu

t (K

bp

s)

The number of video flows

EDCACWAdap

Figure 4.13: Comparison of throughput for video flows in Scenario2

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

2 4 6 8 10 12 14

De

lay (

s)

The number of voice flows

EDCACWAdap

Figure 4.14: Comparison of delay for voice flows in Scenario2

86

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

2 4 6 8 10 12 14

De

lay (

s)

The number of video flows

EDCACWAdap

Figure 4.15: Comparison of delay for video flows in Scenario2

data traffic, the number of STAs having data traffic is varied.Figure 4.16, 4.17 and 4.18 present comparisons of throughput for voice, video and data traffic,

respectively. Figure 4.19, 4.20 and 4.21 show comparisons of delay of voice, video and data traffic,respectively. It is clear from Fig.4.16 and 4.17 that throughput of both voice and video flows cannotbe maintained in EDCA when the number of data traffic increases. Contentions between real-timeand data traffic decrease throughput of real-time traffic. Furthermore, as shown in Fig.4.19 and4.20, delay of real-time traffic becomes large in EDCA when data traffic increases.

Compared to EDCA, throughput of real-time traffic in CWAdap is not much degraded. Whenthe number of data traffic is large, real-time data in STA’s transmission queue increases due toincreased contentions between real-time and non real-time traffic. On the receipt of data framewhose QoS control field indicates buffered frame exists in the sender’s transmission buffer, theAP will increase CW sizes to reduce contentions between real-time flows. However, as previouslymentioned, since CWAdap cannot reduce contentions between real-time and best effort traffic,throughput of real-time flows declines when there are large number of data traffic. This is becauseall of traffic contends for channel access in each transmission even when CW sizes are increased.As shown in Fig.4.19 and 4.20, delay of real-time flows also increases in case where large numberof data traffic exists.

On the other hand, we can find that AdaptiveTXOP can maintain throughput and delay ofreal-time traffic. Since in AdaptiveTXOP, real-time flows are given an appropriate value of TXOPLimit from AP, they maintain a certain level of desired throughput. When network loads arehigh and contentions often occur, the value of TXOP Limit is updated every beacon interval toachieve desired throughput of real-time flows. Consequently, throughput of real-time flows inAdaptiveTXOP is better than CWAdap and EDCA. From Fig.4.19 and 4.20, it is obvious that delayof real-time flows in AdaptiveTXOP is lower than other schemes. If a STA using AdaptiveTXOPhas the right to transmit a frame, it can continuously transmit several frames. Therefore delay

87

20

30

40

50

60

70

80

90

4 5 6 7 8 9 10 11 12

Th

rou

gh

pu

t (K

bp

s)

The number of flows in AC0

EDCACWAdap

AdaptiveTXOP

Figure 4.16: Comparison of throughput for voice flows in Scenario3

of real-time flows is also maintained. From these simulation results, we find that AdaptiveTXOPprotects real-time traffic from data traffic and maintain throughput and delay of real-time traffic.

From simulation results for data traffic in Fig.4.18 and 4.21, we find that throughput and delayof data traffic in CWAdap is higher than AdaptiveTXOP. However, CWAdap cannot fully protectthroughput and delay of real-time traffic from data traffic. In other words, channel resource is notadequately allocated to real-time traffic but to data traffic. On the other hand, in AdaptiveTXOP,real-time traffic is given appropriate TXOP and channel access of data traffic is suppressed. There-fore, throughput of real-time traffic is maintained although throughput of data traffic is lower thanCWAdap.

4.3.6 Conclusion of Real-time Traffic under IEEE802.11e EDCA via EDCAParameters

We discussed throughput degradation problem under IEEE 802.11e EDCA. The problem comesfrom collisions between real-time flows and between real-time and non real-time flows. We there-fore proposed two algorithms to overcome the problem. First, we focused on reducing collisionsbetween real-time flows. In IEEE 802.11e EDCA [2], CW sizes set as default values are notpreferable for an AP to accommodate many real-time flows. Therefore we proposed a method todynamically control CW sizes to enhance the performance of throughput and delay of real-timetraffic. Simulation results showed it could increase the number of real-time flows accommodatedby the AP. Since the simulations did not consider channel error, we will do simulations consideringit as one of our future works. If channel error is taken into account, throughput in EDCA will bemore significantly degraded than CWAdap because transmission cycle specified in Eq.(4.8) willbecome large and the number of STAs, which are ready to transmit frames in a give time, will be

88

120

140

160

180

200

220

240

260

4 5 6 7 8 9 10 11 12

Th

rou

gh

pu

t (K

bp

s)

The number of flows in AC0

EDCACWAdap

AdaptiveTXOP

Figure 4.17: Comparison of throughput for video flows in Scenario3

200

220

240

260

280

300

320

340

360

380

400

4 5 6 7 8 9 10 11 12

Th

rou

gh

pu

t (K

bp

s)

The number of flows in AC0

EDCACWAdap

AdaptiveTXOP

Figure 4.18: Comparison of throughput for data flows in Scenario3

89

0

0.5

1

1.5

2

2.5

4 5 6 7 8 9 10 11 12

De

lay (

s)

The number of flows in AC0

EDCACWAdap

AdaptiveTXOP

Figure 4.19: Comparison of delay for voice flows in Scenario3

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

4 5 6 7 8 9 10 11 12

De

lay (

s)

The number of flows in AC0

EDCACWAdap

AdaptiveTXOP

Figure 4.20: Comparison of delay for video flows in Scenario3

90

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

4 5 6 7 8 9 10 11 12

De

lay (

s)

The number of flows in AC0

EDCACWAdap

AdaptiveTXOP

Figure 4.21: Comparison of delay for data flows in Scenario3

increased. We will confirm this phenomenon for one of our future works.Next, we focused on the use of continuous of TXOP to reduce collisions between real-time

and non real-time flows, and proposed a method for calculating an appropriate value of TXOPLimit, monitoring performances of admitted flows. Through computer simulations we found thatour proposed AdaptiveTXOP could protect real-time flows from non real-time flows.

In the simulations, admission control mechanisms are not evaluated and admission controlmethod presented in [2] is just used. However, they influence performance measures such asthroughput and delay. In our future works, we plan to tackle it.

91

4.4 Quick Data-retrieving for U-APSD in IEEE802.11e WLANNetworks

4.4.1 Introduction

Power consumption is a critical issue when WLAN is used in handheld devices. In IEEE802.11[1], PSM (Power Saving Mechanism), which is called LegacyPS, is supported. The general ideaof LegacyPS is for STAs to keep their radio off when they do not have to send or receive frames.For details, see [1]. But LegacyPS is not suitable for real-time applications with periodicity suchas VoIP because STAs with LegacyPS need to be awake for more than an actual active voice call.Therefore, in IEEE802.11e, a new PSM, which is referred to as U-APSD (Unscheduled AutomaticPower Save Delivery), is proposed [2], [29]. U-APSD works under EDCA. In U-APSD, an APbegins transmissions of downlink frames to a STA operating with U-APSD when it receives anuplink data frame from the STA. In the next section, we explain U-APSD in detail. U-APSD issuitable for real-time applications like VoIP, and especially for bi-directional traffic generated atconstant bit rate. However, if U-APSD is used for real-time traffic with silence period, e.g. ON-OFF traffic [30], downlink frames are caused large buffering delay at an AP. This is because aSTA may have no uplink frame and then an AP cannot start downlink transmissions. Actually, inU-APSD the STA is allowed to send a null-data frame1 in order to start transmissions of downlinkframes buffered at the AP if it has no uplink data frame [2]. However, how often or when the STAsends a null-data frame is not specified in IEEE802.11e [2].

We therefore propose two mechanisms to alleviate buffering delay generated when U-APSDis used for ON-OFF real-time traffic. The first one uses periodic transmissions of uplink null-dataframes to initiate downlink transmissions during OFF period of an uplink flow and it can be im-plemented within the range of IEEE802.11e [2]. Actually, in U-APSD, transmissions of periodicuplink frames are needed to maintain small delay and desired throughput of real-time traffic. How-ever, periodic transmissions of uplink null-data frames become overhead and cause a waste of thewireless bandwidth. When there is much real-time traffic operating with U-APSD, collisions oftenoccur between uplink null-data frames or between an uplink null-data frame and other traffic. Toovercome these problems, we propose the second mechanism. It can allow the AP to periodicallystart downlink transmissions without the reception of uplink null-data frames. Although it needsa little modifications in U-APSD defined in IEEE802.11e [2], reduction of overhead caused byuplink null-data frames results in that it can accommodate more traffic than the first proposal.

The remainder of this research is organized as follows. The next section introduces U-APSD. InSection 4.4.3, our two proposals are presented. Performance evaluations of them are carried in thesection 4.4.4. comparing with other schemes. Finally in section 4.4.5, we present the conclusion.

4.4.2 IEEE802.11e U-APSD

In general, when an AP delivers downlink frames to STAs operating with power saving mode in-cluding U-APSD, it has to confirm that they are awake. If they are awake, it can transmit downlink

1Note that a null-data frame is a data frame whose frame body is null.

92

frames to them. Otherwise it has to buffer those frames until they are awake. U-APSD is definedin IEEE802.11e [2], [29] as a rule of downlink frame delivery and is used under EDCA. On theother hand, uplink frames can be transmitted based on EDCA from STAs at any time because theAP is always awake.

The basic idea of U-APSD is to use a specific interval called U-SP (Unscheduled ServicePeriod) for an AP to deliver downlink frames to a STA. Figure 4.22 shows operations of U-APSD.Since during an U-SP the AP knows that the STA is awake and is ready to receive downlinkframes, it can deliver downlink frames to the STA. In other words, the AP cannot transmit anydownlink frames to the STA except for an U-SP. The STA using U-APSD can start an U-SP witha transmission of an uplink data or null-data frame. An uplink data or null-data frame used toinitiate an U-SP is called a trigger frame. After the AP responds with ACK, an U-SP starts and theAP begins to deliver downlink frames destined for the STA under EDCA. The length of U-SP isdecided by MAX SP (Service Period) length, which indicates the maximum number of frames theAP can transmit during an U-SP, and it is determined through (re) association procedure betweenthe AP and STA. During an U-SP, the AP has to deliver at least one downlink frame to the STA,but no more than the number specified in MAX SP length. Even though the AP receives uplinkframes from the STA after an U-SP starts, they are not treated as trigger frames. If the AP receivesa trigger frame from the STA and has no buffered frame destined for the STA, it transmits a null-data frame to the STA because the AP has to send at least one frame to it. At the end of an U-SP,the AP informs the STA about the end of the U-SP using EOSP (end of service period) field setto 1 in QoS control field in a data frame as shown in Fig.4.22. On receiving a data or null-dataframes with EOSP set to 1, the STA can enter into the doze mode and the AP cannot deliver furtherdownlink frames until a new U-SP starts. In addition to the use of EOSP field, if there are morebuffered data destined for the STA at the end of an U-SP, the AP sets MoreData field in framecontrol field in MAC header to 1 and informs the STA of that more frames are buffered as shownin Fig.4.23. On the receipt of a downlink frame containing EOSP set to 1 and MoreData set to 1,the STA has to send uplink data frames to quickly retrieve frames buffered at the AP. However, ifthere is no uplink data frame buffered at the STA, it can send a null-data frame to start a new U-SP.

As mentioned earlier, a STA operating with power saving mode can transmit an uplink frame atany time. Therefore, it wakes up before starting a transmission of the uplink frame and it can enterinto the doze mode after finishing the transmission. But, in U-APSD, the transmitted uplink framebecomes a trigger frame, and so after finishing the transmission of the uplink frame, the STA hasto wait for a downlink data or null-data frame and continue to be awake until the downlink framewith EOSP set to 1 is delivered from the AP. This is because the AP recognizes the STA is awakeafter the AP receives a trigger frame and an U-SP ends after the AP transmits to the STA a data ornull-data frame with EOSP set to 1.

Actually, when a STA uses U-APSD for bi-directional traffic generated at constant bit rate,it can efficiently achieve power saving because it can periodically wake up and an U-SP can beconstantly initiated by uplink data frames. Moreover it can retrieve downlink frames from the APduring the U-SP initiated by periodically transmitted uplink data frames. Thus, the STA is awakeonly for an actual active voice call and does not have to be awake for a long period [29]. However,since U-APSD does not carefully consider the property of ON-OFF real-time traffic [30], large

93

UL data or

null-data

ACK

U-SP

STA wakes upSTA goes to doze

STA

AP

ACK

DL data or

null-data

ACK

the number of DL frames is no more than MAX SP length

Trigger frame

EOSP=1

DL: Down LinkUL: Uplink Link

DL data or

null-data

Figure 4.22: U-APSD operations

UL data or

null-data

ACK

U-SP

STA wakes up

STA goes to doze

STA

AP DL dataor

null-data

ACK

Trigger frame

EOSP=1,MoreData=1

ACK

ACK

Trigger frame

EOSP=1,MoreData=0

U-SP

UL data or

null-data

DL dataor

null-data

Figure 4.23: STA operations on the receipt of EOSP=1 and MoreData=1

94

ULdata

DLdataACK

ACK

U-SP

STA wakes upSTA knows the end of U-SP and goes to doze mode

DLdata

Many DL frames are buffered at AP

STA

AP

ACK

OFF Period of UL flow

ON Period of UL flow

Figure 4.24: Large buffering delay in the use of ON-OFF traffic

buffering delay at an AP will be generated when U-APSD is used for ON-OFF traffic as describedin Fig.4.24. This is due to the fact that in U-APSD an U-SP and transmissions of downlink framescannot start until the AP receives an uplink trigger frame. During the OFF period of an uplink flow,the STA may not have any uplink frames and cannot start an U-SP. On the other hand, downlinkdata frames arriving at the AP are buffered and cannot be delivered to the STA. Thus if the OFFperiod is long, large buffering delay with regard to downlink frames is generated at the AP andfurthermore buffered frames may be discarded when its buffer overflows or when they have beenbuffered over pre-defined period.

4.4.3 Periodic U-SP for quick data retrieving

In order to reduce buffering delay generated at an AP, a STA operating with U-APSD has to pe-riodically start an U-SP. However, even though an U-SP is frequently initiated, there may be noframes buffered at the AP. Thus, energy will be wasted in the STA. On the other hand, if an U-SP israrely initiated, buffering delay increases even though energy is saved. In fact, the relation betweensaving energy and small delay is the trade-off.

In the followings, there are two proposals to reduce buffering delay, considering the trade-off.They are focusing on how to and when to periodically start an U-SP in case where U-APSD isused for ON-OFF real-time traffic. And both of them operate based on U-APSD. While the firstone can be achieved within the area defined in IEEE802.11e [2], the second one needs a little ofmodifications with regard to initiation of U-SP but can reduce overhead generated in the first one.Hence it can efficiently utilize the wireless bandwidth.

95

U-APSD with Periodic Transmissions of Trigger Frame

When a STA begins an U-SP to retrieve buffered frames, it needs to transmit a trigger frame (uplinkdata or null-data frame) as specified in U-APSD [2]. If there is no uplink frame, it has to send anull-data frame as a trigger frame in order to initiate an U-SP. Therefore the basic idea behindthe first proposal, which is called UPTT (U-APSD with Periodic Transmissions of Trigger frame),is that a STA maintains a timer, which is called triggering timer, in order to periodically initiatetransmissions of a null-data frame as a trigger frame and that if the STA has no uplink frame atthe expiration of the triggering timer, it transmits a null-data frame to begin an U-SP. With the useof null-data frames, UPTT is fully compatible with U-APSD defined in [2]. Since the trade-offbetween saving energy and small delay has to be balanced, an interval (called triggering interval)between consecutive triggering timers has to be carefully chosen. However, it is very difficult forSTAs to predict arrivals of downlink packets at an AP and so it is very difficult to completelyoptimize the triggering interval.

In order to balance the trade-off, the triggering interval can be varied in UPTT. The basicstrategy to decide a triggering interval is that when there are buffered frames at the AP, the STA setsthe triggering interval to minimum one and frequently sends null-data frames to quickly retrievethem, and that when there is no buffered frame at the AP, the STA sets the triggering interval tolarger one in order to continue the doze mode for a long time and to save energy. Besides when theSTA consecutively notices that the AP has no downlink frame, it increases the triggering interval.When increasing the triggering interval, it is incremented in binary exponential fashion like a back-off counter. IfTtr denotes a triggering interval, it is incremented like the following:T tr = 2× Ttr.Hence, if there is no buffered frame at the AP, the triggering interval fast becomes a large value,and so the STA can continue to sleep for a long time. Figure 4.25 and 4.26 show examples offrame exchange sequence and triggering timer operations in UPTT. The triggering interval,T tr,can be varied from minimum to maximum value. In the bellow, detailed operations of UPTT aredescribed.

If a STA decides that U-APSD is used for a flow, a triggering timer is set to maximum intervalwhen WLAN device starts to work. As shown in Fig.4.25, at the expiration of a trigger timer,the STA becomes awake and a null-data frame is transmitted. However, if it has buffered uplinkdata frames, it does not send a null-data frame since those buffered data frames can be triggerframes. Both when a null-data frame is transmitted and not transmitted, the triggering interval isincremented by double but is not beyond the maximum value, and the STA sets the triggering timerand waits for downlink frames. The reason why the triggering timer is set just after the expirationof the timer is the following. If it is set after interpreting EOSP and MoreData fields in receiveddownlink frames, it cannot be set when transmissions of downlink frames are failed. Therefore,in UPTT, the triggering timer is just incremented by double and set just after the expiration of thetimer, and it is, again, reset after interpreting EOSP and MoreData fields in downlink frames. Howto reset the triggering interval and timer when interpreting these fields is shown in the following.

On the receipt of downlink frames from the AP, the STA checks EOSP and MoreData field inMAC header. Since EOSP set to 0 means that U-SP does not end and further downlink frames comefrom the AP, the STA continues to be awake and waits for downlink frames. In case of EOSP=1and MoreData=1, the STA transmits a trigger frame to the QAP to quickly retrieve buffered frames.

96

ULnull-data

ACK

U-SP

STA wakes upSTA continues to be awake

STA

AP

DLnull-data

ACK

Trigger frame

EOSP=1,MoreData=0

ACK DL data

ACK

Trigger frame

EOSP=1,MoreData=1

U-SP

ULnull-data

STA wakes up

Timer expirationIf Ttr is less than the maximum value, then Ttr = 2×TtrTriggering Timer is set

Timer expirationIf Ttr is less than the maximum value, then Ttr = 2×TtrTriggering Timer is set

Data frame is received even though EOSP=1 and MoreData =0.Triggering Timer is canceled and is set with min. interval

TriggeringTimer

STA goes to doze

ULnull-data

Trigger frame

ACKDL data

ACK

EOSP=1,MoreData=0

U-SP

Triggering Timer is canceled. Triggering interval is set to min. interval.Triggering Timer is set

STA goes to doze

Figure 4.25: UPTT operations

Before transmitting the trigger frame, the STA confirms whether there are buffered uplink frames.If there is no buffered data frame, the STA sends a null-data frame as a trigger frame. Otherwise abuffered data frame is transmitted as a trigger frame. In the example shown in Fig.4.25, a null-dataframe is sent as a trigger frame. And then the STA clears a pending timer and resets the timer withthe triggering interval set to minimum one. This is because it is better to retrieve frames bufferedat the AP as soon as possible.

In case of EOSP=1 and MoreData=0, the STA can enter into the doze mode if there is nouplink frame. However, if the received frame is a data frame, the STA sets the triggering intervalto minimum one and resets the triggering timer as shown in Fig.4.25. This is due to the fact thatdata frames of a real-time flows periodically arrive at the AP. On the other hand, if the receivedframe is a null-data frame (i.e., even though the AP receives a trigger frame and an U-SP starts,the AP has no buffered data destined for the corresponding STA. In this case, the AP sends a null-data frame with EOSP=1 and MoreData=0 to end the U-SP.), the STA can go to doze mode andcontinue to sleep until a triggering timer expires or it has uplink a data frame.

As mentioned earlier, STAs can begin uplink data transmissions at any time in U-APSD. There-fore, when an uplink data frame comes into the transmission buffer at a STA and can be transmittedas a trigger frame, a new U-SP is initiated by the uplink data frame. In this case, the triggeringtimer is accordingly canceled and just reset toTtr, as presented in Fig.4.26. Furthermore,Ttr is notupdated and is the same as the value which was used by the canceled triggering timer because thetriggering interval should be set based on whether the AP has buffered downlink frames.

Since, in UPTT, a STA maintains a triggering timer to periodically transmit a null-data frameas a trigger frame, it can quickly retrieve data frames buffered at an AP even though U-APSD is

97

ULnull-data

ACK

U-SP

STA wakes up

STA goes to doze

STA

AP

DLnull-data

ACK

Trigger frame

EOSP=1,MoreData=0

ACKDL null-data

ACK

Trigger frame

EOSP=1,MoreData=0

U-SP

ULdata

UL data is ready to be sent. STA wakes up

Timer expiration.If Ttr is less than the maximum value, then Ttr = 2×TtrTriggering Timer is set

Although Triggering timer is pending, UL data is ready to be sent as a trigger frame.Triggering timer is canceled and is reset to Ttr

TriggeringTimer

STA goes to doze

Figure 4.26: Timer cancellation in UPTT when sending an uplink data frame

used for ON-OFF traffic. Furthermore, since the triggering interval is increased in binary expo-nential manner when the STA consecutively notices there is no frame buffered at the AP, energy isefficiently saved. In fact, UPTT adopts the way to increase the triggering interval in binary expo-nential fashion in order to balance the trade-off between saving energy and small delay. However,if the STA operating with UPTT can have a knowledge about behavior of a specific real-time appli-cation, the triggering interval can be increased following the specific behavior. In this case, moreenergy will be able to be saved. Since in this research we do not assume any specific real-timeapplications, we take the binary exponential triggering interval to balance the trade-off. Besidesabove things, in UPTT, a null-data frame is transmitted only when a triggering timer expires andthe STA has no uplink frame, and thus the STA does not have to transmit a lot of null-data frames.However, a null-data and the corresponding ACK frames consume the wireless bandwidth eventhough UPTT is fully compatible with U-APSD defined in IEEE802.11e [2] and periodic trans-missions of null-data frames are necessary in the U-APSD. Moreover, since null-data frames maycontend with other transmissions of a data or management frame and may be suffered from re-transmissions due to channel error, overhead generated by transmissions of null-data frames willbe increased.

U-APSD with Timer-based Wake-up Rule

In UPTT, a STA has to send a null-data frame as a trigger frame to initiate an U-SP and an APknows that the STA becomes awake on the receipt of the trigger frame. In other words, in UPTT,the STA notifies the AP by using a null-data frame that it becomes awake. Compared to UPTT, in

98

the second proposal which is referred to as UTWR (U-APSD with Timer-based Wake-up Rule), aSTA does not have to send a null-data frame to start an U-SP and an AP can recognize that the STAbecomes awake without receiving the null-frame. Hence UTWR can remove overhead generateddue to uplink null-data frames.

As similar to UPTT, UTWR operates based on U-APSD. Moreover a triggering timer is alsoused in UTWR in order to reduce buffering delay and also for energy-saving the triggering intervalis increased in binary exponential manner when a STA operating with UTWR consecutively no-tices that there is no buffered frame at the AP. However, in UTWR, both the STA and AP maintaintriggering timers in order to quit uplink transmissions of null-data frames. Therefore, at the expi-ration of triggering timers, the STA becomes awake and the AP can know that it is awake. Thatis, UTWR enables the STA and AP to periodically start an U-SP without transmissions of uplinknull-data frames and they use the same rule to determine a triggering interval. In addition to that anU-SP is initiated at the expiration of the triggering timer, as similar to U-APSD and UPTT, it alsostarts when the AP receives uplink data frames as trigger frames. Besides, as similar to U-APSDand UPTT, STAs operating with UTWR can send uplink data frames at any time. In the following,UTWR operations both in STA and AP are explained in detail.

Figure 4.27 and 4.28 show examples of frame exchange sequence and triggering timer oper-ations. In UTWR, the AP maintains a triggering timer for each flow operating with UTWR andthe STA also maintains a triggering timer for its flow with UTWR. The triggering timers set tomaximum interval start to work at a time pre-determined by both the AP and STA through the(re) association procedure. And also they are set when the AP sends and the STA receives a dataframe or null-data frame with EOSP=1 and MoreData=0 as shown in Fig.4.27. In case where theAP sends a data frame with EOSP=1 and MoreData=0, the triggering interval at the AP is set tominimum one and it sets its triggering timer after sending the data frame because downlink real-time data periodically arrives at the AP. On the receipt of the data frame, the STA accordinglysets its triggering interval to minimum one and sets its triggering timer. On the other hand, incase of null-data frame with EOSP=1 and MoreData=0, triggering intervals both in the AP and theSTA are incremented by double (in binary exponential manner) and the triggering timers are setas presented in Fig.4.27. In fact, there is a lag between time when the AP sends the data frameand time when the STA completes the transmissions of the ACK frame Therefore, the triggeringtimer in the AP should be set considering the lag. The lag can be obtained taking into accountthe transmission duration of data or null-data frame, SIFS period and the transmission period ofACK frame. Furthermore, when the STA receives a data or null-data frame with EOSP=1 andMoreData=0, it may fail to transmit an ACK frame to the AP and go to the doze mode. In thiscase, the STA cannot receive the frame retransmitted by the AP. This phenomenon also occurs inU-APSD [2] and UPTT. However, in UTWR, triggering timers can be successfully set both at theAP and the STA since the AP considers the lag and sets its triggering timer. Of course, if the frameis retransmitted by the AP, its triggering timer is accordingly reset.

In case where the AP sends a data frame with EOSP=1 and MoreData=1, it immediately startsa new U-SP in order to quickly deliver buffered frames to the corresponding STA as presented inFig.4.27. And the STA also knows the new U-SP begins on the receipt of data frame with EOSP=1and MoreData=1, and continues to be awake. Since in UTWR a new U-SP can be initiated withoutuplink data or null-data frames, transmissions of those frames needed in U-APSD and UPTT is

99

eliminated.On the expiration of the trigger timer, the STA wakes up and also the AP knows that the

corresponding STA wakes up, and if the AP has no buffered frame destined for the correspondingSTA, it sends a null-data frame with EOSP=1 and MoreData=0 to the STA and both the AP andSTA increment the triggering interval by double as shown in Fig.4.27. Sending a null-data frameto the STA is for ending an U-SP. This operation is the same as one used when in U-APSD andUPTT the AP receives a trigger frame and there is no buffered frame destined for the STA. On theother hand, if there are buffered frames at the AP, data delivery operations are normally processedand there is no need to send a null-data frame.

As mentioned earlier, STAs can transmit uplink frames at any time. Therefore if the AP re-ceives an uplink frame from the STA as a trigger frame, a new U-SP is initiated and thus the timeris canceled as shown in Fig.4.28. When the U-SP is started by the uplink frame and the AP hasno buffered frame destined for the STA, the AP sends and the STA receives a downlink null-dataframe with EOSP=1 and MoreData=0. In this case, they do not increment their triggering inter-vals by double but they just reset their triggering timer using the current value ofT tr as shown inFig.4.28. Resetting the current value ofTtr is for avoiding a situation where triggering intervalsare fast increased to a larger value when there are many uplink transmissions from the STA.

UTWR removes overhead of uplink transmissions of null-data frames. Hence the wirelessbandwidth is more efficiently used in UTWR than in UPTT. Even though the triggering intervalin UTWR is incremented in binary exponential manner, as explained in the previous section, thetriggering interval in UTWR can follow specific behaviors of real-time applications. However,since in UTWR, triggering timers at the AP and STA have to synchronize with each other, simplerules are preferable for determining the triggering interval. Since in this research we do not assumeany specific applications, binary exponential manner is adopted to balance the trade-off betweensaving energy and small delay.

4.4.4 Performance Evaluation

We performed simulations using NS2 [18]. In these simulations, UPTT and UTWR are evaluatedand compared to U-APSD operating without periodic initiation of an U-SP, which is referred toas OnlyMore. STAs operating with OnlyMore send uplink null-data frames as trigger frames onlyif they receive downlink frames with EOSP=1 and MoreData=1 and they do not have any uplinkdata frames. This means that they do not start an U-SP periodically unlike UPTT and UTWR. Inall OnlyMore, UPTT and UTWR, MAX SP length is set to 2. In UPTT and UTWR, minimumand maximum triggering intervals are set to 6.25ms and 100ms, respectively. We assume that bothAP and STA operate with IEEE802.11b[4], basic rate is 2.0Mbps, and data rate is 11.0Mbps.For each AC, we have the following parameters [2]:CWmax[0] = 1023, CWmin[0] = 31;CWmax[1] = 1023, CWmin[1] = 31; CWmax[2] = 31, CWmin[2] = 15; CWmax[3] = 15,CWmin[3] = 7; AIFSN [0] = 7, AIFSN [1] = 3, AIFSN [2] = 2, AIFSN [3] = 2. EachAC operates with the drop-tail algorithm and can accommodate 100 packets. In the followingsimulations, two simulation scenarios are considered. One is performed under only voice traffic(Scenario1) and another is done under voice and Data traffic configurations (Scenario2). In eachScenario, simulations are performed through varying the number of voice STAs and in Scenario2

100

U-SP

STA wakes up STA goes to doze

STA

AP

DLnull-data

ACK

EOSP=1,MoreData=0

DL data

ACK

EOSP=1,MoreData=1

U-SP

STA wakes up

If Ttr is less than the maximum value, then Ttr = 2×TtrTriggering Timer is set

TriggeringTimer　at STA

TriggeringTimer　at AP

AP knows that the corresponding STA wakes up

AP knows that the corresponding STA wakes up

If Ttr is less than the maximum value, then Ttr = 2×TtrTriggering Timer is set

DL data

EOSP=1,MoreData=0

U-SP

ACK

STA goes to doze

Ttr is set to minimum valueTriggering Timer is set

Ttr is set to minimum valueTriggering Timer is setTimer expiration

Timer expiration

Timer expiration

Timer expiration

Figure 4.27: UTWR operations

U-SP

STA wakes up STA goes to doze

STA

AP

DLnull-data

ACK

EOSP=1,MoreData=0

DL null-data

ACK

EOSP=1,MoreData=0

U-SP

STA wakes up

If Ttr is less than the maximum value, then Ttr = 2×TtrTriggering Timer is set

TriggeringTimer　at STA

TriggeringTimer　at AP

AP knows that the corresponding STA wakes up

On receiving UL DATAAP knows that the corresponding STA wakes up

If Ttr is less than the maximum value, then Ttr = 2×TtrTriggering Timer is set

STA goes to doze

Triggering Timer is　reset to Ttr

ACK

UL data

Triggering Timer is canceled

Triggering Timer is canceled

Triggering Timer is　reset to Ttr

Timer expiration

Timer expiration

Figure 4.28: Timer cancellation in UTWR when sending an uplink data frame

101

there are 6 STAs with uplink Data traffic. Voice STAs operate with one of power saving mecha-nisms (OnlyMore, UPTT or UTWR) but STAs with uplink Data traffic work with active mode (i.e.they are always awake). Each voice flow is bi-directional ON-OFF traffic and both busy and idleintervals follow exponential distribution with mean 352ms and 650ms, respectively [30]. Duringbusy period, voice traffic is generated at a constant interval, 20 ms and has a fixed payload sizeof 208 bytes. Its rate is 83.2Kbps. Data traffic has a fixed payload size of 750 bytes. It is UDPtraffic and is generated by a constant interval, 10 ms. In each simulation run, simulation runs for60s. Simulation traffic is generated in the interval from 0s to 30s. We use simulation outputs inthe interval from 30s to 60s in each simulation. We evaluate packet dropping probability of voicetraffic both in Scenario1 and 2. As for delay evaluation, average delay of downlink voice traffic isconsidered in both scenarios and is expressed as averaged delay by using only successfully trans-mitted packets. In Scenario2, average throughput of Data traffic is also evaluated. Furthermore,energy efficiency is evaluated in both scenarios. The efficacy of power consumption can be oftenevaluated only by using the amount of consumed energy. However, power consumption is closelyrelated to how much STAs transmit or receive traffic. If a protocol can promote high throughput,power consumption of the STA on which it is implemented will be large. Therefore, we useenergygoodputdescribed in the below to evaluate power efficiency [32]:

energy goodput =total bits transmitted

total energy consumed

where the total bits transmitted are calculated for application layer packets only. The unit of energygoodput is Kbits/J.

Simulation results in Scenario1

Figure 4.29 and 4.30 show packet dropping probability of voice traffic and average delay of down-link voice traffic, respectively. It is obvious that the dropping probability in OnlyMore increaseswhen the number of voice STAs increases and that average delay of downlink voice traffic in On-lyMore is extremely large. This is due to the fact that when there is no uplink data frame duringOFF period of uplink voice traffic, STAs with OnlyMore cannot start an U-SP. Consequently manydownlink voice packets are buffered at the AP and are discarded. And also large buffering delay isgenerated at the AP. When the buffer for downlink voice packets overflows, they are discarded. InFig.4.30, we calculate delay of downlink voice traffic using only successfully transmitted packets.Therefore, when the number of voice traffic increases, delay of downlink voice traffic is not soincreased even though dropping probability is much increased. On the other hand, since UPTTand UTWR can periodically start an U-SP, delay of downlink voice traffic is much smaller anddropping probability is much lower than those in OnlyMore. From these results, it is obvious thateven if a STA does not have uplink data frame, an U-SP has to be periodically initiated in order tomaintain the performance of ON-OFF real-time applications. However, sending null-data framesto start an U-SP in UPTT is really overhead and a waste of the wireless bandwidth. Even if UPTTis implemented within the range of IEEE802.11e [2], dropping probability in UPTT becomes highwhen the number of voice STAs is large. Accordingly delay of downlink voice traffic in UPTTis increased when there are many voice STAs. Meanwhile, STAs with UTWR can periodicallystart an U-SP without transmissions of uplink null-data frames and UTWR can get rid of overhead

102

0

5

10

15

20

25

30

5 10 15 20 25 30

Dro

pp

ing

Pro

ba

bili

ty (

%)

Number of Voice STAs

OnlyMoreUPTT

UTWR

Figure 4.29: Dropping probability of voice traffic

generated by transmissions of them in UPTT. Hence, in UTWR, low dropping probability is main-tained and small delay of downlink voice traffic is also achieved even when the number of voiceSTAs is large. Figure 4.31 presents a comparison ofenergy goodput. In all mechanisms,energygoodputdeclines when the number of voice STAs is increased. This is due to the fact that whenthe number of voice STAs is large, a STA has to be awake for long time because of contentionswith other voice STAs and so it consumes more energy. Since STAs with OnlyMore do not initiatean U-SP except in case where they have uplink data frames or they receive downlink data frameswith EOSP=1 and MoreData=1, OnlyMore can save energy and has the highestenergy goodputwhen the number of voice STAs is small. However, delay of downlink traffic is extremely high.Therefore the quality of voice traffic must be really bad. Besides, when the number of voice STAsis increased,energy goodputis significantly decreased because the buffer of downlink voice trafficoverflows and downlink voice packets cannot be delivered to them. On the contrary, since UPTTand UTWR can deliver more downlink voice traffic when the number of voice STAs is high,energygoodputin both mechanisms is higher than OnlyMore. Moreover, compared to UPTT, UTWR canreduce overhead and deliver more downlink voice traffic because it does not use uplink null-dataframes as trigger frames, and further it can accommodate more voice traffic than UPTT. Thus,UTWR has higherenergy goodputthan UPTT, especially when there are large number of voiceSTAs.

Simulation results in Scenario2

Figure 4.32 and 4.33 present dropping probability of voice traffic and average delay of downlinkvoice traffic in Scenario2. Compared to those in Scenario1, dropping probability in all mechanismsin Scenario2 is slightly higher. This is because data traffic disturbs voice traffic under contention-based access even though voice traffic has more opportunities to access the wireless medium. As

103

0

0.05

0.1

0.15

0.2

0.25

5 10 15 20 25 30

Ave

rag

e D

ela

y o

f D

ow

nlin

k V

oic

e T

raff

ic (

s)

Number of Voice STAs

OnlyMoreUPTT

UTWR

Figure 4.30: Average delay of downlink voice traffic

100

110

120

130

140

150

160

170

180

190

200

5 10 15 20 25 30

En

erg

y G

oo

dp

ut

(Kb

its/J

)

Number of Voice STAs

OnlyMoreUPTT

UTWR

Figure 4.31: Comparison of energy goodput

104

similar to the results in Scenario1, dropping probability in OnlyMore is really high and averagedelay of downlink voice traffic in OnlyMore is very large. In OnlyMore, an U-SP cannot be startedeven when downlink voice packets arrive at the AP. Thus large buffering delay occurs and theyare discarded at the AP. On the contrary, since STAs operating with either UPTT or UTWR canperiodically start an U-SP and retrieve frames buffered at the AP, delay of downlink voice traffic isextremely small. However, UPTT uses transmissions of uplink null-data frames to start an U-SPand thus these uplink null-data frames cause a waste of the wireless bandwidth. Therefore, whenthere are large number of voice STAs, dropping probability in UPTT is more increased and alsodelay of downlink voice traffic is larger than UTWR.

Figure 4.34 shows a comparison ofenergy goodput. When the number of voice STAs is small,energy goodputin OnlyMore is the highest among all mechanisms because STAs with OnlyMoredo not wake up periodically. However, since it has much large delay of downlink voice traffic,quality of voice traffic is not so good. Furthermore, when the number of voice STAs is increased,energy goodputin OnlyMore significantly decreases because a lot of downlink voice packets arediscarded at the AP and they are not delivered to voice STAs. On the other hand, UPTT and UTWRcan reduce overhead and deliver more downlink voice traffic even when the number of voice STAsis large. Therefore,energy goodputin both schemes is not so different from OnlyMore whenthere are many voice STAs. Moreover, since in UTWR uplink null-data frames are not used astrigger frames, overhead due to transmissions of uplink null data frames are reduced and UTWRcan maintain smaller dropping probability of voice packets than UPTT. Thus, in UTWR moredownlink frames are delivered than UPTT andenergy goodputin UTWR is higher than UPTT.

Finally, Fig.4.35 shows average throughput of Data traffic. Throughput of Data traffic is main-tained in OnlyMore. This is because downlink voice traffic is infrequently retrieved from voiceSTAs and thus Data traffic can have much chance to access the wireless medium. When the num-ber of voice STAs increases, throughput of Data traffic in UPTT declines. Since in UPTT the wire-less bandwidth is consumed by transmissions of null-data frames, bandwidth is not sufficientlyallocated to Data traffic. On the other hand, in UTWR, throughput of Data traffic is maintained.Since UTWR does not use transmissions of uplink null-data frames to initiate an U-SP, UTWR canefficiently utilize the wireless bandwidth. Hence, even when the number of voice STAs is large, itcan maintain throughput of Data traffic and also achieve small dropping probability of voice traffic.

4.4.5 Conclusion

When a STA uses U-APSD provided in IEEE802.11e [2] for real-time traffic with silence period,it is suffered from large delay of downlink traffic. To solve this problem we proposed two mech-anisms, UPTT and UTWR. From simulation results of voice ON-OFF traffic, it was obvious thatboth of them achieved reducing buffering delay and could maintain small dropping probability ofvoice packets. As a result, they could accommodate more voice traffic. Furthermore, comparingUTWR with UPTT, UPTT can be implemented within the area of IEEE802.11e [2] but generateslarge overhead due to transmissions of uplink null-data frames. On the other hand, UTWR couldremove overhead generated in UPTT although it needs a little modifications of U-APSD defined in[2]. Thus, UTWR could efficiently utilize the wireless bandwidth and could achieve smaller delay

105

0

5

10

15

20

25

4 6 8 10 12 14 16 18 20

Dro

pp

ing

Pro

ba

bili

ty (

%)

Number of Voice STAs

OnlyMoreUPTT

UTWR

Figure 4.32: Dropping probability of voice traffic in Scenario2

0

0.05

0.1

0.15

0.2

0.25

4 6 8 10 12 14 16 18 20

Ave

rag

e D

ela

y o

f D

ow

nlin

k V

oic

e T

raff

ic (

s)

Number of Voice STAs

OnlyMoreUPTT

UTWR

Figure 4.33: Average delay of downlink voice traffic in Scenario2

106

90

100

110

120

130

140

150

160

4 6 8 10 12 14 16 18 20

En

erg

y G

oo

dp

ut

(Kb

its/J

)

Number of Voice STAs

OnlyMoreUPTT

UTWR

Figure 4.34: Comparison of energy goodput in Scenario2

440

460

480

500

520

540

560

580

600

620

4 6 8 10 12 14 16 18 20

Ave

rag

e T

hro

ug

hp

ut

of

Da

ta t

raff

ic (

Kb

ps)

Number of Voice STAs

OnlyMoreUPTT

UTWR

Figure 4.35: Average throughput of Data traffic in Scenario2

107

and dropping probability of voice traffic. Besides, it could accommodate Data traffic.Since U-APSD operates under EDCA, a scheduling algorithm implemented in AP influence on

power consumption of STAs. Therefore in future research we will work on it to reduce time periodfor which they are awake.

4.5 Conclusion of this Chapter

This chapter focused on infrastructure networks consisting of IEEE802.11e [2] WLAN and con-tained two researches to enhance the performance of IEEE802.11e WLAN networks.

The first research was on performance improvement of real-time traffic in IEEE802.11e WLANnetworks dynamically configuring EDCA parameters. Since EDCA provides contention basedchannel access, throughput of real-time traffic in an STA is affected by real-time traffic and nonreal-time traffic transmitted by other STAs. Therefore we proposed two methods to overcomeperformance degradation of real-time traffic due to collisions. In the first method, we proposed adynamic adaptation of CW sizes to reduce collisions between real-time flows. In the second, wefocused on the use of continuous TXOP and proposed a method to adaptively select TXOP Limit.Simulation results showed that using these methods could help AP to accommodate more real-timetraffic.

The second research was on performance improvement of downlink traffic operating with U-APSD in IEEE802.11e and it proposed UPTT and UTWR to overcome the problem that real-time traffic with silence period is caused large buffering delay in an AP. Both UPTT and UTWRprovided a mechanism to periodically start U-SP and to enable periodic transmissions of downlinkframes. Simulation results showed that they could reduce delay of downlink traffic operating withU-APSD.

108

Chapter 5

Performance Improvement of LargeIEEE802.11e WLAN Networks Consistingof Multiple APs

109

5.1 Overview of This Chapter

In IEEE802.11 infrastructure networks, it is possible to configure large WLAN networks consistingof multiple APs. In the environment of large WLAN networks, wireless resource is inefficientlyutilized and traffic load is unbalanced among APs if there is no strategy for resource allocation.Thus in the worst case WLAN users associating with the network are perhaps suffered from per-formance degradation and the network is able to accommodate less traffic than it is expected todo. In this sense, wireless resource management plays an important role for load balancing amongAPs. Therefore, in this chapter, we research on radio resource management and propose an APselection mechanism to achieve load balancing in IEEE802.11e networks.

5.2 Access Point Selection Strategy in IEEE802.11e WLAN net-works toward Load Balancing

5.2.1 Introduction

WLAN is currently used in office and public area such as hotel and airport. In these large WLANnetworks, RRM (Radio Resource Management) is important for the efficient use of wireless band-width. For example, since three channels (channel number 1, 6, 11) can be used simultaneouslyin IEEE802.11b [4], APs have to be deployed to minimize overlaps between the same channelsand to reduce interferences between STAs using the same channel[33]. Besides AP deployment astrategy, strategy for STAs to select an AP with which they will associate is also much importantin large WLAN networks. Suppose that there are two APs as shown in Fig.5.1. If STAs have nostrategy, they may associate with one of the APs. As a result, throughput in an AP will decline dueto congestion even though radio resource in another AP is available.

To overcome such problem presented in Fig.5.1, there are two approaches. One is centralizedway proposed in [34]. In [34], a server is used to control STAs’ association and works as admissioncontroller. In fact, admission control is supported in IEEE802.11e, but it’s difficult for admissioncontroller provided in IEEE802.11e to be used for non real-time traffic because such kind of traf-fic has no periodicity. Moreover, IEEE802.11 does not standardize any specific servers to controlSTAs’ association and thus if those servers are required, it is difficult to maintain compatibilityamong WLAN devices provided from different vendors. In fact, IEEE802.11 specifies only proce-dures required before a STA connects to an AP [1]. For example, when Passive Scan takes place, aSTA receives beacon frames transmitted from APs and decides an AP with which it will associate.That is, STAs have the right to select an AP in IEEE802.11 standard [1]. In addition to limitationsin the current standard, if specific servers are needed to control STAs’ association, latency takenduring the HO (Hand-Over) procedure will increase. To reduce latency in HO, it is better for a STAto have the right to select an AP. From these reasons, decentralized approaches were proposed in[36] and [37] and they focused on AP selection mechanisms implemented in STAs. Based on thesame reason as prior arts in [36] and [37], this research discusses and proposes an AP selectionmechanism for load balancing and the efficient use of radio resource.

In [36] and [37], AP selection mechanisms were proposed for load balancing among APs.

110

Many STAs are associating with AP 1 !!AP 1

AP 2

Figure 5.1: Congested AP

Their proposed algorithms considered RSSI (Received Signal Strength Indicator) and the numberof STAs connecting to an AP, and also took into account only BE (Best Effort) traffic. However,in 802.11e WLAN networks, traffic types and their traffic loads should be considered rather thanthe number of STAs. Therefore, it is difficult to apply their algorithms to IEEE802.11e networks.Moreover, they assumed that STAs could use the same transmission rate even when they connectedto APs that were very far from it. However, in general, if a STA communicates with an AP that isfar from it, it needs to use low transmission rate. Therefore, taking into account only the number ofSTAs for load balancing results in increasing the number of STAs that use low transmission rate.Consequently radio resource is inefficiently utilized and throughput of WLAN networks declines.To overcome problems presented above, we propose an AP selection mechanism, called HRFA(High Rate First Association). It considers channel load in an AP for load balancing and also trans-mission rate, which a STA uses to communicate with an AP, for the efficient use of radio resource.Furthermore HRFA considers types of traffic so that it can be applied to IEEE802.11e networks. Infact, HRFA can be implemented without any modifications in IEEE802.11 and 802.11e standard,and so it is fully compatible with IEEE802.11 standard.

The remainder of this research is organized as follows. In Section 5.2.2, related researches andtheir problems are presented. Section 5.2.3 shows our proposed algorithm, HRFA. Performanceevaluations of our proposal are carried in the section 5.2.4, comparing with other schemes. Finallyin section 5.2.5, we present the conclusion.

5.2.2 Prior Arts

RSSI is a metric widely used to decide an AP with which a STA will associate. In this mechanism,the STA receives beacon frames from APs through scanning procedure and monitors RSSI ofreceived frames. At the end of the scanning procedure, it selects an AP with maximum RSSI.

111

In general, if the STA selects an AP with larger RSSI, it can use higher transmission rate andcan efficiently utilize radio resource. Consequently the AP can accommodate more traffic andthroughput in the AP becomes high [38]. However, the consideration of only RSSI results intraffic congestion in an AP as shown in Fig.5.1. In order to achieve load balancing among APs, in[36] and [37], AP selection mechanisms, which considered the number of associating STAs, wereproposed. In [37], the following metric was proposed.

Scorei =1 − Pi

Ni(5.1)

Pi denotes PER (Packet Error Rate) andNi is the number of associating STAs with APi (0 ≤ i ≤L − 1). In [37], the way to calculate PER was not presented, but since it showed that PER wascalculated from RSSI, PER is probably considered as the same as RSSI. After calculating Eq.(5.1)for all APs with which a STA can communicate, an AP with maximum value ofScore i is selected.In [36] and [37], all STAs were assumed to transmit only BE traffic. However, in fact, channelload in the AP may be low, even if the number of associating STAs is large. Therefore, it is betterto consider channel load in an AP. And, in 802.11e networks, types of traffic (e.g. real-time ornon real-time) have to be considered when channel load is taken into account. Since in 802.11enetworks real-time traffic have better chance to access the WM than non real-time traffic, it isbetter for a STA with real-time traffic to connect to an AP whose real-time traffic load is low evenif its non real-time traffic load is slightly high. Moreover, algorithms proposed in [36] and [37]select an AP that is far from a STA and the STA has to use low transmission rate even if other APsthat are near to the STA can accommodate its traffic. As a result, radio resource is inefficientlyutilized and throughput in an AP declines. From these observations, algorithms presented in [36]and [37] possibly degrade the performance of WLAN networks even though they tried to achieveload balancing. In the next section, we present our proposed algorithm, HRFA.

5.2.3 HRFA

Overview of HRFA

This subsection shows an AP selection strategy in HRFA. We assume that in 802.11e networksadmission control is taken for real-time traffic (traffic belonging to AC3 and AC2). The strategy ofHRFA is to use as high transmission rate as possible in order to efficiently utilize radio resource,and only after traffic load in an AP is high, it selects an AP whose traffic load is low. ThereforeHRFA is focus on transmission rate as one of metrics.

Here, load balancing strategy taken by HRFA is presented. The strategy is classified into twotypes, one for real-time traffic and another for non real-time traffic. First, we show load balancingstrategy for real-time. When a STA having real-time traffic selects an AP, the traffic must beaccepted by admission controller implemented in the AP. If it can be accepted, real-time trafficload in the AP has to be considered since real-time traffic can more often access the WM than nonreal-time traffic. Therefore, in HRFA, a STA with real-time traffic selects an AP whose real-timetraffic load is low. However, if there is little difference in real-time traffic load among APs, the STAselects an AP, with which it can communicate using higher transmission rate, in order to efficientlyutilize radio resource.

112

Element ID Length StationCount

ChannelUtilization

AvailableAdmissionCapacity

1 octet 2 octets 2 octets1 octet 1 octet

Figure 5.2: QBSS Load element

Next, load balancing strategy for non real-time traffic is presented. When channel load in anAP is low, a STA having non real-time traffic can have more opportunities to access the WM.Therefore, the STA with non real-time traffic selects an AP whose channel load is low. However,as similar to the strategy for real-time traffic, if there is little difference in channel load amongAPs, the STA selects an AP with which it can communicate at higher transmission rate.

Algorithm of HRFA

This subsection shows a detailed algorithm of HRFA. HRFA considers the following metrics.

• transmission rate

• channel load in an AP

• real-time traffic load in an AP

Transmission rate used to communicate with an AP is determined by LA (Link Adaptation) algo-rithm implemented in STAs. For example, in [39], transmission rate is decided based on the valueof RSSI. Therefore this research also assumes that transmission rate is decided based on RSSI.However, in fact, any LA algorithms are applicable to HRFA. Channel load and real-time trafficload in an AP are announced to STAs by QBSS load element in beacon frame [2]. Figure 5.2shows QBSS load element [2]. Therefore, if a STA executes a scanning procedure, it can get theseinformation [1], [2].

First, the case where a STA with real-time traffic selects an AP is presented. As mentionedin the previous subsection, real-time traffic load in an AP and transmission rate have to be takeninto account for AP selection mechanism of STAs with real-time traffic. In HRFA a STA havingreal-time traffic calculates Eq.(5.2) and selects an AP with maximum value ofScoreRTi.

ScoreRTi = AACi ×Ri (5.2)

AACi shows Available Admission Capacity, which is announced by QBSS load element. Whenthe number of real-time traffic accommodated in APi is n i, AACi presents the remaining channeltime for real-time traffic per a second as shown in Eq.(5.3).

AACi = 1 −ni−1∑

j=0

medium timej (5.3)

113

As explained in section 4.2.2 (Admission Control),medium timej denotes time allocated to flowj by the AP. In Eq.(5.2),Ri is a weight factor determined based on transmission rate. As shown inEq.(5.4) and (5.5), it is expressed by using time,Trh

, needed to transmit at transmission rate ofrh

a data frame whose payload size isS. When calculatingTrh, headers both in physical and MAC

layer are considered.Tmax denotes time needed to transmit a data frame at minimum transmissionrate. We assume that in Eq.(5.5) the number of transmission rates supported in a STA isH andthat in Eq.(5.4)rh is selected for the STA to communicate with APi

Ri =Tmax

Trh

(5.4)

Tmax = max{Trh| 0 ≤ h ≤ H − 1} (5.5)

Suppose that we use 802.11b and can use three transmission rates, 2.0, 5.5, and 11.0Mbps. If wechoose 1024 bytes as the value ofS, we can calculateRi as about 1.0, 2.6, and 4.6 for 2.0, 5.5,and 11.0Mbps, respectively. When we useAACi andRi in Eq.(5.2), we can consider influenceof transmission rate on the remaining capacity to accommodate real-time traffic. Moreover, inEq.(5.2), the current real-time traffic load is not considered but the remaining capacity is taken intoaccount because of the following reason. Suppose that STAa can communicate with APl andmusing transmission rate, 11.0 and 2.0Mbps, respectively, and that the current real-time traffic loadin AP l is 4.7 times larger than one in APm. In this case, if we use in Eq.(5.2) the current real-timetraffic load in place of remaining capacity for real-time traffic, STAa selects the APm and useslow transmission rate even if there is a lot of remaining capacity to accommodate real-time trafficin AP l. Thus, radio resource is inefficiently utilized. Therefore, to avoid the inefficient use ofradio resource, we consider the remaining capacity for real-time traffic in Eq.(5.2). Taking intoaccount the remaining real-time traffic capacity and transmission rate the Eq.(5.2) encourage STAswith real-time traffic to select an AP with which it can communicate at higher transmission rate.Moreover, only after real-time traffic load in an AP is high, Eq.(5.2) allows the STA to select anAP, with which it communicates at low transmission rate, for the purpose of load balancing. Notethat even if the value ofScoreRTi is the highest, a STA cannot associate with APi and connectsto other APs when admission controller implemented in the AP does not accept its request.

Next, we explain an algorithm for non real-time traffic. A STA having non real-time trafficcalculates the Eq.(5.6) and selects an AP with maximum value ofScoreNRTi.

ScoreNRTi = (256 − CLi) ×Ri (5.6)

CLi indicates channel load in APi given from channel utilization field in QBSS load element[2]. It is measured at the AP and presented as a value ranging from 0 to 255. As mentioned inthe previous section , non real-time traffic can have a better chance to access the WM when itcommunicates with an AP whose channel load is low. Therefore, in Eq.(5.6), we consider channelload as one of metrics. And, as similar to the algorithm presented for real-time traffic, in order toefficiently utilize radio resource, weight factorRi and remainging channel load are considered inEq.(5.6). As a result, Eq.(5.6) encourages STAs with non real-time traffic to associate an AP withwhich they can communicate at high transmission rate. If we subtractCLi from 255 in Eq.(5.6),

114

ScoreNRTi will become 0 and transmission rate cannot be considered in Eq.(5.6). Therefore wesubtractCLi from 256 in Eq.(5.6).

Although we separately explain algorithms for real-time and non real-time, if a STA has bothreal-time and non real-time traffic, it uses the algorithm for real-time traffic. Moreover, sincechannel load and wireless link condition changes over time, HO is necessary. However, sinceHRFA is the algorithm to select an AP, it is applicable to HO algorithms proposed in [35] or [37].

5.2.4 Performance Evaluation

We implemented our proposed algorithm in NS-2 [18] and performed simulations. We assumethat both AP and STA operate with IEEE802.11b[4], basic rate is 2.0Mbps, and data rate is chosenfrom 2.0, 5,5, and 11.0Mbps. For simplicity, data rate is determined by distance between theSTA and AP. STAs within 60ms from AP use 11.0Mbps, 5.5Mbps within 120ms, and 2.0Mbpswithin 200ms, respectively. For each AC, we have the following parameters [2]:CWmax[0] =1023, CWmin[0] = 31; CWmax[1] = 1023, CWmin[1] = 31; CWmax[2] = 31, CWmin[2] =15; CWmax[3] = 15, CWmin[3] = 7; AIFSN [0] = 7, AIFSN [1] = 3, AIFSN [2] = 2,AIFSN [3] = 2. Simulation area is200m × 200m. Three APs are placed in (42,50), (100, 150),and (158, 50), respectively. These APs operate in different channels and so they do not interferewith each other. Two types of deployment of STAs are considered. First STAs are uniformlydistributed in the simulation area (Case1). Next they are deployed in a place shaded in Fig.5.3(Case2). For both STA deployments, simulations are performed. As simulation traffic, voice,video, and data traffic are used. Each voice flow is 83.2 Kbps, which is generated by a constantinterval, 20 ms and has a fixed payload size of 208 bytes. This flow corresponds to G.711-codedVoIP [26]. Each video flow is 770 Kbps, which is generated by a constant interval, 13.33 ms andhas a fixed payload size of 1280 bytes. As background traffic, UDP traffic is 800Kbps, whichis generated by a constant interval, 10 ms and has a fixed payload size of 1024 bytes. In eachsimulation run, simulation runs for 200s Simulation traffic is generated in the interval from 0s to100s. We use simulation outputs in the interval from 100s to 200s in each simulation. We evaluatethroughput for all voice, video, and data traffic, and also evaluate delay of voice and video traffic.When we generate simulation traffic, first we randomly select a STA and next traffic which theSTA transmits is chosen from voice, video, and data traffic. Finally when simulation time is 100s,there are the same number of voice, video, and data traffic in each simulation. That is, if flow-setis 5 in simulation results in the following subsection, it means there are 5 voice, 5 video, and 5data flows. In each simulation, HRFA is compared to two AP selection algorithms. One of themis AP selection algorithm taking care only about RSSI (called RSSI). Another is one consideringthe number of associating STAs, which was proposed in [37] (called NumSTA). In the followingsimulations, all AP performs admission control and minimum physical rate is set to 5.5Mbps.Therefore, regardless of AP selection algorithms, a STA having real-time traffic cannot associatewith an AP with which it communicate at 2.0Mbps.

115

200m

200m

AP

AP

AP

100m

100m

Figure 5.3: Simulation Area

Simulation Results in Case1

Figure 5.4 and 5.5 show average throughput of voice and video traffic, and Fig.5.7 and 5.8 presentaverage delay of voice and video traffic. In case where STAs are uniformly deployed in the simula-tion area, the number of traffic and traffic types are also in a way distributed in the area. Therefore,with regard to throughput and delay of voice and video traffic, there is little difference amongRSSI, NumSTA, and HRFA. This is because real-time traffic can have better chance to access theWM under IEEE802.11e EDCA. Thus throughput and delay of real-time traffic are maintained.However, in case of NumSTA, throughput of video traffic declines and also its delay increaseswhen the number of traffic increases. NumSTA distributes the number of associating STAs butdoes not take into account traffic types and load. And, using NumSTA, the number of data traffictransmitted at low transmission rate increases. Furthermore, even if STAs having real-time trafficcan connect to an AP using transmission rate of 11.0Mbps, they may associate with other APswith which they communicate at 5.5Mbps. As a result, radio resource is inefficiently utilized andthroughput and delay of video traffic is degraded when there are much traffic in the simulationarea.

Figure 5.6 shows average throughput of data traffic. Especially in NumSTA, throughput of datatraffic is low. As mentioned above, NumSTA distributes only the number of associating STAs andso the number of traffic transmitted at low transmission rate is increased. Therefore, the inefficientuse of radio resource result in low throughput of data traffic. Since RSSI and HRFA select an APwith which STAs can communicate at high transmission rate, radio resource is efficiently utilized.But in case where only RSSI is considered, traffic congestion happens in a specific AP. On thecontrary, STAs with HRFA can associate with APs whose channel load is low when channel loadincreases in an AP. Therefore, in HRFA, radio resource is efficiently used and load balancing isalso achieved. As a result, throughput of data traffic in HRFA is higher than in RSSI.

116

30

40

50

60

70

80

3 4 5 6 7 8 9 10 11 12

Avera

ge T

hro

ughput(

Kbps)

The number of flow-set (Voice, Video and Data)

RSSINumSTA

HRFA

Figure 5.4: Average throughput of voice traffic in Case1

680

690

700

710

720

730

740

750

760

770

3 4 5 6 7 8 9 10 11 12

Ave

rag

e T

hro

ug

hp

ut

(Kb

ps)

The number of flow-set (Voice, Video and Data)

RSSINumSTA

HRFA

Figure 5.5: Average throughput of video traffic in Case1

117

0

100

200

300

400

500

600

700

800

3 4 5 6 7 8 9 10 11 12

Ave

rag

e T

hro

ug

hp

ut(

Kb

ps)

The number of flow-set (Voice, Video and Data)

RSSINumSTA

HRFA

Figure 5.6: Average throughput of data traffic in Case1

0.001

0.002

0.003

0.004

0.005

0.006

0.007

0.008

0.009

0.01

3 4 5 6 7 8 9 10 11 12

Ave

rag

e D

ela

y(s

)

The number of flow-set (Voice, Video and Data)

RSSINumSTA

HRFA

Figure 5.7: Average delay of voice traffic in Case1

118

0

0.005

0.01

0.015

0.02

0.025

0.03

0.035

0.04

3 4 5 6 7 8 9 10 11 12

Ave

rag

e D

ela

y(s

)

The number of flow-set (Voice, Video and Data)

RSSINumSTA

HRFA

Figure 5.8: Average delay of video traffic in Case1

Simulation Results in Case2

This subsection shows simulation results in case of simulations performed in Case2 scenario. Fig-ure 5.9 and 5.10 present average throughput of voice and video traffic, and Fig.5.12 and 5.13 showaverage delay of voice and video traffic, respectively. When using IEEE802.11e, real-time trafficis given a better chance to access the WM. Therefore in case where traffic load is low, throughputof voice and video traffic is not degraded even in RSSI and NumSTA. However, when traffic loadis high, throughput of voice and video in RSSI and NumSTA declines and accordingly their delayincrease. Since RSSI does not consider load balancing, traffic load becomes high in a specific APand contentions between real-time flows or between real-time and non real-time flows increase.Consequently throughput of voice and video declines even when IEEE802.11e is used. In case ofNumSTA, since its load balancing strategy does not care about traffic types, real-time traffic will bepossibly congested in a specific AP. Thus, throughput of real-time traffic is degraded. Furthermore,in NumSTA, STAs not only with real-time but also with non real-time traffic may communicatewith an AP, which is far from them, at low transmission rate. As a result, radio resource is ineffi-ciently used and throughput of real-time traffic is degraded when traffic load is high. Accordinglyits delay increases. On the contrary, in HRFA, throughput and delay of real-time traffic are main-tained even when traffic load is increased. This is due to the fact that HRFA encourages STAs toconnect to an AP with which they can communicate at high transmission rate and it takes them toother APs only after traffic load is high in an AP. That is, HRFA achieves the efficient use of radioresource and load balancing. Furthermore, since load balancing strategy in HRFA considers traffictypes, real-time traffic is not congested in an specific AP.

Figure 5.11 presents average throughput of data traffic. Since in HRFA radio resource is effi-ciently utilized and load balancing is also achieved, throughput of data traffic is higher than othermethods. When traffic load is increased, difference between throughput in HRFA and RSSI be-

119

20

30

40

50

60

70

80

90

3 4 5 6 7 8 9 10 11 12

Ave

rag

e T

hro

ug

hp

ut(

Kb

ps)

The number of flow-set (Voice, Video and Data)

RSSINumSTA

HRFA

Figure 5.9: Average throughput of voice traffic in Case2

comes little. This is due to the fact that when traffic load is high, throughput of real-time traffic inRSSI declines and on the other hand one of data traffic is maintained to some extent.

5.2.5 Conclusion of Access Point Selection Strategy in IEEE802.11e WLANnetworks

We proposed HRFA as an AP selection mechanism implemented in STAs. It considered trafficload in an AP to perform load balancing and further took into account transmission rate, whicha STA uses to communicate with the AP, in order for STAs participating in a WLAN networkto efficiently share and utilize wireless resource. Thus it achieved both load balancing and theefficient use of radio resource, and it also obtained higher throughput compared to other schemesproposed in prior arts. Since HRFA was evaluated in a simple WLAN network in this research,we will further evaluate the performance of HRFA when it is used in a WLAN network which isdynamically changing and when some HO algorithms are applied to HRFA.

5.3 Conclusion of This Chapter

This chapter focused on large WLAN networks consisting of multiple APs. In the environment oflarge WLAN networks, load balancing among APs becomes very important for congestion avoid-ance and the efficient use of radio resource. Therefore, considering the fact that in IEEE802.11WLAN STAs have the right to choose an AP with which they associate we proposed an access pointselection mechanism towards load balancing, called HRFA. HRFA could work in IEEE802.11eWLAN network and achieved more efficient use of radio resource, compared to prior arts. Sim-ulation results showed that HRFA can help WLAN networks to accommodate more traffic than

120

300

350

400

450

500

550

600

650

700

750

800

3 4 5 6 7 8 9 10 11 12

Ave

rag

e T

hro

ug

hp

ut(

Kb

ps)

The number of flow-set (Voice, Video and Data)

RSSINumSTA

HRFA

Figure 5.10: Average throughput of video traffic in Case2

0

100

200

300

400

500

600

700

800

3 4 5 6 7 8 9 10 11 12

Ave

rag

e T

hro

ug

hp

ut(

Kb

ps)

The number of flow-set (Voice, Video and Data)

RSSINumSTA

HRFA

Figure 5.11: Average throughput of data traffic in Case2

121

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

3 4 5 6 7 8 9 10 11 12

Ave

rag

e D

ela

y(s

)

The number of flow-set (Voice, Video and Data)

RSSINumSTA

HRFA

Figure 5.12: Average delay of voice traffic in Case2

0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

3 4 5 6 7 8 9 10 11 12

Ave

rag

e D

ela

y(s

)

The number of flow-set (Voice, Video and Data)

RSSINumSTA

HRFA

Figure 5.13: Average delay of video traffic in Case2

122

previous researches.

123

Chapter 6

Summary

124

6.1 Concluding Remarks

IEEE802.11 is well known as de-facto WLAN standard and provides MAC and PHY layers spec-ification. It can provide wireless access and mobility to small mobile devices, and is thereforewidely used in many places such as office, hotel and airport. In IEEE802.11 standard specifica-tion, ad hoc and infrastructure modes are supported. STAs using ad hoc mode can form ad hocnetworks. Ad hoc networks are able to operate without any backbone networks and enable STAsto communicate with each other in distributed manner. Besides, if STAs support routing functions,they can perform wireless multihop communication. Therefore ad hoc networks are consideredas useful networks for disaster relief. On the other hand, an AP and STAs using infrastructuremode can form infrastructure networks. They can access a backbone network and further inter-net through the AP in infrastructure networks. In both WLAN networks, power consumption ofWLAN devices is a critical issue when they are embedded into mobile devices. To prolong thelifetime of them, WLAN device must reduce power consumption when they transmit data andadditionally have to operate in low power when they are not in actual communication. Besides,taking into account the fact that voice and videoconference are identified as promising applicationsin WLAN networks and the demand for the use of multimedia applications over WLAN, multime-dia applications have to be smoothly supported in WLAN in order for IEEE802.11 WLAN to bemore widely used.

These challenging issues in IEEE802.11 WLAN networks motivated us to work on severalresearches which enabled to enhance the performance of IEEE802.11 WLAN networks in termsof power consumption and multimedia support both in ad hoc and infrastructure networks. Thisthesis presented each of them and consisted of three research parts.

In the fist part, we had two proposals in ad hoc networks. We first proposed IPSM to improvethe performance of throughput in ad hoc networks operating with IEEE802.11 PSM. Since sendernodes operating with IEEE802.11 PSM have to exchange ATIM frames with receiver nodes, theyhave to buffer packets to transmit ATIM frames and consequently delay increases due to bufferingat sender nodes. On the other hand, nodes operating with IPSM can relay packets without bufferingthem immediately after the nodes receive them because a sender and a receiver nodes exchangeawake period with each other and they know how long the counter part is awake. Finally simulationresults showed that we could achieve high throughput and energy efficient packet transmissionsusing IPSM.

In ad hoc networks, routing protocols impact on power consumption since nodes in a routerelay packets from source to destination nodes. The first research part next proposed an energyefficient routing protocol in ad hoc networks. Proposed routing scheme simultaneously achieved toextend the lifetime of ad hoc networks and to fairly use battery among nodes because it consideredas routing metrics transmission power for reducing power consumption and remaining batterycapacity of each node for fairly utilizing battery among nodes. Simulation results showed that ourproposal could prolong the lifetime of ad hoc networks.

In the second part we had two researches in infrastructure networks. As for infrastructure networks,we focused on IEEE802.11e WLAN since it defined a new MAC protocol to support multimedia

125

applications over WLAN and it is considered as an important function in IEEE802.11 WLANstandard family.

Even though EDCA defined in IEEE802.11e can provide differential channel access and real-time application can get more opportunities to access the wireless medium earlier than non real-time traffic, real-time traffic is suffered from performance degradation due to collisions betweenreal-time flows or between real-time and non real-time flows. The first research in this part there-fore proposed Dynamic Adaptation of EDCA Parameters in IEEE802.11e WLAN networks. Sinceour proposed adaptation can help reduce collisions not only between real-time flows but also be-tween real-time and non real-time flows, it is able to achieve to keep throughput of real-timetraffic high even when traffic load is high in a IEEE802.11e WLAN. Simulation results showedthat Dynamic Adaptation of EDCA Parameters helped reduce collisions between real-time flowsand between real-time and non real-time flows, and allowed APs to accommodate more real-timetraffic.

We secondly researched on U-APSD, which was defined in IEEE802.11e as a power manage-ment function, in this second part. U-APSD assumes that it is used with real-time traffic generatedat a constant period and it can achieve energy efficient real-time data delivery. However, when it isapplied to real-time traffic with silence period like voice traffic, downlink traffic is suffered fromlarge delay since STAs operating with U-APSD cannot periodically start U-SP. Hence we pro-posed UPTT and UTWR. Both of them have a mechanism for STAs to periodically start U-SP andto retrieve downlink data from APs. UPTT can be implemented within the range of IEEE802.11especification. Although UTWR needs some modifications of IEEE802.11e specification, it canreduce overhead of frame exchanges needed in U-APSD and UPTT and achieve lower packetdropping ratio. Compared to U-APSD, our proposed UPTT and UTWR achieved smaller delay ofdownlink traffic and lower packet dropping ratio in computer simulations.

IEEE802.11 WLAN can configure large WLAN networks consisting of multiple APs. Radio re-source in large WLAN networks is inefficiently utilized since radio resource is perhaps inappro-priately allocated to each traffic among APs. In this sense, radio resource management plays animportant role for WLAN applications to efficiently utilize radio resource. In the last researchpart, we researched on radio resource management towards load balancing among APs in largeWLAN networks. Since IEEE802.11 specification gives STAs the right to select an AP with whichthey will associate, an AP selection strategy implemented in a STA can be a key to achieve loadbalancing. AP selection algorithms proposed in the prior arts inefficiently utilized radio resourceand so traffic load was not appropriately balanced among APs. This is because they did not con-sider traffic type (real-time or non real-time traffic) and a principle of radio communication, i.e.,radio wave decays according to the increase of distance between a sender and receiver. Hence weproposed a novel AP selection algorithm, called HRFA to achieve the efficient utilization of radioresource and to be used in the environment of IEEE802.11e WLAN networks. Since STAs withHRFA take into account both transmission rate and channel load for AP selection, they can selectan AP with which they communicate at higher transmission rate and further can take care aboutload balancing among APs. Consequently, simulation results convinced us that WLAN networkswith HRFA was able to achieve more efficient use of radio resource and accommodate more trafficthan prior researches.

126

6.2 Future Works

There are a number of areas with potential for future research in IEEE802.11 WLAN becauseIEEE802.11 standard is always seeking for new technologies to enhance the performance of IEEE802.11 WLAN. The standard is continuing standardization for IEEE802.11 WLAN to be muchmore widely used, adding new features and usecases. Even now there are several standardiza-tion works ongoing under task groups. For example, IEEE802.11n is for enhancements of higherthroughput and IEEE802.11s is for standard in wireless mesh networks [8], [9]. To achieve highthroughput in IEEE802.11n, not only increasing transmission speed in PHY layer but also reduc-ing MAC over head in MAC layer are very important. Recently how to compose MAC frameswith less overhead has been often considered.

Wireless mesh networks defined in IEEE802.11s are considered as wireless backbone networkswhich can be set up at low cost. However, taking into account of wireless features such as pathloss, multipath, contention and etc., wireless mesh networks will provide less capacity than expec-tations if radio resource is inappropriately used. Hence radio resource management must be oneof the important research topics in wireless mesh networks and therefore wireless nodes have toefficiently utilize limited radio resource in distributed manner.

Taking into account advantages of different wireless access technologies and installation cost, fu-ture wireless networks are expected to consist of several wireless access technologies, e.g. cellularnetworks, WLAN networks and other broadband wireless networks. In this environment, capabil-ities of WLAN networks will complement weak points of other kinds of wireless networks sinceWLAN can provide broadband access at low cost. In fact, some of cellular phone now include aWLAN device and provide the capability of wireless broadband access to users. The current prod-ucts does not provide functionalities to seamlessly handover between different wireless systems,but most of users must want to do so.

In order to enable the seamless handover, there are many things to overcome. We have to have apolicy to decide when to perform handover based on several performance measures, such as radiosignal strength, power consumption, communication cost and etc. Besides, signaling processes,which are used in different wireless systems and take place when a device is associating witha base station, have to be merged and made simple because they become time-consuming tasksand generate large delay during handover. Therefore, IEEE802.11 WLAN standard will have toprovide simple signaling processes to enable fast handover with different wireless systems such as3GPP and 3GPP2 cellular systems.

127

Publication

1. “An Improved Power Saving Mechanism for MAC Protocol in Ad Hoc Networks”Shojiro Takeuchi, Kaoru Sezaki and Yasuhiko YasudaIEICE Trans. on COMMUN, Vol.E88-B, NO.7, July 2005

2. “Access Point Selection Strategy in IEEE802.11e WLAN networks towards Load Balancing(written in Japanese)”Shojiro Takeuchi, Kaoru Sezaki and Yasuhiko YasudaIEICE Trans. on COMMUN, to apper in April 2006

3. “A Proposal of Battery Cost Routing in Consideration of Transmission Power”Shojiro Takeuchi, Kosuke Yamazaki, Kaoru Sezaki, Yasuhiko YasudaAPCC, Sep. 2002

4. “Geographical Forwarding with Adaptive Transmission Power Control in Mobile Ad HocNetworks”Shojiro Takeuchi, Yamazaki Kosuke, Kaoru Sezaki and Yasuhiko YasudaProc. of 8th Mobile Multimedia Communications (MoMuC 2003), Oct. 2003

5. “An Improved Power Saving Mechanism for MAC Protocol in Ad Hoc Networks”Shojiro Takeuchi, Yamazaki Kosuke, Kaoru Sezaki and Yasuhiko YasudaProc. of 47th IEEE Global Telecommunications Conference (Globecom 2004), Nov. 2004

6. “Dynamic Adaptation of Contention Windown Sizes in IEEE802.11e Wireless LAN”Shojiro Takeuchi, Kaoru Sezaki and Yasuhiko YasudaProc. of 5th IEEE International Conference on Information, Communications and SignalProcessing (ICICS 2005), Dec. 2005

7. “Heterogeneous Co-simulation with SDL and SystemC for Protocol Modeling (invited pa-per)”Toshiaki Jozawa, Huang Leping, Sakai Tomokazu, Shojiro Takeuchi and Mika KasslilnProc. of IEEE Radio and Wireless Symposium 2006 (RWS 2006), Jan. 2006

8. “Group Mobility Modeling in Mobile Ad Hoc Networks using Pedestrian Tracked Data”Werner Creixell, Kaoru Sezaki and Shojiro TakeuchiProc. of 2nd International Workshop on Mobility Aware Technologies and Applications(MATA 2005), Oct. 2005

128

9. “Quick Data-retrieving for U-APSD in IEEE802.11e WLAN networks”Shojiro Takeuchi, Kaoru Sezaki and Yasuhiko YasudaProc. of IEEE Wireless Communications and Networking Conference (WCNC 2006), April2006

10. “Access Point Selection Strategy in IEEE802.11e WLAN networks”Shojiro Takeuchi, Kaoru Sezaki and Yasuhiko YasudaProc. of IEEE Wireless Communications and Networking Conference (WCNC 2006), April2006

11. “送信電力を考慮した Battery Cost Routingの提案”竹内彰次郎,山崎浩輔,瀬崎薫,安田靖彦電子情報通信学会信学技法 NS2001-241, 2002年 3月

12. “地理的経路制御における適応的電力制御手法” 竹内 彰次郎,山崎 浩輔,瀬崎 薫,安田靖彦電子情報通信学会信学技法 NS2002-231, 2003年 3月

13. “An Improved Power Saving Mechanism for MAC Protocol in Ad Hoc Networks”竹内彰次郎,山崎浩輔,瀬崎薫,安田靖彦電子情報通信学会信学技法 IN2003-118, 2003年 11月

14. “Practical Implementation of Geographic Routing for Mobile Ad Hoc Networks”竹内彰次郎,ベルネルクレイセル,山崎浩輔,瀬崎薫,安田靖彦電子情報通信学会信学技法 NS2003-274, 2004年 3月

15. “TXOP Limit Selection Mechanism to Protect High Priority Traffic in IEEE802,11e”Shojiro Takeuchi, Kaoru Sezaki and Yasuhiko Yasuda電子情報通信学会次世代ネットワークソフトウェア研究会, 2004年 11月

16. “センサーネットワークにおけるパケット衝突抑制型MACプロトコル”関根理敏,竹内彰次郎,瀬崎薫電子情報通信学会信学技法 NS2004-65, 2004年 7月

17. “低消費電力MACプロトコルにおけるスロット予約期間適応的制御”関根理敏,竹内彰次郎,瀬崎薫電子情報通信学会信学技法 NS2005-2, 2005年 5月

18. “送信電力を考慮した Battery Cost Routingの提案”竹内彰次郎,山崎浩輔,瀬崎薫,安田靖彦電子情報通信学会全国大会 2002年 3月

19. “送信電力を考慮した Battery Cost Routingにおけるパケット伝送の一検討”竹内彰次郎,山崎浩輔,瀬崎薫,安田靖彦電子情報通信学会ソサエティ大会シンポジウム 2002年 9月

129

20. “Mobiel Ad Hoc Network Routing:A Proposal of a Routing Algorithm for Mobile Multi-Hop Wireless Networks”Werner Creixell, Kosuke Yamazaki, Shojiro Takeuchi, Kaoru Sezaki電子情報通信学会ソサエティ大会シンポジウム 2002年 9月

21. “地理的経路制御における適応的電力制御手法”竹内彰次郎,山崎浩輔,瀬崎薫,安田靖彦電子情報通信学会全国大会 2003年 3月

22. センサーネットワークにおけるパケット衝突抑制アクセス制御手法”関根理敏,竹内彰次郎,瀬崎薫電子情報通信学会全国大会 2004年 3月

23. “MAC層におけるスロット割り当て制御期間の適応的制御に関する一検討”関根理敏,竹内彰次郎,瀬崎薫電子情報通信学会全国大会 2005年 3月

24. “低消費電力スケジューリングベースMACプロトコルの実装”関根理敏,竹内彰次郎,瀬崎薫電子情報通信学会ソサエティ大会 2005年 9月

130

Bibliography

[1] “Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications”,IEEE Std 802.11 -1999 (Reaff 2003), 2003.

[2] “Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifica-tions: Medium Access Control (MAC) Quality of Service (QoS) Enhancements”, IEEEP802.11e/D12.0,Nov. 2004.

[3] “Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications:high-speed physical layer in the 5 GHz band”, IEEE Std. 802.11a-1999, 1999.

[4] “Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications:Higher-Speed Physical Layer Extension in the 2.4 GHz Band”, IEEE Std. 802.11b-1999,1999.

[5] “Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications:Further Higher Data Rate Extension in the 2.4 GHz Band”, IEEE Std. 802.11g-2003, 2003.

[6] “Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications:Medium Access Control (MAC) Security Enhancements”, IEEE Std. 802.11i-2004, 2004.

[7] “Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications:Radio Resource Measurement”, IEEE P802.11k/D1.0, July 2004, 2004.

[8] “IEEE802.11n Working Group “, http://www.grouper.ieee.org/groups/802/11/, Dec. 2005.

[9] “IEEE802.11s Working Group”, http://www.grouper.ieee.org/groups/802/11/, Dec. 2005.

[10] M. Stemm and R. H. Katz,“Measuring and reducing energy consumption of network inter-faces in hand held devices”, IEICE Transactions on Communications, E80-B(8), pp. 1125 -1131, Aug. 1997.

[11] Laura Marie Feeney and Martin Nilsson,“Investigating the Energy Consumption of a Wire-less Network Interface in an Ad Hoc Networking Environment”, IEEE INFOCOM 2001.

[12] The editors of IEEE802.11,“Wireless LAN Medium Access Control (MAC) and PhysicalLayer (PHY) Specification”1999.

131

[13] Benjie Chen, Kyle Jamieson, Hari Balakrishnan, Robert Morris,“Span: An Energy-EfficientCoordination Algorithm for Topology Maintenance in Ad Hoc Wireless Networks”, ACMMobiCom 2001

[14] Y. Xu, J. Heidemann, and D. Estrin,“Geography-informed energy conservation for adhoc routing”, ACM/IEEE 7th Int l Conf. on Mobile Computing and Networking (Mobi-Com2001), July 2001.

[15] Rong Zheng and Robin Kravets“On-demand Power Management for Ad Hoc Networks”,IEEE INFOCOM 2003.

[16] Eun-Sun Jung and Nitin H. Vaidya,“A Power Saving MAC Protocol for Wireless Networks”,Tech. Rep. University of Illinois at Urbana Champaign, 2002.

[17] H. Woesner, J.-P. Ebert, M. Schlager and A. Wolisz,“Power Saving Mechanisms in Emerg-ing Standards for Wireless LANs : The MAC Level Perspective”, IEEE Personal Communi-cations, June 1998.

[18] http://www.isi.edu/nsnam/ns/index.html, Dec. 2005.

[19] M. Papadopouli and Henning Schulzrinne,“Network Connection Sharing in an Ad Hoc Wire-less Network among Collaborative Hosts”, NOSSDAV, June 1999.

[20] J Gomez, A. Campbell, M. Naghsineh and C. Bisdikian,“Power-aware routing in wirelesspacket networks”, In Proc. of Sixth IEEE International Workshop on Mobile MultimediaCommunications, San Diego, CA, November 1999.

[21] R. Wattenhofer, L. Li, P. Bahl and Y.-M. Wang,“Distributed topology control for powerefficient operation in multihop wireless ad hoc networks”, In Proc. of IEEE INFOCOM 2001.

[22] S. Sihgh, M. Woo and C. S. Raghavendra,“Power-Aware Routing in Mobile Ad Hoc Net-works”, Proc. of Mobicom 1998.

[23] C.-K. Toh, “Maximum Battery Life Routing to Support Ubiquitous Mobile Computing inWireless Ad Hoc Networks”, IEEE Communications Magazine, pp. 138-147, June 2001.

[24] S. Singh and C. S. Raghavendra,“Pamas-power aware multi-access protocol with signalingfor ad hoc networks”, ACM Commun. Rev. July 1998.

[25] David B. Johnson and David A. Maltz,“Dynamic Source Routing in Ad Hoc Wireless Net-works”, ,Kluwer Academic Punlishers pp. 153 - 181, 1996.

[26] Yang Xiao, Haizhon Li, and Sunghyun Choi“Protection and Guarantee for Voice and VideoTraffic in IEEE 802.11e Wireless LANs”, IEEE INFOCOM 2004, March 2004.

[27] Yu-Kiang Kuo, Chi-Hung Lu, Eric Hsiao-Kuang, and Gen-Huey Chen,“An Admission Con-trol Strategy for Differentiated Services in IEEE 802.11”, IEEE Globecom 2003, Dec. 2003.

132

[28] G. Bianchi,“Performance analysis of the IEEE 802.11 distributed coordination function”,IEEE Journal of Selected Areas on Communication Vol.18, no.3, pp. 785-799, Dec. 2000.

[29] Y. Chen, N. Smavatkul and S. Emeott,“Power Management for VoIP over IEEE 802.11WLAN”, in Proc. of IEEE WCNC 2004, vol.5, no.1, March 2004.

[30] K. SRIRAM and W. Whitt,“Characterizing Superposition Arrival Process in Packet Mul-tiplexes for Voice and Data”, IEEE Journal on Selected Areas in Communications, Vol.4,No.6, Sep. 1986.

[31] J. W. Robinson and T. S. Randhawa,“Saturation Throughput Analysis of IEEE 802.11e En-hanced Distributed Coordination Function”, in IEEE Journal on Selected Areas in Commu-nications, Vol.22, No.5, June, 2004.

[32] S. Takeuchi, K. Sezaki and Y. Yasuda,“An Improved Power Saving Mechanism for MACProtocol in Ad Hoc Networks”, in Proc. of IEEE Globecom 2004, Nov. 2004.

[33] M. Kappes, A.S. Krishnakumar and P. Krishnan, “Estimating Signal Strength Coverage for aWireless Access Point”, in Proc. of IEEE Globecom 2004, Nov. 2004.

[34] A. Balachandran, P. Bahl and G.M. Voelker, “Hop-Spot Congestion Relief in Public-areaWireless Networks,” in Proc. of WMCSA, June 2002.

[35] I. Ramani and S. Savage,“SyncScan: Practical Fast Handoff for 802.11 Infrastructure Net-works”, in Proc. of IEEE INFOCOM 2005, March 2005

[36] I. Papanikos and M. Logothetis,“A study on dynamic load balance for IEEE 802.11b wirelessLAN,” in Proc. of COMCON 2001, 2001.

[37] Y. Fukuda and Y. Oie,“Decentralized Access Point Selection Architecture for Wireless LANs-Deployability and Robustness”, in Proc. of IEEE VTC2004-fall, Sep. 2004.

[38] K. Saitoh, Y. Inoue, M. Iizuka and M. Morikura,“An Effective Data Transfer Method ByIntegrating Priority Control into Multirate Mechanisms for IEEE802.11 Wireless LANs”, inProc. of IEEE VTC2002-spring, March 2002.

[39] J.D.P.Pavon and S.Choi, “Link Adaptation Strategy for IEEE802.11 WLAN via ReceivedSignal Strength Measurement”, in Proc. of IEEE ICC2003, June 2003.

133