<%BANNER%>

Improving Fairness and Throughput in Wireless Systems

Permanent Link: http://ufdc.ufl.edu/UFE0042502/00001

Material Information

Title: Improving Fairness and Throughput in Wireless Systems
Physical Description: 1 online resource (109 p.)
Language: english
Creator: Zhang, Ming
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2010

Subjects

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

Notes

Abstract: With the advancement of wireless technologies, wireless systems have been widely used in today's world. Especially, the IEEE 802.11 Wireless LANs (WLANs) have covered a large portion of the urban areas to provide anytime, anywhere Internet service. In addition, the Radio-Frequency Identification system (RFID) is another important wireless network which promises to revolutionize the inventory management in large warehouses, retail stores, hospitals, transportation systems, etc. In this dissertation, we first propose novel solutions for improving fairness and throughput in WLANs. We then introduce a new method to improve reading throughput in large RFID systems. Our first work focuses on achieving MAC-layer time fairness among contending WLANs. The WLANs may overlap and contend with each other. We show that the contention among nearby WLANs is location-sensitive, which makes some hosts much more capable than others to obtain the channel for their transmissions. Another reality is that wireless hosts use different transmission rates to communicate with the access points due to attenuation of their signals. We show that location-sensitive contention aggravates the throughput anomaly caused by different transmission rates. It can cause throughput degradation and host starvation. Achieving time fairness across multiple WLANs is a very difficult problem because the hosts may perceive very different channel conditions and they may not be able to communicate and coordinate their operations due to the disparity between the interference range and the transmission range. In this work, we design a MAC-layer time fairness solution based on two novel techniques: channel occupancy adaptation, which applies AIMD on the channel occupancy of each flow, and Queue Spreading (QS), which ensures that all hosts and only those hosts in a saturated channel detect congestion and reduce their channel occupancies in response. The proposed solution is called AIMD/QS+k. We show that AIMD/QS+k approximates the generic adaptation algorithm for proportional fairness. Our second work focuses on achieving transport-layer fairness among contending WLANs. TCP is the dominating transport-layer protocol used by many applications over WLANs. Contention among multiple nearby WLANs in urban areas may cause severe TCP unfairness, where some TCP flows can achieve very high throughput at the expense of starving others. This unfairness results from the fact that different physical nodes conveying TCP flows at a wireless bottleneck may have different channel observations and consequently they may provide inconsistent feedbacks to the TCP sources. Existing solutions to this problem try to synchronize channel observations of contending nodes by exchanging control messages among them. They rely on the assumption that these nodes are within each other's transmission range, which however may not always hold. In this work, we design a new protocol, called Wireless Probabilistic Drop (WPD), to improve TCP fairness without requiring direct communication among nodes. In WPD, when a node detects congestion, it probabilistically chooses to either drop some packets to resolve the congestion, or aggressively spread the congestion signal to other contending nodes. Each node makes the choice with a probability that is proportional to its flow rate. Henceforth, high-rate flows tend to perform rate reduction more often, and low-rate flows are more likely to increase their flow rates. Eventually, all flows passing the bottleneck are expected to get a fair share of the channel bandwidth. Our third work focuses on improving the RFID reading throughput. In large RFID systems, periodically reading the IDs of the tags is an important function to guard against administration error, vendor fraud and employee theft. Given the low-speed communication channel in which a RFID system operates, the reading throughput is one of the most important performance metrics. The current protocols have reached the physical throughput limit that can possibly be achieved based on their design methods. To break that limit, we have to apply fundamentally different approaches. In this work, we investigate how much throughput improvement the analog network coding can bring when it is integrated into the RFID protocols. The idea is to extract useful information from collision slots when multiple tags transmit their IDs simultaneously. Traditionally, those slots are discarded. With analog network coding, we show that a collision slot is almost as useful as a non-collision slot in which exactly one tag transmits. We propose the Framed Collision-Aware Tag identification protocol (FCAT) that optimally applies analog network coding to maximize the reading throughput, which is 51.1%-70.6% higher than the best existing protocols.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Ming Zhang.
Thesis: Thesis (Ph.D.)--University of Florida, 2010.
Local: Adviser: Chen, Shigang.

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2010
System ID: UFE0042502:00001

Permanent Link: http://ufdc.ufl.edu/UFE0042502/00001

Material Information

Title: Improving Fairness and Throughput in Wireless Systems
Physical Description: 1 online resource (109 p.)
Language: english
Creator: Zhang, Ming
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2010

Subjects

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

Notes

Abstract: With the advancement of wireless technologies, wireless systems have been widely used in today's world. Especially, the IEEE 802.11 Wireless LANs (WLANs) have covered a large portion of the urban areas to provide anytime, anywhere Internet service. In addition, the Radio-Frequency Identification system (RFID) is another important wireless network which promises to revolutionize the inventory management in large warehouses, retail stores, hospitals, transportation systems, etc. In this dissertation, we first propose novel solutions for improving fairness and throughput in WLANs. We then introduce a new method to improve reading throughput in large RFID systems. Our first work focuses on achieving MAC-layer time fairness among contending WLANs. The WLANs may overlap and contend with each other. We show that the contention among nearby WLANs is location-sensitive, which makes some hosts much more capable than others to obtain the channel for their transmissions. Another reality is that wireless hosts use different transmission rates to communicate with the access points due to attenuation of their signals. We show that location-sensitive contention aggravates the throughput anomaly caused by different transmission rates. It can cause throughput degradation and host starvation. Achieving time fairness across multiple WLANs is a very difficult problem because the hosts may perceive very different channel conditions and they may not be able to communicate and coordinate their operations due to the disparity between the interference range and the transmission range. In this work, we design a MAC-layer time fairness solution based on two novel techniques: channel occupancy adaptation, which applies AIMD on the channel occupancy of each flow, and Queue Spreading (QS), which ensures that all hosts and only those hosts in a saturated channel detect congestion and reduce their channel occupancies in response. The proposed solution is called AIMD/QS+k. We show that AIMD/QS+k approximates the generic adaptation algorithm for proportional fairness. Our second work focuses on achieving transport-layer fairness among contending WLANs. TCP is the dominating transport-layer protocol used by many applications over WLANs. Contention among multiple nearby WLANs in urban areas may cause severe TCP unfairness, where some TCP flows can achieve very high throughput at the expense of starving others. This unfairness results from the fact that different physical nodes conveying TCP flows at a wireless bottleneck may have different channel observations and consequently they may provide inconsistent feedbacks to the TCP sources. Existing solutions to this problem try to synchronize channel observations of contending nodes by exchanging control messages among them. They rely on the assumption that these nodes are within each other's transmission range, which however may not always hold. In this work, we design a new protocol, called Wireless Probabilistic Drop (WPD), to improve TCP fairness without requiring direct communication among nodes. In WPD, when a node detects congestion, it probabilistically chooses to either drop some packets to resolve the congestion, or aggressively spread the congestion signal to other contending nodes. Each node makes the choice with a probability that is proportional to its flow rate. Henceforth, high-rate flows tend to perform rate reduction more often, and low-rate flows are more likely to increase their flow rates. Eventually, all flows passing the bottleneck are expected to get a fair share of the channel bandwidth. Our third work focuses on improving the RFID reading throughput. In large RFID systems, periodically reading the IDs of the tags is an important function to guard against administration error, vendor fraud and employee theft. Given the low-speed communication channel in which a RFID system operates, the reading throughput is one of the most important performance metrics. The current protocols have reached the physical throughput limit that can possibly be achieved based on their design methods. To break that limit, we have to apply fundamentally different approaches. In this work, we investigate how much throughput improvement the analog network coding can bring when it is integrated into the RFID protocols. The idea is to extract useful information from collision slots when multiple tags transmit their IDs simultaneously. Traditionally, those slots are discarded. With analog network coding, we show that a collision slot is almost as useful as a non-collision slot in which exactly one tag transmits. We propose the Framed Collision-Aware Tag identification protocol (FCAT) that optimally applies analog network coding to maximize the reading throughput, which is 51.1%-70.6% higher than the best existing protocols.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Ming Zhang.
Thesis: Thesis (Ph.D.)--University of Florida, 2010.
Local: Adviser: Chen, Shigang.

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2010
System ID: UFE0042502:00001


This item has the following downloads:


Full Text

PAGE 1

IMPROVINGFAIRNESSANDTHROUGHPUTINWIRELESSSYSTEMSByMINGZHANGADISSERTATIONPRESENTEDTOTHEGRADUATESCHOOLOFTHEUNIVERSITYOFFLORIDAINPARTIALFULFILLMENTOFTHEREQUIREMENTSFORTHEDEGREEOFDOCTOROFPHILOSOPHYUNIVERSITYOFFLORIDA2010

PAGE 2

c2010MingZhang 2

PAGE 3

Tomyfather 3

PAGE 4

ACKNOWLEDGMENTS Iwouldliketothankmyadvisor,Prof.ShigangChen,forhisconstantguidance,support,andencouragement.Iamprivilegedtohavesuchawonderfuladvisor,whoisatalltimesenthusiastic,optimistic,patient,helpful,andencouraging.Hegavemeextensiveadviceandinsightduringthecourseofmyresearchwork.MyspecialthanksgotoProf.SartajSahni,Prof.Jih-KwonPeir,Prof.AhmedHelmy,andProf.TanWong,fortheirinstructivecommentsandsupportduringmywork.IwouldalsoliketothankallmycolleaguesinProf.Chen'sresearchgroup,includingZhanZhang,LiangZhang,MyungKeunYoon,YingJian,TaoLi,WenLuo,YanQiao,ZhenMoandShuangLin,forprovidingahighlevelofresearchsupport.Iwanttoexpressmydeepestlovetomydarlingwife,XuelianXiao,mymotherXingguoLi,myfatherXianchangZhang,andmyangelOliviaLuoqiaoZhang.Theirlove,understanding,andencouragementhavealwaysbeenthestrongestsupporttome. 4

PAGE 5

TABLEOFCONTENTS page ACKNOWLEDGMENTS .................................. 4 LISTOFTABLES ...................................... 8 LISTOFFIGURES ..................................... 9 ABSTRACT ......................................... 11 CHAPTER 1INTRODUCTION ................................... 14 1.1MAC-layerTimeFairnessacrossMultipleWirelessLANs ......... 15 1.2TCPFairnessacrossMultipleWirelessLANs ................ 18 1.3ImprovingReadingThroughputinLargeRFIDSystems .......... 20 2MAC-LAYERTIMEFAIRNESSACROSSMULTIPLEWIRELESSLANS .... 25 2.1NetworkModelandProblemDenition .................... 25 2.1.1NetworkModel ............................. 25 2.1.2ProblemDenition ........................... 27 2.2Time-AllocationAnomalyandLocation-SensitiveContention ....... 29 2.2.1Time-AllocationAnomaly ........................ 29 2.2.2Location-SensitiveContention ..................... 30 2.3StateoftheArt ................................. 30 2.3.1Type-I:AssumeKnowledgeofContendingFlows .......... 31 2.3.2Type-II:AssumeSameChannelPerception ............. 31 2.3.3Type-III:AssumenoKnowledgeofContendingNodesandDifferentChannelPerception ........................... 32 2.3.4OtherRelatedWork .......................... 34 2.4ASolutionforMAC-layerTimeFairnessacrossMultipleWLANs ..... 35 2.4.1Overview ................................ 35 2.4.2ReleaseRateandChannelOccupancyAdaptation ......... 36 2.4.3QueueSpreading ............................ 38 2.4.4AIMD/QS+k ............................... 40 2.4.5ProportionalFairnessandInterpretationofAIMD/QS+k ...... 42 2.5Simulation .................................... 43 2.5.1OneContentionGroup ......................... 44 2.5.2TwoContentionGroups ........................ 45 2.5.3MultipleContentionGroups ...................... 46 2.6Summary .................................... 46 5

PAGE 6

3TCPFAIRNESSACROSSMULTIPLEWIRELESSLANS ............ 54 3.1NetworkModel ................................. 54 3.2TCPUnfairnessamongContendingWLANs ................. 56 3.3StateoftheArt ................................. 57 3.3.1RandomEarlyDetection(RED) .................... 57 3.3.2NeighborhoodRED(NRED) ...................... 57 3.3.3ProportionalIncreaseSynchronizedMultiplicativeDecrease(PISD) ...................................... 58 3.3.4OtherRelatedWork .......................... 59 3.4WirelessProbabilisticDrop(WPD) ...................... 60 3.4.1BasicIdea ................................ 60 3.4.2PeriodicalMeasurementofStateInformation ............ 61 3.4.3AggressiveMode ............................ 61 3.4.4ProbabilisticDropping ......................... 62 3.4.5MinimumRateAssurance ....................... 62 3.4.6AdaptiveIntermittentRelease ..................... 63 3.4.7WPDProtocol .............................. 64 3.5Simulation .................................... 65 3.5.1SimulationSetup ............................ 66 3.5.2FairnessIndex ............................. 66 3.5.3CaseStudy:ABaseScenario ..................... 66 3.5.4ScalabilityStudy:ThreeContentionGroups ............. 67 3.5.5AStreetScenario ............................ 68 3.6Summary .................................... 68 4IMPROVINGREADINGTHROUGHPUTINLARGERFIDSYSTEMS ..... 74 4.1Background ................................... 74 4.1.1Motivation ................................ 74 4.1.2AnalogNetworkCoding(ANC) .................... 76 4.2TerminologyandProblemDenition ..................... 78 4.2.1Terminology ............................... 78 4.2.2ResolvableCollisionSlots ....................... 78 4.2.3ProblemDenition ........................... 78 4.3StateoftheArt ................................. 80 4.4SlottedCollision-AwareTagIdenticationProtocol(SCAT) ......... 82 4.4.1ProtocolDescription .......................... 82 4.4.2CollisionResolution .......................... 83 4.4.3DeterminingtheOptimalValueforpi ................. 83 4.4.4PseudoCode .............................. 85 4.4.5UnresolvableCollisionSlotsandChannelError ........... 86 4.5FramedCollision-AwareTagIdenticationProtocol(FCAT) ......... 87 4.5.1InefcienciesofSCAT ......................... 87 4.5.2UsingFrames .............................. 88 4.5.3EstimatingtheNumberofTagswithinFCAT ............. 89 6

PAGE 7

4.5.4EstimationVariance,V(^Ni Ni) ...................... 92 4.6SimulationResults ............................... 94 4.6.1ReadingThroughputComparison ................... 94 4.6.2EffectivenessofCollisionResolution ................. 95 4.6.3ReportProbability ........................... 95 4.6.4ImpactofFrameSize .......................... 96 4.7Summary .................................... 96 5CONCLUSION .................................... 101 REFERENCES ....................................... 102 BIOGRAPHICALSKETCH ................................ 109 7

PAGE 8

LISTOFTABLES Table page 2-1Deliveryrate(inpacketspersecond)andchanneloccupancyunderPISDontheNetworkofFig. 2-9 ................................ 47 2-2Comparingthedeliveryrates(inpackets/sec)andthechanneloccupanciesoftheowsinthenetworkofFig. 2-9 under802.11DCF,CWSP,IdleSense,andAIMD/QS+k. ................................... 50 2-3Comparingthedeliveryrates(inpackets/sec)andthechanneloccupanciesinthenetworkofFig. 2-13 under802.11DCF,CWSP,andAIMD/QS+k. .... 53 3-1Comparingtheowrates(inpackets/sec)oftheowsinthenetworkofFig. 3-6 underPISD. ...................................... 71 3-2Comparingtheowrates(inpackets/sec)oftheowsinthenetworkofFig. 3-3 underDropTail,NREDandWPD. .......................... 71 3-3ComparingtheFairnessIndexinTermsofProportionalFairnessinthenetworkofFig. 3-3 underDropTail,NREDandWPD. .................... 71 3-4Comparingtheowrates(inpackets/sec)oftheowsinthenetworkofFig. 3-7 underDropTail,NREDandWPD. .......................... 72 3-5Comparingtheowrates(inpackets/sec)oftheowsinthenetworkofFig. 3-8 underDropTail,NREDandWPD. .......................... 73 4-1ReadingthroughputcomparisonwhenNvariesfrom1,000to20,000 ...... 98 4-2Empty,SingletonandCollisionTimeSlotswhenN=10000 ........... 99 4-3TagIDsResolvedfromCollisionSlots ....................... 99 4-4Thecomputedvalueof!matchescloselywiththeoptimalvalueof!obtainedbysimulation. ..................................... 99 8

PAGE 9

LISTOFFIGURES Figure page 1-1ThreepartiallyoverlappingWLANsformtwocontentiongroupsg1andg2,wheres1,s2ands3arethreeaccesspoints.ATCPowfromaserverontheInternetpassesthroughanaccesspointtoawirelessclientineachWLAN.Notethat,onlythewirelesspartofaow(fromanaccesspointtoawirelessclient)isdrawninthegure. .................................. 24 2-1NetworkModel. .................................... 47 2-2Twocontentiongroups.Flows(w,z)and(u,v)cantransmitsimultaneously. 47 2-3TwoMACowsinthesameWLAN. ........................ 47 2-4DCFonthenetworkinFig. 2-3 withow(x,y)at11Mbpsandow(x,z)at1Mbps.A)Thedeliveryrates.B)Thechanneloccupancies. ........... 48 2-5TwoMACowsindifferentWLANscontend. .................... 48 2-6ChanneloccupanciesoftheowsinFig. 2-5 underDCF.A)Bothowstransmitat11Mbps.B)Flow(x,y)transmitsat11Mbpsandow(u,v)at1Mbps. .. 48 2-7IdleSenseonthenetworkinFig. 2-5 withow(x,y)at11Mbpsandow(u,v)at1Mbps.A)thedeliveryrates.B)thechanneloccupancies. .......... 49 2-8CWSPonthenetworkinFig. 2-5 withow(x,y)at11Mbpsandow(u,v)at1Mbps.A)thedeliveryrates.B)thechanneloccupancies. .......... 49 2-9ThreeWLANsformtwocontentiongroupsg1andg2. ............... 49 2-10Comparingthechanneloccupanciesofows(x,y)and(u,v)inthenetworkofFig. 2-5 .Thetransmissionratesof(x,y)and(u,v)are11Mbpsand1Mbps,respectively.A)Under802.11DCF.B)UnderCWSP.C)UnderIdleSense.D)UnderAIMD/QS+k. ................................. 50 2-11SameasthecaptionofFig. 2-10 ,butthistimeRTS/CTSisturnedoff.A)Under802.11DCF.B)UnderCWSP.C)UnderIdleSense.D)UnderAIMD/QS+k. .. 51 2-12Comparingthechanneloccupanciesofows(x,y)and(u,v)inthenetworkofFig. 2-5 .Thetransmissionratesof(x,y)and(u,v)areare11Mbpsand5.5Mbps,respectively.RTS/CTSisturnedoff.A)Under802.11DCF.B)UnderCWSP.C)UnderIdleSense.D)UnderAIMD/QS+k. ............... 52 2-13SixteenWLANsaredeployedalongtwostreets.Unlessspeciedinthegure,thedefaulttransmissionrateofaowis11Mbps. ................ 53 3-1WhyTCPAIMDdoesn'tworkinwirelessnetworks?A)TCPsynchronizedAIMDinwirednetworks.B)TCPunfairAIMDinwirelessnetworks. ....... 69 9

PAGE 10

3-2TwocontendingWLANs,wheres1ands2aretwoaccesspoints.ATCPowfromaserverontheInternetpassesthroughanaccesspointtoawirelessnodeineachWLAN.Notethat,onlythewirelesspartofeachow(fromanaccesspointtoawirelessclient)isdrawninallguresthroughoutthischapter. 69 3-3ThreepartiallyoverlappingWLANsformtwocontentiongroupsg1andg2,wheres1,s2ands3arethreeaccesspoints.ATCPowfromaserverontheInternetpassesthroughanaccesspointtoawirelessclientineachWLAN. ....... 69 3-4Thechannelidletimesensedbythethreesenderss1,s2ands3inFig. 3-3 underNRED. ..................................... 70 3-5TheowratesofthethreeTCPows,f1,f2andf3inFig. 3-3 underNRED. .. 70 3-6FiveTCPowsformtwocontentiongroups. .................... 70 3-7SixWLANsformthreecontentiongroupsandeachWLANcontainsoneTCPow. .......................................... 71 3-8TwentyfourWLANsarerandomlydeployedalongtwocrossingstreetsandeachWLANcontainsoneTCPow. ........................ 72 4-1Thisexampleshowsthatacollision-resolutionprotocolmayreducethenumberoftimeslotsusedtoidentifyfourtagsfrom11timeslotsto6timeslots.A)Contention-basedprotocol.B)Collision-resolutionprotocol. ........... 97 4-2Alice-BobexampleforAnalogNetworkCoding. .................. 97 4-3Therelativebiasof^Niwithrespecttothenumberoftags. ............ 97 4-4Thenumberoftags,Ni,isnotamonotonicfunctioninE(n1).Parameters:pi=1.414=Niandf=30. .............................. 98 4-5FCATreadingthroughputwithrespectto!. .................... 99 4-6ThereadingthroughputofFCATisstabilizedwhenf10. ........... 100 10

PAGE 11

AbstractofDissertationPresentedtotheGraduateSchooloftheUniversityofFloridainPartialFulllmentoftheRequirementsfortheDegreeofDoctorofPhilosophyIMPROVINGFAIRNESSANDTHROUGHPUTINWIRELESSSYSTEMSByMingZhangDecember2010Chair:ShigangChenMajor:ComputerEngineering Withtheadvancementofwirelesstechnologies,wirelesssystemshavebeenwidelyusedintoday'sworld.Especially,theIEEE802.11WirelessLANs(WLANs)havecoveredalargeportionoftheurbanareastoprovideanytime,anywhereInternetservice.Inaddition,theRadio-FrequencyIdenticationsystem(RFID)isanotherimportantwirelessnetworkwhichpromisestorevolutionizetheinventorymanagementinlargewarehouses,retailstores,hospitals,transportationsystems,etc.Inthisdissertation,werstproposenovelsolutionsforimprovingfairnessandthroughputinWLANs.WethenintroduceanewmethodtoimprovereadingthroughputinlargeRFIDsystems. OurrstworkfocusesonachievingMAC-layertimefairnessamongcontendingWLANs.TheWLANsmayoverlapandcontendwitheachother.WeshowthatthecontentionamongnearbyWLANsislocation-sensitive,whichmakessomehostsmuchmorecapablethanotherstoobtainthechannelfortheirtransmissions.Anotherrealityisthatwirelesshostsusedifferenttransmissionratestocommunicatewiththeaccesspointsduetoattenuationoftheirsignals.Weshowthatlocation-sensitivecontentionaggravatesthethroughputanomalycausedbydifferenttransmissionrates.Itcancausethroughputdegradationandhoststarvation.AchievingtimefairnessacrossmultipleWLANsisaverydifcultproblembecausethehostsmayperceiveverydifferentchannelconditionsandtheymaynotbeabletocommunicateandcoordinatetheiroperations 11

PAGE 12

duetothedisparitybetweentheinterferencerangeandthetransmissionrange.Inthiswork,wedesignaMAC-layertimefairnesssolutionbasedontwonoveltechniques:channeloccupancyadaptation,whichappliesAIMDonthechanneloccupancyofeachow,andQueueSpreading(QS),whichensuresthatallhostsandonlythosehostsinasaturatedchanneldetectcongestionandreducetheirchanneloccupanciesinresponse.TheproposedsolutioniscalledAIMD/QS+k.WeshowthatAIMD/QS+kapproximatesthegenericadaptationalgorithmforproportionalfairness. Oursecondworkfocusesonachievingtransport-layerfairnessamongcontendingWLANs.TCPisthedominatingtransport-layerprotocolusedbymanyapplicationsoverWLANs.ContentionamongmultiplenearbyWLANsinurbanareasmaycausesevereTCPunfairness,wheresomeTCPowscanachieveveryhighthroughputattheexpenseofstarvingothers.ThisunfairnessresultsfromthefactthatdifferentphysicalnodesconveyingTCPowsatawirelessbottleneckmayhavedifferentchannelobservationsandconsequentlytheymayprovideinconsistentfeedbackstotheTCPsources.Existingsolutionstothisproblemtrytosynchronizechannelobservationsofcontendingnodesbyexchangingcontrolmessagesamongthem.Theyrelyontheassumptionthatthesenodesarewithineachother'stransmissionrange,whichhowevermaynotalwayshold.Inthiswork,wedesignanewprotocol,calledWirelessProbabilisticDrop(WPD),toimproveTCPfairnesswithoutrequiringdirectcommunicationamongnodes.InWPD,whenanodedetectscongestion,itprobabilisticallychoosestoeitherdropsomepacketstoresolvethecongestion,oraggressivelyspreadthecongestionsignaltoothercontendingnodes.Eachnodemakesthechoicewithaprobabilitythatisproportionaltoitsowrate.Henceforth,high-rateowstendtoperformratereductionmoreoften,andlow-rateowsaremorelikelytoincreasetheirowrates.Eventually,allowspassingthebottleneckareexpectedtogetafairshareofthechannelbandwidth. 12

PAGE 13

OurthirdworkfocusesonimprovingtheRFIDreadingthroughput.InlargeRFIDsystems,periodicallyreadingtheIDsofthetagsisanimportantfunctiontoguardagainstadministrationerror,vendorfraudandemployeetheft.Giventhelow-speedcommunicationchannelinwhichaRFIDsystemoperates,thereadingthroughputisoneofthemostimportantperformancemetrics.Thecurrentprotocolshavereachedthephysicalthroughputlimitthatcanpossiblybeachievedbasedontheirdesignmethods.Tobreakthatlimit,wehavetoapplyfundamentallydifferentapproaches.Inthiswork,weinvestigatehowmuchthroughputimprovementtheanalognetworkcodingcanbringwhenitisintegratedintotheRFIDprotocols.TheideaistoextractusefulinformationfromcollisionslotswhenmultipletagstransmittheirIDssimultaneously.Traditionally,thoseslotsarediscarded.Withanalognetworkcoding,weshowthatacollisionslotisalmostasusefulasanon-collisionslotinwhichexactlyonetagtransmits.WeproposetheFramedCollision-AwareTagidenticationprotocol(FCAT)thatoptimallyappliesanalognetworkcodingtomaximizethereadingthroughput,whichis51.1%70.6%higherthanthebestexistingprotocols. 13

PAGE 14

CHAPTER1INTRODUCTION IEEE802.11WirelessLANs(WLANs)havebeendenselydeployedinmanyurbanareas[ 2 60 ]tosatisfythedemandforInternetaccessfromanywhereatanytime.InatypicalWLAN,awirelessrouterservingasanaccesspointisconnectedtotheInternetviaacablemodem.Severalwirelessclientssuchaslaptops,mobilephonesandotherwireless-enableddevicesareconnectedtotheaccesspointwirelessly.SomanyWLANsarelocatednearoneanother.IthasbeenobservedthatseverefairnessproblemsmayoccuramongnearbyWLANssharingthesamechannel.WewillstartfromtheMAC-layertostudytheintriguinginterplaybetweenlocation-sensitivecontentionandtime-allocationanomalyin802.11DCFnetworks.Then,wewillmovetothetransport-layerandproposeanewprotocolforimprovingTCPfairnessamongnearbycontendingWLANs. Radio-FrequencyIdenticationsystems(RFID)havebroughtrevolutionarychangetotheinventorymanagementinlargewarehouses,retailstores,hospitals,transportationsystems,etc.ARFIDsysteminvolvesafewreaders(interrogators)andalargenumberoftags(labels).Thetagcontainsanintegratedcircuitforstoringandprocessingdataandanantennaforcommunicatingwiththereader.AuniqueIDisassignedtoeachtagandatagisattachedtoanitem.ByaccessingtheIDsofthetags,thereadercanwirelesslyidentifyandtracktheitemsfromadistanceevenwithoutlineofsight.Inalargewarehouse,itisrequiredtoperiodicallyreadtheIDsofthetagstoguardagainstadministrationerror,vendorfraudandemployeetheft.BecausetheRFIDisoperatedinalow-speedwirelesschannelandthenumberoftagsisexpectedtobelarge,thereadingthroughputbecomesacriticalperformancemetric.WewillinvestigateafundamentallydifferentapproachtoboosttheRFIDreadingthroughput. 14

PAGE 15

1.1MAC-layerTimeFairnessacrossMultipleWirelessLANs Citiesarenowcrowdedwithwirelessaccesspoints.Atairportterminals,ofcebuildings,homesandshops,alaptopcanoftenndseveraltoafewdozensofaccesspointsinusablerange,mostofwhichruntheIEEE802.11b/gprotocolandeachsupportsaWLAN.Withonlythreenon-overlappingchannelsin802.11b/g,nearbyWLANswillinevitablyinterferewitheachother,givingrisetotwointriguingproblemsattheMAC-layer. Therstproblemislocation-sensitivecontention[ 75 ].Thecontentionresolutionprotocol,802.11DCF(DistributedCoordinationFunction),workswellinasymmetricenvironmentwhereallhostsaredownloadingcontentviathesameaccesspoint.ButitdoesnotperformwellinasymmetricsettingsthatarecommonwhenhostsinnearbyWLANscontendinthesamechannel.Dependingontheirrelativespatiallocations,somehostsmaygainhugeadvantageinoccupyingthechannelfortheirtransmissions.Astheyobtainmostofthechannelbandwidth,hostsinotherWLANsarestarved. Location-sensitivecontentionisverydifculttodealwithbecauseoftwofundamentalwirelessproperties.Therstpropertyisthatcontentionisdenedbytheinterferenceorcarrier-sensingrange,whereascommunicationhappenswithinthetransmissionrange.Consequently,contendinghostsindifferentWLANsmaynotbeabletoexplicitlyexchangeorimplicitlyoverhearnecessaryinformationtocoordinatetheiroperations.Theymaynotevenknowwhomtheycontendwith.ThesecondpropertyisthathostsindifferentWLANsmaysenseverydifferentchannelconditions(intermsofchannelidletime,transmissionfailurerate,orbufferoccupancy)evenwhentheycontendinthesamechannel.Whentheyobservethesamechannelindifferentstates,iftheycannotcommunicate(duetotherstproperty),theirreactionsareboundtobedifferent,causingunfairness.Anextensivediscussiononthisissuecanbefoundin[ 34 ].Uptodate,theresearchonWLANfairnesshaslargelyignoredtheabovetwoproperties.Mostexistingsolutionsrelyontheassumptionthatthecontendinghostsareable 15

PAGE 16

tocommunicateoroverheareachother'stransmissionorthattheysensethesamechannelconditions.TherecentworkofPISD[ 34 ]breaksawayfromsuchassumptions.However,itrequiresthatallwirelesshostsmustbeinthesamecontentiongroup,whichisoftenuntrueinreality.ImaginealargenumberofWLANsaredeployedinacitycenter.Theypartiallyoverlaponeanothertoprovideafullcoverageofthewholearea.Thereisnowaythattheywillallmutuallycontendwitheachother.PISDcancausesevereproblemsinsuchascenario. Thesecondproblemistime-allocationanomaly.IEEE802.11ballowsfourdifferenttransmissionrates,11Mbps,5.5Mbps,2Mbpsand1Mbps.IEEE802.11gor802.11aallowseightdifferentrates.Thetransmissionrateofawirelesshostisdeterminedthroughautoratefallbackbasedonhowreliablythehostcancommunicatewiththeaccesspointatacertainrate.Thetransmissionratewillbelowerifthehostisfurtherawayfromtheaccesspointorthereareobstacles(suchaswalls)betweenthem.Itiswellknownthat,ifasinglehostinaWLANchoosesalowtransmissionrate,allotherhostsinthesameWLANsufferwithlowthroughput[ 30 ].Toaddresstheaboveanomaly,researchershaveproposedtoreplacethroughputfairnesswithtimefairness,inwhichallhostsoccupythechannelforthesamefractionoftime.However,mostpriorworkdoesnotconsidertheimpactoflocation-sensitivecontentionthatexistsamongnearbyWLANs. Ononehand,location-sensitivecontentioncanpushtime-allocationanomalytotheextreme,allowingalow-ratehostinoneWLANtoobtainanexcessiveamountofchanneltimeandstarvethehigh-ratehostsinanearbyWLAN.Ontheotherhand,location-sensitivecontentionmakestime-allocationanomalyamuchharderproblemtosolvebecausecontendinghostsindifferentWLANsmaybeoutsideofeachother'stransmissionrange(butwithintheinterferenceorcarrier-sensingrange).Unlikeamultihopwirelessnetworkthathasacommunicationpathbetweenanytwonodes,inourWLANsetting,theremaynotexistanynodebetweentwocontendinghoststorelay 16

PAGE 17

theirinformation.Mostpriorsolutionsfortimefairnesshavelargelyignoredtheimpactoflocation-sensitivecontention.Theyeitherrelyonacentralcoordinator[ 30 66 67 ],orassumethateachhosthascertainknowledgeaboutitscontendinghosts[ 35 ],thatthecontendinghostswillsensethesamechannelcondition[ 31 52 ],orthatallhostsinallWLANsmutuallycontend[ 34 ].Theseassumptionsdonotholdingeneral. Inthiswork,wesolvethetimefairnessproblemunderlocation-sensitivecontentionamongmultipleWLANs.Looselyspeaking,oursolution,calledAIMD/QS+k,ensuresthateachwirelesshostwillreceiveafairshareofthechanneltimeevenwhenithasnowaytoknowexactlywhomitcontendswith.Preciselyspeaking,AIMD/QS+kapproximatelyachievesproportionalfairnessinchanneltimeallocationforthehostsacrossmultipleWLANs.ThedesignofAIMD/QS+kisbasedontwonoveltechniques,calledchanneloccupancyadaptationandqueuespreading.Theformerappliesadditiveincreasemultiplicativedecreaseoneachhost'schanneloccupancy.Thelatteraccuratelyidentieswhichhostssaturatethechannelduringthetimeofcongestion.Forexample,supposetwohosts,xandy,indifferentWLANscausecongestion.Duetolocationadvantage,xcansendoutallitspacketsattheexpenseofloweredthroughputaty.Hence,onlyydetectscongestion.Toresolvecongestion,bothyandxshouldperformmultiplicativedecrease.Ifyreducesitschanneloccupancyalone,xwillsimplypickuptheextrabandwidthandwidentheunfairness.Withoutanymeansofcommunication,howcanymakesurethatx(butnoneofitsothercontendinghoststhatdonotcontributetothecurrentcongestion)willjointhechannel-occupancyreduction?Remarkably,thisproblemcanbesolvedinafullydistributedwaywithoutrequiringthenodestoexchangeinformation,overhear,orevenknoweachother'sexistence.WedemonstratethatAIMD/QS+kachievesalmostperfectproportionalfairnessinscenarioswheretheexistingsolutionsfail. 17

PAGE 18

1.2TCPFairnessacrossMultipleWirelessLANs TCPisthedominatingtransportlayerprotocolusedbymanyapplicationsoverWLANs[ 18 21 ].Peoplesurfwebsites,chatbyusinginstantmessengersanddownloadlesthroughTCPconnections.However,ithasbeenobservedthatTCPmaydemonstrateseveralfairnessproblemsinWLANs.Oneistheupstream/downstreamTCPunfairnesswithinasingleWLANsuchthatupstreamowsmaygainmuchmorechannelbandwidththandownstreamows[ 4 5 54 57 ].AnotheroneistheTCPunfairnessamongmultiplecontendingWLANs,wheretheTCPowsinsomeWLANsmayachieveveryhighthroughputattheexpenseofstarvingothersinnearbyWLANs[ 75 78 ].Inthiswork,wewillfocusonthelatterproblem. Basedonnetworktrafcconditions,TCPdynamicallyadjuststhesourcesendingratebyperformingAIMD(AdditiveIncreaseMultiplicativeDecrease)onitscongestionwindow.AIMDworkswellinwirednetworks,whereabottleneckrouterdetectscongestionandnotiestheTCPsourcesofthecongestion.However,itmaynotalwaysworkwellinwirelessnetworksbecauseTCPowsmaypassthroughdifferentphysicalnodesthatcontendinabottleneckchannel.Thesenodesmayhavedifferentchannelobservationsanddifferentcapabilitiesofobtainingthechannelfortransmissions[ 26 34 ].Consequently,whenonlysomenodes(butnotall)detectchannelcongestion,theTCPowsthatpassthosenodeswillreceivecongestionfeedbackbuttheowspassingothernodeswillnot.ForTCPtoachievefairness,congestionfeedbackmustbeconsistentacrossallowsthatcontendinthebottleneck.InconsistentfeedbacktotheowsourceswillrenderTCPineffectiveinitsratecontrolamongcontendingows. Fig. 1-1 showsanexampleofthreepartiallyoverlappingWLANs,wheres1,s2ands3areaccesspoints.ATCPowfromaserverontheInternetpassesthroughanaccesspointtoawirelessclientineachWLAN.NotethatweonlydrawthewirelesspartofeachTCPow(fromtheaccesspointtothewirelessclient)inallguresthroughoutthedissertation.Supposes1ands2contendandtheyformacontentiongroupg1. 18

PAGE 19

Nodess2ands3alsocontendandtheyformanothercontentiongroupg2.Buts1ands3donotcontend;theycantransmitsimultaneously.Becauses2contendswithboths1ands3,itsensesabusierchannelandislesscapableofaccessingthechannelduetoheaviercontention.Hence,whenthechannelissaturated,asitsqueuebuildsupandexceedsathreshold,s2islikelytodetectthecongestionrst.Atthetimewhens2detectscongestion,s1ands3maystillhavesmallqueuessincetheyaremorecapableofsendingouttheirpacketsduetolesscontention(asweobserveconsistentlyinourns-2simulations).Ifs2immediatelydropspacketstoinformthesourceoff2toreduceitssendingrate,thecongestionwillberesolvedwhiletheowspassings1ands3arestillperformingadditiveincrease.Inthiscase,f1andf3willseizeupthechannelbandwidthgivenupbys2.Hence,theTCPrateadaptationactuallypromotesunfairnessamongtheratesoff1,f2andf3becauseitpenalizestheowthatexperiencesheaviercontention(andthusperformsmultiplicativedecreasemorefrequently). Tosolvethisproblem,Xuetal.[ 75 ]proposedNRED(NeighborhoodRandomEarlyDetection).Thebasicideaistomakesurethatallcontendingnodesobservethesamechannelcondition.Forinstance,inFig. 1-1 ,s1,s2ands3periodicallysynchronizetheirchannelobservationsthroughexchangingcontrolmessages.Therefore,theycandetectchannelcongestionsimultaneouslyandthendroppacketstonotifytheTCPsourcestoproperlyperformratereduction,whichleadstoenhancedfairness.However,duetothedisparityamongtransmission/interference/carrier-sensingranges[ 74 ],nodesthatarenotabletodirectlycommunicatemaystillcontendinthesamechannel.InFig. 1-1 ,ifs1,s2ands3areoutsideofeachother'stransmissionrange,NREDwillfailbecausethesenodescannotsynchronizetheirchannelobservationsandprovideconsistentfeedbackstotheTCPsources. Inthiswork,weproposeanewsolution,WPD(WirelessProbabilisticDrop),forimprovingTCPfairnesswithoutrequiringanymeansofdirectcommunicationamongcontendingnodes.Itisafullydistributedsolutionthatdoesnotrequireanymodication 19

PAGE 20

totheoperationalprotocolsofTCPand802.11DCF(exceptforcertainMACparameterchangesduringcongestion).WeusetheexampleinFig. 1-1 toillustratethebasicidea.Supposes2detectscongestionrst.Insteadofdroppingpacketsimmediately,oursolutionrequiress2tomakeaprobabilisticchoicebetweentwostates:resolutionorsignaling.Intheresolutionstate,s2willdroppacketstoresolvethecongestion.Inthesignalingstate,s2willnotdroppackets,butinsteaditwillsignalothernodesaboutthecongestionbyaggressivelypushingmorepacketsintothechannel.Ass2grabsmorechannelbandwidth,s1ands3willobservethattheirqueuesbuildupandeventuallypassthethreshold.Oncetheydetectcongestion,theywillperformsimilaroperations.Inoursolution,eachnodethatdetectscongestionmakesachoicebetweenresolutionandsignalingstates;theprobabilityforchoosingtheresolutionstateisproportionaltothenode'schanneloccupancy(i.e.,thefractionoftimeduringwhichthenodeoccupiesthechannelforitstransmission).Hence,ifs2hasasmallerchanneloccupancy,itwillhaveahigherprobabilitytochoosethesignalingstate,whiles1ands3aremorelikelytochoosetheotherstateandperformpacketdropping.Consequently,f2willincreaseitssendingratemoreoftenwhilef1andf3performratereduction,whichshrinksthegapbetweentherateoff2andtheratesofotherows.Undersuchdynamics,oursimulationsdemonstratethatallcontendingowsreceivefairsharesofthechannelbandwidthoverthelongruninFig. 1-1 aswellasinothermorecomplexscenarios. 1.3ImprovingReadingThroughputinLargeRFIDSystems Thebarcodesystembringsnumerousbenetsfortheretailstores.Itspeedsupthecheckoutprocess,makesthepricechangeeasier,andallowsquickaccessforthepropertiesofeachmerchandizeitem.Italsohasseriouslimitation.Abarcodecanonlybereadincloserange.Supposeaninventorymanagementpolicyrequiresperiodicalreadingofallitemsinordertoguardagainstadministrationerror,vendorfraudandemployeetheft.Onewillhavetouseaportablelaserscannerandmanuallyreadthebarcodesoneafteranother,whichistediousanderror-prone.RFIDtags, 20

PAGE 21

whichcanbereadwirelessly,provideanidealsolutiontothisproblem[ 19 79 ].Eachtagcarriesauniqueidenticationnumber(ID),andaRFIDreadercanretrievetheIDofatagevenwhenthereareobstaclesbetweenthem.Althoughthepassivetagsaremostpopular,theyarenotsuitableforautomatedinventorymanagementinalargeareabecausetheycanonlybereadinafewmeters.Inordertoreadalltags,wehavetoeitherdeploynumerousreaders,eachcoveringasmallarea,ormanuallymoveareaderaround,whichisagaininefcientanderror-prone.Thisworkconsidersthebattery-poweredactive(orsemi-passive)tagsthatcanbereadinalongdistanceandhavemoresoftware/hardwareresourcesthanthepassivetags. ThecommunicationbetweentheRFIDreaderandthetagsisoperatedinalow-speedchannel.YetthenumberoftagsinalargeRFIDsystemisexpectedtobeverylarge.Therefore,oneofthemostcriticalperformancemetricsisthereadingthroughput,whichistheaveragenumberofuniquetagIDsthatthereadercancollectinasecond.Thecurrentprotocolshavereachedthephysicalthroughputlimitthatcanbeachievedbasedontheirdesignmethods.Inthetime-slottedALOHA-basedprotocols[ 16 42 56 63 68 71 82 ],atagtransmitsitsIDineachtimeslot(orsomeslotinaframe)withacertainprobabilitypuntilthereceiptofitsIDisacknowledgedbytheRFIDreader.ThereadingthroughputisfundamentallylimitedbytheprobabilisticcollisionthatoccursinALOHA-basednetworks.Theoptimalthroughputis1 eT,whereeisthenaturalconstantandTisthelengthofatimeslot[ 61 ].Itisachievedwhenpischosensuchthattheprobabilityforexactlyonetagtransmittingineachslotis36.8%.Theothermajorclassisthetree-basedprotocols,whichorganizethereadingprocessinabinarytreestructure[ 9 51 83 ]andimprovethereadingthroughputbybalancingthetree[ 3 51 72 ].Analyticalandsimulationresultshaveshownthatthebestperformanceofthetree-basedprotocolsiscomparabletothebestoftheALOHA-basedprotocols. TobreakthefundamentallimitoftheALOHA-basedprotocols,wehavetoresorttofundamentallydifferentapproaches.Inthiswork,weapplytherecently-proposedanalog 21

PAGE 22

networkcodingscheme[ 36 ]toRFIDsystemsandinvestigatehowsignicantlyitcanimprovethereadingthroughput. WhatlimitsthethroughputoftheALOHA-basedprotocols?Radiocollision,whichhappenswhenmorethanonetagtransmitsinaslot.Theconventionalwisdomisthatcollisionslotsdonotcarryusefulinformationandthereforethoseslotsarewasted.Thatishowevernottrue.Recentresearchshowsthat,byembracingtheinterferenceofwirelesscommunication,physical-layernetworkcodingcansignicantlyimprovethenetworkthroughput[ 81 ].Inparticular,theanalognetworkcodingscheme[ 36 ]hasbeenexperimentallyimplemented.However,itsusefulnesshasonlybeendemonstratedundertoyexamples. Thecontributionsinthisworkaretwo-fold:First,weoptimallyintegrateanalognetworkcodingintotheRFIDsystemtomaximizethereadingthroughputbymakingsomecollisionslotsalmostasusefulasnon-collisionslots(inwhichonlyonetagtransmits).ThedifferenceisthattheformerallowtheRFIDreadertolearnnewtagIDsaftersometime,whilethelatterletthereaderlearnnewIDsrightaway.Second,wedemonstratethepracticalvalueoftheanalognetworkcodingresearchbyprovidinganinterestingapplicationscenario. Technically,wedesigntherstcollision-resolutiontagidenticationprotocolthatestablishestheengineeringandtheoreticalfoundationforintegratinganalognetworkcodingintotheprocessoftagreading.Wederivetheoptimalsystemparametersforimprovingthereadingthroughput.Wealsoreducetheprotocoloverheadthroughaframedstructureandanembeddedestimatorforthenumberoftagsthatarecurrentlyparticipatingintheprotocol.TheproposedprotocolisabletoefcientlyutilizetheinformationcarriedincollisionslotsandthusbreakthefundamentallimitofALOHA-basedprotocolsthatdonotuseanalognetworkcoding.Ourworkanswerstwoimportantquestions:HowtooptimallyapplyanalognetworkcodingforRFIDreading?Howmuchthroughputgaincananalognetworkcodingbring?Thesimulationresults 22

PAGE 23

showthatthereadingthroughputcanbeimprovedby51.1%70.6%whenusingtoday'sanalognetworkcodingmethodandthethroughputcanbemuchhigherifthecodingmethodisimprovedinthefuture. Therestofthisdissertationisorganizedasfollows.Chapter 2 proposesAIMD/QS+k,theprotocolforachievingMAC-layertimefairnessunderlocation-sensitivecontention.Chapter 3 proposesWPD,theprotocolforimprovingTCPfairnessamongnearbyWLANs.Chapter 4 proposesFCAT,theprotocolforimprovingreadingthroughputinlargeRFIDsystems.Chapter 5 concludesourstudy. 23

PAGE 24

Figure1-1. ThreepartiallyoverlappingWLANsformtwocontentiongroupsg1andg2,wheres1,s2ands3arethreeaccesspoints.ATCPowfromaserverontheInternetpassesthroughanaccesspointtoawirelessclientineachWLAN.Notethat,onlythewirelesspartofaow(fromanaccesspointtoawirelessclient)isdrawninthegure. 24

PAGE 25

CHAPTER2MAC-LAYERTIMEFAIRNESSACROSSMULTIPLEWIRELESSLANS Inthischapter,westudytheMAC-layertime-allocationanomalyacrossmultipleWLANsunderlocation-sensitivecontention. Withonlyalimitednumberofnon-overlappingchannels,hostsinnearbyWLANsinevitablycompeteforthesamechannel.Whenahosttransmitsatalowtransmissionrate,allothercontendingnodesmaysufferwithlowthroughput,whichisthewell-knowntime-allocationanomaly.Inaddition,basedontheirrelativespatiallocations,contendinghostsmayhavedifferentcapabilitiesofobtainingthechannelfortheirtransmissions.Somehostsmaygrabmostofthechannelbandwidthwhiletheothersarestarved.Thisiscalledlocation-sensitivecontention.Weobservethatthelocation-sensitivecontentionmaypushthetime-allocationanomalytotheextreme,whichcausesunfairchannelbandwidthallocationanddegradestheoverallnetworkthroughput.WewillproposeanovelsolutiontoachieveMAC-layertimefairness. Therestofthischapterisorganizedasfollows.Section 2.1 presentsthenetworkmodelandproblemdenition.Section 2.2 describesthetime-allocationanomalyandthelocation-sensitivecontention.Section 2.3 discussestherelatedwork.Section 2.4 proposesourAIMD/QS+ksolution.Section 2.5 showsthesimulationresults.Section 2.6 givesthesummary. 2.1NetworkModelandProblemDenition Inthissection,wegivethenetworkmodelandtheproblemdenition. 2.1.1NetworkModel Consideranumberofwirelessaccesspointsthataredeployedinanarea.EachaccesspointconnectsoneormorewirelesshoststotheInternet.Eachhosthasatransceiverthatcaneithertransmitorreceiveatatime.AnaccesspointanditshostsformaWLAN.Theaccesspointselectsachannel,i.e.,asub-bandoftheavailablefrequencyrange,tocommunicatewithitshosts.IEEE802.11b/ghas11channels, 25

PAGE 26

amongwhichonlythreearenon-overlapping.TransmissionsinnearbyWLANsthatselectthesameoroverlappingchannelsmayinterferewithoneanother.However,spatialchannelreusehappensamongdistantWLANsthatcanusethesamechanneltotransmitsimultaneouslywithoutinterference. Fig. 2-1 showsanexample,whereeachcirclerepresentsanaccesspoint,eachsquarerepresentsahost,andeachsolidlinerepresentsawirelessconnectionbetweenahostandanaccesspoint.Adashedlinebetweentwonodes(whichcanbeeitheraccesspointorhost)meansthattheyarewithineachother'scarrier-sensingrange. Weuseanodetorefertoeitheranaccesspointorawirelesshost.ThesequenceofdatapacketssentfromanodextoanodeyconstituteaMACow(x,y).Thetransmissionrangeofnodexdenesthedistancewithinwhichanothernodeisabletodecodethedatatransmittedbyx.Theinterferencerangeofnodeydenesthedistancewithinwhichanothernode'stransmissionwillinterferewithy'sreceptionofapacketfromx.Thecarrier-sensingrangeofnodexdenesthedistancewithinwhichanothernode'stransmissionwillcausextosenseabusychannel.Toavoidcollision,itshouldbesetnolessthanthemaximuminterferencerange,whichcanbe1.78timesthetransmissionrangeassuggestedin[ 74 ](alsothedefaultvalueusedinns-2).1 WeassumeaDCF-likeMACprotocol.TwoMACowscontendifthefollowingconditionsaremet:(1)theirtransmissionsaremadeinthesamechannelorinchannelswithoverlappingfrequencybandsthatcauseinterference,and(2)thesenderofaowcancarrier-sensethetransmissionoftheotheroworthereceiverofaowisinterferedbythetransmissionoftheotherow.IfRTS/CTSisturnedon,thencontentionalso 1Notethatthecarrier-sensingrangecanbearticiallycongured.Itcanbemadetoequalthetransmissionrange[ 12 ].Thishoweverdoesnotalterthefactthatcontentiongoesbeyondthetransmissionrangebecausetheinterferencerangeisnotaquantitythatcanbearticiallycongured.Reducingthecarrier-sensingrangetobesmallerthantheinterferencerangeincreasesthechanceofcollision. 26

PAGE 27

happensifthereceiverofaowcancarrier-sensethetransmissionoftheotherow.Whenthetwoows(x,y)and(u,v)contend,wealsosaythatnodexhasacontendingnodeu. ThetransmissionrateofaMACow(x,y)isdenedasthemodulationrateofthesenderx,atwhichxtransmitsitsbitsinthechannel.Thechanneloccupancy(oroccupancyforshort)ofaowisdenedasthefractionofaunittimeduringwhichtheowoccupiesthechannelfortransmission.ItincludesboththetimefortransmittingdatapacketsandthetimeforcontrolpacketssuchasACKs.Thedeliveryrateof(x,y)isdenedastheaveragerateatwhichycansuccessfullyreceivedatafromx.Itisboundedbythetransmissionratemultipliedbythechanneloccupancy. Amaximalsetofmutuallycontendingowsiscalledacontentiongroup,alsoknownasacontentioncliqueinthepriorwork[ 32 33 ].Theowsinacontentiongrouphavetotaketurntotransmit.Thesumofthechanneloccupanciesforallowsinagroupiscalledtheaggregateoccupancyofthegroup,whichisboundedbyone.Acontentiongroupissaidtobesaturatedorcongestedifnoowcanfurtherincreaseitsdeliveryratewithoutdecreasingtherateofanotherowinthegroup.Twoowsindifferentcontentiongroupsmaybeabletotransmitsimultaneouslyduetochannelspatialreuse.AnexampleisgiveninFig. 2-2 ,whichhastwocontentiongroups:g1consistsof(w,z)and(x,y);g2consistsof(x,y)and(u,v).Flows(w,z)and(u,v)cantransmitatthesametime. 2.1.2ProblemDenition Whenthereisonlyonecontentiongroup,thetimefairnessproblemistoequalizethechanneloccupanciesofallMACowswhilefullyutilizingthechannelcapacity.However,whentherearemorethanonecontentiongroup,itisnotalwayspossibletoequalizetheows'channeloccupanciesbecauseeachowexperiencesdifferentcontention.Proportionalfairness[ 37 ]hasbeenintroducedforsuchcases. 27

PAGE 28

LetFbethesetofallMACows,Gbethesetofcontentiongroups,andq=(qf,f2F)beavectoroffeasiblechanneloccupancies,suchthatqf0,8f2F,andPf2gqfQg,8g2G,whereQgisthemaximumaggregateoccupancythatgcanhavebeforeitiscongested.QgwillbesmallerthanoneduetotheprotocoloverheadofDCF.Afeasiblevectorqisproportionallyfairifforanyotherfeasiblevectorq,thesumoftheproportionalchangesisnon-negative[ 37 ],i.e., Xf2Fqf)]TJ /F14 10.909 Tf 10.9 0 Td[(qf qf0.(2)Ithasbeenshownin[ 37 ]thatthisisequivalenttondafeasiblevectorqthatmaximizesthesumofautilityfunctionU(qf)=ln(qf). maximizeXf2FU(qf)subjecttoXf2gqfQg,8g2Gqf0,8f2F.(2)Exactlysolvingtheprobleminarun-timeenvironmentisextremelyhardduetothecomplexinteractionamongthecontendingows,aswewillelaborateinthiswork.Ourgoalistoapproximatetheproportionaltime-fairnessthroughafullydistributed,DCF-compatiblesolution.DCF-compatibilitymeansthatthesolutiondoesnotrequireanymodicationtotheDCFprotocoloritsrandombackoffalgorithm,butitmayretrievesomestateinformationfromtheMAClayerandmodifysomeMACparameters(suchasthesizeoftheminimumcontentionwindow)onthey. Westressthattheconceptofcontentiongroupisintroducedonlytohelpusdescribeoursolution.TheoperationsinourdistributedsolutionneverneedtoactuallyidentifywhichowsbelongtoeachcontentiongrouporknowthevalueofQg.Infact,Qgdoesnotevenhavetobeaconstant.Asitmaydriftovertime,theoptimalvalueforqfwillalsoevolve.Oursolutionwilldynamicallyconvergethechanneloccupanciesof 28

PAGE 29

theowstowardstheircurrentoptimalvalues.Unlikethepreviousworkthatconstructsthecontentiongroupsexplicitly[ 32 33 ],oursolutiondoesnotdosobecausethisisnotalwaysfeasibleinourgeneralsettingwherenodesthatcontendmaynotknoweachother'sidentities.Forexample,inFig. 2-2 ,ifthedistancebetweenzandxisgreaterthanthetransmissionrangebutsmallerthantheinterferencerange,thenxhasnowaytoknowifitcontendswithonelinkortenlinks. 2.2Time-AllocationAnomalyandLocation-SensitiveContention 2.2.1Time-AllocationAnomaly Thetime-allocationanomalyamonghostsinasingleWLANisawellknownfact[ 30 ].DCFensuresthattheMACowsinaWLANhaveequalchancetosendtheirpackets.Inotherwords,eachMACowwillsendroughlythesamenumberofpacketsoverlongrun.Supposetheowshaveroughlythesameaveragepacketsize.Itwilltakemoretimeforaowwithasmallertransmissionratetotransmitapacket.Hence,DCFendsupgivingmorechanneltimetoaMACowwithasmallertransmissionrate,whicheffectivelykeepsthechanneloperateatthesmallerratemoreoften,reducingtheWLAN'soverallthroughput. WedemonstratethisanomalythroughasimulationontheWLANinFig. 2-3 ,whichhastwoMACows,(x,y)and(x,z),whosetransmissionratesare11Mbpsand1Mbps,respectively.Allsimulationsintheworkareperformedinns-2[ 1 ].Thelengthofeachwirelesslinkis150m,theinterferencerangeis1.78timesthelengthofthewirelesslink[ 74 ],thetransmissionrangeis250m,andthecarrier-sensingrangeis550m.Thepacketsizeis1000bytes.Without(x,z),when(x,y)istheonlyactiveow,itsdeliveryrateis622packetspersecond.Nowwith(x,z),onewouldexpectthedeliveryrateof(x,y)willbecutbyhalf.However,Fig. 2-4A showsthatthedeliveryrateof(x,y)ismerely96packetpersecond.Fig. 2-4B showsthatitschanneloccupancyisjust0.12.(Thesumofthetwoows'channeloccupanciesisbelowonebecauseoftheprotocoloverheadattheMACandphysicallayers.) 29

PAGE 30

2.2.2Location-SensitiveContention WhenMACowsinnearbyWLANscontendinthesamechannel,therelativepositionsoftheWLANsandtheirhostscangivesomehostsmuchgreatercapabilityofoccupyingthechannelthanothers.Thisiscalledlocation-sensitivecontention.Duetothelocation-sensitivecontention,weobservethattime-allocationanomalyexistsamonghostsinnearbyWLANsevenwhenthetransmissionratesofallhostsarethesame.WeperformsimulationonthenetworkinFig. 2-5 ,whichmodelsthescenariooftwoneighboringhomeswhoseWLANsoperateinthesamechannel.Whenthetransmissionratesofbothlinksare11Mbps,thesimulationresultinFig. 2-6A demonstratesthatthechanneloccupanciesofthetwoowsareverydifferentanddependentonthedistancebetweenyandu.Thehidden-terminalproblemarisinginFig. 2-5 wasrstdocumentedandanalyzedin[ 10 ],which,however,doesnotconsiderthefactthatthecarrier-sensingrangeandinterferencerangearegreaterthanthetransmissionrange.Readersarereferredto[ 34 ]foramuchmoredetailedexplanationonthecauseoflocation-sensitivecontentioninthisnetwork. Ifthehoststhataremorecapableofobtainingthechannelhavelowtransmissionrates,theirchanneloccupanciescanbecomesohighthatotherhostswillbetotallystarved.Ifweadddifferenttransmissionratesontopoflocation-sensitivecontention,theydrivethetime-allocationanomalyproblemtotheextreme.InFig. 2-5 ,whenthetransmissionrateof(x,y)is11Mbpsbutthetransmissionrateof(u,v)is1Mbps,thesimulationresultinFig. 2-6B showsthatthechanneloccupancyof(x,y)isalmostzerowhenthedistancebetweenyanduislessthan250m. 2.3StateoftheArt Someexistingtime-fairnesssolutions[ 30 66 67 ]relyonacentralcoordinator,whichisnotpracticallyfeasiblewhenwedealwithmultipleWLANsthatcontendbutmaynotbeabletocommunicateduetothedisparitybetweenthecarrier-sensingrangeand 30

PAGE 31

thetransmissionrange.Belowwesurveythedistributedsolutions,whichareclassiedintoseveraltypes. 2.3.1Type-I:AssumeKnowledgeofContendingFlows AType-Isolutionassumesthateachnodehascertainknowledgeaboutitscontendingnodes.AnexampleistherecentpaperofTFCSMA[ 35 ].Itassumesthateachnodeknowsthenumberofcontendingnodes,whichisneededincomputingthenode'stargetthroughput(equivalenttothedeliveryrateinthiswork). However,inawirelessnetwork,itisdifculttolearnthesetofcontendingowsandkeeptrackofwhichowsarecurrentlyactiveinsendingdata.Explicitexchangeofcontrolpacketsorimplicitoverhearingrequiresthenodestobewithineachother'stransmissionrange.However,contentionisdenedbythecarrier-sensingrange,whichismuchgreaterthanthetransmissionrange.Consideranexamplewherethecarrier-sensingrangeistwiceofthetransmissionrange.Becausetheareaoutsideofthetransmissionrangebutwithinthecarrier-sensingrangeisthreetimestheareawithinthetransmissionrange,thenumberofcontendingnodesthatcannotbeidentiedcanbemuchgreaterthanthenumberofowsthatcanbeidentied,whichseriouslyaffectstheperformanceofTFCSMA.Thesimilarproblemexistsforthesolutions[ 32 46 47 69 75 ],whicharedesignedforthroughputfairnessinsteadoftimefairness.Theyrequireinformationcollectedthroughoverhearing,butnotallcontendingnodescanbeoverheard. 2.3.2Type-II:AssumeSameChannelPerception AType-IIsolutiondoesnotrequireanodetohaveanyknowledgeaboutitscontendingnodes,butassumesthatthecontendingnodeswillsensethesamechannelcondition.AnexampleisIdleSense[ 28 31 ],whichusesanon-DCFprotocoltoadapteachnode'scontentionwindow,suchthatthemeannumberofidleslotsbetweentwotransmissionsinthechannelismaintainedatacertaindesirablevalue,e.g.,5.6for802.11b.ItworkswellforasingleWLAN,butnotformultipleWLANs.First,IdleSense 31

PAGE 32

requiresallnodestosensethesameidletimeinthechannel.InFig. 2-5 ,supposenodeuiswithintheinterferencerangeofybutoutsidethecarrier-sensingrangeofx.Evenwhenthechannelissaturatedbythetransmissionon(u,v),nodexwillsenseanidlechannel!Second,evenifallnodessensethesameidletime,IdleSensemaystillworkpoorlyformultipleWLANsduetolocation-sensitivecontention.WesimulateIdleSenseonthenetworkinFig. 2-5 ,wherexanduareoutsideeachother'stransmissionrangebutwithinthecarrier-sensingrange.TheresultisshowninFig. 2-7 .IdleSenseactuallystarvesow(x,y)inthisscenario. TheworkbyNandagopaletal.[ 52 ]alsofallsinthiscategory.Itusesanon-DCFtime-slottedcontentionresolutionprotocolandassumesthatthesendersofallowsinacontentiongroupwillhavethesametransmissionfailureprobability,whichisalsonottrueamongcontendinghostsindifferentWLANs[ 34 ].ConsiderthenetworkinFig. 2-5 ,whereuisoutsideofthecarrier-sensingrangeofxbutwithintheinterferencerangeofy.Whenutransmitsadatapacket,xwillsenseanidlechannelandtransmit,whichalwaysresultinatransmissionfailure.Whenxtransmitsadatapacket,ifualsotransmits,itwillnotleadtoatransmissionfailure(unlesstheACKsfromyandvhappentobetransmittedatthesametime,whichisararecase). 2.3.3Type-III:AssumenoKnowledgeofContendingNodesandDifferentChan-nelPerception AType-IIIsolutionneitherrequiresanodetohaveanyknowledgeaboutitscontendingnodes,norassumesthatthecontendingnodeswillsensethesamechannelcondition.Oneexampleistofragmentthepackets[ 20 ].Flowswithsmallertransmissionrateswillhavesmallerfragmentsizes.IftheMACowstransmitforroughlythesamenumberoftimesandeachtimetransmitonefragment,thentheirchanneloccupanciescanbeequalized.Theoppositesolutionistoaggregatethepackets[ 59 62 ].Flowswithlargertransmissionrateswillhavelargerpacketaggregates.Again,iftheMACowstransmitforroughlythesamenumberoftimesandeachtimetransmitone 32

PAGE 33

aggregate,thentheiroccupanciescanbeequalized.Bothsolutionsusethenumberofbitsineachtransmissiontocompensatethedifferenceinthetransmissionrate.TheirassumptionthatDCFgivesallMACowsequalchancetotransmitistrueinthesymmetricsettingwhentheowsbelongtothesameWLAN.However,itisnottrueinasymmetricsettingswhentheowsbelongtodifferentWLANs.ConsiderthenetworkinFig. 2-5 .Supposethetransmissionratesoftwoowsareboth11Mbpsandthepacketshavethesamesize.TheabovesolutionsarereducedtoDCF.ThesimulationresultisshowninFig. 2-6B .Thechanneloccupanciesoftheowsareverydifferent.Flow(u,v)musthavetransmittedformanymoretimesthanow(x,y)inordertoproduceitslargeoccupancy. Anothersolution,CWSP(ContentionWindowScalingProtocol)[ 38 ],istomakethesizeoftheminimumcontentionwindowinverselyproportionaltothetransmissionrate.Itdecreasestheprobabilityforaowwithasmalltransmissionratetoobtainthechannelandconsequentlyreducesitschanneloccupancy.However,thisheuristicapproachcannotbringquantitativeprecisionanddoesnotalwaysworkwell.WesimulateCWSPonthenetworkinFig. 2-5 ,wherethetransmissionratesof(x,y)and(u,v)are11Mbpsand1Mbps,respectively.TheresultsareshowninFig. 2-8 ,whichexhibitsreversediscriminationthechanneloccupancyofthe11-Mbpsow(x,y)isabouttwicetheoccupancyofthe1-Mbpsow(u,v). PISD[ 34 ]appliesproportionalincreasesynchronizedmultiplicativedecreasetocontroltherateofeachow.Itsmainobjectiveistoachieveweightedthroughputfairnessalthoughitmaybeextendedtoapproximatetimefairness.However,itmakesanassumptionthatallowsbelongtoasinglecontentiongroup.Whenthecontentiongroupissaturated,anynodethatdetectsthecongestionwilljamthechanneltohelpothernodesalsodetectthecongestion.Thisensuresthatthenodeswillperformmultiplicativedecreasesimultaneously.However,becauseallnodeswillparticipateinjamming,foralargedeploymentofWLANsthatformmanypartially-overlapping 33

PAGE 34

contentiongroups,jammingwillspillouttoothercontentiongroupsthatarenotsaturated.Itleadstocascadedjamming.Itbeginsfromonenodeatacongestedhotspot.Itsjammingcausesthenearbynodestodetectcongestionbecausetheycanhardlysendoutpackets.Whenthesenodesstarttheirjamming,thenodesfurtherawaywillfeeltheimpact.Astheprocessrepeats,nodesoutsideofthecongestedcontentiongroupwillfalselydetectcongestionandunnecessarilyreducetheirrates. TheaboveobservationisconrmedbythesimulationonthenetworkinFig. 2-9 ,whichhastwooverlappingcontentiongroups:g1hastwomutuallycontendingowsandg2hasveows.Thetwogroupsshareacommonow(x,y).Onewouldexpectthedeliveryrateof(w,z)willbehighbecausechannelspatialreuseoccursbetween(w,z)and(u,vi),1i4,sincetheydonotcontend.Letthetransmissionratesofallowsbe11Mbps.ThesimulationresultisshowninTable 2-1 .Duetocascadejamming,thedeliveryrateof(w,z)iscomparabletotheratesoftheowsthatbelongtothebottleneckg2.Thetotaloccupancyofg2ishigh.Thetotaloccupancyofg1ishoweververylow.Thereasonisthat,wheng2iscongestedandxperformsjamming,notonlyuing2willfeelit,butwing1willalsofeelit. 2.3.4OtherRelatedWork Thereisawealthoftheoreticalstudyonutilityoptimizationandproportionalfairness[ 37 40 45 48 ].Tremendousprogresshasbeenmadetoapplythistheoryinwirelessnetworks.However,theexistingworkassumesdifferentnetworkmodelsandthuscannotbedirectlyappliedtosolveourproblem.Someexistingsolutionsrequireacentralized,NP-hardschedulingalgorithmforallwirelesslinks[ 23 43 53 ].Othersassumeatime-slottedcellularnetworkmodel[ 22 ]orassumeanode-exclusiveinterferencemodel[ 11 17 44 ]wherelinkscantransmitsimultaneouslyaslongastheydonotshareacommonnode. Thereisalargebodyofworkonratefairnessinmultihopwirelessnetworks(MWN).ItmayappearthattheWLANnetworksstudiedinthisworkarespecialcasesofthe 34

PAGE 35

generalMWNs.Thisisnottrue.TheyaredifferentproblemsandtheWLANnetworksmayposegreaterchallenges.Forone,neighborsinanMWNcanexchangeinformation(suchasthecasein[ 58 75 ]),whereasthecontendinghostsinnearbyWLANsmaybetoofartocommunicateandtheWLANscanbeviewedascomponentsofapartitionednetwork. 2.4ASolutionforMAC-layerTimeFairnessacrossMultipleWLANs Inthissection,weproposeourtimefairnesssolutionforWLANs,whosedesignisbasedontwonoveltechniques,calledchanneloccupancyadaptationandqueuespreading,whichtogetherapproximatelyachieveproportionalfairnessamonghostsincontendingWLANs. 2.4.1Overview ThebasicideaofoursolutionistoapplytheAIMD(additiveincreasemultiplicativedecrease)adaptationonchanneloccupancyineachcontentiongroup.Morespecically,allowsinacontentiongroupperformAIMDontheirchanneloccupancies.Whentheaggregateoccupancyofthegroupdoesnotcausechannelcongestion,eachowinthegroupwillincreaseitschanneloccupancybyaconstantamount,whichisadditiveincrease.Whentheaggregateoccupancyofthegroupbecomestoolargeandcauseschannelcongestion,allowsinthegroupwilldecreasetheirchanneloccupanciesbyacertainpercentage,whichismultiplicativedecrease. Ifthereisonlyonecontentiongroup,thenAIMDwillindeedequalizethechanneloccupanciesofallows.Thegapbetweenthelargestoccupancylandthesmallestoccupancysamongallowsstaysthesameaftereachadditiveincreasebutshrinksaftereachmultiplicativedecrease.Thatisbecause,asbothlandsarereducedbythesamepercentage,lisreducedbyalargerabsoluteamount,andhencetheirgapshrinks.Thatgapdiminishesovertimeaftermultiplicativedecreaseisperformedperiodically. 35

PAGE 36

Whentherearemultiplecontentiongroupsandeachnodemayparticipateinmorethanonegroup,thenAIMDwhichisperformedindependentlyineachgroupmaynotequalizethechanneloccupanciesofallows.AswewilldiscussinSection 2.4.5 ,itapproximatelyachievesproportionalfairness.Forexample,inFig. 2-9 ,ow(x,y)participatesintwocontentiongroups.Inproportionalfairness,thechanneloccupancyof(x,y)willbemodestlysmallerthanthoseofotherowsing2because(x,y)alsoparticipatesinmultiplicativedecreaseing1(whichhappenslessfrequentlybecauseg1onlyhastwoowsandittakeslongertimeforadditiveincreasetocongestthechannel).Thisisgenerallyregardedasapositivefeaturebecauseitstrikesabalancebetweenfairnessandthroughputimprovement.Asmaller(butnottoosmall)channeloccupancyfor(x,y)improvesthechannelspatialreusebetween(w,z)and(u,vi),1i4,astheycantransmitsimultaneously. Ourideafortimefairnessfacestwomajortechnicalproblems.TherstproblemishowtoactuallyperformAIMDonchanneloccupancy.Thesecondproblemismostchallenging:Whenacontentiongroupiscongested,wewantallowsinthegroupandonlytheowsinthegrouptoperformmultiplicativedecrease.ConsidertheexampleinFig. 2-9 .Supposeg2iscongested.Amongallowsing2,(x,y)residesatthemostdisadvantageouslocation.As(x,y)isleastcapableofobtainingthechannel,itwillbethersttodetectthechannelcongestion.Ifxperformsmultiplicativedecreasetoresolvethecongestion,otherowsing2willcontinueperformingadditiveincreasetopickupthebandwidthithasgivenup.Hence,nodexmustperformcertainoperationsthatcauseu(butnotw)toalsodetectthecongestion.Recallthatthenodes,x,uandw,maynotbeabletodirectlycommunicateoroverheareachother.Therestofthesectionwillsolvetheseproblems. 2.4.2ReleaseRateandChannelOccupancyAdaptation Thecontrolfunctionoftheproposedsolutionresidesatthenetworklayer.ItdoesnotrequireanymodicationtotheDCFprotocol,butmustbeabletoqueryforthe 36

PAGE 37

queuelengthattheMAClayerandmodifythesizeoftheminimumcontentionwindow.Itadaptsthechanneloccupancyofaowbycontrollingthereleaserate,atwhichthenetworklayerreleasespacketsintotheMAClayer. Whenthecombinedreleaserateofallowsinacontentiongroupissmall,thechannelcanforwardallpacketsreleasedtotheMAClayer.Hence,thedeliveryrateisequaltothereleaserate,andtheMAC-layerqueueremainsempty.TheowscanimprovetheirchanneloccupanciesbyreleasingmorepacketsintotheMAClayer.Whenthecombinedreleaserateoftheowsinthegroupbecomestoohighandtheaggregatechanneloccupancyistoolargesuchthatthechannelissaturated,notallpacketsreleasedtotheMAClayercanbeforwarded.Excesspacketshavetobebufferedatthenodes'MAC-layerqueues.Eventually,onenodewillobservethatitsqueuelengthpersistentlygrowsandexceedsathreshold.Thenodewillperformtheoperationofqueuespreading(inthenextsubsection),whichensuresthatthequeuelengthsatallothernodesinthecongestedcontentiongroupwillalsoexceedthethreshold(suchthattheyalldetectthecongestion).Then,thenodesshouldreducetheirchanneloccupanciesbydecreasingthereleaserates,andfewerpacketsaremadeavailabletotheMAClayerinordertorelievethecongestion. Intuitively,thequeue-lengththresholdforcongestiondetectionshouldbeproportionaltothetransmissionrateoftheow.Thatisbecause,undertimefairness,aowwithalowtransmissionratewillhaveasmallreleaserate,whichwouldmakeitsqueuehardertopassthethreshold(thushardertodetectcongestion)ifthethresholdwasaconstant.Hence,wedenethethresholdasHr,whereHisasystemwideconstantandristheow'stransmissionrate. Theprotocolforchanneloccupancyadaptationisgivenasfollows:ConsideranarbitraryMACow(x,y)withatransmissionrater.ThesenderxadaptsitschanneloccupancyaftereachtimeperiodofT.IfitsqueuelengthattheMAClayerisbelowthethresholdHr,itwilladditivelyincreaseitschanneloccupancybyaconstant. 37

PAGE 38

Ifthequeuelengthreachesthethreshold,itwillmultiplicativelydecreaseitschanneloccupancybyapercentage.Toimplementmultiplicativedecreaseonthechanneloccupancy,xsimplyreducesitsreleaseratebyapercentage.Toimplementadditiveincreaseonthechanneloccupancy,xincreasesitsreleaseratebyr(becauseittakestimetotransmitrbitsatrater).Hence,whileallowssharethesameparameter,theincrementsfortheirreleaserates,r,willbeproportionaltotheirtransmissionrates. 2.4.3QueueSpreading Weproposeanewtechnique,calledqueuespreading,whichmakessurethatthesendersofallowsinacongestedcontentiongroupwilldetectthecongestionandtheywillperformmultiplicativedecreasetogether. TheaggregatereleaserateRg(t)ofallowsinacontentiongroupgisafunctionoftimet.Rg(t)=Pf2gRf(t),whereRf(t)isthereleaserateofowf.LetC(t)bethemaximumthroughputthatthegroupcanpossiblyobtainfromthechannelattimet.WestressthatC(t)isonlyneededtodescribeouridea.TheoperationofqueuespreadingdoesnotrelyontheknowledgeofC(t).WhenRg(t)exceedsC(t),ifwelookattheowsasawhole,therearemorepacketsreleasedtotheMAClayerthanitcansendout.TheexcesspacketsincreasethequeuelengthsoftheowsatacombinedrateofRg(t))]TJ /F3 11.955 Tf 12.71 0 Td[(C(t).Theproblemisthat,duetolocation-sensitivecontention,mostexcesspacketsmaybequeuedupatoneowthatisleastcapableofaccessingthemedium.Whilethatowobservesitsqueuelengthexceedsthethresholdandperformsmultiplicativedecrease,otherowsmorecapableofobtainingmediummaystillndtheirqueuesemptyandthuscontinuewithadditiveincrease,whichwillenlargethegapamongtheowratesandresultinworseunfairness. Oursolutiontotheaboveproblemistospreadtheexcesspacketsamongthequeuesofallowsinthegroup.Foranarbitraryow(x,y)whosetransmissionrateisr,wheneverthepacketqueueatthesenderxexceedsHr,xwilltemporarilymodify 38

PAGE 39

itsMACparameterstoincreaseitsabilityofobtainingthemedium,suchthatitsqueuelengthcanbereducedbacktoHr.WhenthequeuelengthbecomesHr,thenodewillrestoretheoriginalMACparameters.Theideabehindqueuespreadingisveryintuitive:Afteranodedetectscongestion,thenodewillkeepitsqueuelengthatHrbydynamicallyadjustingitsMACparameters.Becauseitsqueuenolongergrows,theexcesspacketsinthechannelwillhavetobebufferedelsewhere,pushingthequeuesatothernodesup.Oncetheirqueuesreachthethreshold,theywilldothesamething.Excesspacketswillalwaysbepushedtothenodesthathavenotdetectedcongestionyet. Anodethatperformschanneljamming[ 34 ]triestooccupythechannelasmuchaspossible,andconsequentlyitwillaffectallneighboringnodes,includingthoseoutsideofthesaturatedgroup.Onthecontrary,anodethatperformsqueuespreadingonlytriestomatchitssendingratewithitsreleaseratesuchthatthelocalqueuedoesnotgrowfurther.Therefore,itwillnotaffecttheneighborsoutsideofthesaturatedgroup.Letrfbethetransmissionrateoff.Wehavethefollowingproposition. Proposition1:Thesendersofallowsinacontentiongroupgwilldetectcongestionbytheendofatimeperiod[t0,t1)ifthefollowingconditionsaresatised:(1)Rg(t))]TJ /F3 11.955 Tf 10.5 0 Td[(C(t)>0,8t2[t0,t1),(2)Rt1t0(Rg(t))]TJ /F3 11.955 Tf 12.69 0 Td[(C(t))dtPf2gHrf,and(3)queuespreadingisperformed. Proof:Toprovebycontradiction,weassumethatasubsetofows,g0g,doesnotdetectcongestionbytimet1.Ononehand,theows'packetqueuesareshorterthantheirrespectivethresholds.Thus,thetotalnumberofpacketsintheirqueuesislessthanPf2g0Hrf.Ontheotherhand,byperformingtheoperationofqueuespreading,theowsing)]TJ /F3 11.955 Tf 12.27 0 Td[(g0aremorecapableofobtainingthemediumthanthoseing0.Hence,theycancontroltheirqueuelengthstothethresholdvaluesattheexpenseoftheowsing0,whosequeueswillbegrowing.Thetotalnumberofpacketsqueuedattheowsing)]TJ /F3 11.955 Tf 11.97 0 Td[(g0isPf2g)]TJ /F4 7.97 Tf 6.58 0 Td[(g0Hrf.Duringthetimeperiod[t0,t1),byCondition(1),thetotalnumberof 39

PAGE 40

excesspacketsthatarequeuedbyallowsmustbeRt1t0(Rg(t))]TJ /F3 11.955 Tf 12.03 0 Td[(C(t))dt.Therefore,thenumberofpacketsqueuedattheowsing0mustbeRt1t0(Rg(t))]TJ /F3 11.955 Tf 10.14 0 Td[(C(t))dt)]TJ /F11 11.955 Tf 10.14 8.96 Td[(Pf2g)]TJ /F4 7.97 Tf 6.59 0 Td[(g0Hrf.ByCondition(2),thisnumberisnolessthanPf2g0Hrf,leadingtothecontradiction.2 Itiseasytoseethat,ifthecongestionisdetectedbythesendersofallowsinagroup,thenthetotalnumberofexcesspacketsintheirqueuesthatthechannelcannotdeliverisatleastPf2gHrf.Wehavethefollowingnecessaryconditionforcongestiondetection. Proposition2:Supposethesendersofallowsinacontentiongroupghaveemptyqueuesattimet0andRg(t))]TJ /F3 11.955 Tf 12 0 Td[(C(t)>0,8t2[t0,t1).Ifallowsdetectthecongestionofgbytimet1,thenwemusthaveRt1t0(Rg(t))]TJ /F3 11.955 Tf 11.95 0 Td[(C(t))dtPf2gHrf. 2.4.4AIMD/QS+k Weneedtointegratequeuespreading(QS)intothechannel-occupancyadaptationprotocolinSection 2.4.2 .However,astraightforwardcombinationofthetwowillnotdothetrick.Therstintegratedprotocol,calledAIMD/QS,isgivenasfollows:Eachow(x,y)performstheAIMDchanneloccupancyadaptationperiodicallyasdescribedinSection 2.4.2 .Inaddition,whenthequeuelengthatthesenderxreachesthethreshold,xperformsqueuespreadinguntiltheendofthecurrentperiodTwhenitdoesmultiplicativedecrease.Duringtheoperationofqueuespreading,ifthequeuelengthisabovethethreshold,thesenderxaggressivelyreducesitsminimumcontentionwindowtoasmallfractionofthedefaultsizeinordertoensurethatithastheprioritytooccupythechannel.Oncethequeuelengthisreducedtothethreshold,xrestoresthedefaultminimumcontentionwindow.Bydoingso,itkeepsthequeuelengthatthethreshold. Proposition2statesthat,inorderforthesendersofallowsinacongestedgroupgtodetectcongestion,thenumberofexcesspacketsbufferedbyallowsmustbeatleastPf2gHrf.However,AIMD/QShasnomeanstoguaranteethat.Let(x,y)betheow 40

PAGE 41

inthegroupthatisleastcapableofobtainingthechannel,andrbeitstransmissionrate.Infact,aslongasthenumberofexcesspacketsisHr,itmayhappenthatallexcesspacketsendupinthequeueofx,whichdetectscongestionandperformsmultiplicativedecrease,whileothernodeshaveemptyqueues. Tosolvethisproblem,wedesignageneralizedprotocol,AIMD/QS+k,wherekisanon-negativeinteger.ThesenderofeachowcarriesouttheoperationsofAIMD/QSexceptthat,afteritndsitsqueuelengthreachesHr,itwillcontinueperformingadditiveincreaseforksubsequentperiodsofTbeforemakingmultiplicativedecrease.Supposeaow'squeuereachesHrduring[t0,t0+T).Itwillincreasethereleaserateattimest0+T,t0+2T,...,t0+kTbyanamountr,andthendecreasethereleaserateattimet0+(k+1)Tbyapercentage.Theideaistomakesurethattherewillbeenoughexcesspacketstoallowallnodestodetectcongestion.ThefollowingpropositiongivestheformulaforpickingtheparametersofAIMD/QS+k,suchthatwhenacontentiongroupiscongested,allowsinthegroupwilldetectthecongestionandthusperformmultiplicativedecrease.(Theowsoutsideofanycongestedgroupwillnotdothatbecausetherearenoexcesspacketstopushtheirqueuesoverthethreshold.) Proposition3:AIMD/QS+kensuresthedetectionofcongestionbyallowsinacongestedgroupif Hk(k)]TJ /F5 11.955 Tf 11.95 0 Td[(1) 2T,fork2.(2) Proof:Consideranarbitrarytimet=0.Letgbetherstcontentiongroupthatbecomescongestedaftert=0.Whengbecomescongested,itsconstituentowsreleasemorepacketsthanthechannelcandeliver.Theexcesspacketswilleventuallypushthequeuelengthofaow(u,v)ingoverthethreshold.Withoutlosinggenerality,supposeithappensduring(iT,(i+1)T].LetCbethechannelcapacitythatcanbemaximallyobtainedbythegroupgatthemoment.Clearly,Rg((i+1)T)>C.Otherwise,therewouldbenoexcesspacketstopushthequeuelengthof(u,v)tothethreshold. 41

PAGE 42

Flow(u,v)willbetherstonetoperformmultiplicativedecreaseandthatwillhappenattimet=(i+k+1)T.Hence,Rg(.)isanon-decreasingfunctionbeforet=(i+k+1)T.BecausethereleaserateofeachowfingincreasesbyrfaftereachperiodT,wemusthaveRg((i+1+j)T)=Rg((i+1)T)+jPf2grf.Hence, Z(i+k+1)T(i+1)T(Rg(t))]TJ /F14 10.909 Tf 10.91 0 Td[(C)dtk)]TJ /F6 7.97 Tf 6.58 0 Td[(1Xj=0(Rg((i+1+j)T))]TJ /F14 10.909 Tf 10.91 0 Td[(C)T=k)]TJ /F6 7.97 Tf 6.59 0 Td[(1Xj=0(Rg((i+1)T))]TJ /F14 10.909 Tf 10.91 0 Td[(C)T+k)]TJ /F6 7.97 Tf 6.59 0 Td[(1Xj=0jXf2grfT>k)]TJ /F6 7.97 Tf 6.59 0 Td[(1Xj=0jXf2grfT=k(k)]TJ /F16 10.909 Tf 10.91 0 Td[(1) 2Xf2grfT.(2)If( 2 )ismet,wewillhaveR(i+k+1)T(i+1)T(Rg(t))]TJ /F3 11.955 Tf 11.79 0 Td[(C)dt>Pf2grfH.Hence,byProposition1,AIMD/QS+kmakessurethatallowsingdetectthecongestion.2 2.4.5ProportionalFairnessandInterpretationofAIMD/QS+k Intheirseminalpaper[ 37 ],Kelly,MaullooandTanshowthatthereexistfullydistributedalgorithmsthatachieveglobaloptimizationofthesystemutility.Belowwerewritetheirprimalalgorithminournotations(withsimplicationofremovingtheweight).LetFbethesetofallowsandGfbethesetofcontentiongroupstowhichowfbelongs. d dtqf(t)=)]TJ /F3 11.955 Tf 11.96 0 Td[(qf(t)1 "2Xg2Gf(Xf02gqf0(t))]TJ /F3 11.955 Tf 11.95 0 Td[(Qg+")+.(2) Thepricefunctions(Pf02gqf0(t))]TJ /F3 11.955 Tf 12.18 0 Td[(Qg+")+forg2Gftaketheformsuggestedin[ 37 ],suchthatwhen"!0,theaboveadaptation,whenperformedindependentlybythesendersofallows,willmaximizethesystemutility,Pf2Flnqf. AIMD/QS+kcanbeinterpretedasadiscreteapproximationof( 2 ).Therstitemontherightsideof( 2 )suggestsadditiveincrease.AIMD/QS+kincreasesqfbyaconstantamountaftereachtimeperiodofT.Theseconditemontherightsidesuggestsmultiplicativedecrease.Pg2Gf(Pf02gqf0(t))]TJ /F3 11.955 Tf 12.69 0 Td[(Qg+")+isthepenalty 42

PAGE 43

factor.AIMD/QS+kdecreasesqfbyapercentagewhentheaccumulatedpenaltyRt1t0Pg2Gf(Pf02gqf0(t))]TJ /F3 11.955 Tf 12.15 0 Td[(Qg+")+dtreachesathreshold,wheret0isthetimewhenthepreviousmultiplicativedecreaseisperformedandt1isthecurrenttime. Thekeyistomeasurethepenaltyforeachcontentiongroupg,Rtt0(Pf02gqf0(t))]TJ /F3 11.955 Tf -440.49 -23.91 Td[(Qg+")+dt,whichbecomespositiveonlywhengiscongested.Itislikelythat,foranyowf,onlyonegroupinGfiscongestedatatimeifischosensmallsuchthatacontentiongroupwillspendmuchmoretimewithoutcongestionthantimewithcongestion. Interestingly,thepricecanbeindirectlymeasuredthroughthelocalqueuelength.LetRf(t)bethereleaserateofowfandrf(t)bethetransmissionrate.InAIMD/QS+k,thechanneloccupancyqfiscontrolledthroughthereleaserateRf.In( 2 ),qfrepresentstheresourcedemandofowf.Itisthechanneloccupancythattheowdemandsinordertotransmitallreleasedpackets.Hence,qfisproportionaltoRf(becauseeachbittakesthesameamountoftimetotransmit).Aftergiscongested,furtheradditiveincreasemadebyAIMD/QS+kwilllinearlyincreasethereleaserate(orthechanneloccupancy)ofeachow.Inthemeantime,thenumberofexcesspacketsthathavetobebufferedwillincreaseproportionally.Thisimpliesthat,approximately,thecombinedqueuelengthforallowsinggrowsataspeedproportionaltotheamountofexcesschanneloccupancy,(Pf02gqf0(t))]TJ /F3 11.955 Tf 12.23 0 Td[(Qg)+.Furthermore,thetechniqueofqueuespreadingensuresthatthesendersofallowswillseethesamepriceastheexcesspacketsarespreadamongthequeuestopushallofthemoverthethreshold.Here,thepricetakesadiscreteform,0ifthethresholdisnotreachedand1ifthethresholdisreached. 2.5Simulation WeperformextensivesimulationsonvariousscenariostoevaluatetheperformanceofAIMD/QS+kinachievingMAC-layertimefairness.WecompareAIMD/QS+kwithsomeexistingwork,including802.11DCF,CWSP[ 38 ],andIdleSense[ 31 ]. 43

PAGE 44

Allsimulationsareperformedonns-2.AIMD/QS+kworksontopofDCF.TheparametersofDCFusethedefaultvaluessetbyns-2accordingtotheprotocolstandards.Forthechanneloccupancyadaptation,=0.03,=0.5,andT=1second.TheparametersforqueuespreadingaredeterminedbasedonProposition3.Inparticular,wechoosek=2andH=0.03seconds,whichsatisfy( 2 )intheproposition.Theparametersofothersolutionsarechosenbasedontheoriginalpapers. Weusethreesimulationscenarioswithincreasingcomplexity.Theresultsfromasimplescenarioareeasiertointerpret,whiletheresultsfromacomplexscenarioareclosertowhatwillbeseeninpractice. 2.5.1OneContentionGroup TherstsimulationscenarioisbasedonthenetworkofFig. 2-5 ,whichcontainsonlyonecontentiongroupoftwoMACows,(x,y)and(u,v).Thelengthofeachwirelesslinkis150m,theinterferencerangeis1.78timesthelengthofthewirelesslink[ 74 ],thetransmissionrangeis250m,andthecarrier-sensingrangeis550m.Thepacketsizeis1000bytes.Theseparameterswillalsobeusedinotherscenarios. Fig. 2-10 comparesthechanneloccupanciesoftheowsachievedunderDCF,CWSP,IdleSense,andAIMD/QS+k,respectively,withthetransmissionrateof(x,y)being11Mbpsandtherateof(u,v)being1Mbps.UnderDCF,thechanneloccupancyof(u,v)ismuchhigherthantheoccupancyof(x,y).UnderCWSP,thesituationisopposite.Theoccupancyof(x,y)isbetter.IdleSenseperformsverywelluntilthedistanceisbeyond250m,where(u,v)istotallystarved. InrealdeploymentofWLANs,RTS/CTSisoftenturnedoff.OursimulationsndthatthenetworkthroughputisconsistentlyhigherwithoutRTS/CTS.Thereasonisthat,accordingtothe802.11standard,RTS/CTSaresentatthelowesttransmissionrate.Hence,theywilltake656sat1Mbps,whichrepresentsignicantoverhead,consideringthatitonlytakes940stotransmitadatapacketof1,000bytesat11Mbps.WeturnoffRTS/CTSandre-runthesimulation.TheresultisshowninFig. 2-11 44

PAGE 45

Indeed,theaggregatechanneloccupancyisimprovedineachcase.AIMD/QS+kcanmaintaintimefairness.TheperformanceofIdleSenseimproveswhenthedistanceisgreaterthan250m,butdegradeswhenthedistanceissmallerthan250masthe11-Mbpsow(x,y)isdepressedbythe1-Mbpsow(u,v). Wechangethetransmissionrateof(u,v)from1Mbpsto5.5Mbpsandre-runtheabovesimulation(withRTS/CTSturnedoff).TheresultisshowninFig. 2-12 .Again,onlyAIMD/QS+kachievestimefairnessunderlocation-dependentcontentionforalldistancevalues. 2.5.2TwoContentionGroups ThesecondsimulationscenariousesthenetworkofFig. 2-9 ,whichhastwocontentiongroups.Groupg1containstwoows,(x,y)and(w,z).Groupg2containsveows,(x,y),(u,v1),(u,v2),(u,v3),and(u,v4).Thelengthofeachwirelesslinkis150m.Thedistancebetweenzandxis200m.Thedistancebetweenyanduisalso200m.Thetransmissionrateof(w,z)is2Mbps.Thetransmissionrateof(u,v3)is1Mbps.Thetransmissionratesofotherowsare11Mbps. ThesimulationresultisshowninTable 2-2 .UnderDCF,(x,y)isstarvedandtheoccupanciesof(u,v1),(u,v2)and(u,v4)areverylow.UndereitherCWSPorIdleSense,(x,y)isstarved.UnderAIMD/QS+k,(x,y)hasadecentchanneloccupancy,eventhoughitissmallerthanothersduetothenatureofproportionalfairness.ComparingwiththeresultofPISDinTable 2-1 ,AIMD/QS+kimprovesthechanneloccupancyof(w,z)from0.161to0.546.Weshallnotcomparethedeliveryratesbecausethetransmissionrateof(w,z)is11Mbpsthereand2Mbpshere.Thetotalthroughputofg1underAIMD/QS+kiscomparabletothethroughputunderDCForIdleSense,butmuchbetterthanthethroughputunderCWSP.Thetotalthroughputofg2underAIMD/QS+kismuchbetterthanthethroughputunderDCForIdleSense,butsmallerthanthethroughputunderCWSP. 45

PAGE 46

2.5.3MultipleContentionGroups ThethirdsimulationscenarioisdesignedbasedonthenetworkofFig. 2-13 ,whereWLANsaredeployedalongtwocrossingstreets.Therelativepositionsofthenodesaredrawninthegure.Thelengthofeachwirelesslinkis150m.ThedistancebetweentheclosestnodesintwoadjacentWLANsis200m.Thecontentionrelationshipamongtheowsisautomaticallydeterminedinns-2basedtheseparametersandthoseinSection 2.5.1 .Thetransmissionratesofsomeowsarespeciedinthegure,andtheratesofothersare11Mbpsbydefault. ThesimulationresultisshowninTable 2-3 .UnderDCF,ows5,11and15haveverylowchanneloccupancies.Afewothershavelowoccupancies.UnderCWSP,ows8and14haveverylowchanneloccupancies.UnderIdleSense,ows1,2,3,4,and9haveverylowchanneloccupancies.ThisisnotasurprisingoutcomebecauseIdleSensewasdesignedtoworkamonghostsinasingleWLAN.UnderAIMD/QS+k,allowshavereasonablechanneloccupancies.Itsoveralldistributionofchanneloccupanciesismuchfairerthanthoseofothers.WebelievethissimulationresultdemonstratesthestrongperformanceofAIMD/QS+kunderacomplexscenario. 2.6Summary ThischapterproposesanewtimefairnesssolutionthatapproximatesthegenericadaptationalgorithmforproportionalfairnessamongmultipleWLANs.Thenewsolutionaddressestheproblemoflocation-sensitivecontention,andconsiderablyoutperformstheexistingsolutions.Itisfullydistributed.Eachnodeonlyperformslocalizedoperations.ItisDCF-compatibleandonlyneedstomodifythesizeoftheminimumcontentionwindow. 46

PAGE 47

Figure2-1. NetworkModel. Figure2-2. Twocontentiongroups.Flows(w,z)and(u,v)cantransmitsimultaneously. Figure2-3. TwoMACowsinthesameWLAN. Table2-1. Deliveryrate(inpacketspersecond)andchanneloccupancyunderPISDontheNetworkofFig. 2-9 owrateoccupancyowrateoccupancy (w,z)123.260.161(u,v2)123.270.161(x,y)112.350.147(u,v3)123.160.161(u,v1)123.170.161(u,v4)123.230.161 47

PAGE 48

A B Figure2-4. DCFonthenetworkinFig. 2-3 withow(x,y)at11Mbpsandow(x,z)at1Mbps.A)Thedeliveryrates.B)Thechanneloccupancies. Figure2-5. TwoMACowsindifferentWLANscontend. A B Figure2-6. ChanneloccupanciesoftheowsinFig. 2-5 underDCF.A)Bothowstransmitat11Mbps.B)Flow(x,y)transmitsat11Mbpsandow(u,v)at1Mbps. 48

PAGE 49

A B Figure2-7. IdleSenseonthenetworkinFig. 2-5 withow(x,y)at11Mbpsandow(u,v)at1Mbps.A)thedeliveryrates.B)thechanneloccupancies. A B Figure2-8. CWSPonthenetworkinFig. 2-5 withow(x,y)at11Mbpsandow(u,v)at1Mbps.A)thedeliveryrates.B)thechanneloccupancies. Figure2-9. ThreeWLANsformtwocontentiongroupsg1andg2. 49

PAGE 50

A B C D Figure2-10. Comparingthechanneloccupanciesofows(x,y)and(u,v)inthenetworkofFig. 2-5 .Thetransmissionratesof(x,y)and(u,v)are11Mbpsand1Mbps,respectively.A)Under802.11DCF.B)UnderCWSP.C)UnderIdleSense.D)UnderAIMD/QS+k. Table2-2. Comparingthedeliveryrates(inpackets/sec)andthechanneloccupanciesoftheowsinthenetworkofFig. 2-9 under802.11DCF,CWSP,IdleSense,andAIMD/QS+k. 802.11DCFCWSPIdleSenseAIMD/QS+k owrateoccupancyrateoccupancyrateoccupancyrateoccupancy (w,z)201.080.865155.640.670205.830.886126.830.546(x,y)0.3470.0015.960.0082.070.00375.840.099(u,v1)66.180.086183.470.239107.870.141138.010.180(u,v2)64.480.084182.530.238112.690.147138.590.181(u,v3)67.670.59414.630.12842.290.37122.620.199(u,v4)67.850.088183.670.240109.370.143138.630.181 50

PAGE 51

A B C D Figure2-11. SameasthecaptionofFig. 2-10 ,butthistimeRTS/CTSisturnedoff.A)Under802.11DCF.B)UnderCWSP.C)UnderIdleSense.D)UnderAIMD/QS+k. 51

PAGE 52

A B C D Figure2-12. Comparingthechanneloccupanciesofows(x,y)and(u,v)inthenetworkofFig. 2-5 .Thetransmissionratesof(x,y)and(u,v)areare11Mbpsand5.5Mbps,respectively.RTS/CTSisturnedoff.A)Under802.11DCF.B)UnderCWSP.C)UnderIdleSense.D)UnderAIMD/QS+k. 52

PAGE 53

Figure2-13. SixteenWLANsaredeployedalongtwostreets.Unlessspeciedinthegure,thedefaulttransmissionrateofaowis11Mbps. Table2-3. Comparingthedeliveryrates(inpackets/sec)andthechanneloccupanciesinthenetworkofFig. 2-13 under802.11DCF,CWSP,andAIMD/QS+k. 802.11DCFCWSPIdleSenseAIMD/QS+k owrateoccupancyrateoccupancyrateoccupancyrateoccupancy ow152.880.10879.790.1643.290.00761.120.125ow251.180.067169.370.2214.540.00681.740.107ow354.450.45213.230.1160.1930.00220.590.181ow452.270.068139.290.1826.250.00877.970.102ow521.410.04448.550.099303.070.62278.030.160ow697.180.85336.030.3166.850.06042.510.373ow737.290.076226.470.464367.190.753169.030.347ow815.110.1331.920.01761.010.53617.310.152ow973.080.09572.820.09523.010.031169.950.222ow10469.730.613538.690.702132.730.173257.310.336ow1110.350.01376.780.100120.990.15886.210.112ow1289.710.7887.510.06645.310.39843.300.380ow13107.070.140549.740.717349.470.456264.330.345ow1431.410.0641.870.004195.930.40243,480.089ow1516.820.022362.290.47263.430.083203.110.265ow16105.770.92940.150.35295.090.83560.010.527 53

PAGE 54

CHAPTER3TCPFAIRNESSACROSSMULTIPLEWIRELESSLANS Inthischapter,westudythetransport-layerTCPfairnessproblemamongmultiplecontendingWLANs. IthasbeenobservedthatTCPowsinsomeWLANsmayachieveveryhighthroughputattheexpenseofstarvingtheowsinothernearbyWLANs.Thisunfairnessmaybarelybenoticeableifusersaccessthenetworkintermittently.However,astheInternethasbecomeincreasinglyvideo-rich,unfairnessatitswirelessperimeterwillbenoticeabletotheusersengaginginlongvideo-streamingsessions,suchaswirelessIPTVandremotecameramonitoring.Asweknow,toensurefairchannelbandwidthallocationinwirednetworks,wehaveCSMA/CDattheMAC-layerforone-hoplinksandTCPatthetransport-layerforend-to-endows.Hence,afterweproposeAIMD/QS+kinChapter 2 forachievingMAC-layertimefairnessinWLANs,wewillcontinuetoproposeanovelsolutionforimprovingfairnessamongcontendingTCPows. Therestofthischapterisorganizedasfollows.Section 3.1 presentsthenetworkmodel.Section 3.2 describestheTCPfairnessproblemamongcontendingWLANs.Section 3.3 discussestherelatedwork.Section 3.4 proposesournewprotocolforimprovingTCPfairness.Section 3.5 showsthesimulationresults.Section 3.6 givesthesummary. 3.1NetworkModel Inthissection,wepresentthenetworkmodelusedinthischapter. WeconsideracommonscenariowheremultipleWLANscoexistinanarea.AtypicalWLANconsistsofanaccesspointandmultiplewirelessclients.WeassumeIEEE802.11a/b/gDCForotherDCF-likeprotocolsattheMAClayer.EachWLANselectsonechannel(asub-bandoffrequencyrange)fordatatransmissionsbetweentheaccesspointandclients.Withonlyalimitednumberofnon-overlappingchannels 54

PAGE 55

in802.11wirelessnetworks,aWLANmaycontendinthesamechannelwithnearbyWLANs. WeconsiderTCPows(orTCPconnections)betweenwirelessclientsandInternetservers.EachowpassesanInternetwiredpathandanalwirelesslinkacrossaWLAN.TCPperformswellforcongestiononthewiredpath,butnotforthewirelesslink,whichisthefocusinthiswork.ItiswellknownthatpacketdropsduetowirelesstransmissionfailurescaninterferewithTCPcongestionsignaling.Thisproblemhasbeenstudiedextensivelyandcanbelargelysolvedbyincreasingthenumberofretransmissionattempts.Thisworkinvestigatesaless-studiedissue,i.e.,howinconsistentchannelobservationsbydifferentcontendingnodeswillcausetheTCPrateadaptationtofailinachievingitsfairnessobjective(seetheintroduction)andhowtosolvethisproblem. TheowrateofaTCPconnectionisdenedasthenumberofpacketsthatthewirelessclientsuccessfullyreceivespersecond.ThesendingrateisdenedasthenumberofpacketsthatthesourceofaTCPow(e.g.,aserverontheInternet)sendsoutpersecond.WhenaTCPowcrossesawirelesslink,thechanneloccupancyofthewirelessnodethatforwardsthepacketsisdenedasthefractionoftimethatthenodeoccupiesthechannelforitsdatadelivery(i.e.,DATA/ACKexchanges). Bothaccesspointsandwirelessclientsarereferredtoasnodes.Twonodescontendifone'sdatatransmissioncanmaketheothersenseabusychannelorcorrupttheother'sdatareception.Agroupofmutuallycontendingnodesformsacontentiongroup.Anodemayparticipateinmultiplecontentiongroups.Weusetheconceptofcontentiongrouponlytosimplifythepresentationofthework;theoperationsinoursolutiondonotneedtoknowtheactualcontentionrelationshipamongwirelessnodes.InthediscussionofawirelessnodecarryingaTCPconnection,weoftenusetheowrateofthewirelessnodetorefertotheowrateoftheTCPconnectionthatthenodecarries. 55

PAGE 56

3.2TCPUnfairnessamongContendingWLANs TCPperformsAIMD(AdditiveIncreaseMultiplicativeDecrease)oncongestionwindowsizetodynamicallyadjustsourcesendingrateforachievinghighthroughputandfairness.Specically,thesenderofaTCPowadditivelyincreasesitssendingratetoexploitresidualchannelbandwidth.Anditmultiplicativelydecreasesitssendingratewhencongestionisdetected. OntheInternet,whenarouteriscongestedandbeginstorandomlydroppackets,packetlossisfeltbyallTCPowsthatpassthroughtherouter.Hence,multiplicativedecreasewillbeperformedatallsenders.WeillustratetheratesoftwoTCPowsovertimeinFig. 3-1A .Theratesarenormalizedsuchthatcongestionoccurswhentheirsumisequalto1.Initially,theratesaredifferent.Ateachmultiplicativedecrease,thetworatesarereducedbythesamepercentageandconsequentlythelargerratewillbereducedbyalargeramount,closingthegapbetweenthetwo.Eventually,theratesconvergetothesamevalueafteraseriesofmultiplicativedecreases.Now,considerthewirelessnetworkinFig. 3-2 ,wheretwowirelessnodesaredownloadingfromtheInternetviaaccesspointss1ands2,respectively.Attime6inFig. 3-1B ,whenthecombinedrateofowf1andf2reachesthechannelcapacity,becausenodes2ismorecapableofobtainingthechannelbandwidththannodes1(duetothehidden-terminalproblem),itcansuccessfullyforwardmostofitspackets,whilenodes1observesbufferbuildupandpacketdrop.Consequently,thesenderofowf1detectscongestionandperformsmultiplicativedecrease,whilethesenderofowf2doesnot.Sinceowf1experiencesmultiplicativedecreasemorefrequently,itsaverageratewillbesmallerthanthatofowf2.Thecomplexityoftheproblemwillbecomeextraordinarywhenwemovetomultiplepartially-overlappingWLANs. 56

PAGE 57

3.3StateoftheArt Inthissection,wediscusstheexistingsolutionsforimprovingTCPfairnessinwirelessnetworks.WerstdiscusswhyRED[ 25 ],NRED[ 75 ]andPISD[ 34 ]cannotsolvetheTCPfairnessprobleminWLANs.Wethenpresentotherrelatedwork. 3.3.1RandomEarlyDetection(RED) RED[ 25 ]hasbeenwidelydeployedinwirednetworkstorelievechannelcongestionandenhanceTCPfairness.ThebasicideaofREDisthat,whenarouterdetectscongestion,itwillprobabilisticallydropeachnewarrivalpackettoinformthesendersofallpassingTCPowstoreducetheirsendingrates.Theowsexperiencepacketdroppingproportionaltotheirowrates,whichimprovesfairnessamongcontendingows.However,amongagroupofcontendingWLANs,thereisnosuchacentralrouteractingasatrafccoordinator.TheTCPowscontendingatthebottleneckmaypassdifferentphysicalnodes.Thesenodesmayhavedifferentchannelobservationsintermsofbufferoccupancy,transmissionfailurerateorchannelidletime.Consequently,theywillprovideinconsistentfeedbackstothesendersofTCPows,causingunfairness. 3.3.2NeighborhoodRED(NRED) TosolveRED'sproblem,Xuetal.proposedNRED[ 75 ]forwirelessnetworks.ThebasicideaofNREDistomakesureallcontendingnodesobservethesamechannelcondition.Specically,allnodesmonitorthechannelbusy/idlestatusandperiodicallysynchronizetheirobservationsbyexplicitmessageexchanges.Consequently,allnodescandetectcongestionsimultaneously.Theywillthendroppacketswithprobabilitiesproportionaltotheirowrates,whichhelpenhanceTCPfairness.However,NREDreliesontheassumptionthatallcontendingnodescanoverheareachother,whichmaynotalwaysbetrue.MultiplecontendingWLANsmaybedeployedinalargeareasuchthattheymaycontendforthesamechannelyetnotbeabletodirectlycommunicateduetothelimitedtransmissionrange. 57

PAGE 58

WeshowanexampleinFig. 3-3 .TherearethreeoverlappingWLANsformingtwocontentiongroups.ThenodesindifferentWLANsareoutsideofeachother'stransmissionrangeandthustheycannotdirectlycommunicate.Fig. 3-4 showsthatthenodess1,s2ands3havequitedifferentchannelobservationsintermsofchannelidletime.Nodess1ands3alwayssensemuchmorechannelidletimethannodes2does.Sincetheycannotsynchronizetheirperceptionsofthechannelconditionsbyexchangingmessages,theywillexperienceunfairpacketdropping.ThisunfairpacketdroppingwillleadtodifferentratesoftheTCPowspassings1,s2ands3.Fig. 3-5 showsthatf1andf3achieveveryhighthroughputwhilef2isalmoststarved. 3.3.3ProportionalIncreaseSynchronizedMultiplicativeDecrease(PISD) Aswecansee,theTCPunfairnessstemsfromtheMAC-layerunfairness.CanwesolvetheTCPfairnessproblembyachievingtheMAC-layerfairness? In[ 34 ],Yingetal.proposedPISDtoachieveMAC-layerratefairnessinasymmetricwirelessnetworksettings.Withoutrequiringdirectcommunications,PISDusesachanneljammingtechniquetoforceallcontendingnodestodetectcongestion.Then,allnodesperformsynchronizedratereductiontoresolvethecongestion.Eventually,allowscanachievethesameowrate.However,PISDonlyfocusesontheMAC-layeranditlacksofinteractionwithTCPcongestioncontrolmechanism.Forexample,PISDneedstoobserveacertainnumberofqueuedpacketstodetectcongestion.Andithastopushoutenoughpacketswithinashortperiodoftimetojamthechannel.ButthesenderofaTCPowmayhaveextremelylowsendingratewhentheowexperiencesheavychannelcontention.Consequently,thenodecarryingthisTCPowatthebottleneckmaynotbeabletoaccumulateenoughpacketstotriggerandperformPISD'sratecontrolfunction. WeperformasimulationonthenetworkofFig. 3-6 ,whereveTCPowsformtwocontentiongroups.Theheavychannelcontentionmaypushtherateofowf2to 58

PAGE 59

extremelylow.Table 3-1 showsthattheowf2isstillstarvedunderPISD.ThisexampledemonstratesthatMAC-layerfairnessdoesnotdirectlyleadtoTCPfairness. 3.3.4OtherRelatedWork Yangetal.[ 78 ]proposedanon-work-conservingschedulingforimprovingTCPfairnessamongowscrossingwirelessadhocnetworksandwirednetworks.ThebasicideaistosetatimertocontrolthespeedofsendingpacketstotheMAClayer.Thelengthofthetimerisdeterminedbytheowrate.Ittriestoenhancefairnessamongcontendingowsbypunishinghigh-rateows.Thisapproachissimple;however,itsignicantlydowngradestheoverallthroughput. Erginetal.[ 21 ]veriedtheTCPperformancedegradationbytestbedtracesinanunplanneddeploymentofWLANs.SomeworkshasbeenproposedtomitigatetheinterferencebetweentwonearbyWLANsbytransceiverparameteroptimization[ 6 70 84 ],channelassignment[ 49 50 ]andassociationcontrol[ 8 49 ].However,theymaycausechannelunderutilizationorincurlongdelays.Moreover,theycannotsolvetheTCPfairnessprobleminadensedeployment. ThereisalargebodyofworkonTCPupstream/downstreamfairnessinWLANs[ 4 5 54 57 ].Pilosofetal.[ 57 ]illustratedtheTCPupstream/downstreamthroughputanomalywithinWLANsthroughanalysis,simulationandexperiment.TheyproposedtocontrolthereceiverwindowofallTCPowsattheaccesspointsidetoprovidemorechannelbandwidthfordownstreamows.Parketal.[ 54 ]introducedchannelaccesscostthatisusedtoinformtheTCPsendertoadjustitssendingratetoachieveper-stationfairness.Aguileraetal.[ 5 ]usedIdleSense[ 31 ]methodtoassignsufcientbandwidthtotheaccesspoint.Abeysekera[ 4 ]dynamicallyadjustedtheminimumcontentionwindowataccesspointsaccordingtotheratioofthetotalrateofdownstreamowstotherateofanupstreamow.AlltheaboveworksfocusontheTCPunfairnesswithinasingleWLANandtheycannotbeappliedtothescenariosofmultipleoverlappingWLANs. 59

PAGE 60

Anotherlineofresearchistoaddressfairnessissuesovermultihopwirelessnetworks[ 58 73 76 77 80 ].However,theTCPunfairnessamongnearbyWLANsstudiedinthisworkhasdifferentemphasis.Inamultihopnetwork,aowcancarrycertaininformationforthenodesalongaroutingpath[ 58 80 ].Contrarily,twoWLANsmaycontendforthesamechannelbuttheyhavenomeansofdirectcommunication. 3.4WirelessProbabilisticDrop(WPD) Inthissection,werstexplainthebasicideabehindoursolutionWPD(WirelessProbabilisticDrop),thenintroduceseveralratecontroltechniques,andnallydescribehowtheyworktogethertoenhanceTCPfairness. 3.4.1BasicIdea EachwirelessnodemonitorsthesizeofthelocalMAC-layerqueueandmeasuresitschanneloccupancy.Anodedetectschannelcongestionwhenitsqueuesizeexceedsacertainthreshold1.Ifthenodethatdetectscongestiondropspacketsimmediately,itmaycauseunfairnessbecausethecontendingnodesinthesamesaturatedchannelmaynotyetdetectcongestionandhencetheymaynotdroppackets.Tosolvethisproblem,insteadofdroppingpacketsimmediately,wemakethenodetotransitfromitsnormalstatetoeitherresolutionorsignalingstateforaperiodoftimeT(suchas1secondinoursimulations). Intheresolutionstate,thenodeemulatesRED(RandomEarlyDetection)[ 25 ]byprobabilisticallydroppingpacketstoresolvecongestion.ItcausestheTCPsourcetoobservetripleduplicateACKsandconsequentlyreducethesendingratebyshrinkingthecongestionwindow.Inthesignalingstate,insteadofdroppingpackets,thenodeaggressivelycompetesforchannelaccessbymodifyingitsMACparameters,suchasreducingtheminimumcontentionwindow.Thisiscalledtheaggressivemode.Our 1Thesamemethodisusedin[ 34 ],whosefocusishoweveronMACows,insteadofTCPows. 60

PAGE 61

previousresearchhasshownthatreducingtheminimumcontentionwindowisveryeffectiveforenhancinganode'sabilitytooccupythechannel.Asthisnodeenterstheaggressivemodeforaperiodoftimeandconsumesmorechannelbandwidth,othercontendingnodessendoutfewerpackets.Theywillobservetheirqueuesbuildupandexceedthethreshold.Inthisway,thecongestionsignalisspreadtothesenodeswithoutexplicitcommunication,causingthemtotransittoeitherresolutionorsignalingstate. Theprobabilityforanodetochoosetheresolutionstateisproportionaltothenode'scurrentchanneloccupancy.Hence,nodeswithlowchanneloccupanciestendtoselectthesignalingstateandgrabmorechannelbandwidth,whilenodeswithhighchanneloccupanciesaremorelikelytoselecttheresolutionstatethatcausesthetraversingTCPowstoreducerates,whichhelpsachievingfairness.Itmayhappenthatthecontendingnodesallchoosethesignalingstate.Inthiscase,somenodeswillobservequeueoverows,andthestandardtaildropisperformedtoresolvecongestion. BelowwegivemoredetaileddescriptionforvarioustechniquesthatareemployedbyWPD. 3.4.2PeriodicalMeasurementofStateInformation Eachwirelessnodeperiodicallymeasuresitsaveragequeuesizeqandaveragechanneloccupancyutolearnthechannelcondition.ItalsomeasurestheaverageraterofeachTCPowthatitcarries.Attheendofeverymeasurementperiodm(suchas0.1second),thenewvalueofqiscomputedasq:=(1)]TJ /F3 11.955 Tf 12.25 0 Td[(w)q+wq,where:=istheassignmentoperator,qisthecurrentqueuesize,and0
PAGE 62

isknownastheminimumcontentionwindow,whichistheinitialwindowsizefortheexponentialbackoffalgorithm.ThedefaultvalueofCWminis31.InWPD,anodeinthesignalingstatewillentertheaggressivemodebytemporarilyreducingCWmintoasmallvalue(suchas3inoursimulations)foraperiodoftime.WithareducedvalueofCWmin,thenodeismorecapableofobtainingthechannelthanothercontendingnodes.Asthisnodeaggressivelyoccupiesthechannel,lessbandwidthisleftforothers,whosequeuelengthsarethenforcedup.Hence,theaggressivemodeisusedbyanodethathasdetectedcongestiontospreadthecongestionsignaltoothercontendingnodes. 3.4.4ProbabilisticDropping Intheresolutionstate,anoderesolvescongestionbydroppingpackets,whichinformstheTCPsourcestoreducetheirsendingrates.WeadoptaprobabilisticdroppingalgorithmthatemulatestheRED[ 25 ]inwirednetworks:Abaseprobabilitypbisrstcomputedaspmaxu,wherepmaxisapredenedmaximumpacketdroppingprobability.Thenthenodedropseacharrivalpacketwithaprobabilitypa:=pb=(1)]TJ /F3 11.955 Tf -435.09 -23.91 Td[(countpb),wherecountisthenumberofpacketsthathavebeenforwardedsincethepreviouspacketdrop.Hence,thedroppingprobabilityisanincreasingfunctionofthenode'schanneloccupancy.Anodewithhigherchanneloccupancywilldropmorepacketsinitsresolutionstate,whichhelpsimprovefairnessamongthecontendingnodes. 3.4.5MinimumRateAssurance Toavoidstarvation,itisdesirabletoensurethateachTCPconnectionhasaminimumratermin,evenforonethatiscarriedbyawirelessnodeundertheheaviestcontention.Asanoptimization,whenawirelessnodedetectsthattheaverageowraterisbelowrmin,itenterstheaggressivemodeuntilrreachesrmin.EnsuringtheminimumowratehasapositiveimpactinWPDbecauseitassuresthatanodeinthesignalingstatewillhaveenougharrivalpacketstooccupythechannelinordertospreadthe 62

PAGE 63

congestionsignal.Withoutenougharrivalpackets,thenodewouldhavetoinjectfakepacketstooccupythechannel,whichwastesbandwidth. 3.4.6AdaptiveIntermittentRelease Whenmultiplewirelessnodescontendinthesamechannel,ifallofthemcontinuouslycompeteforchannelaccess,collisioncanhappenfrequently.Wedesignamethodcalledadaptiveintermittentreleasetointerruptthepatternofcontinuouschannelaccess.Anodestoresthereceivedpacketsinanetwork-layerqueueandreleasesthepacketstoaMAC-layerqueuefortransmission.Basedonthechannelcondition,wecontrolthetimetips(calledInter-PacketSpacing)betweentwoconsecutivereleases.AftertheMAClayertransmitsallitspackets,itmayhavetowaitforthenetworklayertoreleasethenextone.Weobservethat,atthecongestiontime,suchwaithelpsreducecollisions. Theadaptationoftipsisdescribedasfollows:Whenanodeperformsprobabilisticdroppinginitsresolutionstate,itperformsmultiplicativeincreaseontipssuchthattips:=(1+)tipsaftereachmeasurementperiodmuntilthecongestionisresolved(i.e.,theaveragequeuelengthfallsbelowthethreshold),whereisthefactorofmultiplicativeincrease.Inallothercases,thenodeperformsadditivedecreaseontipssuchthattips:=tips)]TJ /F8 11.955 Tf 12.09 0 Td[(,whereisaconstantvalue.Iftipsbecomeslessthan50s,wesetittobe50s.Therationalebehindtheaboveadaptationistokeeptipssmallunlesscongestionoccurs.Whenthathappens,weincreasetheinter-packetspacingfornodesintheresolutionstate.Aswewillseeinthenextsubsection,nodeswithhigherchanneloccupanciesaremorelikelytobeinthisstate.Theirlargerinter-packetspacingsmakeiteasierforothernodeswithlowerchanneloccupanciestoaccessthechannelduringcongestion. Whenadaptiveintermittentreleaseisused,thequeuelengthqismeasuredbasedonthecombinedlengthofthenetwork-layerqueueandtheMAC-layerqueue. 63

PAGE 64

3.4.7WPDProtocol WenowassembletheaboveratecontroltechniquestoformWPD.Eachnodecanbeinoneofthreestates:normal,resolutionandsignaling. Inthenormalstate,anodereleasespacketsfromthenetworklayertotheMAClayerthroughadaptiveintermittentrelease.Aftereachmeasurementperiodm,thenodecomputesq,uandr.WhenqisbelowathresholdH,itadditivelydecreasestips.OnceqexceedsH,itdrawsarandomnumberrdfrom[0,1].Ifrdu,thenodetransitstotheresolutionstatetoperformprobabilisticdropping;otherwise,ifrd>u,ittransitstothesignalingstatetoentertheaggressivemode.Itisclearthattheprobabilityfortheresolutionstateisproportionaltothechanneloccupancy. Intheresolutionstate,anodeperformsprobabilisticdroppingtoinformtheTCPsourcetoreduceitssendingrate.Inthemeantime,itperformsmultiplicativeincreaseontipstogiveothercontendingnodesmorechanceofaccessingthechannel.AfteraperiodoftimeT,thenodetransitstothenormalstate. Inthesignalingstate,anodeusestheaggressivemodetoconsumemorechannelbandwidthinordertospreadthecongestionsignaltoothernodes.AfteraperiodoftimeT,thenodetransitstothenormalstate.Whentheminimumrateassuranceisimplemented,anodemayalsotemporarilyentertheaggressivemodetobringupitsowrateifitistoolow. ThepseudocodefortheoperationsofWPDisgivenbelow. WirelessProbabilisticDrop1.attheendofeachmeasurementperiodm2.q:=(1)]TJ /F14 10.909 Tf 10.9 0 Td[(w)q+wq3.u:=(1)]TJ /F14 10.909 Tf 10.9 0 Td[(w)u+wu4.r:=(1)]TJ /F14 10.909 Tf 10.91 0 Td[(w)r+w(pks=m)5.ifq>Hthen6.rd1:=random(0,1)7.ifrd1>uthen 64

PAGE 65

8.duringaperiodTdo9.CWmin:=3foraggressivemode10.endduring11.restoreCWmintotheoriginalvalue12.else13.tips:=(1+)tips14.duringaperiodTdo15.pb:=pmaxu16.count:=017.foreacharrivalpacketdo18.count:=count+119.pa:=pb=(1)]TJ /F14 10.909 Tf 10.91 0 Td[(countpb)20.ifpa<0thenpa=121.rd2:=random(0,1)22.ifrd2
PAGE 66

3.5.1SimulationSetup Allsimulationsareperformedinns-2[ 1 ].WeuseTCPNewRenoatthetransportlayer.ThesizeofaTCPdatapacketis1,000bytes.FTPisusedtogenerateeachTCPow.WeuseIEEE802.11DCFattheMAClayer.RTS/CTSisturnedoffbydefaultduetoitshighoverhead,whichistoday'scommonpractice.Forthetransmissionrangeandthecarrier-sensingrange,weusens-2defaultvalues,whichare250mand550m,respectively.Theinterferencerangeisequaltothelengthofawirelesslinktimes1.78,whichisalsothedefaultsettinginns-2.Thetransmissionrateis11Mbps.WecompareWPDwithDropTail(whichdropspacketswhenthequeueisoverowed)andNRED[ 75 ],whoseparametersarechosenbasedontheoriginalpaper.WesettheparametersforWPDasfollows:theratemeasurementperiodmis0.1second,thethresholdHis5packets,thestateperiodTis1second,theminimumraterateminis25packetspersecond,themaximumdropprobabilitypmaxis0.03,theweightingfactorwis0.20,theis2andtheis50s.Eachsimulationisexecutedfor150seconds,andtheaverageTCPratesarereported. 3.5.2FairnessIndex Weusethetheoretically-computedratesundertheproportionalfairnessmodel[ 37 ]asthebenchmarksforcomparison.Proportionalfairnessstrivestobalancebetweenthefairnessrequirementandtheoverallnetworkthroughput.Besidesthesebenchmarkowrates,wealsocomputeanoverallfairnessindex,whichisthesummationofautilityfunction,Pf2Flnrf,whererfistherateofaTCPowfandFisthesetofows.Ourcomputedbenchmarkratesmaximizethisindex.Amongthefairnesssolutionsundercomparison,onethatachievesahighervalueofthefairnessindexhasbetterperformanceintermsofproportionalfairness. 3.5.3CaseStudy:ABaseScenario WerstperformacasestudyonabasescenarioinFig. 3-3 ,wherethreeaccesspointss1,s2ands3formtwocontentiongroupseventhoughtheyareplacedoutsideof 66

PAGE 67

eachother'stransmissionrange.WeobservethatunfairnessarisesamongthethreeTCPows,f1,f2andf3.AsshowninTable 3-2 ,whenDropTailisusedasthequeuemanagementscheme,f1andf3canbothobtainowratesabove430packetspersecond,whereasf2isalmoststarved.Inthiscase,withouttheabilityfortheaccesspointstoexplicitlysynchronizetheirchannelobservations(duetobeingoutsideofeachother'stransmissionrange),NREDmakeslittleimprovement.ForWPD,weincrementallydeploythetechniquesproposedinSection 3.4 todemonstratetheirindividualimpactontheperformance.Werstadopttheaggressivemode(Section 3.4.3 )andprobabilisticdropping(Section 3.4.4 ).Table 3-2 showsthattheowrateoff2isincreasedto35.83packetspersecond.Aftertheminimumrateassurance(Section 3.4.5 )isincorporated,theowrateoff2isincreasedto74.75packetspersecond,whichdemonstratesitspositiveimpactontheperformanceofWPD.Finally,weapplytheadaptiveintermittentrelease(Section 3.4.6 ),whichfurtherimprovesfairness.Theowrateoff2isincreasedto126.46packetspersecond.ThenalresultsofWPDareclosetothetheoreticalowratesundertheproportionalfairnessmodelthatareshowninthelastrowofthetable. WecompareWPDwithDropTailandNREDintermsoftheirfairnessindicesinTable 3-3 ,whichshowsthatWPDhasthehighestindexvalue,closetotheoptimalindexvaluethatisachievableundertheproportionalfairnessmodel. 3.5.4ScalabilityStudy:ThreeContentionGroups NextweevaluatetheperformanceofWPDonascenariowithmorethantwocontentiongroups.Fig. 3-7 hassixTCPowsbelongingtothreecontentiongroups.ThesimulationresultsinTable 3-4 showthatf3hasverylowthroughputvaluesunderDropTailandNRED.WPDimprovesitto75.30packetspersecond.ItisworthnotingthatcomparingwithDropTailandNRED,WPDalsoachievesbetterfairnessamongtheotherveows. 67

PAGE 68

3.5.5AStreetScenario Fig. 3-8 showsamorecomplicatedscenario,wheretwenty-fourTCPowsarecarriedbyWLANsrandomlydeployedalongtwocrossingstreets.Therelativepositionsofthenodesaredrawninthegure.Thelengthofeachwirelesslinkis150m.Thecontentionrelationshipamongtheows,whichismuchmorecomplicatedthanthoseinthepreviousscenarios,isautomaticallydeterminedbyns-2. Table 3-5 showsthat,underDropTail,sixTCPowsarestarved(lessthan10packetspersecond)andveTCPowshavelowowrates(lessthan40packetspersecond).NREDperformsbetterthanDropTail.ButitstillhasthreestarvedTCPowsandthreelow-rateows.UnderWPD,allowscanachievedecentthroughputvalues.WebelievethatthissimulationdemonstratestheeffectivenessofWPDincomplexsettings. 3.6Summary ThischapterproposesWPDforachievingTCPfairnessamongmultiplecontendingWLANs.Itisafullydistributedsolutiononlybasedonlocalinformation.ItcanworkseamlesslywithcurrentTCPandMACstandards.Itimplicitlyspreadscongestioninformationtoallcontendingowsinabottleneckwithoutrequiringdirectcommunications.ExtensivesimulationsshowthatitcansignicantlyimproveTCPfairnessinvariousscenarioswhenotherexistingsolutionsfail. 68

PAGE 69

A B Figure3-1. WhyTCPAIMDdoesn'tworkinwirelessnetworks?A)TCPsynchronizedAIMDinwirednetworks.B)TCPunfairAIMDinwirelessnetworks. Figure3-2. TwocontendingWLANs,wheres1ands2aretwoaccesspoints.ATCPowfromaserverontheInternetpassesthroughanaccesspointtoawirelessnodeineachWLAN.Notethat,onlythewirelesspartofeachow(fromanaccesspointtoawirelessclient)isdrawninallguresthroughoutthischapter. Figure3-3. ThreepartiallyoverlappingWLANsformtwocontentiongroupsg1andg2,wheres1,s2ands3arethreeaccesspoints.ATCPowfromaserverontheInternetpassesthroughanaccesspointtoawirelessclientineachWLAN. 69

PAGE 70

Figure3-4. Thechannelidletimesensedbythethreesenderss1,s2ands3inFig. 3-3 underNRED. Figure3-5. TheowratesofthethreeTCPows,f1,f2andf3inFig. 3-3 underNRED. Figure3-6. FiveTCPowsformtwocontentiongroups. 70

PAGE 71

Table3-1. Comparingtheowrates(inpackets/sec)oftheowsinthenetworkofFig. 3-6 underPISD. f1f2f3f4f5 owrate356.835.34141.4589.43166.75 Table3-2. Comparingtheowrates(inpackets/sec)oftheowsinthenetworkofFig. 3-3 underDropTail,NREDandWPD. f1f2f3 DropTail433.341.60430.58NRED379.9717.48377.00aggressivemode+probabilisticdropping379.7635.83378.75aggressivemode+probabilisticdropping+minimumrateassurance339.4374.75343.20aggressivemode+probabilisticdropping+minimumrateassurance+adaptiveintermittentrelease(WPD)252.34126.46260.54OptimalProportionalFairness288.67144.33288.67 Table3-3. ComparingtheFairnessIndexinTermsofProportionalFairnessinthenetworkofFig. 3-3 underDropTail,NREDandWPD. DropTailNREDWPDOptimal 12.6114.7315.9316.30 Figure3-7. SixWLANsformthreecontentiongroupsandeachWLANcontainsoneTCPow. 71

PAGE 72

Table3-4. Comparingtheowrates(inpackets/sec)oftheowsinthenetworkofFig. 3-7 underDropTail,NREDandWPD. owDropTailNREDWPD f1221.65203.17153.03f2234.10196.89189.98f37.5022.2975.30f4205.91154.11136.07f5205.28154.71135.42f655.58114.38145.17 Figure3-8. TwentyfourWLANsarerandomlydeployedalongtwocrossingstreetsandeachWLANcontainsoneTCPow. 72

PAGE 73

Table3-5. Comparingtheowrates(inpackets/sec)oftheowsinthenetworkofFig. 3-8 underDropTail,NREDandWPD. owDropTailNREDWPDowDropTailNREDWPD f1216.4235.4163.9f130.83.545.6f2215.9174.484.2f14191.5181.4162.6f31.16.8163.9f15242.3226.7181.1f4205.2189.5165.0f160.40.852.9f5218.6205.699.8f17217.4207.3194.5f69.825.5106.4f18214.2209.5167.4f710.291.989.4f19388.9198.9125.9f823.595.4124.4f2017.0117.0168.0f9382.3183.0144.6f21402.9237.4155.1f107.138.288.2f224.026.184.9f1130.6104.0147.7f2326.293.6146.7f12397.7286.3218.6f24403.2301.5217.4 73

PAGE 74

CHAPTER4IMPROVINGREADINGTHROUGHPUTINLARGERFIDSYSTEMS Inthischapter,wefocusonimprovingreadingthroughputinlargeRFIDsystems. InaRFID-equippedlargewarehouse,theRFIDreaderneedstoperiodicallycollectinformationfromthetagsforthepurposeofinventorymanagement.Hence,thereadingthroughputbecomesaveryimportantperformancemetric.Duringthereadingprocess,multipletagsmayreporttothereadersimultaneously,causingcollision.Existingreadingprotocolsstrivetoincreasethereadingthroughputbyminimizingthechanceofcollision.Ithasbeenprovedthatthetraditionalmethodshavereachedtheirphysicallimit.Wewilldemonstrateafundamentallydifferentapproachtoimprovethereadingthroughputbyextractingusefulinformationfromthecollision. Therestofthischapterisorganizedasfollows.Section 4.1 presentsthemotivationofourwork.Section 4.2 givestheproblemdenition.Section 4.3 discussestherelatedwork.Section 4.4 proposesacollision-resolutiontagidenticationprotocolandderivestheoptimalsystemparameters.Section 4.5 improvestheprotocolforlessoverhead.Section 4.6 showsthesimulationresults.Section 4.7 givesthesummary. 4.1Background 4.1.1Motivation ConsideraRFIDsystemwithalargenumberofactive(orsemi-active)tagsdeployedinaregion.WeassumethattheRFIDreaderandthetagstransmitwithsufcientpowersuchthattheycancommunicateoveralongdistance.TheproblemisforthereadertocollecttheIDsofalltagswithinthecommunicationrange.Ifthecommunicationrangecannotcoverthewholedeploymentregion,thereadermayhavetoperformthereadingprocessatseverallocationsandremovetheduplicateIDswhensometagsarecoveredbymultiplereadings.Inthischapter,wefocusonthereadingoperationatasinglelocation,Ourgoalistooptimizethereadingthroughput,whichistheaveragenumberoftagIDsthatthereaderisabletocollectineachsecond. 74

PAGE 75

Duringthereadingprocess,multipletagsmaytransmittheirIDssimultaneously,causingcollision.Somecollision-avoidancemethodssuchasFDMAorCDMArequiresophisticatedschedulingmethodstominimizebandwidthwasteduetoidlesub-channelsorunusedcodes[ 65 ].TheoverheadforsophisticatedschedulingcanbetoocostlyforaRFIDsystemwhereeachtagonlyneedstodeliveronepieceofinformation(i.e.,itsID)tothereader.Hence,contention-basedtime-slottedprotocolshavebecometheindustrialstandards[ 7 ]. Inacontention-basedprotocol,eachtagtransmitsitsIDinatimeslotwithareportprobabilitypthatistunedtoreducecollision.AtagstopswhenitreceivestheacknowledgementfromthereaderthatitsIDhasbeensuccessfullyreceived.Itcanbeshownthattheoptimalreadingthroughputistheoreticallyboundedby1 eT,whereeisthenaturalconstantandTisthelengthofatimeslot[ 61 ].Insuchaprotocol,36.8%ofthetimeslotswillbeidleand26.4%oftheslotswillhavecollision. Canwedobetterthan1 eT?Weobservethatthereadingthroughputcanbeimprovedifwemakegooduseofthecollisionslots.Supposethereaderreceivesamixedsignalinacollisionslotwhenbothtagt1andtagt2transmittheirIDs.Inalaterslot,ifthereaderreceivestheindividualsignalfortheIDoftagt1,itcanremovethissignalfromthemixedsignalandrecovertheindividualsignalfortheIDoftagt2. ConsidertheexampleinFig. 4-1 ,wherefourtagstransmittheirIDstothereader.InFig. 4-1A ,whenacontention-basedprotocolisused,ittakes11slotsforthereadertocollectallfourIDs.InFig. 4-1B ,whenacollision-resolutionprotocolisusedtoresolvecollision,only6slotsarenecessary.Inparticular,whenthereaderreceivesthesignalfromt1inthethirdslot,itremovesthissignalfromthemixedsignalreceivedintherstslotandrecoverstheIDoft4.Similarly,whenitreceivesthesignalfromt3inthesixthslot,italsolearnstheIDoft2fromthefourthslot. 75

PAGE 76

4.1.2AnalogNetworkCoding(ANC) Canweremoveanindividualsignalfromamixedsignaltorecovertheotherconstituentsignal?Thisquestionhasrecentlybeenbroughtupinthecontextofphysical-layernetworkingcode.Signicantprogresshasbeenmadeinboththeoryandimplementation[ 36 81 ]. Kattietal.implementedtheAnalogNetworkCoding(ANC)anddemonstrateditseffectivenessintheAlice-BobnetworkshowninFig. 4-2 .Traditionally,fourtimeslotsareneededforAliceandBobtoexchangeapairofmessages:AlicesendsamessagetotherouterandtherouterforwardsittoBob,andviceversa.However,byusingANC,onlytwotimeslotsarenecessary:AliceandBobtransmitsimultaneouslytotherouter.Insteadofdroppingthemixedsignal,theroutersimplyampliesandbroadcastsittobothAliceandBob.AlicesubtractsherownsignalfromthemixedsignalanddecodesBob'smessage.Similarly,BobcanextractAlice'smessage. WebrieydescribethemethodusedbyKattietal.Readersarereferredto[ 36 ]formoredetails.TheANCprotocolisdesignedbasedonMSK(MinimumShiftKeying)[ 55 ]andhasbeenimplementedusingsoftwaredenedradios.ThesignaltransmittedbyAlicecanberepresentedas s[n]=Aseis[n],(4) whereAsistheamplitudeofthenthsampleands[n]isitsphase.Similarly,Bob'ssignalcanberepresentedas s[n]=Bseis[n].(4) IfAliceandBobtransmitsimultaneously,therouterwillreceivethesumofthetwosignals,whichcanberepresentedas y[n]=h0Asei(s[n]+0)+h00Bsei(s[n]+00),(4) whereh0andh00arethechannelattenuationand0and00arethephaseshift.Werewriteitforsimplicityas 76

PAGE 77

y[n]=Aei[n]+Bei[n],(4) whereA=h0As,B=h00Bs,[n]=s[n]+0,and[n]=s[n]+00.Uponreceivingthemixedsignalfromtherouter,AlicecanresolveAandBfromthefollowingtwoenergyequations[ 29 36 ], =E[jy[n]j2]=A2+B2,(4) =2 WXjy[n]j2>jy[n]j2=A2+B2+4AB=,(4) whereE[.]istheexpectationandWisasamplingwindowsize.InMSK,abit`1'isrepresentedasaphasedifferenceof=2overatimeintervalt,whereasabit`0'isrepresentedasaphasedifferenceof)]TJ /F8 11.955 Tf 9.29 0 Td[(=2overt.Forexample,ifthephasedifferencebetweenthe(n+1)thsampleandthenthsample,[n+1])]TJ /F8 11.955 Tf 12.08 0 Td[([n],is=2,thenabitistransmitted.SinceAliceknowsherownsignal,from( 4 ),shecanestimatethephasedifferencesofBob'ssignal,[n+1])]TJ /F8 11.955 Tf 12.36 0 Td[([n],whichcanbetranslatedintothebitstreamsentbyBob[ 36 ]. ThetaskofresolvingthemixedsignalinacollisionslotinaRFIDsystemissimplerthanthesametaskinthewirelessnetworkshowninFig. 4-2 .First,Aliceknowstheamplitudeofhersignalwhenitistransmittedout,butshedoesnotknowtheamplitudeofhersignalwhenitreachestherouterandmixedwithBob'ssignal.WhenAlicereceivedtheampliedmixedsignalfromtherouter,itbecomesdifcultforhertoremoveherownsignalfromthemixedone.IntheRFIDsystem,supposethereaderreceivesthemixedsignalfromt1andt2inoneslotandtheindividualsignaloft1inanotherslot.Becausethesamesignaloft1appearsinthetwoslots,itbecomeseasiertoremoveitfromthemixedsignal. Second,itisverydifculttosynchronizetransmissionsbetweenwirelessnodes,andthustheproposedANCprotocolhastointroduceacomplicatedmechanismtorelievethisproblem,whereastransmissionsinaRFIDsystemcanbesynchronizedbythereader'ssignal. 77

PAGE 78

Giventhatthetechnologyforcollisionresolutionexists,thenextquestionishowtooptimallyuseittomaximizetheperformanceofawirelesssystem.WewillanswerthequestioninthecontextofRFIDsystems. 4.2TerminologyandProblemDenition 4.2.1Terminology Duringtheexecutionofatime-slottedcontention-basedprotocol,ifnotagtransmitsinatimeslot,wecallitanemptyslot.Ifonetagtransmits,itiscalledasingletonslot.Ifmorethanonetagtransmits,itisacollisionslot.Inparticular,ifktagstransmitsimultaneously,theslotiscalledak-collisionslot,wherek2.Inordertoguardagainstchannelerror,eachIDcarriesaCRCcode.Inasingletonslot,theRFIDreaderreceivestheIDsignalfromasingletag.ItwillverifythecorrectnessofthereceivedIDbycheckingtheCRCcode. 4.2.2ResolvableCollisionSlots Anemptyslotiseasytoidentifybecausenosignalisreceived.ThereadercandistinguishasingletonslotfromacollisionslotbyrstconvertingthesignalintoanIDandthenverifyingthecorrectnessoftheCRCcode.Foracollisionslot,thereaderrecordsamixedsignalthatcombinestheindividualsignalsofthetagsthattransmitsimultaneously.Inlatersingletonslots,thereaderwillreceivetheindividualIDsignalsfromsomeofthosetags.WhenthereadereventuallyreceivestheIDsignalsfromallbutoneofthosetags,wesaythek-collisionslotisresolvableifwecanderivetheIDsignalofthelasttagbyremovingthe(k)]TJ /F5 11.955 Tf 12.83 0 Td[(1)IDsignalsfromthemixedsignal.TheexperimentalstudyofAnalogyNetworkCodingbyKattietal.in[ 36 ]hasshownthat2-collisionslotsareresolvable. 4.2.3ProblemDenition ThemainproblemwewanttosolveinthisworkishowtooptimallyapplyanalognetworkcodingtomaximizetheRFIDreadingthroughput.Wedesignacollision-awaretagidenticationprotocolandderivetheoptimalreportprobability(atwhichatag 78

PAGE 79

transmitsitsIDineachslot)thatmaximizesthenumberofsingletonand2-collisionslots(fromwhichIDscanbeextractedbyANC). In[ 36 ],theauthorsstatethatANCcanbeappliedtoresolvecollisioninvolvingmorethantwosignals.Ononehand,aswewilldemonstrateinSection 4.6 ,resolving2-collisionslotsbasedontoday'stechnologywillalreadyprovideapracticallysignicantboosttothereadingthroughout.Ontheotherhand,insteadofrestrictingourworkto2-collisionslots,wedecidetogeneralizeourprotocolsothatitcanworkwithanyfutureANCmethodthatresolves-collisionslots,where(2)isaninputparameter.Suchgeneralizationshedslightontheamountofthroughputgainthatcanpossiblybeobtainedthroughanalognetworkcoding.Inparticular,theresultsinSection 4.6 showthatthereadingthroughputwillbehigherwhenislarger(becausemorecollisionslotsbecomeuseful).However,therateofthroughputimprovementdiminishesquicklyasincreases.Hence,itisnotnecessarytomaketoolarge.Inpractice,weexpecttobeasmallnumber(suchas2,3or4). Clearly,ANCandotherphysical-layernetworkcodingmethodscanbeappliedinvariousdifferentcommunicationcontexts,eachofwhichhasitsuniquetechnicalchallenges.Forexample,collisionresolutionhasbeenusedinsatelliteaccessnetworks,whereeachterminaltransmitsasinglepackettwiceattworandomly-selectedtimeslotsineachMACframe[ 15 ].Thethroughputupperboundcanbepredictedifthenumberofpacketsperslotisknown(whichrequirestheknowledgeofthenumberoftransmittingterminals).Inourcontext,wedonotderiveathroughputupperboundforagivensetofsystemparameters.Instead,wedeterminethebestsystemparameterthatoptimizesthethroughput.Wedonotassumetheknowledgeforthenumberoftagsthatisparticipatingintheprotocol.Infact,thisnumberchangesovertimebecauseafteratagsuccessfullydeliversitsIDtothereader,itwillstoptransmitting.Atagmaytransmitforone,twoormoretimesatanytimeslotsduringthereadingprocess.Moreover, 79

PAGE 80

becausethenumberofparticipatingtagschanges,theoptimalsystemparameteralsochangesovertime. 4.3StateoftheArt Allexistingcontention-basedtagreadingprotocolsarecalledanti-collisionprotocolsbecausetheytreatcollisionaswasteandtrytoavoidit[ 24 ].Mostoftheseprotocolsfallintotwoclasses:theALOHA-basedprotocols[ 16 42 56 63 68 71 82 ]andthetree-basedprotocols[ 3 9 51 72 83 ]. IntheALOHA-basedprotocols,thereaderbroadcastsarequestandeachtagrandomlyselectsatimeslottoreportitsID.Ifexactonetagreports,thereaderretrievesitstagIDandthistagwillremainsilentfortherestofreadingprocess.Simultaneousreportsinaslotwillleadtocollision.Therefore,theALOHA-basedprotocolstrytomaximizetheprobabilitythatexactonetagreportsinaslot.TheALOHA-basedprotocolsdifferinhowthereadersendstherequestandhowthetagselectsaslottoreport.IntheslottedALOHA[ 68 ],thereadersendsoutacontentionprobabilityatthebeginningofeachslotandeachunreadtagwiththisprobabilitytoreplywithitsID.InthebasicframedslottedALOHA[ 42 ],slotsaregroupedintoframeswiththesamexedframesize.Eachunreadtagpicksuparandomslotwithineachframetoreport.Itispossiblethatthenumberoftagsfarexceedsthenumberofslotsinaframesothattheframeisfullofcollision.Toovercomethisproblem,thedynamicframedslottedALOHA(DFSA)[ 16 ]introducesframeswithdynamicframesize.Itisprovedthatthemaximalreadingthroughputisachievedwhentheframesizeisequaltothenumberofunreadtags[ 16 ].DFSAdeterminesthesizeofthenextframebyestimatingthenumberofunreadtagsaftereachframe.However,inpractice,itmaybeimpracticaltosettheframesizeindenitelyhighconsideringthereexistalargenumberoftags[ 42 ].TheenhanceddynamicframedslottedALOHA[ 42 ]usesframeswithlimitedframesizebyrestrictingthenumberofrespondingtagsinaframe.Themaximalreadingthroughput 80

PAGE 81

oftheALOHA-basedprotocolsisboundedby1 eT[ 61 ].Inotherwords,foreachslot,theprobabilityofsuccessfullyreadinganewtagis36.8%. Inthetree-basedprotocols,thetagreadingprocedurecanbeinterpretedasarecursivesplittingprocedure.Thegeneralschemaworksasfollows:Inaslot,thereadersendsaquerywithacertainconditionandeachtagthatmeetstheconditionwillrespond.Ifasetoftagsrespondconcurrently,thereadersplitthemintosmallersubsets.Theprocedurerepeatsuntileverysubsetonlycontainsasingletagwhichcanbeidentiedbythereader.Differentsplittingcriterialeadtodifferentprotocols.Thebinary-treeprotocols[ 3 51 72 ]splitasetoftagsusingarandombinarynumber.Specically,eachtaghasacounterinitializedto0.Uponreceivingaquery,eachtagthathasacountervalue0willrespond.Oncecollisionhappens,thereadersendsanewquerywithanindicationofthecollision.Eachcollidingtagdrawsarandombinarynumber(i.e.0or1)andaddsittoitscounter.Thesetofcollidingtagsisthusdividedintotwosubsets:oneisthesetoftagswhosecountersremain0andtheotheroneisthesetoftagswhosecountersincreaseto1.Whencollisionhappens,allothertagsthatdonottransmitalsoincreasetheircountersbyone;otherwise,theydecreasetheircountersbyone.Ananalysisshowsthatthemaximalreadingthroughputofthebinary-treeprotocolsis1 2.88T[ 13 ].Thequery-treeprotocols[ 9 51 83 ]usethetagIDforsplitting.AtagIDisauniquebitstring.Eachquerycontainsaprexp1..piwhereiisthelengthoftheprex.Eachtag,whoseIDcontainsthisprex,transmitsitsIDasaresponse.Ifmultipleresponsescollide,thereaderwillgeneratetwonewprexesp1..pi0andp1..pi1byattachingabit0and1,respectively.Thesetofcollidingtagsisdividedintotwosubsets:onesubsetisthegroupoftagswhoseIDscontaintheprexp1..pi0andtheothersubsetisthegroupoftagswhoseIDscontaintheprexp1..pi1.Aquery-treeprotocolcanhavequitedifferentreadingthroughputsdeterminedbythetagIDdistribution.Itisshownthatthemaximalreadingthroughputisboundedby1 2.88TforasetofuniformlydistributedtagIDs[ 41 ]. 81

PAGE 82

4.4SlottedCollision-AwareTagIdenticationProtocol(SCAT) Inthissection,weproposetheSlottedCollision-AwareTagidenticationprotocol(SCAT).Inthenextsection,wewilloptimizetheprotocolforlessoverhead. 4.4.1ProtocolDescription SCATisatime-slottedprotocol.Thetimeslotsaresynchronizedbythereader'ssignal.Eachtimeslotconsistsofthreesegments:theadvertisementsegment,thereportsegment,andtheacknowledgementsegment. Intheadvertisementsegment,theRFIDreadersendsoutaslotindexiandareportprobabilitypi,whereibeginsfromzeroandincreasesbyoneaftereachslot. Inthereportsegment,eachtagdecidestotransmititsIDwithprobabilitypi.Toactuallybroadcastthereportprobability,thereadermaysendoutanl-bitintegerbpi2lcinsteadofarealnumberpi.AtagcomputesahashfunctionH(IDji)whoserangeis[0..2l).IfH(IDji)bpi2lc,thetagtransmitsitsID. Foranemptyslot,thereadertransmitsanegativeacknowledgement.Foracollisionslot,thereaderwillnotbeabletotellhowmanytagshavetransmittedsimultaneouslyinthereportsegment.Itwillrecordthemixedsignalandtransmitanegativeacknowledgement.Themixedsignalandtheslotindexformacollisionrecord.Overtimethereaderwillcollectagroupofsuchrecords.Theoperationforasingletonslotismorecomplicated.ThereaderlearnstheIDofataginthereportsegment.UsingthisID,ittriestoresolvesomecollisionrecordstolearnmoretagIDs(seethenextsubsection).Itthentransmitsapositiveacknowledgement,togetherwiththeIDsthatarelearnedfromtheresolutionofthepreviouscollisionrecords. Whenthetagthattransmitsinthereportsegmentreceivesthepositiveacknowledgement,itwillstopparticipatingintheSCATprotocolasitsIDhasbeensuccessfullydeliveredtothereader.Similarly,whenatagthattransmitteditsIDatanearlierslotbuthasnotreceivedapositiveacknowledgementyetreceivesitsownIDintheacknowledgementsegment,itwillstopparticipatinginSCAT. 82

PAGE 83

TheSCATprotocolstopswhennotagtransmitsanymore.Whenthereaderndsacertainnumberofconsecutiveemptyslots,itsetspi=1foroneslotandifitstillndsanemptyslot,itknowsthattheIDsofalltagshavebeencollected. 4.4.2CollisionResolution WhentheCRCreceivedinthereportsegmentisveriedtobecorrect,thereaderlearnstheIDofatagfromthecurrentsloti.KnowingtheID,theRFIDreadercaneasilygureoutthepreviousslotsinwhichthistaghastransmitted.Foranarbitrarycollisionrecordwithslotindexj,thetagmusthavetransmittedifH(IDjj)bpi2lc.Ifthatisthecase,thereaderremovesthesignalreceivedinthecurrentslotfromthemixedsignalinthecollisionrecord,treatstheresultasifitwastheIDsignalofasingletag,andextractstheCRCcode.IftheCRCcodeisveriedtobecorrect,thecollisionrecordisresolvedandthereaderlearnsanadditionaltagID.ThesignalforthatIDcanbeusedtoresolveothercollisionrecordsinasimilarprocessasdescribedabove. Resolvingthecollisionslotsincurscomputationoverhead.Hence,weexpectthereadertobecomputationallycapableorconnectedtoapowerfulcomputingdevice.ItisworthnotingthattheRFIDsystemworksinalowspeedchannel(53KbpsforthePhilipsI-Codesystem),whiletheoriginalANC[ 36 ]andthefollow-upwork[ 27 ]aredesignedfor11Mbpsorhigherthroughputchannels,whichisfarmoredemanding,yetexperimentally-demonstratedfeasible. 4.4.3DeterminingtheOptimalValueforpi WewanttodeterminetheoptimalreportprobabilitypiforeachslotsuchthatthenumberofslotsforcollectingtheIDsofalltagsisminimized.Consideranarbitrarytimeslotwithindexi.Whenthereisonlyonetagtransmitting,theRFIDreaderwilllearntheIDofthetag.Iftherearetwotagstransmitting,thereaderwillnotlearnanyIDnowbutwilllearnoneIDlaterwhentheotherIDisknown(suchthatthecollisionrecordofthisslotcanberesolved).Similarly,whenktagstransmitinthisslotfork,thereaderwilllearnoneIDfromthecollisionrecordwhentheother(k)]TJ /F5 11.955 Tf 12.04 0 Td[(1)IDsareknown.Essentially, 83

PAGE 84

asingletonslotorak-collisionslotwillallowthereadertolearnoneIDnoworlater.Hence,weshallchoosethevalueofpithatmaximizestheprobabilityforone,two,...,ortagstotransmitinthecurrentslot. LetNbethenumberoftagsinthesystem.Itsvaluecanbeestimatedtoanarbitraryaccuracy[ 39 ]inapre-stepofSCAT.Thispre-stepwillberemovedinthenextsection.Beforesloti,thereadermayhavesuccessfullycollectedandacknowledgedanumbernioftagIDs,andthosetagswillnolongerparticipateintheprotocolofSCAT.LetNibethenumberoftagsthatthereaderhasnotidentiedyetbeforesloti.Sinceniisknowntothereader,Niisalsoknown. Aseachtagdecidestotransmitwithprobabilitypi,thenumberoftagsthattransmitwillbearandomvariableXithatfollowsthebinomialdistribution.TheprobabilityforXi=k,8k2[0..Ni]is)]TJ /F4 7.97 Tf 5.48 -4.38 Td[(Nikpki(1)]TJ /F3 11.955 Tf 12.04 0 Td[(pi)Ni)]TJ /F4 7.97 Tf 6.59 0 Td[(k.OurobjectiveistomaximizetheprobabilityofXi2[0..],whichisXk=1ProbfXi=kg=Xk=1Nikpki(1)]TJ /F3 11.955 Tf 11.95 0 Td[(pi)Ni)]TJ /F4 7.97 Tf 6.59 0 Td[(k. (4) Weexpecttobesmall.Inthefollowing,weconsider=2,3,or4.When=2,( 4 )becomes2Xk=1ProbfXi=kg=Nipi(1)]TJ /F3 11.955 Tf 11.95 0 Td[(pi)Ni)]TJ /F6 7.97 Tf 6.59 0 Td[(1+Ni(Ni)]TJ /F5 11.955 Tf 11.96 0 Td[(1) 2p2i(1)]TJ /F3 11.955 Tf 11.96 0 Td[(pi)Ni)]TJ /F6 7.97 Tf 6.59 0 Td[(2'Nipie)]TJ /F4 7.97 Tf 6.59 0 Td[(Nipi+N2ip2i 2e)]TJ /F4 7.97 Tf 6.58 0 Td[(Nipi. (4) Let!=Nipi.SubstitutingNipiby!in( 4 ),wehave2Xk=1ProbfXi=kg'(!+!2 2)e)]TJ /F26 7.97 Tf 6.59 0 Td[(!. (4) 84

PAGE 85

Tondthevalueof!thatmaximizestheaboveformula,wedifferentiatetherightsideandletitbezero. d(!+!2 2)e)]TJ /F26 7.97 Tf 6.59 0 Td[(! d!=(1)]TJ /F8 11.955 Tf 13.15 8.09 Td[(!2 2)e)]TJ /F26 7.97 Tf 6.59 0 Td[(!=0.(4) Solvingtheaboveequation,wehave!=1.414.Hence,theoptimalreportprobabilityispi=1.414=Ni. Followingthesameprocess,wederivethat,when=3,theoptimalreportprobabilityispi=1.817=Ni,andwhen=4,itispi=2.213=Ni. Resolvingthecollisionslotsrequiresasufcientnumberofsingletonslots.Otherwise,ifallslotshavecollision,noneofthemwillberesolved.Fortunately,whenissmall(whichshouldbethecaseaswehavediscussedinSection 4.2.3 andwillfurtherelaborateinSection 4.6.1 ),therearesufcientsingletonslotstoresolvemostcollisionslots.OursimulationresultsinSection 4.6.3 showthattheoptimalreportprobabilitiesobtainedbyexhaustivesearchmatchcloselywiththeabovecomputedvalues. 4.4.4PseudoCode ThepseudocodefortheoperationoftheRFIDreaderduringtheithslotisgivenbelow.LetSbethesetofnewlyknownIDs(togetherwiththeirsignals)thatcanbeusedtoresolvesomeofthecollisionrecords.LetIbethesetofIDsthatarelearnedbyresolvingthecollisionrecords.LetRjbethecollisionrecordforslotj. Reader'sOperationatSloti1.broadcastanadvertisementhi,pii2.recordthesignalsiinthereportsegment3.extractIDifromsi4.ifthechannelisidleduringthereportsegmentthen5.broadcastanegativeacknowledgement6.elseifCRCinIDiisveriedtobecorrectthen7.S:=fhIDi,siig 85

PAGE 86

8.I:=;9.whileS6=;do10.removeanelementhID,sifromS11.foreachcollisionrecordRjdo12.ifH(IDjj)pjthen13.addstothesetofknownindividualsignalsinRj14.removeknownsignalsfromthemixedsignalinRj15.extractID0fromtheresultingsignals016.ifCRCinID0isveriedtobecorrectthen17.S:=S+fhID0,s0ig18.I:=I+fID0g19.removethecollisionrecordRj20.endfor21.endwhile22.broadcastapositiveacknowledgementandtheIDsinI23.else24.addhi,siiasacollisionrecord25.broadcastanegativeacknowledgement 4.4.5UnresolvableCollisionSlotsandChannelError Thereadingprocessnormallytakesashortperiodoftime(minutesfortensofthousandsoftags).Duringthistime,weexpectthetagstobestaticallylocated.TheMSKemployedbyANCcantolerateacertainlevelofnoiseandchannelvariation.However,ifthespontaneousnoiseistoolarge,acollisionslotmaynotberesolvable.Theonlyimpactisthattheslotisnotuseful,andthereadercanstilllearntheIDsfromotherslots.Atagwillstoptransmittingonlyafteritreceivespositiveconrmationfromthereader.Aslongasmost2-collisionslotscanberesolved,theproposedprotocolstillachievesmuchhigherreadingthroughput. 86

PAGE 87

Channelerrormaycorruptthesignaltransmittedbyatagortheacknowledgementtransmittedbythereader.ThisproblemiscommontoallRFIDreadingprotocols.Thesolutionisalsocommon:ThetagwillkeeptransmittingitsIDuntilitreceivesthepositiveconrmationfromthereader.Inthiscase,thereadermayreceiveanIDmorethanonceandtheduplicateswillbediscarded. Theproposedprotocolisnotsuitableforanenvironmentwherethechannelnoiseissosevereorthetagsmovesomuchandsofastduringthereadingoperationthatmostcollisionslotsarenotresolvable.Inthiscase,weshoulduseacontention-basedprotocolwithoutcollisionresolution.Itisbeyondthescopeofthisworktoinvestigatethenoiselevelofeachspecicenvironment.Instead,wearemoreinterestedinknowingwhatistheupperlimitofthroughputimprovementthatANCcanbring(inanenvironmentwheremost2-collisionslotsareresolvable). 4.5FramedCollision-AwareTagIdenticationProtocol(FCAT) Inthissection,weproposeaframedversionofthepreviousprotocoltoimproveitsefciency. 4.5.1InefcienciesofSCAT SCATutilizestheinformationcarriedinthecollisionslots.However,itisnotpracticallyefcientduetoanumberofreasons. First,tocalculatepi,theRFIDreaderhastoknowNi,whichinturnrequiresittoknowN.Itincursconsiderableoverheadtoaccuratelyestimatethenumberoftagsinthesystemasapre-steptoSCAT.Wewanttoremovesuchapre-stepandestimateNasabyproductduringtheprotocolexecution. Second,theadvertisementsegmentofeachslotrepresentssignicantoverheadwhichisnotalwaysnecessary.Forconsecutiveslots,theslotindexchangesfromitoi+1andthereportprobabilitychangesfrom!=Nito!=Ni+1,whereNiandNi+1atmostdifferbyone.AsthereportprobabilitychangeslittlewhenNiisreasonablylarge,the 87

PAGE 88

readerdoesnothavetomakeadvertisementineachslot.Itmayadvertiseonceeverycertainnumberofslots,andthetagswillusethesamereportprobabilityinthoseslots. Third,afterresolvingacollisionrecord,thereaderlearnsanextraIDanditbroadcaststheIDinordertoinformthecorrespondingtagtostopparticipatingintheprotocol.However,insteadoftransmittingthewholeID(whichis96bitsforGEN2tags),thereadermaytransmittheslotindexofthecollisionrecord.Atagstorestheindicesoftheslotsinwhichithastransmitted.Ifthetagreceivesaslotindexthatmatchesastoredone,itknowsthatthereadermusthavecollecteditsID.Aslotindexcanbemuchsmallerthan96bits.Ifweuse23-bitslotindices,morethan8millionslotsareallowed.Inoursimulations,thenumberofslotsrequiredneverexceeds2N. 4.5.2UsingFrames WeproposetheFramedCollision-AwareTagidenticationprotocol(FCAT),whichimprovesSCATbyeliminatingtheinefciencydescribedinSection 4.5.1 .FCATsharesmuchoftheprotocoldetailswithSCAT.Belowwewillfocusondescribingtheirdifferences. InFCAT,timeisdividedintoframesofsizef.Thatis,eachframeconsistsofftimeslots.Eachframehasanindex,startingfromzero.Theindexofthejthslotintheithframeisif+j.Beforeaframebegins,theRFIDreaderbroadcastsapre-frameadvertisement,includingtheframeindexiandthereportprobabilitypi.Eachslotoftheframeconsistsofareportsegment,duringwhichthetagstransmittheirIDs,andanacknowledgementsegment,duringwhichthereadertransmitseitherapositiveacknowledgementoranegativeacknowledgement. Inanyslotoftheithframe,eachtagtransmitsitsIDwithprobabilitypi.Afterreceivingthesignalinthereportsegment,thereaderperformsthesameoperationsasinSCAT,exceptthatitdoesnottransmittheIDslearnedfromresolvingthecollisionrecordsintheacknowledgementsegment.Instead,ittransmitstheslotindicesoftheresolvedcollisionrecords,whichareshorterthantheIDsthemselves.Ifatagreceives 88

PAGE 89

aslotindexthatmatchesaslotinwhichithastransmitteditsID,itstopsparticipatinginFCAT.Certainly,ifatagreceivesapositiveacknowledgementforitsIDjusttransmittedinthereportsegment,itwillalsostopparticipatinginFCAT. 4.5.3EstimatingtheNumberofTagswithinFCAT Finally,weaddresstheproblemofhowtolearnthevalueofN.Thereexistefcientmethodsforestimatingthenumberoftags.However,usingthemasapre-stepofFCATincursconsiderableoverhead.Inthefollowing,weembedanestimationmethodwithinFCAT. Consideranarbitraryframewithindexi.Letn0,n1andncbetherandomvariablesforthenumbersofempty,singletonandcollisionslots,respectively.WecanestimatethestatisticalrelationshipbetweentheserandomvariablesandthenumberNioftagsthatarecurrentlyparticipatingintheprotocol.Basedonthatrelationship,wecanestimateNifromthemeasuredvaluesofn0andnc.Ourapproachsharessomesimilaritywith[ 39 ].However,in[ 39 ],eachtagtransmitsatmostonceintheframe.InFCAT,eachtagparticipatesprobabilisticallyineveryslotoftheframe. LetXjbetheindicatorrandomvariablefortheeventthatthejthslotintheframeisempty,i.e.,Xj=1meansthejthslotisemptyandXj=0meansitisnotempty.Similarly,letYjbetheindicatorrandomvariablefortheeventthatthejthslotisasingletonslot.Becauseeachtagdecidestotransmitwithprobabilitypiineveryslotintheframe,wehaveProbfXj=1g=(1)]TJ /F3 11.955 Tf 11.95 0 Td[(pi)Ni,8j2[1..f]. (4) Theexpectedvalueofn0isE(n0)=fXj=1(1)]TJ /F3 11.955 Tf 11.95 0 Td[(pi)Ni=f(1)]TJ /F3 11.955 Tf 11.96 0 Td[(pi)Ni. (4) 89

PAGE 90

TheprobabilityforthejthslotintheframetobeasingletonisProbfYj=1g=Ni1pi(1)]TJ /F3 11.955 Tf 11.95 0 Td[(pi)Ni)]TJ /F6 7.97 Tf 6.58 0 Td[(1=Nipi(1)]TJ /F3 11.955 Tf 11.95 0 Td[(pi)Ni)]TJ /F6 7.97 Tf 6.58 0 Td[(1. (4) Theexpectedvalueofn1isE(n1)=fXi=1Nipi(1)]TJ /F3 11.955 Tf 11.96 0 Td[(pi)Ni)]TJ /F6 7.97 Tf 6.59 0 Td[(1=fNipi(1)]TJ /F3 11.955 Tf 11.95 0 Td[(pi)Ni)]TJ /F6 7.97 Tf 6.58 0 Td[(1. (4) Obviously,E(n0)+E(n1)+E(nc)=f.HenceE(nc)=f)]TJ /F3 11.955 Tf 11.96 0 Td[(E(n0))]TJ /F3 11.955 Tf 11.96 0 Td[(E(n1)=f(1)]TJ /F5 11.955 Tf 11.96 0 Td[((1)]TJ /F3 11.955 Tf 11.96 0 Td[(pi)Ni)]TJ /F3 11.955 Tf 11.95 0 Td[(Nipi(1)]TJ /F3 11.955 Tf 11.96 0 Td[(pi)Ni)]TJ /F6 7.97 Tf 6.59 0 Td[(1)=f(1)]TJ /F5 11.955 Tf 11.96 0 Td[((1)]TJ /F3 11.955 Tf 11.96 0 Td[(pi)Ni)]TJ /F6 7.97 Tf 6.59 0 Td[(1(1)]TJ /F3 11.955 Tf 11.95 0 Td[(pi+!)). (4) TheaboveequationcanberewrittenasNi=ln(1)]TJ /F4 7.97 Tf 13.15 5.47 Td[(E(nc) f))]TJ /F5 11.955 Tf 11.95 0 Td[(ln(1)]TJ /F3 11.955 Tf 11.96 0 Td[(pi+!) ln(1)]TJ /F3 11.955 Tf 11.96 0 Td[(pi)+1. (4) Attheendoftheithframe,thereadercountsthevalueofnc.SubstitutingE(nc)bytheinstancevaluenc(obtainedintheithframe),thereaderobtainsanestimationofNibythefollowingformula:^Ni=ln(1)]TJ /F4 7.97 Tf 13.15 4.7 Td[(nc f))]TJ /F5 11.955 Tf 11.95 0 Td[(ln(1)]TJ /F3 11.955 Tf 11.96 0 Td[(pi+!) ln(1)]TJ /F3 11.955 Tf 11.96 0 Td[(pi)+1. (4) Next,wederiveE(^Ni).Tosimplifytheequations,letC1=1 ln(1)]TJ /F4 7.97 Tf 6.59 0 Td[(pi),C2=)]TJ /F6 7.97 Tf 10.5 5.47 Td[(ln(1)]TJ /F4 7.97 Tf 6.59 0 Td[(pi+!) ln(1)]TJ /F4 7.97 Tf 6.58 0 Td[(pi)+1,andfunctiong(nc)=ln(1)]TJ /F4 7.97 Tf 13.31 4.71 Td[(nc f).Weexpandtherighthandsideof( 4 )byitsTaylorseriesaboutq=E(nc).^Ni=C1g(q)+(nc)]TJ /F3 11.955 Tf 11.95 0 Td[(q)g0(q)+1 2(nc)]TJ /F3 11.955 Tf 11.95 0 Td[(q)2g00(q)+1 6(nc)]TJ /F3 11.955 Tf 11.95 0 Td[(q)3g000(q)+...+C2. (4) 90

PAGE 91

Sinceq=E(nc),themeanofthesecondtermin( 4 )is0.Therefore,wekeeptherstthreetermswhencomputingtheapproximatedvalueofE(^Ni).E(^Ni)'C1g(q)+1 2E((nc)]TJ /F3 11.955 Tf 11.95 0 Td[(q)2)g00(q)+C2. (4) WehaveE((nc)]TJ /F3 11.955 Tf 12.03 0 Td[(q)2)=V(nc)bydenitionandg00(q)=)]TJ /F6 7.97 Tf 21.56 4.71 Td[(1 (q)]TJ /F4 7.97 Tf 6.58 0 Td[(f)2sinceg(q)=ln(1)]TJ /F4 7.97 Tf 13.23 5.04 Td[(q f).ThevarianceV(nc)isderivedinSection 4.5.4 .Applying( 4 )inSection 4.5.4 ,wehaveE(^Ni)'Ni)]TJ /F3 11.955 Tf 36.33 8.09 Td[(e!)]TJ /F5 11.955 Tf 11.96 0 Td[(1)]TJ /F8 11.955 Tf 11.96 0 Td[(! 2fln(1)]TJ /F3 11.955 Tf 11.96 0 Td[(pi)(1+!). (4) Therefore,Bias(^Ni Ni)=E(^Ni Ni))]TJ /F5 11.955 Tf 11.95 0 Td[(1=1+!)]TJ /F3 11.955 Tf 11.95 0 Td[(e! 2fNiln(1)]TJ /F3 11.955 Tf 11.96 0 Td[(pi)(1+!). (4) Fig. 4-3 showstheabsolutevalueofBias(^Ni Ni)withrespecttothenumberoftagsNi.ThethreelinesshowthattheabsolutevaluesofBias(^Ni Ni)are0.0082,0.011and0.014,for!=1.414,1.817and2.213,respectively.Theyareallverysmall. AddingthenumberoftagswhoseIDsarealreadyknown,thereaderhasanestimationforthetotalnumberoftagsinthesystem,denotedas^Ni.Thevarianceof^Niisthesameasthevarianceof^Ni,i.e.,V(^Ni)=V(^Ni).BecauseNi
PAGE 92

estimatorislarger.AsshowninFig. 4-4 ,Niisnotamonotonicfunctionwithrespecttothenumberofsingletonslots.Hence,wecannotusen1toestimatethevalueofNi. 4.5.4EstimationVariance,V(^Ni Ni) Consideranarbitraryframewithindexi.LetZjbetheindicatorrandomvariablefortheeventthatthejthslotintheframeisacollisionslot.Sincenoslotisspecial,Zj,8j2[1..f],followsthesamedistribution.Theyareindependentrandomvariables.Becausenc=Pfj=1Zj,wehaveV(nc)=fXj=1V(Zj)=fV(Z1). (4)E(Z1)=1)]TJ /F5 11.955 Tf 12.42 0 Td[((1)]TJ /F3 11.955 Tf 12.43 0 Td[(pi)Ni)]TJ /F3 11.955 Tf 12.42 0 Td[(Nipi(1)]TJ /F3 11.955 Tf 12.42 0 Td[(pi)Ni)]TJ /F6 7.97 Tf 6.59 0 Td[(11)]TJ /F3 11.955 Tf 12.42 0 Td[(e)]TJ /F4 7.97 Tf 6.58 0 Td[(Nipi)]TJ /F3 11.955 Tf 12.42 0 Td[(Nipie)]TJ /F4 7.97 Tf 6.58 0 Td[(Nipi.E(Z21)=E(Z1)becauseZ1isanindicatorrandomvariable.Hence,wehaveV(Z1)=E(Z21))]TJ /F5 11.955 Tf 11.95 0 Td[((E(Z1))2=(1+Nipi)e)]TJ /F4 7.97 Tf 6.59 0 Td[(Nipi(1)]TJ /F5 11.955 Tf 11.96 0 Td[((1+Nipi)e)]TJ /F4 7.97 Tf 6.59 0 Td[(Nipi). (4) Therefore,V(nc)=f(1+Nipi)e)]TJ /F4 7.97 Tf 6.58 0 Td[(Nipi(1)]TJ /F5 11.955 Tf 11.95 0 Td[((1+Nipi)e)]TJ /F4 7.97 Tf 6.59 0 Td[(Nipi). (4) Accordingtothecentrallimittheorem,iffislarge,ncisapproximatelynormallydistributed.Whenf!1,ncconvergestothenormaldistribution,ncD!Norm(,2),whereisE(nc)asgivenin( 4 ),2isV(nc)asgivenin( 4 ),andD!meansconvergenceindistribution. Accordingtothe-method[ 14 ],wehave h(nc)D!Norm(h(),2[h0()]2)(4) foranyfunctionh(.)suchthath0()existsandtakesanon-zerovalue. InSection 4.5.3 ,theestimationformulaisdesignedbasedon( 4 ),whichiscopiedhereE(nc)=f(1)]TJ /F5 11.955 Tf 11.48 0 Td[((1)]TJ /F3 11.955 Tf 11.49 0 Td[(pi)Ni)]TJ /F3 11.955 Tf 11.49 0 Td[(Nipi(1)]TJ /F3 11.955 Tf 11.49 0 Td[(pi)Ni)]TJ /F6 7.97 Tf 6.59 0 Td[(1).Letg(.)bethemappingfunction 92

PAGE 93

fromNitonc.TheaboveequationcanberewrittenasE(nc)=g(Ni).Fig. 4-4 showsthatg(.)isamonotonicfunction,andhenceithasauniqueinversefunction,denotedash(.). AccordingtoSection 4.5.3 ,^Niiscomputedfrom( 4 )bysubstitutingE(nc)withtheinstancevalueofnc(obtainedaftertheithframe).nc=f(1)]TJ /F5 11.955 Tf 11.95 0 Td[((1)]TJ /F3 11.955 Tf 11.95 0 Td[(pi)^Ni)]TJ /F5 11.955 Tf 13.52 2.66 Td[(^Nipi(1)]TJ /F3 11.955 Tf 11.95 0 Td[(pi)^Ni)]TJ /F6 7.97 Tf 6.58 0 Td[(1)f(1)]TJ /F3 11.955 Tf 11.95 0 Td[(e)]TJ /F6 7.97 Tf 7.66 1.77 Td[(^Nipi)]TJ /F5 11.955 Tf 13.52 2.66 Td[(^Nipie)]TJ /F6 7.97 Tf 7.66 1.77 Td[(^Nipi). (4) Clearly,nc=g(^Ni)and^Ni=h(nc).Applying^Ni=h(nc)to( 4 ),wehave ^NiD!Norm(h(),2[h0()]2).(4) Weknowthath(g(Ni))=Ni.Differentiatingbothsides,wehaveh0(g(Ni))g0(Ni)=1.Hence, h0()=h0(E(nc))=h0(g(Ni))=1 g0(Ni).(4) Therefore,from( 4 ),thevarianceof^NiisV(^Ni)=2[h0()]2=V(nc) [g0(Ni)]2=(1+Nipi)eNipi)]TJ /F5 11.955 Tf 11.95 0 Td[((1+2Nipi+N2ip2i) fN2ip4i, (4)V(^Ni Ni)=(1+Nipi)eNipi)]TJ /F5 11.955 Tf 11.96 0 Td[((1+2Nipi+N2ip2i) fN4ip4i. (4) Belowweperformapproximatecomputationtogivearoughideaonhowbigthisvarianceis.InSCATorFCAT,^Nipi=!,where!is1.414,1.817or2.213for=2,3or4,respectively.Oursimulationsshowthat^NireliablyconvergestoNiwheniislarge.Hence,wesubstituteNipiwith!in( 4 ),andthevarianceV(^Ni Ni)is0.0342,0.0287or0.0265respectivelyfordifferent!values. 93

PAGE 94

4.6SimulationResults Inthissection,wepresentsimulationresultstoevaluatetheperformanceofourmainprotocolFCAT.WecompareFCATwiththeexistingwork,includingtheDynamicFramedSlottedALOHA(DFSA)[ 16 ],EnhancedDynamicFramedSlottedALOHA(EDFSA)[ 42 ],AdaptiveBinarySplitting(ABS)[ 51 ]andAdaptiveQuerySplitting(AQS)[ 51 ].ThersttwoareALOHA-basedandthenexttwoaretree-based. WeuseFCAT-todenotetheFCATprotocolinwhichk-collisionslotswithkareresolvable,where=2,3,4.ThereportprobabilitypiisdeterminedbasedontheformulagiveninSection 4.4.3 .Specically,piissettobe1.414=Ni,1.817=Niand2.213=NiinFCAT-2,FCAT-3andFCAT-4,respectively.OthervaluesofpiarealsoinvestigatedinSection 4.6.3 .Theframesizefissetto30timeslots;theperformanceofFCATunderdifferentfvalueswillalsobestudied.Theparametersusedinotherprotocolsareselectedbasedontheiroriginalpaperswheneverpossible. Inthesimulations,wesetthetimeslotlengthbasedonthePhilipsI-Codespecication[ 64 ].Thetransmissionrateis53kbit/sec.Hence,ittakes18.88stotransmiteachbit.WesettheIDlengthtobe96bits(includingthe16bitsCRCcode),whichtakes1812s.Thereader'sacknowledgementconsistsof20bits,(includingtheCRCcode),whichtakes378s.Thewaitingtimebeforethereportsegmentortheacknowledgementsegmentis302stoseparatetransmissions.Therefore,eachslotisabout2.8ms.Thesimulationresultsaretheaverageoutcomeof100runs. 4.6.1ReadingThroughputComparison Werstcomparetheprotocolsintermsofthereadingthroughput,whichistheaveragenumberoftagIDsthattheRFIDreadercancollectineachsecondduringtheprotocolexecutiontimebeforeallIDsareread.Table 4-1 showsthereadingthroughputsoftheprotocolswhenthenumberoftagsvariesfrom1,000to20,000.Duetocollisionresolution,FCAT-2achieves51.1%55.6%throughputimprovementoverDFSA, 94

PAGE 95

54.8%70.6%improvementoverEDFSA,59.6%62.9%improvementoverABS,64.1%67.7%improvementoverAQS. Asexpected,FCAT-3performsbetterthanFCAT-2,andFCAT-4performsbetterthanFCAT-3.However,theimprovementofFCAT-4overFCAT-3ismuchsmallerthanthatofFCAT-3overFCAT-2.FCAT-5(whoseresultsarenotshowninthetable)performsonlyslightlybetterFCAT-4.Forexample,whenN=10,000,itsreadingthroughputis270.9tagIDspersecond,whichisslightlybetterthan265.1ofFCAT-4.Thisindicatesaquicklyshrinkingmarginofimprovementasincreasesandsuggeststhatalargevalueofispracticallyunnecessary. Wealsoevaluatethereadingtimeintermsoftimeslots.Table 4-2 showsthenumbersofempty,singletonandcollisionslotsusedtoread10,000tags.WecanseethatfeweremptyslotsarewastedinFCATthaninallothercomparedprotocols.FCATalsousesmuchfewersingletonslotstocollectalltagIDsbecauseFCATcanextracttaginformationfromthecollisionslots,whileotherprotocolshavetoreadtagssolelyinthesingletonslots.FCAT-4hasmorecollisionslotsthanFCAT-2.ThereasonisthatFCAT-4canutilizeacollisionslotinwhichuptofourtagscollide,andhenceFCAT-4encouragesmoretagstotransmitsimultaneously. 4.6.2EffectivenessofCollisionResolution InTable 4-3 ,weshowthenumberoftagIDsthatareresolvedfromthecollisionslots.FCAT-2obtainsabout40%oftagIDsfromthecollisionslots.Thepercentageisabove57%forFCAT-3andabove68%forFCAT-4.Forexample,whenthereare10,000tagsinthesystem,FCAT-2willreadmorethan4,000ofthemfromthecollisionslots,whichareignoredbythepreviousprotocols. 4.6.3ReportProbability Thereportprobabilitypiiscalculatedas!=Ni.NiisthenumberoftagsparticipatinginslotiandthemethodinSection 4.5.3 isusedtoestimateNiaftereachframe.Theoptimalvalueof!issetinSection 4.4.3 .Weusesimulationtoconrmouranalytical 95

PAGE 96

resultanddemonstratehowthevalueof!affectstheperformanceofFCAT.Fig. 4-5 showsthereadingthroughputwithrespectto!whenthereare10000tags.If!issettoosmall,thereadingthroughputdecreasesbecausemanyslotsareemptyandthuswasted.If!issettoolarge,italsohurtstheperformancebecausetherearetoomanycollisionslotsandtoomanytagscollideinthoseslots,makingthemunresolvable. Bytryingallpossiblevaluesof!,wecanusesimulationtondthetrueoptimal!(andthecorrespondingoptimalreportprobability)thatmaximizesthereadingthroughput.AsshowninTable 4-4 ,theoptimalvalueof!observedinthesimulationmatchescloselywiththevaluecomputedinSection 4.4.3 ,i.e.,1.414when=2,1.817when=3,and2.213when=4.Alsoshowninthesametable,thereadingthroughputachievedbyFCATusingthecomputedreportingprobabilityisalmostthesameasthemaximum-achievablethroughputundertheoptimalreportingprobabilityobtainedbysimulationthroughexhaustivesearch. 4.6.4ImpactofFrameSize Fig. 4-6 showstheimpactoftheframesizefinasystemwith10,000tags.Wecanseethatthereadingthroughputisstabilizedwhenf10. 4.7Summary Thischapterconcludesthatthephysical-layernetworkcodingcanindeedsignicantlyimprovethespeedatwhichaRFIDreadercollectsinformationofthetags.Thereasonisthattheinformationcarriedinmanycollisionslots,whichwaspreviouslydiscarded,canbeutilizedalmostaseffectivelyastheinformationcarriedinthesingletonslots.ThecurrentanalognetworkcodingmethodcanimprovethereadingthroughputofaRFIDsystemby51.1%70.6%.Asthetechnologiesofphysical-layernetworkcodingareimproved,thereadingthroughputcanpotentiallybedoubled. 96

PAGE 97

A B Figure4-1. Thisexampleshowsthatacollision-resolutionprotocolmayreducethenumberoftimeslotsusedtoidentifyfourtagsfrom11timeslotsto6timeslots.A)Contention-basedprotocol.B)Collision-resolutionprotocol. Figure4-2. Alice-BobexampleforAnalogNetworkCoding. Figure4-3. Therelativebiasof^Niwithrespecttothenumberoftags. 97

PAGE 98

Figure4-4. Thenumberoftags,Ni,isnotamonotonicfunctioninE(n1).Parameters:pi=1.414=Niandf=30. Table4-1. ReadingthroughputcomparisonwhenNvariesfrom1,000to20,000 NFCAT-2FCAT-3FCAT-4DFSAEDFSAABSAQS 1000197.7234.8238.8130.8115.9123.9117.92000199.5237.2257.5131.8121.5123.7119.43000200.2239.7261.4132.1122.9123.8120.44000201.0240.1262.1132.8124.8123.9120.55000201.3240.4262.3130.1126.1123.8120.86000201.3241.5263.7132.4126.3123.6120.97000201.3241.2264.9131.1126.4123.8121.18000201.4241.8265.1131.9127.1123.6121.19000201.2241.5265.4131.0127.8123.7121.110000201.3241.8265.1131.4127.8123.9121.211000201.7241.5266.0130.0127.6123.9121.112000200.8241.8265.9130.3126.8123.8121.213000201.0241.7265.9129.2127.3123.8121.214000200.4241.3266.2130.9127.6123.5121.315000200.8241.2266.0131.7127.7124.2121.316000200.9241.8265.9131.3128.2123.8121.317000200.2241.3265.5130.5128.1124.1121.318000199.7240.7265.9130.0128.2123.6121.319000199.1240.9266.4129.2128.2123.7121.320000199.1241.3266.1129.1128.6123.9121.3 98

PAGE 99

Table4-2. Empty,SingletonandCollisionTimeSlotswhenN=10000 FCAT-2FCAT-3FCAT-4DFSAEDFSAABSAQS empty418922571345100761070544104737singleton58614055293510000100001000010000collision701674978050720872341440914735total17066138091233027284279392881929472 Table4-3. TagIDsResolvedfromCollisionSlots NFCAT-2FCAT-3FCAT-4 10004236007075000210230083561100004139594570651500060628819104822000079051150713656 Figure4-5. FCATreadingthroughputwithrespectto!. Table4-4. Thecomputedvalueof!matchescloselywiththeoptimalvalueof!obtainedbysimulation. Optimal!MaximumThroughputcomputed!FCATThroughput 21.42202.11.41201.331.90241.91.82241.842.12266.22.21265.1 99

PAGE 100

Figure4-6. ThereadingthroughputofFCATisstabilizedwhenf10. 100

PAGE 101

CHAPTER5CONCLUSION Inthisdissertation,wefocusonfairnessandthroughputissuesinwirelesssystems. WeproposenovelsolutionsforimprovingfairnessandthroughputincontendingWLANs.WestartfromtheMAC-layerandrevealthatthelocation-sensitivecontentioncanexaggeratethetime-allocationanomalyin802.11DCFnetworks.WeproposeanewprotocolAIMD/QS+kwhichachievesproportionalMAC-layertimefairnessinmultiplecontentiongroupsunderlocation-sensitivecontention.Wethenmovetothetransport-layerandshowthatasevereTCPfairnessproblemmayoccuramongnearbyWLANs.Mostexistingsolutionsrelyontheassumptionthatallcontendingowscanexplicitlyexchangemessages,whichhowevermaynotalwayshold.WepresentoursolutionWPDwhichimplicitlyspreadsthecongestioninformationamongcontendingows.ThissolutioncansignicantlyimproveTCPfairnesswhenotherexistingsolutionsfail. WeintroduceanewmethodtoboosttheRFIDreadingthroughput.Wedemonstratethatthephysical-layernetworkcodingcanbeeffectivelyintegratedintoRFIDreadingprotocols.Webelievethisistherstworkthatappliesphysical-layernetworkcodingtohelpimprovethereadingthroughputoflargeRFIDsystems.TheproposedprotocolFCATcanincreasethereadingthroughputbyover50%usingthecurrentanalognetworkcodingtechnology. 101

PAGE 102

REFERENCES [1] TheNetworkSimulator-ns-2. http://www.isi.edu/nsnam/ns/ (2008). [2] IEEE802.11WG,Part11:WirelessLANMediumAccessControl(MAC)andPhysicalLayer(PHY)Specication.IEEE(1999). [3] InformationtechnologyautomaticidenticationanddatacapturetechniquesCradiofrequencyidenticationforitemmanagementairinterface-part6:parametersforairinterfacecommunicationsat860-960MHz.(2003). [4] Abeysekera,B.A.H.S.,Matsuda,T.,andTakine,T.DynamicContentionWindowControlMechanismtoAchieveFairnessbetweenUplinkandDownlinkFlowsinIEEE802.11WirelessLANs.IEEETransactionsonWirelessCommunications7(2008).9:3517. [5] Aguilera,E.L.,Heusse,M.,Grunenberger,Y.,Rousseau,F.,Duda,A.,andCasademont,J.AnAsymmetricAccessPointforSolvingtheUnfairnessProbleminWLANs.IEEETransactionsonMobileComputing7(2008).10:1213. [6] Akella,A.,Judd,G.,Seshan,S.,andSteenkiste,P.Self-managementinChaoticWirelessDeployments.InProc.ofACMMobiCom(2005). [7] Azambuja,M.C.,Marcon,C.A.M.,andHessel,F.P.SurveyofStandardizedISO18000-6RFIDAnti-collisionProtocols.InProc.ofIEEEInternationalConferenceonSensorTechnologiesandApplications(SENSORCOMM)(2008). [8] Bejerano,Y.,Han,S.-J.,andLi,L.E.FairnessandLoadBalancinginWirelessLANsUsingAssociationControl.InProc.ofACMMobiCom(2004). [9] Bhandari,N.,Sahoo,A.,andIyer,S.IntelligentQueryTree(IQT)ProtocoltoImproveRFIDTagReadEfciency.IEEEInternationalConferenceonInformationTechnology(ICIT)(2006). [10] Bharghavan,V.,Demers,A.,Shenker,S.,andZhang,L.MACAW:AMediaAccessProtocolforWirelessLANs.InProc.ofACMSIGCOMM(1994). [11] Bui,L.,Eryilmaz,A.,Srikant,R.,andWu,X.JointAsynchronousCongestionControlandDistributedSchedulingforMulti-hopWirelessNetworks.InProc.ofIEEEINFOCOM(2006). [12] Camp,J.,Mancuso,V.,Gurewitz,O.,andKnightly,E.W.AMeasurementStudyofMultiplicativeOverheadEffectsinWirelessNetworks.InProc.ofIEEEINFOCOM(2008). [13] Capetenakis,J.I.TreeAlgorithmsforPacketBroadcastChannels.IEEETransactionsonInformationTheory25(1979).5:505. 102

PAGE 103

[14] Casella,G.andBerger,R.L.StatisticalInference.2ndedition,DuxburyPress(2002). [15] Casini,E.,Gaudenzi,R.D.,andHerrero,O.R.ContentionResolutionDiversitySlottedALOHA(CRDSA):AnEnhancedRandomAccessSchemeforSatelliteAccessPacketNetworks.IEEETransactionsonWirelessCommunications6(2007).4:1408. [16] Cha,J.R.andKim,J.H.DynamicFramedSlottedALOHAAlgorithmsUsingFastTagEstimationMethodforRFIDSystems.IEEEConsumerCommunicationsandNetworkingConference(CCNC)(2006). [17] Chen,L.,Low,S.H.,Chiang,M.,andDoyle,J.C.Cross-layerCongestionControl,Routing,andSchedulingDesigninAdHocWirelessNetworks.InProc.ofIEEEINFOCOM(2006). [18] Chen,X.,Zhai,H.,Wang,J.,andFang,Y.ASurveyonImprovingTCPPerformanceoverWirelessNetworks.ResourceManagementinWirelessNetworking16(2005):657. [19] Das,R.GlobalRFIDMarketTops$5.5Billion. http://www.convertingmagazine.com/article/CA6653688.html (2009). [20] Dunn,J.,Neufeld,M.,Sheth,A.,Grunwald,D.,andBennett,J.APracticalCross-LayerMechanismforFairnessin802.11Networks.InProc.ofInternationalConferenceonBroadbandNetworks(BROADNETS)(2004). [21] Ergin,M.A.,Ramachandran,K.,andGruteser,M.AnExperimentalStudyofInter-cellInterferenceEffectsonSystemPerformanceinUnplannedWirelessLANDeployments.ACMComputerNetworks52(2008).14:2728. [22] Eryilmaz,A.andSrikant,R.FairResourceAllocationinWirelessNetworksUsingQueue-lengthbasedSchedulingandCongestionControl.InProc.ofIEEEINFOCOM(2005). [23] .JointCongestionControl,RoutingandMACforStabilityandFairnessinWirelessNetworks.IEEEJournalonSelectedAreasinCommunications,specialissueonNonlinearOptimizationofCommunicationSystems24(2006).8:1514. [24] Finkenzeller,K.RFIDHandbook:FundamentalsandApplicationsinContactlessSmartCardsandIdentication.(2003). [25] Floyd,S.andJacobson,V.RandomEarlyDetectionGatewaysforCongestionAvoidance.IEEE/ACMTransactionsonNetworking1(1993).4. [26] Garetto,M.,Salonidis,T.,andKnightly,E.W.ModelingPer-owThroughputandCapturingStarvationinCSMAMulti-hopWirelessNetworks.InProc.ofIEEEINFOCOM(2006). 103

PAGE 104

[27] Gollakota,S.and(MIT),D.Katabi.ZigZagDecoding:CombatingHiddenTerminalsinWirelessNetworks.InProc.ofACMSIGCOMM(2008). [28] Grunenberger,Y.,Heusse,M.,Rousseau,F.,andDuda,A.ExperiencewithanImplementationoftheIdleSenseWirelessAccessMethod.InProc.ofCoNEXT(2007). [29] Hamkins,J.AnAnalyticTechniquetoSeparateCochannelFMSignals.IEEETransactionsonCommunications48(2000).4:543. [30] Heusse,M.,Rousseau,F.,Berger-Sabbatel,G.,andDuda,A.PerformanceAnomalyof802.11b.InProc.ofIEEEINFOCOM(2003). [31] Heusse,M.,Rousseau,F.,Guillier,R.,andDuda,A.IdleSense:AnOptimalAccessMethodforHighThroughputandFairnessinRateDiverseWirelessLANs.InProc.ofACMSIGCOMM(2005). [32] Huang,X.andBensaou,B.OnMax-MinFairnessandSchedulinginWirelessAdHocNetworks:AnalyticalFrameworkandImplementation.InProc.ofACMMOBIHOC(2001). [33] Jain,K.,Padhye,J.,Padmanabhan,V.N.,andQiu,L.ImpactofInterferenceonMulti-hopWirelessNetworkPerformance.InProc.ofACMMobiCom(2003). [34] Jian,Y.andChen,S.CanCSMA/CANetworksBeMadeFair?InProc.ofACMMobiCom(2008). [35] Joshi,T.,Mukherjee,A.,Yoo,Y.,andAgrawal,D.AirTimeFairnessforIEEE802.11MultirateNetworks.IEEETransactionsonMobileComputing7(2008).4:513. [36] Katti,S.,Gollakota,S.,andKatabi,D.EmbracingWirelessInterference:AnalogNetworkCoding.InProc.ofACMSIGCOMM(2007). [37] Kelly,F.,Maulloo,A.,andTan,D.RateControlinCommunicationNetworks:ShadowPrices,ProportionalFairnessandStability.JournaloftheOperationalResearch49(1998).3:237. [38] Kim,H.,Yun,S.,Kang,I.,andBahk,S.Resolving802.11PerformanceAnomaliesthroughQoSDifferentiation.IEEECommunicationsLetters9(2005).7:655. [39] Kodialam,M.andNandagopal,T.FastandReliableEstimationSchemesinRFIDSystems.InProc.ofACMMobiCom(2006). [40] Kunniyur,S.andSrikant,R.End-to-endCongestionControl:UtilityFunctions,RandomLossesandECNMarks.IEEE/ACMTransactionsonNetworking11(2003).5:689702. 104

PAGE 105

[41] Law,C.,Lee,K.,andSiu,K.Y.EfcientMemorylessProtocolforTagIdentication.InProf.ofInternationalWorkshoponDiscreteAlgaorithmsandMethodsforMobileComputingandCommunications(2000):75. [42] Lee,S.R.,Joo,S.D.,andLee,C.W.AnEnhancedDynamicFramedSlottedALOHAAlgorithmforRFIDTagIdentication.IEEEInternationalConferenceonMobileandUbiquitousSystems:NetworksandServices(MOBIQUITOUS)(2005). [43] Lin,X.andShroff,N.JointRateControlandSchedulinginMultihopWirelessNetworks.InProc.ofIEEECDC(2004). [44] Lin,X.andShroff,N.B.TheImpactofImperfectSchedulingonCross-LayerCongestionControlinWirelessNetworks.IEEE/ACMTransactionsonNetworking14(2006).2:302. [45] Low,S.H.andLapsley,D.E.OptimizationFlowControl,I:BasicAlgorithmsandConvergence.IEEE/ACMTransactionsonNetworking7(1999).6:861874. [46] Luo,H.,Cheng,J.,andLu,S.Self-CoordinatingLocalizedFairQueueinginWirelessAdHocNetworks.IEEETransactionsonMobileComputing3(2004).1:86. [47] Luo,H.,Lu,S.,andBharghavan,V.ANewModelforPacketSchedulinginMultihopWirelessNetworks.InProc.ofACMMobiCom(2000). [48] Massoulie,L.andRoberts,J.BandwidthSharing:ObjectivesandAlgorithms.InProc.ofIEEEINFOCOM(1999). [49] Mishra,A.,Brik,V.,Banerjee,S.,Srinivasan,A.,andArbaugh,W.AClient-drivenApproachforChannelManagementinWirelessLANs.InProc.ofIEEEINFOCOM(2006). [50] Mishra,A.,Shrivastava,V.,Agarwal,D.,Banerjee,S.,andGanguly,S.DistributedChannelManagementinUncoordinatedWirelessEnvironments.InProc.ofACMMobiCom(2006). [51] Myung,J.andLee,W.AdaptiveSplittingProtocolsforRFIDTagCollisionArbitration.InProc.ofACMMobiHoc(2006). [52] Nandagopal,T.,Kim,T.,Gao,X.,andBharghavan,V.AchievingMACLayerFairnessinWirelessPacketNetworks.InProc.ofACMMobiCom(2000). [53] Neely,M.,Modiano,E.,andLi,C.FairnessandOptimalStochasticControlforHeterogeneousNetworks.IEEE/ACMTransactionsonNetworking16(2008).2:396. [54] Park,E.C.,Kim,D.Y.,Kim,H.,andChoi,C.H.ACross-LayerApproachforPer-StationFairnessinTCPoverWLANs.IEEETransactionsonMobileComputing7(2008).7:898. 105

PAGE 106

[55] Pasupathy,S.MinimumShiftKeying:ASpectrallyEfcientModulation.IEEECommunicationsMagazine(1979). [56] Peng,Q.,Zhang,M.,andWu,W.VariantEnhancedDynamicFrameSlottedALOHAAlgorithmforFastObjectIdenticationinRFIDSystem.IEEEInternationalWorkshoponAnti-counterfeiting,Security,Identication(2007). [57] Pilosof,S.,Ramjee,R.,Raz,D.,Shavitt,Y.,andSinha,P.UnderstandingTCPFairnessoverWirelessLAN.InProc.ofIEEEINFOCOM(2003). [58] Rangwala,S.,Jindal,A.,Jang,K.Y.,Psounis,K.,andGovindanL,R.UnderstandingCongestionControlinMulti-hopWirelessMeshNetworks.InProc.ofACMMobiCom(2008). [59] Razandralambo,T.,Guerin-lassous,I.,Iannone,L.,andFdida,S.DynamicandDistributedPacketAggregationtoSolvethePerformanceAnomalyin802.11WirelessNetworks.ComputerNetworks52(2008).1:77. [60] Release,RSAPress.AnnualRSAWirelessSecuritySurvey:WithEncryptedWi-FiVulnerable,HowCompaniesandIndividualsareRiskingtheirAssetsandReputations. http://www.rsa.com/press_release.aspx?id=9725 (2008). [61] Roberts,L.G.ALOHAPacketSystemwithandwithoutSlotsandCapture.ACMSIGCOMMComputerCommunicationReview5(1975).2:28. [62] Sadeghi,B.,Kanodia,V.,Sabharwal,A.,andKnightly,E.OpportunisticMediaAccessforMultirateAdHocNetworks.InProc.ofACMMobiCom(2002). [63] Sarangan,V.,Devarapalli,M.R.,andRadhakrishnan,S.AFrameworkforFastRFIDTagReadinginStaticandMobileEnvironments.TheInternationalJournalofComputerandTelecommunicationsNetworking52(2008).5:1058. [64] Semiconductors,Philips.I-CODESmartLabelRFIDTags. http://www.nxp.com/acrobat_download/other/identification/SL092030.pdf (2004). [65] Shih,D.,Sun,P.L.,Yen,D.C.,andHuang,S.M.TaxonomyandSurveyofRFIDAnti-collisionProtocols.ComputerandCommunications29(2006).11:2150. [66] Tan,G.andGuttag,J.Time-basedFairnessImprovesPerformanceinMulti-rateWirelessLANs.InProc.ofUSENIXAnnualTechnicalConference(2004). [67] .The802.11MACProtocolLeadtoInefcientEquilibria.InProc.ofIEEEINFOCOM(2005). [68] Telatar,I.E.andGallager,R.G.CombiningQueuingTheoryandInformationTheoryforMultiaccess.IEEEJournalonSelectedAreasCommunication13(1995).6:963. 106

PAGE 107

[69] Vaidya,N.H.,Bahl,P.,andGupta,S.DistributedfairschedulinginawirelessLAN.InProc.ofACMMobiCom(2000). [70] Vasan,A.,Ramjee,R.,andWoo,T.Y.C.ECHOS-EnhancedCapacity802.11Hotspots.InProc.ofIEEEINFOCOM(2005). [71] Vogt,H.EfcientObjectIdenticationwithPassiveRFIDTags.InProc.ofInternationalConferenceonPervasiveComputing(2002). [72] Weis,S.A.,Sarma,S.E.,Rivest,R.L.,andEngels,D.W.SecurityandPrivacyAspectsofLow-costRadioFrequencyIdenticationSystems.InProc.ofInterna-tionalConferenceonSecurityinPervasiveComputing(SPC)(2003). [73] Xu,K.,Bae,S.,LEE,S.,andGerla,M.TCPBehavioracrossMultihopWirelessNetworksandtheWiredInternet.InProc.ofACMWoWMoM(2002). [74] Xu,K.,Gerla,M.,andBae,S.HowEffectiveistheIEEE802.11RTS/CTSHandshakeinAdHocNetworks?InProc.ofIEEEGLOBECOM(2002). [75] Xu,K.,Gerla,M.,Qi,L.,andShu,Y.EnhancingTCPFairnessinAdHocWirelessNetworksusingNeighborhoodRED.InProc.ofACMMobiCom(2003). [76] Xu,S.andSaadawi,M.DoestheIEEE802.11MACProtocolWorkWellinMultihopWirelessAdHocNetworks?IEEECommunicationsMagazine39(2001).6:130. [77] .RevealingTCPUnfairnessBehaviorin802.11BasedWirelessMulti-hopNetworks.InProc.ofIEEEPIMRC(2001). [78] Yang,L.,Seah,W.,andYin,Q.ImprovingFairnessamongTCPFlowsCrossingWirelessAdHocandWiredNetworks.InProc.ofACMMOBIHOC(2003). [79] Zecca,G.,Couderc,P.,Banatre,M.,andBeraldi,R.SwarmRobotSynchronizationUsingRFIDTags.InProc.ofIEEEPerCom(2009). [80] Zhai,H.,Chen,X.,andFang,Y.ImprovingTransportLayerPerformanceinMultihopAdHocNetworksbyExploitingMACLayerInformation.IEEETransac-tionsonWirelessCommunications6(2007).5:1692. [81] Zhang,S.,Liew,S.,andLam,P.P.PhysicalLayerNetworkCoding.InProc.ofACMMobiCom(2006). [82] Zhen,B.,Kobayashi,M.,andShimizu,M.FramedALOHAforMultipleRFIDObjectIdentication.IEICETransactionsonCommunications(2005). [83] Zhou,F.,Chen,C.,Jin,D.,Huang,C.,andMin,H.EvaluatingandOptimizingPowerConsumptionofAnti-collisionProtocolsforApplicationsinRFIDSystems.InProc.ofACMInternationalSymposiumonLowPowerElectronicsandDesign(ISLPED)(2004). 107

PAGE 108

[84] Zhu,J.,Metzler,B.,Guo,X.,andLiu,Y.AdaptiveCSMAforScalableNetworkCapacityinHigh-DensityWLAN:AHardwarePrototypingApproach.InProc.ofIEEEINFOCOM(2006). 108

PAGE 109

BIOGRAPHICALSKETCH MingZhangwasborninChangsha,China,in1977.HereceivedtheB.S.degreefromSunYat-senUniversity,Guangzhou,China,in2000,andtheM.S.degreefromPekingUniversity,Beijing,China,in2004,bothincomputerscience.In2004,hejoinedtheDepartmentofComputerandInformationScienceandEngineeringattheUniversityofFlorida,topursuehisPh.D.degree.HisadvisorwasDr.ShigangChen.HisresearchinterestsincludedQoSandsecurityinwirelessnetworks. 109