Citation
Fair Packet Scheduling and Bandwidth Management in Wireless Networks

Material Information

Title:
Fair Packet Scheduling and Bandwidth Management in Wireless Networks
Creator:
Jian, Ying
Place of Publication:
[Gainesville, Fla.]
Florida
Publisher:
University of Florida
Publication Date:
Language:
english
Physical Description:
1 online resource (108 p.)

Thesis/Dissertation Information

Degree:
Doctorate ( Ph.D.)
Degree Grantor:
University of Florida
Degree Disciplines:
Computer Engineering
Computer and Information Science and Engineering
Committee Chair:
Chen, Shigang
Committee Members:
Fortes, Jose A.
Sahni, Sartaj
Chow, Yuan-Chieh R.
Wong, Tan F.
Graduation Date:
12/19/2008

Subjects

Subjects / Keywords:
Bandwidth ( jstor )
Channel capacity ( jstor )
Flow velocity ( jstor )
Jamming ( jstor )
Radio ( jstor )
Scheduling ( jstor )
Signals ( jstor )
Simulations ( jstor )
Topology ( jstor )
Transmitters ( jstor )
Computer and Information Science and Engineering -- Dissertations, Academic -- UF
assurance, csma, differentiation, fairness, mac, rate, service, wireless
Genre:
bibliography ( marcgt )
theses ( marcgt )
government publication (state, provincial, terriorial, dependent) ( marcgt )
born-digital ( sobekcm )
Electronic Thesis or Dissertation
Computer Engineering thesis, Ph.D.

Notes

Abstract:
Our study focused on fair packet scheduling and bandwidth management in CSMA/CA based wireless networks. We address the fairness problem for MAC-layer links and study end-to-end service differentiation and rate assurance for multihop flows. Fine-level rate control, particularly meeting rate requirements and differentiating various types of end-to-end traffic, remains an open problem for multihop wireless networks. Traditionally, rate assurance in wired networks is achieved through resource reservation and admission control, which can be efficiently implemented since the bandwidth capacity of each communication link is known and the sender of a link has the information of all flows that compete for the bandwidth of the link. In a wireless network, however, the capacity of each wireless link can change unpredictably over time due to contention from nearby links and dynamic channel conditions. An end-to-end flow consumes available bandwidth not only at links on its route but also at all nearby contending links, which makes resource reservation extremely complicated. We propose a new adaptive rate control function based on two novel techniques, called proportional packet scheduling (PPS) and dynamic weight adaptation with floor and ceiling (DWA). PPS distributes channel bandwidth among MAC (one-hop) flows in proportion to their weights. DWA adapts flows' weight according to their rate requirements and priorities. End-to-end traffic is classified into two categories: best-effort flows and QoS flows with rate requirements. The QoS flows are assigned with different priorities. PPS and DWA together achieve three important objectives without resource reservation and admission control. First, when bandwidth contention arises, the rate requirements of the QoS flows are satisfied in the order of priorities. Second, beyond the rate requirements, the rest bandwidth is allocated to the flows in a differentiated manner, taking both bandwidth demand and priority into consideration. Third, no flow is starved and all bandwidth is effectively utilized. Another important problem is on how to achieve fairness for the MAC-layer links in multiple contending WLANs or multihop networks, where the carrier sensing range and the interference range are much larger than the transmission range. We demonstrate that CSMA/CA networks, including IEEE 802.11 networks, exhibit severe fairness problem in many scenarios. Most existing solutions require nodes to overhear transmissions made by contending nodes and, based on the overheard information, adjust local rates to achieve fairness among all contending links. Their underlying assumption is that transmissions made by contending nodes can be overheard. However, this assumption holds only when the transmission range is equal to the carrier sensing range, which is not true in most real networks. As our study reveals, the overhearing-based solutions, as well as several non-overhearing AIMD solutions, cannot achieve MAC-layer fairness in various settings. We propose a new rate control protocol, called PISD (Proportional Increase Synchronized multiplicative Decrease). Without relying on overhearing, it provides fairness in CSMA/CA networks, particularly IEEE 802.11 networks, by using only local information and performing localized operations. It combines several novel rate control mechanisms, including synchronized multiplicative decrease, proportional increase, and background transmission. PISD works precisely for scenarios where all MAC flows mutually contend but has limitations when applied to networks consisting multiple contention groups. We develop PISD further and propose two new schemes to overcome the limitations. We prove that flows' rates attained under the two new schemes approximate proportional fairness. ( en )
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.
Thesis:
Thesis (Ph.D.)--University of Florida, 2008.
Local:
Adviser: Chen, Shigang.
Statement of Responsibility:
by Ying Jian.

Record Information

Source Institution:
UFRGP
Rights Management:
Copyright Jian, Ying. Permission granted to the University of Florida to digitize, archive and distribute this item for non-profit research and educational purposes. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder.
Resource Identifier:
664679596 ( OCLC )
Classification:
LD1780 2008 ( lcc )

Downloads

This item has the following downloads:


Full Text

PAGE 1

1

PAGE 2

2

PAGE 3

3

PAGE 4

Iwouldliketothankmyadvisor,Prof.ShigangChen,forhisconstantguidance,support,andencouragement.Iamprivilegedtohavesuchawonderfuladvisor,whoisatalltimesenthusiastic,optimistic,patient,helpful,andencouraging.Hegavemeextensiveadviceandinsightduringthecourseofmyresearchwork.MyspecialthanksgotoProf.SartajSahni,Prof.RandyChow,Prof.YeXia,Prof.JoseFortes,andProf.TanWong,fortheirinstructivecommentsandsupportduringmywork.IwouldalsoliketothankallmycolleaguesinProf.Chen'sresearchgroup,includingZhanZhang,LiangZhang,MyungKeunYoon,MingZhangandTaoLi,forprovidingahighlevelofresearchsupport.IwanttoexpressmydeepestlovetomydarlingwifeQianqian,myparents,andmybrother.Theirlove,understanding,andencouragementhavealwaysbeenthestrongestsupporttome. 4

PAGE 5

page ACKNOWLEDGMENTS ................................. 4 LISTOFTABLES ..................................... 7 LISTOFFIGURES .................................... 8 ABSTRACT ........................................ 10 CHAPTER 1INTRODUCTION .................................. 12 1.1ServiceDierentiationandRateAssuranceBasedonWeightedFairPacketScheduling .................................... 12 1.2FairBandwidthAllocationinCSMA/CANetworks ............. 16 2PROPORTIONALPACKETSCHEDULINGAMONGMACFLOWS ..... 21 2.1RelatedWork .................................. 22 2.2ContentionamongMACFlows ........................ 24 2.3ProportionalPacketScheduling ........................ 25 2.4Properties .................................... 28 2.5EnhancingPPSbyRequest-for-RTSPackets ................. 31 2.6StateoftheArtandOurContribution .................... 32 2.7PerformanceEvaluation ............................ 33 2.8Summary .................................... 35 3END-TO-ENDSERVICEDIFFERENTIATIONANDRATEASSURANCE .. 38 3.1RelatedWork .................................. 38 3.2Objectives .................................... 39 3.3Challenge .................................... 42 3.4DesignOverview ................................ 43 3.5DWA:DynamicWeightAdaptationwithFloorandCeiling ......... 45 3.6Properties .................................... 46 3.7AvoidingPacketDrops ............................. 48 3.8FlowDynamicsandChannelDynamics .................... 49 3.9Intra-owContentionandInter-owContention ............... 50 3.10PerformanceEvaluation ............................ 50 3.11Summary .................................... 53 4FAIRBANDWIDTHALLOCATIONINCSMA/CANETWORKS ....... 56 4.1RelatedWork .................................. 57 4.2NetworkModel ................................. 58 4.3FairnessProblemRemainsOpeninCSMA/CANetworks .......... 59 5

PAGE 6

............................ 59 4.3.2LimitationofOverhearing-BasedSolutions .............. 61 4.3.3AIMDDoesNotWorkEither ..................... 62 4.4ProportionalIncreaseSynchronizedmultiplicativeDecrease ......... 64 4.4.1SynchronizedMultiplicativeDecrease ................. 64 4.4.2AISD:AdditiveIncreaseSynchronizedMultiplicativeDecrease ... 65 4.4.3PISD:ProportionalIncreaseSynchronizedMultiplicativeDecrease 68 4.4.4PISDwithBackgroundTransmission ................. 68 4.4.5Discussion ................................ 70 4.5Analysis ..................................... 71 4.5.1WeightedFairnessandConvergenceTime ............... 71 4.5.2ChannelCoverage ............................ 74 4.5.3ConvergenceAccuracy ......................... 74 4.6AdditionalSimulations ............................. 75 4.7LimitationofPISD ............................... 78 4.8ProportionalFairScheduling .......................... 79 4.8.1ReducingChannelJammingStrength ................. 80 4.8.2ProblemofPISD-RS .......................... 82 4.8.3PFS(ProportionalFairScheduling) .................. 83 4.9AnalysisonProportionalFairness ....................... 86 4.10SimulationsonProportionalFairness ..................... 89 4.11Summary .................................... 91 5CONCLUSION .................................... 103 REFERENCES ....................................... 104 BIOGRAPHICALSKETCH ................................ 108 6

PAGE 7

Table page 2-1Flowrates(inpacketspersecond)onthetwo-owtopology ........... 36 4-1Flowratesachievedbydierentprotocols. ..................... 92 4-2Underdierentweightassignments,owratesarealwaysproportionaltoowweights. ........................................ 92 4-3FlowratesachievedbyPISD. ............................ 92 4-4FlowratesachievedbyPISD-RS,comparedwithproportionalfairnessandbaseline802.11DCF. ...................................... 93 4-5FlowratesachievedbyPFS. ............................. 93 7

PAGE 8

Figure page 2-1Fivetypesofcontendingowsoffi;j. ........................ 36 2-2Packetlabelpriority ................................. 36 2-3Request-for-RTShelpspreemption ......................... 36 2-4Two-owtopology .................................. 37 2-5Five-owtopology .................................. 37 2-6ProtocolofAdjConWin ............................... 37 2-7ProportionalPacketScheduling ........................... 37 2-8ProportionalPacketSchedulingsupportsdynamicweight ............. 37 3-1Networktopology ................................... 54 3-2SettingA,IEEE802.11DCF ............................ 54 3-3SettingA,DWA ................................... 54 3-4SettingA,DWA,doublingthenumberofows .................. 54 3-5SettingB,DWA ................................... 54 3-6SettingB,EDCA ................................... 54 3-7SettingC,DWA ................................... 55 3-8SettingD,DWA ................................... 55 4-1Networkoftwoows,(a;b)and(c;d). ....................... 94 4-2Ratesoftwoowswithrespecttothedistancebetweenbandc.Thedistancefromatobandthatfromctodareboth150m. .................. 94 4-3Manycontendingnodescannotbeoverheardbyi. ................. 94 4-4Ingeneral,Huang-Bensaouprotocoldoesnotworkifanyoneofthecontendinglinkscannotbeoverheard. .............................. 95 4-5AdditiveIncreaseMultiplicativeDecrease:multiplicativedecreaseoccursupontransmissionfailure. ................................. 95 4-6AdditiveIncreaseMultiplicativeDecrease:multiplicativedecreaseoccurswhenbueroccupancypassesacertainthreshold. .................... 95 4-7Idlesense:Thesamecontentionwindowsizedoesnotensurefairness. ...... 95 8

PAGE 9

.......... 96 4-9UnsynchronizedmultiplicativedecreaseinCSMA/CAcannotachievefairness. 96 4-10SynchronizedmultiplicativedecreaseequalizestheowratesforthenetworkinFig. 4-1 ........................................ 96 4-11RatesoftwocontendingowsunderAISDwithrespecttotime. ......... 96 4-12Givenwa;b=3andwc;d=1,underPISD,therateofow(a;b)isaboutthreetimesthatofow(c;d). ............................... 97 4-13RatesoftwocontendingowsunderPISDwithrespecttotime.wa;b=3,wc;d=1,andthedistancefrombtocis100m. ....................... 97 4-14Backgroundtransmissionwillutilizesomeunusedchannelbandwidthforpackettransmission. ..................................... 97 4-15Networktopology ................................... 98 4-16(a):FlowratesunderIEEE802.11DCF;(b):FlowratesunderHuang-Bensaouprotocol. ........................................ 98 4-17ProportionalIncreaseSynchronizedMultiplicativeDecreaseachievesfairnessamongallows. .................................... 99 4-18Flowratesareslightlyimprovedwithbackgroundtransmission. ......... 99 4-19Increasingthevalueofreducesbothconvergencetimeandchannelcoverage. 99 4-20Increasingthevalueofreducesbothconvergencetimeandconvergenceaccuracy. 99 4-21Hostsh2andh6arechangedtoservers. ...................... 99 4-22(a)Whentheserverseachhaveweight3andtheclientseachhaveweight1,therateofaserveristhreetimesthatofaclient;(b)Downwardspikesarereducedwhenisdecreased. ................................. 100 4-23FourWLANsnetworktopology ........................... 100 4-24Adhocnetworktopology ............................... 100 4-25Multihopnetworkwithsixows. .......................... 101 4-26Multihopnetworkwithveows. .......................... 101 4-27Multihopwirelessnetworkwith23MACows. .................. 101 4-28Flowratesattainedunderdierentapproaches ................... 102 9

PAGE 10

OurstudyfocusedonfairpacketschedulingandbandwidthmanagementinCSMA/CAbasedwirelessnetworks.WeaddressthefairnessproblemforMAC-layerlinksandstudyend-to-endservicedierentiationandrateassuranceformultihopows. Fine-levelratecontrol,particularlymeetingraterequirementsanddierentiatingvarioustypesofend-to-endtrac,remainsanopenproblemformultihopwirelessnetworks.Traditionally,rateassuranceinwirednetworksisachievedthroughresourcereservationandadmissioncontrol,whichcanbeecientlyimplementedsincethebandwidthcapacityofeachcommunicationlinkisknownandthesenderofalinkhastheinformationofallowsthatcompeteforthebandwidthofthelink.Inawirelessnetwork,however,thecapacityofeachwirelesslinkcanchangeunpredictablyovertimeduetocontentionfromnearbylinksanddynamicchannelconditions.Anend-to-endowconsumesavailablebandwidthnotonlyatlinksonitsroutebutalsoatallnearbycontendinglinks,whichmakesresourcereservationextremelycomplicated. Weproposeanewadaptiveratecontrolfunctionbasedontwonoveltechniques,calledproportionalpacketscheduling(PPS)anddynamicweightadaptationwithoorandceiling(DWA).PPSdistributeschannelbandwidthamongMAC(one-hop)owsinproportiontotheirweights.DWAadaptsows'weightaccordingtotheirraterequirementsandpriorities.End-to-endtracisclassiedintotwocategories:best-eortowsandQoSowswithraterequirements.TheQoSowsareassignedwithdierent 10

PAGE 11

AnotherimportantproblemisonhowtoachievefairnessfortheMAC-layerlinksinmultiplecontendingWLANsormultihopnetworks,wherethecarriersensingrangeandtheinterferencerangearemuchlargerthanthetransmissionrange.WedemonstratethatCSMA/CAnetworks,includingIEEE802.11networks,exhibitseverefairnessprobleminmanyscenarios.Mostexistingsolutionsrequirenodestooverheartransmissionsmadebycontendingnodesand,basedontheoverheardinformation,adjustlocalratestoachievefairnessamongallcontendinglinks.Theirunderlyingassumptionisthattransmissionsmadebycontendingnodescanbeoverheard.However,thisassumptionholdsonlywhenthetransmissionrangeisequaltothecarriersensingrange,whichisnottrueinmostrealnetworks.Asourstudyreveals,theoverhearing-basedsolutions,aswellasseveralnon-overhearingAIMDsolutions,cannotachieveMAC-layerfairnessinvarioussettings. Weproposeanewratecontrolprotocol,calledPISD(ProportionalIncreaseSynchronizedmultiplicativeDecrease).Withoutrelyingonoverhearing,itprovidesfairnessinCSMA/CAnetworks,particularlyIEEE802.11networks,byusingonlylocalinformationandperforminglocalizedoperations.Itcombinesseveralnovelratecontrolmechanisms,includingsynchronizedmultiplicativedecrease,proportionalincrease,andbackgroundtransmission.PISDworkspreciselyforscenarioswhereallMACowsmutuallycontendbuthaslimitationswhenappliedtonetworksconsistingmultiplecontentiongroups.WedevelopPISDfurtherandproposetwonewschemestoovercomethelimitations.Weprovethatows'ratesattainedunderthetwonewschemesapproximateproportionalfairness. 11

PAGE 12

Wirelesstechnologieshavecomealongway,fromtheradiotechnologiesthatallowdatabitstobeexchangedbetweentwophysically-disconnecteddevices,tothemultipleaccesstechnologiesthatallowagroupofdevicestoshareacommonwirelesscommunicationchannel,andtotheroutingtechnologiesthatallowout-of-rangedevicestocommunicateviamultihopwirelesspaths.FollowingtheenormoussuccessofWLAN,multihopwirelessnetworks,includingmeshnetworks,ad-hocnetworks,sensornetworks,areexpectedtoleadinthenextwaveofdeployment.Toimprovetheirapplicabilityinpractice,notonlymustthesenetworksprovidearobustandecientcommunicationservice,butalsotheyshouldprovideexibletoolsfortracengineeringinordertosupportdiverseuserapplications. 12

PAGE 13

Resourcereservationinmultihopwirelessnetworksalsohasproblems.Anend-to-endowconsumesbandwidthnotonlyatlinksonitsroutebutalsoatallnearbycontendinglinks,whichmakesresourcereservationextremelycomplicated.Withspatialchannelreuse,thelocalchannelperceivedbyeachwirelesslinkisdierentbecauseeachlinkhasadierentsetofcontendinglinks.Twocontendinglinkswillconsumebandwidthineachother'sperceivedchannel.Consideranewowwhoseraterequirementisrandroutingpathisa!b!c!d.Inordertosupporttheow,thechannelwherelink(a;b)residesshouldhave3rresidual(unused)bandwidthbecause(a;b),(b;c)and(c;d)mutuallycontendandtheywilleachconsumerbandwidthinthechannelwhencarryingtheow(assumingIEEE802.11DCF).Similarly,thechannelswhereotherlinksoftheroutingpathresidealsoneedmorethanrresidualbandwidthfortheow.Evenlinksoutsideofthepathneedresidualbandwidthtosupporttheow.Consideranearbylink(x;y)thatcontendswith(a;b).Supposeitschannelisalreadysaturatedduetoheavytraconsomeothercontendinglinks.Nowifweaddthenewow,astherateon(a;b)isincreased,therateon(x;y)willbedrivendown,causingtheviolationofthepreviousresourcereservationmadeon(x;y).Determininghowmuchbandwidth(x;y)needsinordertosupportthenewowisnotaneasytask.Itdependsonhowmuchchannelspatialreusecanbedonebetween(a;b)andotherlinkscontendingwith(x;y).Therefore,resourcereservationrequirescoordinationamonglinksontherouteandallotherlinksthatcontendwiththem. 13

PAGE 14

1 { 4 ].Wheneverylinkseesthesamechannelwiththesamesetofcontendinglinks,manyoftheaboveproblemsareeitheravoidedormuchsimplied.Adierentrestrictioncanbeassumingeachnodetransmitsatadierentfrequency[ 5 ].Theseconddirectionistoworkoncoarse-levelservicedierentiationthatdoesnotproviderateassurance[ 6 { 9 ].Forexample,dierentbackopolicies[ 7 ]ordierentcontentionwindowsizes[ 6 ]areassignedtopacketsofdierentclassestoprovidequalitativeschedulingpreferences.IEEE802.11e[ 10 ]belongstothiscategorywhenitisappliedinamultihopwirelessnetwork.Thethirddirectionistodesignheuristicstoaddressthehardproblemsinresourcereservationandadmissioncontrol.Mostworkfocusesonestablishingaheuristicapproachforeachnodetoestimateitschannel'sresidualbandwidth,whichwillbeusedtoguideadmissioncontrol.Thebandwidthestimationismadebasedonchannelidletime[ 6 ; 11 ],averagepackettransmissiondelay[ 12 ; 8 ; 2 ],orchannel-accessprobabilisticmodels[ 13 { 16 ].Asdetailedanalysisin[ 16 ]pointsout,noneofthemconsiderstheimpactofhiddenterminalsinmultihopwirelessnetworks,andeachwillperformpoorlyundercertainscenarios.Moreover,theresidualbandwidthmeasuredmaycontinuouslychangeduetodynamicchannelconditions,andestimatingbandwidthdoesnotsolvethecomplicatedresourcereservationproblemdiscussedpreviously. Insteadoftakingahead-onapproachtoaddressthedicultproblemsofresourcereservationandadmissioncontrol,wewanttotakeastepbackandaskwhetherresourcereservationandadmissioncontrol,legacyfromwirednetworks,aresuitableformultihopwirelessnetworks.Wewanttondanalternativesolutionforwirelessnetworksthatcannotonlysolvetheproblemsbutalsohaveamuchsimplerdesign.Inthisstudy,weproposetoreplaceadmissioncontrolandresourcereservationwithasimpleyeteectiveadaptiveratecontrolfunctionsuitedforhandlingnetwork/tracdynamics.Itautomaticallyadaptsthebandwidthdistributiontosatisfytheraterequirementsofas 14

PAGE 15

Weclassifyend-to-endtracintotwocategories:best-eortowsandQoSowswithminimumraterequirements.TheQoSowsareassignedtoserviceclassesofdierentpriorities.Wehavethefollowingthreeobjectivesforratecontrol.TherateassuranceobjectiverequiresQoSowstobesupportedintheorderofpriorities.Ahigher-priorityowcanpreemptthebandwidthofalower-priorityow.Followingthepriorityorder,thenetworkshouldsupportasmanyQoSowsaspossible.Beyondmeetingtheminimumraterequirements,thebandwidthdierentiationobjectiverequirestheremainingbandwidthtobeallocatedtoend-to-endowsbasedontheirprioritiesaswellasbandwidthdemand.Aowwithahigherminimumraterequirementandahigherpriorityshouldreceivealargeramountofextrabandwidth.Theno-starvation/maximum-utilizationobjectiverequiresthatnoowisstarvedandallnetworkbandwidthisutilizedwhenpossible. Toachievethethreeobjectives,wedesignouradaptiveratecontrolfunctionbasedontwonoveltechniques[ 17 ].First,weenhanceCSMA/CAprotocolsforweightedbandwidthallocationthroughanewtechnique,calledproportionalpacketscheduling(PPS),whichdistributeschannelbandwidthamongMAC(one-hop)owsinproportiontotheirweights.Comparingwiththeexistingschemes,PPSrealizesweightedbandwidthallocationinanegranularityandachievesbetterthroughputduetoreducedradiocollision.ItisalsotherstfullydistributedsolutionthatachievesprovableweightedmaxminfairnessinCSMA/CAnetworksofdynamicchannelconditions,withaboundederrorthatcanbemadearbitrarilysmall. Second,workingontopofPPS,anewtechniquecalleddynamicweightadaptationwithoorandceiling(DWA)isproposed,whichallowseachMACowtoindependently 15

PAGE 16

Whenwirelesshostssharethesamecommunicationchannel,theyshouldbegivenafairchanceofaccessingthewirelessmedium.FairnessisoneofthecoreproblemsthatanyMACprotocolmustaddress.Itpreventsthesituationthatsomehostsobtainmostofthechannel'sbandwidthwhileothersstarve.Amoregeneralproblemisthatofweightedfairness,wherethechannel'sbandwidthobtainedbyahostisproportionaltoitsweight,whichisassignedbytheuserbasedonapplicationrequirements.Forexample,whenawebserverandaclienthostsharethesamelocalchannel(e.g.inaWLAN),theservermaybegivenahigherweightbecauseitmayhavetouploadcontenttomultipleusersontheInternetsimultaneously. RandombackointheIEEE802.11DCFachievesfairnessinaWLANwhereallhostsaredownloadingcontentfromtheInternetviathesameaccesspoint.However,asobservedinourstudy,itcannotachievefairness(letaloneweightedfairness)inmanyotherscenarios.Forexample,whenaserverthatuploadscontenttotheInternetsharestheaccesspointofaWLANwithaclienthostthatdownloads,theclientmayobtainmostofthechannel'sbandwidthwhiletheserverisslowedtocrawl.Whentheaccesspointsat 16

PAGE 17

ThefairnessprobleminIEEE802.11networksismostlyduetothefundamentallimitationofCSMA/CA,whichgivespreferenceinmediaaccesstosomelinksoverothers,dependingontheirspatiallocations.Asthisproblemiswellrecognized,manyfairnesssolutionshavebeenproposedinthepastdecade[ 18 { 25 ].Theyfallintwocategories:overhearing-basedsolutionsandnon-overhearingsolutions. Theoverhearing-basedsolutionsrequireeachnodetomonitortheactivityofallcontendingnodesandcollecttheirlinks'information(suchasrate,schedulingtag,orbuerstatus).Basedonthecollectedinformation,anodedecidesitsownmediacontentionpolicy:i)increase/decreaseminimumcontentionwindowifthelocalrateisabove/belowtheaveragerateofallcontendinglinks[ 18 ; 19 ; 22 ],ii)serializetransmissionsamongcontendinglinksbasedontheirschedulingtags[ 20 ; 21 ],oriii)emulateTDMAbycomputingacontention-freeslottedscheduleamongthelinks[ 23 ].Thekeyquestionishowtocollectinformationforthecontendinglinks,whichmaybemultiplehopsaway.Onenaiveapproachisforeachnodetooodtheinformationdescribingitsadjacentlinkstoallnodeswithinacertainnumberofhops.Thisapproachisnotonlycostlybutalsoawedbecause,asisobservedin[ 23 ; 16 ],hopcountisnotareliablemeanstoidentify 17

PAGE 18

Mostnon-overhearingsolutionsusetheclassicalAIMD(AdditiveIncreaseMultiplicativeDecrease)forratecontrol.Ononehand,anodecannotoverheartheexactinformationintransmissionsmadeoncontendinglinkswhoseradiosignalisstrongenoughtocauseinterferencebuttooweaktodecode.Ontheotherhand,withoutoverhearing,thenodecanstillsensetheaggregateimpactofinterferencefromthoselinksbymonitoringhowbusythechannelis,howfrequentlyitsowntransmissionsfail[ 24 ; 26 ],orhowfastitslocalbuerislledup.Basedonsuchinformation,emulatingthebehaviorofTCPinsomesense,eachnodemaysetathresholdtodecidewhenthechanneliscongestedsuchthatmultiplicativedecreaseshouldbeperformed.Thisdirectionlooksreasonable.However,oursimulationsinns-2showthatAIMDfailstoachievefairness,too,notbecausetherationalebehindAIMDisawed,butbecausetheinteractionbetweenAIMDandCSMA/CAneutralizestheeectivenessofAIMD.AIMDmayalsobeappliedtothecontentionwindowbasedonthenumberofidleslotsbetweentwotransmissions[ 25 ; 27 ](whichcanbemeasuredthroughcarriersenseinsteadofoverhearing).Wewillshowlaterthatthisapproachalsohaslimitations. Weproposeanewratecontrolprotocol,PISD(ProportionalIncreaseSynchronizedmultiplicativeDecrease)[ 28 ],thatprovidesfairnessinCSMA/CAnetworks,particularlyinIEEE802.11networks.WithinthedesignofPISD,wemakethreecontributions.First,ourstudyrevealsthefundamentalreasonsexactlywhytheexistingfairnesssolutions,aswellasAIMD,donotworkunderrealisticcontentionconditions.Particularly,forAIMD, 18

PAGE 19

WealsoenhancePISDfurther.PISDworkspreciselyforscenarioswhereallMACowsmutuallycontend,butwhenitisappliedtoanetworkconsistingofmultiplecontentiongroups,thenetwork'sthroughputcanbedegraded.WedevelopPISDfurtherandproposetwonewschemes,PISD-RSandPFS,toovercomethelimitation.PISD-RSisasimplesolution,whilePFSisabetteryetmoresophisticatedsolution.BothofthemimprovePISDbyrevisingthewaytheyjamthewirelesschannel,wherechanneljammingisthekeytechniquethatPISDexploitstoachievesynchronizedmultiplicativedecrease.ThetwonewproposedschemesareabletoprovidefairnessinCSMA/CAnetworkswithmultiplecontentiongroups.Weconductboththeoreticalanalysisandsimulationstostudytheireectiveness.Byusingconvexoptimizationtheory,weprovethatows'ratesachievedbyPFS(andPISD-RS)approximateproportionalfairness. Therestofthisstudyisorganizedasfollows.Chapter 2 proposesPPS,thetechniqueforweightedbandwidthallocationamongMACowsinmultihopwirelessnetworks.Chapter 3 proposesDWA,whichworksontopofPPSandachievesend-to-endservicedierentiationandowrateassurance.Chapter 4 proposesPISDanditsenhanced 19

PAGE 20

5 concludesourstudy. 20

PAGE 21

Inthischapter,wepresentanewMAC-layerschedulingtechniquecalledproportionalpacketscheduling(PPS)thatachievesweightedbandwidthallocationamongsingle-hopMAC-layerowsinamultihopwirelessnetwork.WedesignthistechniquetoserveasabasisfortheimplementationofDWA,anupperlayerQoSprotocolthatwillbeintroducedinChapter 3 DWAinChapter 3 needstoworkontopofaMAC-layerprotocolthatsupportsweightedbandwidthallocation,andtheowweightsshouldbeabletobedynamicallychanged.Withaddedcomplexity,someMACprotocolsinthecurrentliteraturecanpotentiallybemodiedtoserveforthispurpose.Protocols[ 29 ; 18 ; 19 ; 21 ]proposedtoachievefairnessamongcontendingMACowsmaybeenhancedtosupportweightedbandwidthallocation,butinsuchprotocolsknowledgeonlocalnetworktopologyoranadditionalfairsharecomputationphaseisneeded,whichmayleadtodicultiesonachievingdynamicweightchange.TheoverlayMAC(OML)[ 23 ]andtheRegulatedContentionMAC(RCMAC)[ 4 ]candirectlysupportweightedbandwidthallocation.However,RCMACwasdesignedtoworkundersingle-hopcontentionsandOMLcannotbeeasilymodiedtosupportfastweightadjustment.Inaddition,mostoftheseworksdonotprovideweightedbandwidthallocationinaverynegranularity. PPS,whichisintroducedinthischapter,achievesweightedbandwidthallocationinanegranularityandisparticularlysuitablefordynamicweightadaptation.Allowingchannelcapacitytoevolvespatiallyandtemporally,thisistherstfullydistributedsolutionthatachievesprovableweightedmaxminfairnessinCSMA/CAnetworkswithaboundederrorthatcanbemadearbitrarilysmall.Webelieveitisastrongresult.Themax-minfairnessachievedin[ 30 ]usesamulti-channelcontentionmodelthatisdierentfromtheCSMA/CAmodelinourstudy.Theworkin[ 19 ]requiresthecomputationofallcontentioncliques(whichisdiculttoimplementdistributedlyandmustberedone 21

PAGE 22

Therestofthischapterisorganizedasfollows.Section 2.1 discussestherelatedwork.Section 2.2 studiesthevetypesofcontentionsinthenetwork.Section 2.3 describesPPSindetails.Sections 2.4 analyzesthepropertiesofPPS.Sectoin 2.5 introducesRRTS,whichisanenhancementtoPPS.Section 2.6 pointsoutthestateoftheartandourcontribution.Section 2.7 evaluatestheperformanceofPPSthroughsimulations.Section 2.8 summarizesthechapter. 29 ]studiedtheproblemoffairdistributionofbandwidthandmaximizationofresourceutilization.Theyrstassureeachowinanetworkwithaminimumchannelshare,thenmaximizeaggregatechannelutilizationbyspatialchannelreuse.Thedistributedimplementationofthisalgorithmneedstopologyinformationtobepropagatedalongaconict-freespanningtree.Intheirfollow-upwork[ 31 ; 21 ],SFQ(Start-timeFairQueueing)isappliedtomultihopwirelessnetworksinadistributedway.Intheapproach,servicetags(representingtransmissiondeadlines)arepiggybackedinpackets,andeachnodemaintainsthestatusofallitscontendingows.Locally,theowwiththeminimumservicetagisscheduledrst.Inaddition,spatialchannelreuseisexploitedthroughsimultaneoustransmissions.Thisapproachneedseachnodetomaintainthestatusofallitscontendingows,andthewayitdealswiththehiddenterminalproblemisonlyheuristic,whichcanonlyalleviatesbutnotsolvetheproblem.In[ 18 ; 19 ; 32 ],toachievefairbandwidthdistributionamongcontendingwirelesslinksinamultihop 22

PAGE 23

33 ],Liexploitedasimilarapproachtoensureaconservative,smallfairshareofbandwidthforeachend-to-endow.In[ 34 ],Nandagopaletal.proposedageneralanalyticalfairnessmodelandaMACprotocoltoapproximateproportionalfairness,whichadjuststhecontentionwindowsizebasedontheoccurrenceofretransmissions.In[ 20 ],Vaidyaetal.achievedfairnessinawirelessLANbyadjustingthecontentionwindowusingafairqueueingalgorithm.In[ 30 ],TassiulasandSarkaraddressedthemax-minfairnesswithamulti-channelcontentionmodelthatisdierentfromtheCSMA/CAmodelusedinourstudy. Manyalgorithmsdescribedaboverequireadditionalproceduresthatcalculateandexchangethefairsharesofthewirelesslinks,orrequirethemaintenanceofinformationaboutlocalnetworktopologyandowstatus.Theserequirementsdiminishtheexibilityofthealgorithms.Inaddition,noneoftheexistingalgorithmpromisestoprovidefairnessinaverynegranularity,whichisessentiallycausedbythehiddenterminalproblem.Inthisstudy,basedonathoroughstudyonalldierenttypesofcontentions,wedesignourweightedbandwidthallocationprotocolthatdoesnothavetheabovedefects,anditisabletoprovideafoundationforimplementingservicedierentiationandrateassuranceintheupperlayer.Inourprotocol,aspecialcontrolpacketcalledRRTS(Request-for-RTS)isusedtohelpprovidefairnessinanegranularity.ItshouldbementionedthatwearenottherstonewhointroducetheRRTScontrolpacket.RRTSisinitiallyintroducedin[ 35 ]tosolvecontentionproblemsinmultihopwirelessnetworks.Inthisstudy,weimproveitfurtheranduseittogetherwithpacketlabelingtoachieveweightedfairscheduling. 23

PAGE 24

2-1 .OperationsinPPSaredesigneddierentlyandspecicallyforthem. Thersttypeofcontendingowsincludesthosewhosesendersarealsoi.Anexampleisfi;kinthegure.fi;kcontendswithfi;jbecausenodeicannotsendtwopacketssimultaneously. ThesecondtypeincludesthoseMACowswhosesendersareinthetransmissionrangeofnodei.Theirreceiversmayormaynotbeinthetransmissionrangeofi.Anexampleisfa;binthegure.Itcontendswithfi;jbecauseRTS/DATAsentbyacanreachiandcauseradiocollisionwheniisreceivingCTS/ACKfromj. ThethirdtypeincludesthoseMACowswhosesendersareinthetransmissionrangeofnodej.Theirreceiversmayormaynotbeinthetransmissionrangeofj.Anexampleisfe;minthegure.Itcontendswithfi;jbecauseRTS/DATAsentbyecanreachjandcauseradiocollisionwhenjisreceivingRTS/DATAfromi. ThefourthtypeincludesthoseMACowswhosesendersareoutsidethetransmissionrangesofbothiandj,butreceiversliewithinthetransmissionrangeofi.Anexampleisfd;cinthegure.Itcontendswithfi;jbecauseCTS/ACKpacketsfromccanreachnodeiandcauseradiocollisionwheniisreceivingCTS/ACKfromj.Itiswellknownthatthemostsevereunfairnessoccursunderthisscenario. ThefthtypeincludesthoseMACowswhosesendersareoutsidethetransmissionrangesofbothiandj,butreceiversliewithinthetransmissionrangeofj.Anexampleisfh;ginthegure.Itcontendswithfi;jbecauseCTS/ACKpacketsfromgcanreachnodejandcauseradiocollisionwhenjisreceivingRTS/DATAfromi.Becausethesendersof 24

PAGE 25

Finally,itiseasytoseethatanyownotbelongingtotheabovevetypesdoesnotcontendwithfi;j. ConsideraMACowfi;joverawirelesslink(i;j).Letwi;jbetheweightoftheow.Themeanrateoftheowisdenedasitsratedividedbyitsweight.ThegoalofPPSistoequalizethemeanratesofcontendingows.Thesenderoffi;jmaintainsaseparatepacketqueuefortheow.Italsokeepsacounter,denotedasci;j,providinganindirect,discretemeasurementforthemeanrateoftheow,whichwewilldiscussindetailsshortly. ThebasicideabehindPPSisthat,inordertoequalizethemeanratesofcontendingows,weshouldalwaysgivethehighestpriorityinmediaaccesstotheowwhosecurrentmeanrateisthesmallest.PPStriestomaximizethesmallestowrateinthenetwork,thenmaximizethesecondsmallest,andsoon.Inordertomaximizespatialchannelreuse,PPSneverleaveslocalchannelidlewhenthereisabackloggedMACow.ThedesignofPPSfollowstworulesbelow.AssumeallMACowsunderdiscussionarebacklogged. 1. AMACowshouldoccupythechannelfortransmissionifithasthesmallestmeanrateamongallcontendingowsorifthechannelisidle. 2. AMACowshouldnotcompeteformediaaccessifitsmeanrateisnotthesmallestamongitscontendingowsandthechannelisbusy. First,wedescribehowthesenderofaMACowmaintainsitscounter.Assumetheclocksatallnodesarelooselysynchronized.Timeisdividedintoperiods.Atthe 25

PAGE 26

Next,wedescribehowthesendersofcontendingowscoordinatetheorderoftransmissionsbasedontheircountervalues.ConsideranarbitraryMACowfi;j.Belowwediscusswhatinformationthesenderandthereceiverwillgatherandwhatoperationstheywillperformbasedonthatinformation. Letni;jbethenumberofbitsyettobetransmittedover(i;j)beforeci;jisincreasedbyone.Whenadatapacketissentfromitoj,RTS/CTS/DATA/ACKpiggybackbothci;jandni;j.Foralldierenttypesofcontendingowsoffi;jshowninFig. 2-1 ,neighborsofeitheriorj,suchasa,c,eandg,willlearnci;jandni;jthroughoverhearing.Similarly,iwilllearnthecountervaluesoffa;bandfd;c,andjwilllearnthecountervaluesoffe;mandfg;h,aswellasthenumberofbitsyettobetransmittedbeforeeachcounterisincreasedbyone.Therefore,thesenderiofowfi;jonlyknowstheinformationforsomecontendingows,andthereceiverknowstheinformationfortherestcontendingows. Tohelpunderstandpacketlabels,Fig. 2-2 showstwoexamples.Whenschedulingpackettransmissionsthegoalistogivepacketswithsmallerlabelshigherpriorities,andatthesametime,thechannelshouldbefullyutilized.InFig. 2-2 (a),aissendingapacket 26

PAGE 27

2-2 (b),whereows1-4mutuallycontendwitheachotherandow5contendsonlywithow4.Flow4cannotsenditspacketwithalabelof4,becauseallthelabelsofitscontendersarestill3.However,forow5,thoughitscontender,ow4,hasasmallerlabel,itshouldignorethisandcontinuetosend,otherwise,thechannelwouldbelocallyidleandspatialchannelreuseisnotutilized.TheoperationsofPPSaredescribedasfollows: Wheniisthesenderofmultipleowssuchasfi;jandfi;k,ialwaysschedulestheowwiththeminimumcounter.Withoutlosinggenerality,letthisowbefi;j.Nodeiwillrefrainfromaccessingmediaifitoverhearsthatacontendingowwithasmallerorequalcounteristransmitting.ItwillattempttoaccessmediawithRTSiffi;jhasthesmallestcounteramongthecontendingowsthatitknowsorifitsensesanidlechannelforacertainperiodoftime.AfterRTSissuccessfullydeliveredtothereceiverjthroughaCSMA/CAprotocol,therearetwopossiblecases. 2-1 )withasmallerorequalcounteriscurrentlyinatransmissionburst,jwillrejecti'sRTSwithanewcontrolmessageREJ,carryingthecontendingow'scounterandthenumberofbitstobetransmittedbeforethatcounterwillbeincreasedbyone.NotethatREJshouldbesentafterj'scurrentNAVexpiresinordertoavoidinterferingwithconcurrenttransmissions.BasedontheinformationreceivedinREJ,isetsanappropriatetimerandwillre-attempttransmissionaftertimeout. 27

PAGE 28

TheabovedesignofPPShasthefollowingproperties. Proof:Weexaminethevetypesofcontendingows.Thesendersoffi;k,fa;bandfe;mwillrefrainfromaccessingmediabecausetheyknowthatfi;jhasasmallercounter.Inparticular,alearnstheinformationbyoverhearingRTS/DATAfromi,andelearnstheinformationbyoverhearingCTS/ACKfromj. Next,considerfd;candfh;g.Theirsenders,dandh,mayattempttoaccessmediabecausetheycannotoverhearthecountervalueoffi;jthatispiggybackedintransmissionsmadebyiandj.ByaCSMA/CAprotocol,theywilleventuallysucceedindeliveringRTStotheirreceivers,candg,whoknowthatfi;j'scounterissmaller.ThereceiverswillrejectRTSbyreplyingREJbacktothesenders.Withoutoverhearingatransmissioncarryingasmallercounter,iwillcontinueitstransmissionburstviaCSMA/CA.2

PAGE 29

Nowlet'sexaminetheinterruptedow.Ifitisfa;borfe;m,itssenderwilloverhearthecountervalueoffi;j,piggybackedinthetransmissionsmadebetweeniandj.Oncethesenderndsthatitscounterislarger,itwillnottrytoresumeitstransmissionburst.Next,iftheinterruptedowisfd;corfh;g,itsreceiverwilloverhearthecountervalueoffi;j,butitssenderwillnot.ThesenderwilltrytoresumethetransmissionburstbysendingRTStothereceiver,whichisdorh.UponreceivingRTS,thereceiverwillreplyREJ,stoppingthesenderfromfurthertrying.2 29

PAGE 30

T,wherelisthenumberofbitstransmittedbeforeacounterisincreasedbyoneandTisthePPSperiod.BothlandTaresystemparameters. Moresignicantly,whenweincreasethePPSperiod,theowratesresultedfromPPSindenitelyapproachtowardsweightedmax-minfairness:themeanratemi;jofanyMACowfi;jcannotbeincreasedwithoutdecreasingthemeanratemi0;j0ofanotherMACowfi0;j0,forwhichmi0;j0mi;j. tmri;j(t)<(ci;j(t)+1)l t(2{1) Nowwerstprovethefollowinglemma. t. t.ByProperty3,underPPS,thereisnoextraunusedbandwidthforfi;jtoincreaseitsrate.Inotherwords,increasingmri;j(t)meansthatthemeanrateofantherowmri0;j0(t)hastobedecreased,andbyourassumption,mri0;j0(t)>mri;j(t)+2l t.Then,byinequality 2{1 ,wehaveci0;j0(t)ci;j(t)+2.However,byProperty2,fi;jwillnotletfi0;j0hasacountervaluethatisgreaterthan 30

PAGE 31

t
PAGE 32

2-3 (b)showsadierentscenariowherefh;gtriestopreemptfi;j.Ifgisnotinthetransmissionrangeofi,evenwhenDATAistransmittingfromitoj,nodegisabletoreceiveRTSfromh.However,gcannotreplyACKbecauseitiswaitingonNAV.Tosolvethisproblem,aftertheNAVexpires,gsendsaRRTSpackettosolicitanewRTSfromh. WhileanumberofMACprotocolsintheliterature[ 29 ; 18 ; 19 ; 21 { 23 ]weredesignedtoachievefairnessamongcontendingMACows,tothebestofourknowledge,ourprotocolisthersttoachieveallthreepropertiesdiscussedabove.Incomparison,EMLM-FQ[ 21 ]requireseachnodetokeeptrackofcertainstateinformationofcontendingowsanddoesnotguaranteeatightboundonitsapproximatefairness.Themax-minfairnessachievedin[ 30 ]usesamulti-channelcontentionmodelthatisdierentfromtheCSMA/CAmodelusedinthisstudy.Theworkin[ 19 ]requireseachnodetocomputefairbandwidthsharesforitsownlinksandthenearbycontendinglinks,whichinturnreliesontheknowledgeofneighborhoodtopology.Tocalculatefairshares,anodemustknowthesetofcontendingowsthatarecurrentlybacklogged,fromwhichthelocalcontentioncliquescanbecomputed.Inaddition,itassumesallcliqueshaveequal,xedcapacity. 32

PAGE 33

PPSalsohasotheradvantages:Itismuchsimpler,easytoimplementandanalyze,andreducesradiocollision.Itdoesnotrequiremodifyingthebackoalgorithmoftheexistingchannelaccessprotocols. 36 ],ontopofIEEE802.11DCF.Forcomparison,wealsoimplementedaninuentialpacketschedulingprotocol[ 19 ](referredasAdjConWin)andIEEE802.11eEDCA[ 10 ].AdjConWinachievesfairnessamongMACowsbydynamicallyadjustingtheirminimumcontentionwindows.IEEE802.11eEDCAhasfouraccesscategories:background,besteort,video,andvoice. Ifnotspeciedotherwise,thedefaultsimulationparametersaregivenasfollows:Thetransmissionrateissettobe11MbpsbasedonIEEE802.11b,andeachpacketis1000byteslong.ThePPSperiodis1second.Theparameterlissettobethelengthofvepackets.Besideswhathavebeenstatedabove,otherparameters(suchthatthosefor802.11DCF)usethedefaultvaluessetbyns2accordingtotheprotocolstandards. SinceAdjConWinisdesignedtoachievefairness,inordertomakethesimulationresultscomparable,wesettheweightsofallowsinPPStobeone.WerstperformsimulationsonthesimpletopologyinFig. 2-4 ,whereaisnotinthetransmissionrangeofc,butbis.SupposetwoMACows,fa;bandfc;d,arebothbacklogged.UnderIEEE802.11DCF,itiswellknownthatsevereunfairnesscanhappeninthisscenariobecausefa;bcanhardlyacquirethechannel[ 19 ; 23 ].WerunsimulationsforftysecondsunderDCF,EDCA,AdjConWin,andPPS,respectively.Theows'ratesarestillmeasuredinpacketspersecond.TheresultsareshowninTable 2-1 .Clearly,DCFcausessevereunfairness,withfc;dacquiringmostofthechannelcapacityandfa;breceivinglittle.ForEDCA,weassignfa;btothevideoaccesscategoryandfc;dtothebest-eortcategoryin 33

PAGE 34

TherearesomeimportantdierencesbetweenAdjConWinandPPS,whichareelaboratedinSection 2.6 .Becauseofitsfullylocalizedoperations,PPSisequallyeectiveunderdynamicsettingwherethesetofMACows,aswellastheweightsoftheows,changeovertime.ThisiscriticalforDWA,thetechniqueintroducedinthenextchapterbasedonPPS,becauseaMACowforaserviceclassonawirelesslinkwillbeinactivewhenitsqueueisemptyandbecomeactiveagainwhenitsqueueisbacklogged,whichhappenswhenend-to-endowsjoinordepartfromthenetwork.UnlikePPS,AdjConWinislesseectivewithadynamicsetofMACowsbecauseitrequireseachnodetocentrallycomputethelocalcontentioncliquesandthefairbandwidthsharesforitsownowsaswellasnearbycontendingows,undertheassumptionthateachcliquehasanequal,xedcapacity.Thesendersofcontendingowscanbethreehopsaway.Inadynamicenvironment,theoverheadwillbeveryhighifallnodesconstantlyexchangetheircurrentowinformationinordertoupdatethecorrectvaluesforfairbandwidthshares.Ifsuchupdateisnotdone,thenetworkperformancesuers.Weperformsimulationsontheve-owtopologyinFig. 2-5 ,whereeachdashedellipsecontainsnodescanheareachother'stransmission.Supposeow1'squeueisemptyfromtime10to20,ow2'squeueisemptyfromtime20to30,ow3'squeueisemptyfromtime30to40,ow5'squeueisemptyfromtime40to50,andallqueuesareotherwisebacklogged.Fig. 2-6 and 2-7 showtheowratesunderAdjConWinandPPS,respectively.TheaverageowratesunderPPSaremuchhigherbecauseAdjConWinrequiresclosecoordinationamongcontendingnodesandsuchcoordinationbreaksdownwithadynamicsetofows,whilePPSreliesonfullylocalizedoperations. 34

PAGE 35

2-5 ,whereallowsbeginwithweightoneandow1changesitsweightto2attime15andthento4attime30.TheresultinFig. 2-8 showsthatPPSmaintainsweightedbandwidthallocationunderdynamicweight.AdjConWindoesnotconsiderowweights,letalonedynamicweight(whichrequiresfullylocalizedoperations). 3 ,wewillintroduceanewtechniqueestablishedontopofPPStoprovideservicedierentiationandrateassuranceamongend-to-endowsinmultihopwirelessnetworks. 35

PAGE 36

Flowrates(inpacketspersecond)onthetwo-owtopology Figure2-1. Fivetypesofcontendingowsoffi;j. Figure2-2. Packetlabelpriority Figure2-3. Request-for-RTShelpspreemption 36

PAGE 37

Two-owtopology Figure2-5. Five-owtopology Figure2-6. ProtocolofAdjConWin Figure2-7. ProportionalPacketScheduling Figure2-8. ProportionalPacketSchedulingsupportsdynamicweight 37

PAGE 38

BasedonPPS,theweightedbandwidthallocationtechniqueintroducedintheChapter 2 ,wedesignanoveltechniquecalleddynamicweightadaptionwithoorandceiling(DWA)toprovideservicedierentiationandrateassuranceamongadynamicsetofend-to-endowsinmultihopwirelessnetworks. 7 ]proposedasimpleapproachforservicedierentiation,inwhichdierentclassesofpacketsareassignedwithdierentbackopolicies.Ahnetal.proposedSWAN[ 8 ],whichprovidesservicedierentiationbetweenbest-eortTCPandreal-timeUDP.InSWAN,alocalAIMDratecontrolschemeisutilizedtoshapebest-eorttrac.Forreal-timetrac,probepacketsareusedtoachieveadmissioncontrolandexplicitcongestionnoticationsareexploitedtodynamicallyregulatetracchanges.Intheirotherworks[ 37 ; 6 ],dierentcontentionwindowsizesareappliedtopacketsfromreal-timeandbest-eorttracrespectivelyinordertoachieveservicedierentiation.AvirtualMACalgorithmisintroducedtoestimatethecurrentchannelcondition,andadmissioncontrolismadeaccordingly.However,thesemechanismsarenotdesignedformultihopnetworks.Karenosetal.[ 9 ]designedaratecontrolframeworkinsensornetworks.Noneoftheseworksachievestherateassuranceobjectiveandthebandwidthdierentiationobjectiveasdenedinourstudy. Someresearchhasbeendedicatedtosupportbandwidthguaranteeateitherthelinklayerorthenetworklayer.LinandGerla[ 38 ]proposedareal-timeprotocolcalledMACA/PRtoprovidebandwidthguaranteeoversingle-hoplinks,whichemulatesTDMAinCSMAandrequireseachnodetomaintainareservationtablethatkeepstrackofdynamictimereservationsofreal-timeows.Thisemulationmayleadtoinecienttransmissionschedules.Shahetal.[ 2 ]proposedanadmissioncontrolanddynamic 38

PAGE 39

3 ].Banchsetal.[ 1 ]proposedamechanismthatprovidesrateassurancetowirelessLANsonaperstaionbasis,wheredesiredbandwidthisallocatedbyadaptivelyadjustingcontentionwindowsize.However,thismechanismcannotbeappliedtomultihopwirelessnetworks.SarkarandTassiulas[ 5 ]designedadistributedalgorithmforend-to-endfairnessunderamulti-channelmodelwhereeachnodetransmitsatadierentfrequency.Intheproposedmechanism,sessionratesarecontrolledbyatokenbucketalgorithm,andtokensaregeneratedaccordingtoone-hopneighborsinsuchawaythattheratelimitinbottleneckscanspreadout. ThemajordierencesofourproposedQoSmechanismfromtheabovemechanismsareasfollows.First,ourmechanismworksformultihopend-to-endowsinadhocwirelessnetworks.Second,wecombinetheconceptsofservicedierentiationandrateassurancetogether,whereowsarecategorizedintoarbitrarymultiplepriorities,andowraterequirementsareassuredbasedontheirpriorities.Thirdly,andthemostimportant,weemphasizethefairnessissuewhenimplementingQoSfeatures.Wearguethat,withoutfairnessproblemssolved,servicedierentiationitselfislessmeaningful.Itisnotacceptablethataowinalowerprioritygetsabetterqualityofservicethananotherowinahigherpriorityjustbecausetheirlink-layercontentionsareunfair.ThatiswhywebuildourQoSmechanismbasedonourweightedfairschedulingprotocol. 39

PAGE 40

Thenetworkprovidesanumberofdierentiatedserviceclasses,eachhavingadierentpriority.Best-eortowsareassignedtothebest-eortserviceclassthathasthelowestpriority.QoSowsareassignedtootherclasses.ThepriorityofaQoSowisequaltothepriorityoftheserviceclasstowhichtheowisassigned.WhentheminimumraterequirementofaQoSowissatised,wesaythenetworksupportstheow.Wehavethreeobjectives. 40

PAGE 41

Itshouldbenotedthatthereareotherwaysofdeningtheobjectives.Forexample,insteadofdistributingtheremainingbandwidth(afterrateassurance)amongend-to-endowsbasedonbandwidthdemandanddierentiatingfactor,onemayprefertoopti-mizetheaggregatethroughputofthenetworkundertheconstraintofrateassurance.Optimizingtheaggregatethroughputwillnaturallyprefershortowsoverlongowsanddisregardthepriorities.Proportionalfairness[ 39 ]canbeusedtoaddressthe 41

PAGE 42

40 ; 41 ]doesnotprovideper-owbandwidthassurance,undertheIntServmodel[ 42 ]wecanreserveacertainamountofbandwidthforaowalongitsroutingpathbyusingper-owweightedfairqueueing[ 43 ].Ateachlink(i;j)onthepath,nodeiassignseachpassingowaweightthatisproportionaltotheow'sraterequirement.Theraterequirementsofthepassingowscanbesatisedaslongastheirsumdoesnotexceedthelinkcapacity. Next,wetrytomaptheabovesolutiontoamultihopwirelessnetwork.WebeginbyconsideringonlyoneQoSserviceclass.Whilewestillperformweightedfairqueueingateachlinkandassigneverypassingowaweightthatisproportionaltotheow'sraterequirement,thereisafundamentaldierencebetweenwirelessnetworksandwirednetworks.Awirelesslinkdoesnothaveaxedcapacity.Itsharesacommonchannelwithnearbycontendinglinks.Eventhecapacityofthewirelesschannelmaybedynamicduetointerferenceandmulti-ratelinks.Thekeyforrateassuranceistodistributebandwidthtowirelesslinksinsuchawaythateachlinkreceivesaportionthatisequaltoorabovethetotalraterequirementofallowspassingthelink. Theproblembecomesmuchmorecomplicatediftherearemultipleserviceclassesandthecombinedbandwidthdemandexceedstheavailablebandwidth.Moreover,asend-to-endowscomeandgoandchannelconditionschangeovertime,wemustadaptthebandwidthallocatedtoeachwirelesslink,aswellasthebandwidthallocatedtoeachserviceclassofthelink,inordertocontiguouslysupportrateassuranceandservice 42

PAGE 43

End-to-endservicedierentiationandrateassurancerequireMAC-layersupport.Inparticular,weneedaprotocolthatallocatesbandwidthtoMACowsinproportiontotheirweights.Considertwocontendinglinks,(i;j)and(a;b),eachreceivingafairshareofthechannelcapacityviaaCSMA/CAprotocolsuchasIEEE802.11DCF.Supposetwoend-to-endowsofhighprioritypassthrough(i;j)andoneend-to-endowoflowprioritypassesthrough(a;b).IEEE802.11DCFwilllimittherateofeitherhigh-priorityowtohalfthatofthelow-priorityow. Forthehigh-priorityowstoreceivemorebandwidththanthelow-priorityone,weneedaMACprotocolthatcanexiblyredistributebandwidthamongMACows.Inparticular,weareinterestedinaprotocolthatallocatesbandwidthtoMACowsinproportiontotheirweights.PPSintroducedinChapter 2 servesforthispurposeperfectly.Intheaboveexample,withPPS,ifwewanttherateofahigh-priorityend-to-endowtobetwicethatofalow-priorityow,wesimplyassign4astheweightoftheMACowon(i;j)and1astheweightoftheMACowon(a;b). Therearetwolevelsofbandwidthdistribution.Attherstlevel,weperformweightedbandwidthallocationbyPPSthatisintroducedinChapter 2 todistributethechannelcapacityamongcontendingMACows.Wheneverpossible,eachMACowshouldacquire 43

PAGE 44

43 ]toassureeveryend-to-endow'sraterequirementismet(locallyatthiswirelesslink).Therefore,wewillfocusontherstlevelintherestofthischapter. Whiletheabovetwo-levelbandwidthdistributionarchitecturemayappeartobearoutinedesign,ournoveltyisinsolvingtheproblemofhowtoassignappropriateweightstoMACows(attherstlevel)suchthattheobjectivesinSection 3.2 canbeachievedinadynamic,fully-distributedenvironment,wherebothowsandchannelconditionsmaychangeandthereisnotanentitythathasglobalnetwork/tracinformation. Tosolvethisproblem,weproposeanewtechniquecalleddynamicweightadaptationwithoorandceiling.TheweightofeachMACowadaptsbetweenalowerbound(calledoor)andanupperbound(calledceiling).TheoorisproportionaltotheraterequirementoftheMACow,andtheceilingisproportionaltotheproductoftheraterequirementandthedierentiatingfactor(whichislargerforaMACowofhigherpriority,givingsuchaowahigherceiling).BecausetheraterequirementofaMACowmaychangeasend-to-endowscomeandgo,theoorandtheceilingmaychange,too.EachMACowperiodicallyadaptsitsweightbetweentheoorandtheceilingasfollow:ThesenderoftheMACowmeasuresitsrateovereachweight-adaptationperiod,whichmaybesetthesameasordierentfromthePPSperiod.IfthemeasuredrateisbelowtherequirementoftheMACow,thesenderincreasestheweightoftheowattheendofeachperioduntiltheceilingisreachedortherequirementissatised.Ifthemeasuredrate 44

PAGE 45

T Tocomputeqki;j,weneedtoexaminetwocases.First,ifanend-to-endowcarriedbyfki;jhasabackloggedqueueati,weshouldallocatesucientbandwidthforfki;jtosupporttheminimumraterequirement.Second,iftheend-to-endowdoesnothaveabackloggedqueueandthearrivalratetothequeueissmallerthanitsraterequirement,theowmusthaveanupstreambottleneckandweonlyneedtoallocateenoughbandwidthtocoverthearrivalrate.Theeectiveraterequirementofanend-to-endowisequaltotheminimumraterequirementintherstcaseandthearrivalrateinthesecondcase.Wedeneqki;jasthesummationoftheeectiveraterequirementsofallend-to-endowscarriedbyfki;j. 45

PAGE 46

Weightwki;jisinitializedtobeLki;janditerativelyadjusted.Attheendofeachweight-adaptationperiod,ifrki;j
PAGE 47

First,considertwoend-to-endowscontendforbandwidthatacommonbottleneck.SupposetheypasstwocontendingMACows,fki;jandfk0i0;j0,inthesamebottleneckchannel,andk>k0.IftheratesoftheMACowsarebelowtheminimumrequirements,theywillbothincreaseweightsunlesstheceilingsarereached.Theirceilingsaredierent.Thelargestnormalizedweightforfki;j(whichisdk)islargerthanthelargestnormalizedweightforfk0i0;j0(whichisdk0).Hence,fki;jisabletoincreaseitsweightfurthertoacquiremorebandwidthperunitofraterequirementthanfk0i0;j0.Consequently,iffk0i0;j0issupported,fki;jmustalsobesupported,butiffki;jissupported,fk0i0;j0mayormaynotbesupported(duetotheconstraintoflowerceiling).Therateassuranceobjectiveismet.Now,supposeneitherfki;jnorfk0i0;j0canbesupported.Theirweightswillbothreachtheceilings,andthebandwidthallocationbetweenthemwillbeproportionaltotheproductoftheraterequirementandthedierentiatingfactor.Duetothetwo-levelbandwidthdistributionarchitecture,theratioofbandwidthallocationsamongMACowswillbeinheritedbytheend-to-endowsthattheMACowscarry.Hence,thesecondhalfofthebandwidthdierentiationobjectiveissatised. Next,westudythecasethatthenetworkhasenoughbandwidthtosupporttheminimumraterequirementsofallend-to-endows.ConsidercontendingMACowsatanarbitrarylocationinthenetwork.Werstshowthatallowswillbesupported.Anyowwhoserateislargerthanitsminimumrequirementwilldecreaseitsweight,givingupsomebandwidth.Ifnoowusesmorebandwidththanitsminimumrequirement,certainlyeveryonewillbesupported.Whathappensifaowhasahigherratethantheminimumrequirementevenwhenitsweightisreducedtotheoor?Sincetheow'snormalizedweight,now,isthelowestamongall,otherowsmustbereceivingthesameormorebandwidthforeachunitofraterequirement.Hence,theirraterequirementsmusthavebeensatisedaswell.Now,ifthereisstillextrabandwidth 47

PAGE 48

Finally,becausetheweightofeachMACowhasaceiling(dkqki;j),itcannotindenitelyincreasethefractionofchannelbandwidththatitconsumes.Inotherwords,thebandwidthconsumedbyend-to-endowsinacertainpriorityclassislimitedatanylocationinthenetwork;themaximumfractionofbandwidththeycanconsumeisproportionaltothedierentiatingfactordk,whichiscongurable.Therefore,theno-starvationobjectiveisalsomet.Tomaximallyutilizetheavailablebandwidth,theunderlyingMAC-layerschedulingprotocolthatimplementsweightedbandwidthallocationmustbework-conserving,i.e.,itmustallowMACowstoconsumebandwidthleftunusedbyotherMACows. 44 ],whichallowstheupstreamnodektosendapackettoionlywhenihasenoughfreespaceinthequeuetoholdthepacket.Supposethebuerspaceforthequeueisslottedwitheachslotstoringonepacket.Theresidualbueratnodeichangeswhenireceivesorsendsapacket.Tokeeptheupstreamnodeupdatedwithi'sbuerstate,wheneveritransmitsapacket(RTS/CTS/DATA/ACK),itpiggybacksitscurrentbuerstateintheframeheader,forexample,usingonebittoindicatewhetherthereisatleastonefreebuerslot.Whentheupstreamnodekoverhearsapacketfromi,itcachesthebuerstateofi.Ifi'sbueris 48

PAGE 49

44 ]fordiscussiononvariousissuessuchasfailedoverhearing. First,letanotherhigh-priorityend-to-endowjoininthenetwork.Theowsourcesignalsalongtheroutingpathtoadditsraterequirementtothehigh-priorityMACows.Theceilingsofthehigh-priorityMACowsareincreasedaccordingly.SupposeatonelocationxthechannelcapacityisnotsucienttosupportallQoSows.Withtheirweightsattheoors,theMACowsoftwoprioritiesbothndthattheyarenotgettingenoughbandwidth.Theywillincreasetheirweights.Thelow-priorityowhasalowerceilingandwillstopincreasingrst.Thehigh-priorityowwillbeabletogetmorebandwidthtosatisfytherequirement. Second,letthenewly-joinedend-to-endowdepartfromthenetwork.Atlocationxthebandwidthforthedepartedowwillbeinheritedbytheotherhigh-priorityend-to-endowsharingthesameMACow.Sinceitacquiresmorebandwidththantheraterequirement,thehigh-priorityMACowwillreduceitsweighttotheoor,givingawaybandwidthtothelow-priorityow. Third,supposethechannelcapacityatlocationxisdecreasedduetoenvironmentalnoise,causingtheactualdataratesofbothMACowstodecreasebelowtheraterequirements.Similartotherstscenario,theirweightswilladaptindividuallyandindependently,butbecausethehigh-priorityowhasahigherceiling,itwillreceivemorebandwidthtomeetitsraterequirement. 49

PAGE 50

33 ]inbandwidthdistributionamongend-to-endows.Consideranend-to-endowthatfollowsapathk!i!jwheretheintra-owcontentionbetweensub-owk!iandsub-owi!jmayleadtoonesub-owgrabbingmorebandwidththantheother,whiletheyshouldeachhavethesamebandwidth.Consideranotherend-to-endowthatfollowsthesamepath.Theinter-owcontentionmayassignthesamebandwidthsharetoeachend-to-endow,whiletheyshouldbeassignedsharesbasedontheirraterequirementsandpriorities.Intra-owandinter-owcontentionsareaproblemforrandom-accesswirelessnetworks.However,whenwehaveaMAC-layerschedulingprotocol(PPS,inChapter 2 )thatachievesweightedbandwidthallocationandacongestionavoidancescheme[ 44 ]thatpreventspacketdrops,thesecontentionscanbesolvedbyassigningappropriateweightstoMACows,aswedidinthissection. 36 ],ontopofIEEE802.11DCF. WeshowhowwellDWAcanachieverateassuranceandbandwidthdierentiation.Ifnotspeciedotherwise,thedefaultsimulationparametersaregivenasfollows:Thetransmissionrateissettobe11MbpsbasedonIEEE802.11b,andeachpacketis1000byteslong.ThePPSperiodis2seconds.Theparameterlissettobethelengthofvepackets.Theweight-adaptationperiodis2seconds.TherearetwoQoSserviceclassesforthelowerpriority1andthehigherpriority2.Theirdierentiatingfactorsared1=2andd2=4,respectively.is10%.Besideswhathavebeenstatedabove,otherparameters(suchthatthosefor802.11DCF)usethedefaultvaluessetbyns2accordingtotheprotocolstandards.WealsocompareDWAwithIEEE802.11eEDCA[ 10 ],whichhasfouraccesscategories:background,besteort,video,andvoice. 50

PAGE 51

3-1 withadynamicsetofows,whoseraterequirementsandprioritieschangeovertime.Thenetworkconsistsof30nodesthatarerandomlydeployedinanarea.Thewirelesslinksformedbetweennodesareshowninthegure,wheretheaveragedegreeofanodeis4.4.Thediameterofthenetworkis7hops.Thereare24multihopend-to-endows,whichcannotbeshowninthegure.Thesource/destinationnodesofeachowarerandomlychosen.Weperformsimulationsunderfourdierentsettings.TheresultsareshowninFig. 3-2 3-8 WerstturnoDWA.TheowratesunderIEEE802.11DCFareshowninFig. 3-2 .Flowrateismeasuredinthenumberofpacketssuccessfullytransmittedpersecond(pps).Thecontentionlevelsexperiencedbytherandomly-generatedowsarevastlydierent.Withoutadditionalmechanismstocompensatesuchdierence,theowratesachievedunder802.11arequiteunpredictablewithsomemuchhigherthanothers. WethenturnonDWA.ThesimulationresultisshowninFig. 3-3 .ThenetworkisabletosatisfytheraterequirementsofallQoSows.Flows4,6and16havemuchhigherratesthanothersbecausetheirroutingpathshappentohavelesscontentionwithotherows.TheratesofallQoSowsvaryfromonetoanotheralsobecausetheowsexperiencedierentlevelsofcontentionontheirpaths.Thevariationamongtheratesofbest-eortowsareduetothesamereason. Nextwecreate8newowsineachserviceclass.FornewQoSows,theirraterequirementsarestill10pps.TheresultisshowninFig. 3-4 .Foreasycomparison,wereassignedowidssuchthattheratesofbest-eortowsareshownunderids0-15,theratesofpriority-1owsareshownunder16-31,andtheratesofpriority-2owsareshownunder32-47.Afterdoublingthenumberofows,theraterequirementsofpriority-2ows 51

PAGE 52

Finally,weletthenewowsdepartfromthenetwork,andtheowratesgobacktoFig. 3-3 WhenDWAisturnedon,thesimulationresultisshowninFig. 3-5 .Theraterequirementsofows16-23(priority2)aresatisedornearlysatised.Someofthemreceiveslightlylowerratesduetomoreintensecontentions.Thenetworkcannotsupporttheraterequirementsofows8-15(priority1)evenwhentheirweightsareadaptedtotheceilings.Duetolowerpriority,theirceilingsarelowerthanthoseofows16-23.ButtheyhavehigherrateswithDWAthanwithoutit(Fig. 3-2 ).Thebest-eortowsreceivetheremainingnetworkbandwidth.Bydesign,nobest-eortowisstarved. Forthepurposeofcomparison,weperformthesamesimulationunderIEEE802.11eEDCA,withpriority-2owsassignedtothevideoaccesscategory,priority-1owsassignedtothebest-eortaccesscategory,andprority-0owsassignedtothebackgroundcategory.TheresultsareshowninFig. 3-6 ,whichshowsthatows16-23(priority2)takesnetworkbandwidthaggressively,starvingtheotherows.EDCAgivesxedpreferencetohigher-priorityowsandlacksane-levelratecontrolmechanismthatnotonlyallocatesbandwidthbasedonprioritiesbutalsobalancetheminimumneedsamongallQoSows. ThenexttwosettingsaredesignedtodemonstratethegreatexibilityofDWAincontrollingthebandwidthdistributionamongend-to-endows. 3-7 .Nowtheraterequirementsofows8-15(priority2)aresatised. 52

PAGE 53

3-8 .Theraterequirementsofthenewpriority-2ows,whoseidsare0-7,aresatised. 2 .Thenewtechniqueiseectiveyetsimpletoimplement,whichisimportantforpracticalwirelesssystems.Theperformanceisextensivelyevaluatedbysimulations,demonstratingthenewcapabilitiesthatcanbemadeavailableinfuturewirelessnetworks. 53

PAGE 54

Networktopology Figure3-2. SettingA,IEEE802.11DCF Figure3-3. SettingA,DWA Figure3-4. SettingA,DWA,doublingthenumberofows Figure3-5. SettingB,DWA Figure3-6. SettingB,EDCA 54

PAGE 55

SettingC,DWA Figure3-8. SettingD,DWA 55

PAGE 56

Inthischapter,wedemonstratethatCSMA/CAnetworks,includingIEEE802.11networks,exhibitseverefairnessprobleminmanyscenarios,wheresomehostsobtainmostofthechannel'sbandwidthwhileothersstarve.Mostexistingsolutionsrequirenodestooverheartransmissionsmadebycontendingnodesand,basedontheoverheardinformation,adjustlocalratestoachievefairnessamongallcontendinglinks.Theirunderlyingassumptionisthattransmissionsmadebycontendingnodescanbeoverheard.However,thisassumptionholdsonlywhenthetransmissionrangeisequaltothecarriersensingrange,whichisnottrueinmostrealnetworks.Asourstudyreveals,theoverhearing-basedsolutions,aswellasseveralnon-overhearingAIMDsolutions,cannotachieveMAC-layerfairnessinvarioussettings.Weproposeanewratecontrolprotocol,calledPISD(ProportionalIncreaseSynchronizedmultiplicativeDecrease).Withoutrelyingonoverhearing,itprovidesfairnessinCSMA/CAnetworks,particularlyIEEE802.11networks,byusingonlylocalinformationandperforminglocalizedoperations.Itcombinesseveralnovelratecontrolmechanisms,includingsynchronizedmultiplicativedecrease,proportionalincrease,andbackgroundtransmission. WealsoimprovePISDfurther.PISDworkspreciselyforscenarioswhereallMACowsmutuallycontend,butwhenitisappliedtoanetworkconsistingmultiplecontentiongroups,thenetwork'sthroughputcanbedegraded.WedevelopPISDfurtherandproposetwonewschemes,PISD-RSandPFS,toovercomethelimitation.Weprovethatows'ratesattainedunderthetwonewschemesapproximateproportionalfairness. Therestofthechapterisorganizedasfollows.Section 4.1 discussestherelatedwork.Section 4.2 givesthenetworkmodel.Section 4.3 describesthefairnessproblem.Section 4.4 proposesourPISDsolution.Section 4.5 analyzestheperformanceofPISD.Section 4.6 presentsadditionalsimulationresultsonPISD.Section 4.7 discoversalimitationofPISD.Section 4.8 introducesPISD-RSandPFSassolutionstothe 56

PAGE 57

4.9 provesthatproportionalfairnessisachievedbyPFS(andPISD-RS).Section 4.10 conductssimulationsonPISD-RSandPFS.Section 4.11 summarizesthechapter. 18 ; 19 ],toachievefairbandwidthdistributionamongcontendingwirelesslinksinamultihopwirelessnetwork,everynodeisrequiredtomeasuretheratesofcontendinglinksthroughoverhearingandthenchangeitsownratebyadjustingeithertheminimumcontentionwindoworthecontentionwindowdirectly.In[ 22 ],ChenandZhangalsorelyonoverhearingamongcontendingnodesforappropriatedistributionofchannelcapacityinordertoachieveaggregatefairness. Luoetal.'sapproach[ 29 ]assignseachMACowabasicfairshareofbandwidthandthenmaximizesaggregatechannelutilizationthroughspatialchannelreuse.Thedistributedimplementationrequireseachsendertoknowallcontendingows(throughpiggybackingandoverhearing),andalsorequirestopologyinformationtobepropagatedthroughaconict-freespanningtree.Infollow-upwork[ 21 ]theyproposeMLM-FQ(Maximize-Local-MinimumFairQueueing),whichrequirescontendingnodestotransmitintheorderofpacketservicetags(representingtransmissiondeadlines).Itreliesoneachnodekeepingtrackofservicetagsatothernodesthroughoverhearing.InVaidyaetal.'searlierDFS(DistributedFairSchedulingProtocol)[ 20 ],anodesetsabackotimerbasedonthenishtagofitsnextpackettobetransmitted.DFSdependsonoverhearingtocorrectlyupdatethelocalvirtualclock,basedonwhichthenaltagiscomputed. OML[ 23 ]emulatesTDMAontopofCSMA/CAtoimplementdistributedweightedfairqueueing.Foreachofitspackets,thesendermustinformthecontendingnodesthatitwillparticipateinthetimeslotcompetition.Thisinformationispiggybackedinthepacketheaderandoverheardbyothernodes.(Alternatively,onecanusecontrolmessagestooodthisinformationtocontendingnodesafewhopsaway,which,however,causessignicantoverhead.) 57

PAGE 58

34 ]proposeageneralanalyticalfairnessmodelandaMACprotocoltoapproximateproportionalfairness.Thisworkassumesthatlinksinthesamecontentionregionwillexperiencethesamelossprobability,whichishowevernotalwaystrue.AswedescribeinSection 4.3.1 ,thelossprobabilitiesoftwolinkscanbeverydierent|theactualvaluesdependontherelativespatiallocationsofthelinks.In[ 30 ],TassiulasandSarkaraddressthemax-minfairnessproblemusingamulti-channelMACmodelthatisdierentfromtheCSMA/CAmodelusedinthisstudy.Theirlaterwork[ 45 ]forend-to-endbandwidthguaranteesisalsobasedonthemulti-channelmodel. AIMDhasbeenextensivelystudiedinthepast[ 46 { 48 ; 39 ],mostlyinthecontextofTCP.CrowcroftandOechslin[ 49 ]modifyAIMDtoachieveweightedproportionalfairnessinTCP.Aswedemonstratelater,theAIMDprotocolsdesignedforCSMA/CAnetworksbyCaietal.[ 24 ],byXueetal.[ 26 ]andbyHeusseetal.[ 25 ; 27 ]canonlyprovidefairnessundercertainsituations.AIMDhasalsobeenusedinwirelessnetworksforcongestioncontrol[ 50 ; 51 ]. 58

PAGE 59

AnexampleisshowninFig. 4-1 ,whereeachofthetwo802.11DCFwirelesslinkscarriesaMAC-layerow.ThegureshowsanadhocnetworkortwonearbyWLANswhoseaccesspoints(aandc)eachsupportawirelesshost(bandd).Whenthedistance 59

PAGE 60

35 ],which,however,doesnotconsiderthesituationswherethecarrier-sensingrangeandinterferencerangearegreaterthanthetransmissionrange. Werunsimulationswithns-2v2.32[ 36 ]tostudytheratesofthetwoows.Thesimulationparametersaregivenasfollows:Thetransmissionrangeofthenodesis250m,thecarriersensingrangeis550m,andthelengthsofbothlinksare150m.Thetransmissionrateis11Mbps,andthepacketlengthis1,000bytes.TheparametersfortheIEEE802.11DCFarethedefaultvaluessetbyns-2accordingtotheprotocolstandards. Fig. 4-2 showstheaveragenumbersofpacketspersecondsentoverthetwolinkswithrespecttothedistancebetweennodebandnodec.Whenthedistanceisbelow250m,ow(c;d)obtainsmostofthechannel'sbandwidth.Whenthedistanceisbetween250mand400m,ow(a;b)obtainsmostofthechannel'sbandwidth.Whenthedistanceisbetween400mand550m,ow(c;d)regainstheupperhand.Whenthedistanceisgreater550m,thetwolinksareoutofeachother'scarriersensingrangeandtheywillbothobtainhighbandwidth.TheexplanationonwhysuchunfairnesshappensisgiveninAppendixA. Higher-layerratecontrolsuchasTCPcannotsubstituteforaMAC-layerfairnesssolution.SupposeaTCPconnectionC1traverseslink(a;b)whileanotherconnectionC2passes(c;d).ThesetwoTCPconnectionscompeteforthesameresource|thewirelesschannelsharedby(a;b)and(c;d).Considerthescenariowherethelengthofthewirelesslinksis150mandthedistancebetweenbandcis100m.Thesimulationinns-2showsthatC1isalmoststarvedwhiletherateofC2isaround280packetspersecond.Thereasonisthat(c;d)isfarmorecapableofobtainingthechannelthan(a;b)underCSMA/CA,whichmakespacketsfromC1pronetomoredropsandlargerdelay.For 60

PAGE 61

18 { 23 ],mostrelyingontracinformationthateachnodecollectsbyoverhearingtransmissionsmadebycontendingnodes.ThemostprevalentratecontrolschemeistomodifytherandombackoalgorithmsuchthataMACowthathasasmallerratethanotherswillsetasmallerbackowindowandthusacquiremorebandwidth[ 18 ; 19 ; 22 ].Howdoesaowlearnthatitsrateissmallerthantheratesofitscontendingows?Thecommonapproachrequiresthesenderoftheowtoestimatetheratesofotherowsbyoverhearing.Otherratecontrolschemes,suchasOML[ 23 ],DFS[ 20 ]andEMLM-FQ[ 21 ],alsodependonoverhearing(seeSection 4.1 ). Theproblemisthatoverhearingislimitedwithinthetransmissionrangebutcontentionisdenedbytheinterferencerangeandthecarriersensingrange.Considerawirelesslink(i;j)inFig. 4-3 ,wherethetransmissionrangeofthesenderiisshownbythesolidcircle,thecarriersensingrangeofiisshownbythedottedcircle,andtheinterferencerangeofthereceiverjisshownbythedashedcircle.Whenanynodeinthecarriersensingrangeofimakesatransmission,iwillsenseabusychannelandwithholditsowntransmission.Whenanynodeintheinterferencerangeofthereceiverjmakesatransmission,itwillinterferewiththesignalfromi.Inthe802.11DCF,ifjsensesabusychannelbeforereceivinganRTS,itwillnotreturnCTS.Inthiscase,anynodeinthecarriersensingrangeofjwillinterferewiththecommunicationon(i;j).Clearly,theinterferencerangeisdeterminedbythesignalstrengthatthereceiver,whichisrelatedtothedistancebetweenthesenderiandthereceiverj.Thecarriersensingrangeistypicallysettobenolessthanthemaximuminterferencerange,whichcanbe1.78timesthetransmissionrangeassuggestedin[ 52 ](alsothedefaultvalueusedinns-2). 61

PAGE 62

WeimplementtheHuang-Bensaouprotocol[ 19 ],wherefairnessisachievedbyeachnodeadjustingitscontentionwindowbasedontheoverheardinformationofthecontendingows.ThesimulationresultforthenetworkofFig. 4-1 isshowninFig. 4-4 .TheHuang-Bensaouprotocolachievesalmostperfectfairnesswhenbandcarewithinthetransmissionrangeofeachother(suchthatccanoverhearb'sCTS/ACK).However,whenthedistancebetweenbandcisbeyond250m,theHuang-Bensaouprotocolistotallyineective.Thesameistrueforallotherschemesrelyingonoverhearing. 50 ; 51 ],butthereisverylimitedresearchonapplyingAIMDtoachievefairnessamongMACows.Inthefollowing,wewillshowthatAIMDisill-ttedtothispurpose. ForAIMDtowork,thesenderofaowmustbeabletodetectwhenthechannelissaturated(congested),whichisthetimeformultiplicativedecrease.Thereareanumber 62

PAGE 63

4-1 andassumethatnodeciswithintheinterferencerangeofbbutoutsidethecarriersensingrangeofa.Inthiscase,evenwhenthechannelissaturatedbytransmissionson(c;d),nodeawillsenseanidlechannel. Second,thesendermaytreateveryfailedtransmissionasasignalofchannelsaturationandperformmultiplicativedecrease[ 24 ; 26 ].WesimulatetheAIMDprotocolin[ 24 ](theonein[ 26 ]issimilar)onthenetworkinFig. 4-1 ,andtheresultisshowninFig. 4-5 .TheprotocolworksneinaWLANenvironmentwherethelinksareallfromacommonaccesspointtodierenthosts,whichcorrespondstothedatapointsfordistancebeing150m(suchthataandcoverlaptoserveastheaccesspointwhilebanddarehosts.)However,itperformspoorlyinasymmetricsettingswhenthedistancebetweenbandcisgreaterthanzero. Third,thesendermaymonitoritsbueroccupancy.Eachsendergeneratespacketsfortransmissionatacertainrate,whichiscontrolledbyAIMD.Itsignalscongestedchannelwhenthebuerlengthexceedsathreshold.ThesimulationresultonthenetworkofFig. 4-1 isshowninFig. 4-6 .Again,fairnessisnotachieved. AIMDmayalsobeusedtoindirectlycontroltheowrates.IdleSense[ 25 ; 27 ]replacesDCF'srandombackobyadaptivelysettingthesameoptimalsizeforthecontentionwindowatallhosts.Itwasshownin[ 25 ]that,ifthemeannumberofidleslotsbetweentwotransmissionsinthechanneliscontrolledtoacertaindesirablevalue,e.g.,5.6for802.11b,thecontentionwindowsizewillbenear-optimalfortracthroughputandfairness.ThealgorithmofIdleSenseisforeachhosttomeasurethemeannumber^niofidleslotsbetweentwotransmissionsinthechannelandtograduallyincreaseitscontentionwindowwhen^niisbelowthedesirablevalueormultiplicativelydecreaseitswindowwhen^niisabovethedesirablevalue.IdleSensemakestheassumptionthat,whenAIMD 63

PAGE 64

4-1 ,wherethedistancebetweenbandcis150m.Supposethecontentionwindowsatthesendersaresettothesamesize.Fig. 4-7 showsthattherateofow(c;d)isfargreaterthanthatof(a;b)becausetheformer'sspatiallocationgivesitabetterchancetoobtainthechannelevenwhenitscontentionwindowisthesame. Whenarouterbecomescongested,packetlossisfeltbyallTCPconnectionsthatpasstherouter.Hence,synchronizedmultiplicativedecreasewillbeperformedatthesenders.WeillustratetheratesoftwoTCPconnectionsovertimeinFig. 4-8 .Theratesarenormalizedsuchthatthecongestionhappenswhentheirsumisequalto1.Initially,theratesaredierent.Ateachmultiplicativedecrease,thetworatesarereducedbythesamepercentageandconsequentlythelargerratewillbereducedbyalargeramount,closingthegapbetweenthetwo,whichwilleventuallyconvergetothesamevalue. 64

PAGE 65

4-1 ,thewirelesslinkshavedierentopportunitiestoobtainthewirelessmediumfortransmission,dependingontheirspatiallocations.FromFig. 4-2 ,weknowthat(c;d)ismorecapableofobtainingthemediumthan(a;b)whenthedistancebetweenbandcisshorterthan250m.Attime6inFig. 4-9 ,whenthecombinedrateofow(a;b)andow(c;d)reachesthechannelcapacity,because(c;d)isabletoobtainthebandwidthitneeds,nodecsendsallitspacketsoutbutnodeaobservesbuerbuildupandtransmissionfailure(withamuchlargerlikelihood).Consequently,adetectschannelcongestionandperformsmultiplicativedecrease,whilecdoesnot.Sincetherateof(a;b)experiencesmultiplicativedecreasemorefrequently,itwillbesmallerthantherateof(c;d). Therstproblemishowtodetectchannelcongestion.Foreachow(i;j),thesenderistoresallarrivalpacketsinarepositorybuerabovetheMAClayer.Itlocallymaintainsatime-dependenttargetrateri;j(t)atwhichpacketsfromtherepositorybuerarereleasedtotheMAClayerfortransmissiontothereceiverj.Theowisbackloggedifthepacketarrivalrateisgreaterthanthetargetratesuchthattherepositoryisnotempty.Thetargetrateofabackloggedowisadditivelyincreasedovertime.TheactualrateatwhichtheMAClayersendsoutpacketsiscalledthesendingrate,whichisboundedbythetargetrate. Whenthesumofthetargetratesofallcontendingowsinthechannelissmallerthanthecapacityofthechannel,all(ormost)packetsreleasedbythesenderstotheMAClayercanbetransmitted.ConsequentlythesenderswillnotobservepersistentlygrowingpacketqueuesattheirMAClayer.However,additiveincreasewilleventuallyimprovethetargetratessuchthattheirsumexceedsthechannelcapacity.Whenthishappens, 65

PAGE 66

Thesecondproblemishowthesenderthatdetectschannelcongestioninformsthecontendingnodessuchthattheycanperformsynchronizedmultiplicativedecrease.Onesolutionisforthesenderanditsreceivertojamthechannelwitharadiosignalforanextendedperiodoftime.Beforejamming,thecontendingnodesareabletotransmitatdecentrates(becausethesumofalltargetrateshasjustpassedthechannelcapacityforasmallamountafterthemostrecentadditiveincrease).Duringjamming,theycanhardlysendoutanypackets,whichgivesthemaclearindicationthatsomeoneisjamming,andtheonlyreasonforjammingisthatchannelcongestionhasbeendetected.Astheirqueuelengthsexceedthethreshold,theywilljoinjamming,whichprovidesadditionalassurancethatallcontendingnodesinthechannelwilllearnthatthechanneliscongested.Althoughthejammingapproachworks,itwastesbandwidth.Insteadofusingadedicatedradiosignal,anodecanjamthechannelwithitsownpackets.Duringjamming,toensurethatthenodeisabletooccupythechannel,wereduceitsminimumcongestionwindowtoasmallfractionofthedefaultsize.Besideswindowreduction,thejammingpacketsareexpectedtofollowthesamecollisionavoidance/resolutionprotocol(suchasDCF)asotherpacketsdo.(Thisiswhatwedoinalloursimulations.) TheAISDprotocolissummarizedasfollows.Aftereachunitoftime,thesenderofabackloggedow(i;j)increasesitstargetrateby Atthisrate,thesenderreleasespacketstoaqueue,fromwhichtheMAClayerpicksuppacketsfortransmission.Inonetimeunit,packetsoftotalsizeri;j(t)willbereleased. 66

PAGE 67

Whenthepacketqueueatnodeiforlink(i;j)exceedsathresholdlength,iclaimsthatthechanneliscongestedandjamsthechannelimmediately.Ifthequotaissucientlylarge,itjamsfortherestofthecurrenttimeunit;otherwise,itjamsforonemoretimeunit.Thejammingisperformedbyreleasingallpacketswithinthequotatothequeueandreducingtheminimumcontentionwindowtoasmallvalue.Multiplicativedecreaseisperformedattheendofthetimeunitduringwhichjammingisperformed. Asasafeguard,multiplicativedecreaseshouldnotbeperformedfortwoconsecutivetimeunits.Theprotocoldoesnotrequiretheclocksofthenodestobesynchronized.Ifanodeisthesenderformultipleows,itperformsmediaaccessandrandombackoindependentlyforeachow.Packetsfordierentowsarequeuedseparately.Considerows(a;b)and(a;c).Suppose(a;b)contendswith(i;j)while(a;c)doesnot.When(i;j)istransmitting,nodeaperformsindependentmediaaccessforitstwoows.Forexample,itmaysendanRTStobandthensetthebackotimerfor(a;b)duetoanRTScollisionatb.Whilewaitingonthetimerfor(a;b),itsendsanRTStocandthendeliversapacketon(a;c). WesimulateAISDonthenetworkofFig. 4-1 withthefollowingadditionalparameters:is5kBps,is25%,thetimeunitisonesecond,thequeue-lengththresholdthattriggersjammingis10packets,andtheminimumcontentionwindowforjammingisonetenthofthedefaultsize.Inthesimulation,eachpacketis1kBlong.WendAISDcanrobustlyensuresynchronizedmultiplicativedecrease.Fig. 4-10 showsthat,usingAISD,theratesofthetwoowsareaboutthesameforanydistancebetweenbandc.Fig. 4-11 showsAISD 67

PAGE 68

ThePISDprotocolissimilartoAISDexceptforhowthetargetratesareincreased:Aftereachunitoftime,thesenderofow(i;j)increasesitstargetrateby TherestoftheprotocolisthesameasAISD.WeproveinthenextsectionthatPISDachievesweightedfairness. WeagainsimulatePISDonthenetworkinFig. 4-1 .Weassignwa;b=3andwc;d=1,andtheresultisshowninFig. 4-12 .Weightedfairnessisachieved.Fig. 4-13 showstheratesofthetwoowswithrespecttotimewhenthedistancebetweenbandcis100m.Clearly,ow(a;b)achievesthreetimestherateofow(c;d)becauseitincreasestherateatthreetimesthespeedofthelatter. 4-8 .Similarly,usingPISD,CSMA/CAwillnotfullyutilizethechannelcapacityrightaftermultiplicativedecrease.Itcanbeeasilyshownthat,intheory,theaveragerateofaowissmallerthantheoptimal 68

PAGE 69

Rightbeforeeachmultiplicativedecrease,thesumofthetargetratesatallcontendingnodesexceedsthechannelcapacitybyasmallamount.Afterthetargetratesaremultiplicativelydecreased,theirsumisbelowthechannelcapacitybyafractionofatmost.Foreachow(i;j),thesenderiremembersitstargetraterightbeforethemostrecentmultiplicativedecrease.Thisrateiscalledthebackgroundrate,whichstaysthesameuntilthenextmultiplicativedecrease.Ourbasicideaisthatwewanttoensurethatallsendersareabletotransmitattheirtargetratesand,ifthereisextrachannelbandwidth,weallowthesenderstocompeteforadditionaltransmissionsuptotheirbackgroundrates.Whenanode'ssendingrateisaboveitstargetrate,itstransmissioniscalledabackgroundtransmission.Whenthesendingrateisbelowthetargetrate,itstransmissioniscalledaregulartransmission.Whenaregulartransmissionofonenodecontendswithabackgroundtransmissionofanother,theformershouldbegivenpriority.Toachievesuchdierentiation,weincreasetheminimumcontentionwindowforbackgroundtransmission. ThePISDprotocolwithbackgroundtransmissionisasfollows.Proportionalincreasesynchronizedmultiplicativedecreaseisperformedonthetargetrateasusual.ButasenderireleasespacketstotheMAClayeratthebackgroundrate(theratebeforethelastmultiplicativedecrease).Thenodealsokeepstrackofthenumberntofbytesthatwouldhavebeenreleasedatthetargetrate.Letbethetimethathaselapsedinthecurrent 69

PAGE 70

WesimulatePISDwithbackgroundtransmissiononthenetworkinFig. 4-1 withaminimumcontentionwindowforbackgroundtransmissiontwicethesizeofthedefaultminimumcontentionwindowforregulartransmission.(Notethatthedefaultvalueforregulartransmissionissetbyns-2basedonthestandardof802.11DCF.)Fig. 4-14 showsthattheowspickuptheextrabandwidthleftbyPISDforadditionalpackettransmission.Thisextrabandwidth,representingonlyasmallfractionofchannelcapacity,isnotregulatedbyproportionalincreasemultiplicativedecrease,andconsequentlyitisunevenlydistributedbetweentheowsbasedontheIEEE802.11DCF. 53 ].Onesolutiontothisproblemistochangethedenitionoffairness.Insteadofensuringafairbandwidthshare,weallocateeachowafairshareofchanneloccupationtime.PISDcanbeadaptedtoservethispurpose.Considerthreecontendingwirelesslinkswhosetransmissionratesare11Mbps,5Mbpsand2Mbps,respectively.Ifwelettheweightofthe11Mbpsowbeone,weshallassigntheweightsofothertwoowstobe5 11and2 11,respectively.Whilethetwoowswillsendatlowerrates,theirtransmissionstakeinverselyproportionallylongertime,resultinginthesamechanneloccupationtimeforthethreeows. 70

PAGE 71

46 ].MuchworkaboutAIMDhasbeenperformedinthecontextofTCP[ 47 ; 48 ; 39 ].Inthissection,weanalyzePISDandshowthatitachievesweightedfairnessafterconvergence.Moreimportantly,wederivetheconvergencetime,thechannelcoverage,andtheconvergenceaccuracywithrespecttoand,andrevealtheperformancetradeothatcanbemadebychangingthesetwoparameters. Whenmultiplicativedecreasehappens,evenifthecombinedtargetrateofallowsmaybegreaterthanthechannelcapacity,itwillbegreateronlybyasmallamountduetothenatureofadditiveincrease.Tosimplifytheanalysis,wetreatthemasequal.APISDperiod,denotedasP,isdenedasthetimebetweentwoconsecutivemultiplicative 71

PAGE 72

W First,wedeterminethevalueofri;j(lP).DuringeachPISDperiod,thetargetrateisrstmultiplicativelydecreasedandthenproportionallyincreased.Hence,forl>0,ri;j(lP)=ri;j((l1)P)(1)+wi;jP 72

PAGE 73

WedenetheconvergencetimeasthetimeittakesforAi;j(l)tobe"-closetoitstargetAi;j.The"-closenessisdenedasfollows:jAi;j(l)Ai;jj

PAGE 74

Wlog1"(Cwi;j Astheows'targetratesareproportionallyincreasedinaPISDperiod,itispossiblethat,rightbeforemultiplicativedecrease,thesumofalltargetratesisslightlygreaterthanthechannelcapacity,inwhichcasenotallpacketscanbedelivered.Westudytheimpactofthiscasebelow. 74

PAGE 75

(1 W+ Puttingalloftheaboveanalysistogether,wecanseethatchoosingthevaluesofandisactuallymakingatradeoamongthreesystemproperties:convergencetime,channelcoverage,andconvergenceaccuracy. 4.3.1 andSection 4.4.2 .Theparametersfor802.11DCFusethedefaultvaluessetbyns-2accordingtotheprotocolstandards.Bydefault,=2kBps,and=25%.WewillstudyhowdierentvaluesforandaectPISD'sperformance.Bydefault,backgroundtransmissionisturnedo. ShowninFig. 4-15 ,ourrstsimulationscenarioconsistsoftwoaccesspoints,aandb,locatedattwonearbybuildings.Nodeasendsdatatothreeclienthosts,h1,h2andh3.Nodebalsosendsdatatothreeclienthosts,h4,h5andh6.Theclientsareevenlyspreadaroundtheaccesspoints.Thelengthofeachwirelesslinkis80m,andthedistancebetweenaandbis480m.Fig. 4-16 (a)showstheowratesundertheIEEE802.11DCF.Whentheratecurvesofseveralowsoverlap,wewillexplainwhichowseachcurverepresentsinbothtextandgurecaption.Underthe802.11DCF,theratesofows(a;h1),(a;h2)and(a;h3)arethesame(thelowercurveintheleftplot)becausenodea

PAGE 76

4.3.1 ),nooverhearing-basedsolutionsworkhere.ThesimulationresultfortheHuang-Bensaouprotocol[ 19 ]isshownFig. 4-16 (b),whichiscomparabletowhatthe802.11DCFproduces. PISDisabletoachievefairnessamongallows,asshowninFig. 4-17 .Startingfromdierentinitialrates,allowsconvergetothesamefairrate.Thetotalthroughputis428.1packetspersecond,comparingwith444.6underthe802.11DCFwithorwithouttheHuang-Bensaouprotocol.Next,weturnonbackgroundtransmission,andtheresultisshowninFig. 4-18 .Someowsachievehigheraveragerates,andthetotalthroughputbecomes455.8packetspersecond.Itishigherthanthethroughputunderthe802.11DCFbecauseofreducedradiocollisions,thankstointermittentreleaseofpacketstotheMAClayeratthebackgroundrate.SincetheadditionalrateacquiredthroughbackgroundtransmissionobscurestheratecurveproducedbyPISD,forpresentationclarity,wewillturnitoinothersimulations. WenowstudyhowandaecttheperformanceofPISD.ThesimulationsconrmtheanalyticalresultsinSection 4.5.1 .First,wedoublethevalueofwhilekeepingthesame,andthesimulationresultinFig. 4-19 showsthattheowratesconvergequicker,whencomparingwithFig. 4-17 .Italsoshowsthattheaverageowrateissmaller,indicatingasmallerchannelcoveragebyPISD(aspredictedinSection 4.5.2 ),andthereforemorebandwidthisallocatedforbackgroundtransmission.Second,wedoublethevalueofwhilekeepingthesame,andthesimulationresultinFig. 4-20 showsthattheowratesconvergequicker,whencomparingwithFig. 4-17 .Italsoshowsthatconvergenceaccuracydecreasesduetothesuddendropintheratesofsomeowsrightbeforemultiplicativedecrease,aspredictedinSection 4.5.3 76

PAGE 77

4-21 ,wemodifythetopologybyturningh2andh6intoserversthatuploaddatatoaccesspointsaandb,respectively.Wesettheservers'weightsto3,andthesimulationresultinFig. 4-22 (a)showsthattherateofaserverisabouttwicetherateofaclient.Onemaynoticethespikesintherateadaptationcurvesinthegure.SuchspikesareexpectedaccordingtoouranalysisinSection 4.5.3 .TheyonlyhappeninthelasttimeunitofaPISDperiodwhentheaggregatedtargetrateofallcontendingowsexceedthechannelcapacity.Whennodesthatdetectchannelcongestionjamthechannelwiththeirpackets,thenodethatdetectscongestionlastmaysendatalowrate,causingadownwardspike.SincethespikesmayonlyhappenattheendofaPISDperiod,itsimpactontheaverageowrateislimited,whichisconrmedbytheanalyticalresultinSection 4.5.3 andthesimulationresultintheleftplot,wheretheaverageratesoftheserverowsare124.0and125.1packetspersecondrespectively,andtheaverageratesoftheclientowsare41.8,42.3,42.4and42.7packetspersecondrespectively.Moreover,Section 4.5.3 showsthatdecreasingwillimproveconvergenceaccuracy(i.e.,reducespikes),whichisconrmedbythesimulationresultinFig. 4-22 (b),whereisreducedbytwothirds. Nextweexpandthenetworktohavefouraccesspointsandtenhosts.Theaccesspointsarelocatedatthecornersofa380m380msquare.Thedistancefromthehoststotheiraccesspointsvariesfrom70mto150m.TheirrelativepositionsareshowninFig. 4-23 .TheratesoftheowsunderPISD,PISDwithbackgroundtransmission(PISD-b),DCF,andtheHuang-Bensaouprotocol(H.-B.)areshowninTable 4-1 .PISDisabletoachievefairnesswhiletheDCFandtheHuang-Bensaouprotocolcannotinthisscenario. OursecondsimulationscenarioisanadhocnetworkshowninFig. 4-24 ,wherevisitorstoacommercialconferencedownloadinformationfromexhibitboothstotheirlaptopsviadirectwirelesslinksthatsharethesamechannel.Thesizeoftheareais400mby600m,andthenodesareplottedintheareabasedontheirassignedcoordinates.The 77

PAGE 78

4-2 .EachrowcontainsoneweightassignmentandthecorrespondingowrateachievedbyPISD.TheresultsdemonstratethegreatexibilityandquantitativeprecisionthatPISDisabletobringintoCSMA/CAnetworks. InPISD,toinsuresynchronizedmultiplicativedecrease,aowjamsitsneighboringowswithfulltransmissioncapabilityafteritdetectsasaturatedchannel.Asaresult,theneighboringowsbeingjammedwillagaindetectasaturatedchannelandjamtheirneighbors.Thisprocessindeedleadstoasynchronizedmultiplicativedecreaseinthewholenetwork.Infact,underPISD,thetransmissionratesoftheowsareboundedbythebottleneckofamultihopnetwork,wherewedenethebottleneckofamultihopnetworkasthecontentiongroupwiththelargestnumberofmutuallycontendingows.Wecanobservethisproblemfromanexamplebelow. AsshowninFig. 4-25 ,thenetworkconsistsofsixows.Thesixowsformtwocontentiongroups.IntherstcontentiongroupS1,ow(h1;h2)onlycontendswithow(h3;h4).InthesecondcontentiongroupS2,theveows(h3;h4),(h5;h6),(h7;h8),(h9;h10),(h11;h12),allcontendwitheachother.Weconductsimulationswithns2tostudytheratesoftheowsachievedbyPISD.Thelengthsofthesixlinksare150m,andothersimulationparametersremainthesameasprevioussections. 78

PAGE 79

4-3 showstheaveragenumbersofpacketspersecondsentoverthesixows.WeseethatPISDachievesfairnessamongalltheows.ContentiongroupS2isthebottleneckofthenetwork,inwhichsixowsmutuallycontend.Forthesesixows,theyallattainequalrates,andthesummationoftheirratesis368.9PPS(packetspersecond),whichisclosetothechannelcapacity WehaveseenthattheproblemofPISDisthatthelowratesinthebottleneckofanetworkwillpropagatetootherpartsofthenetwork.Weexpectanewschedulingschemethatsolvesthisproblemandworksproperlyfornetworkswithmultiplecontentiongroups.Inthefollowingsections,weextendPISDintonewschemesthatareabletoapproximateproportionalfairness. 79

PAGE 80

4.9 ,throughtheoreticalanalysisweshowthattheessencebehindPFSisproportionalfairness. 4-25 getlowrateunderPISD.Supposeatacertaintimeunit,contentiongroupS2issaturatedbutcontentiongroupS1isnot.Notlosinggenerality,supposeow(h5;h6)istherstonethatdetectschannelsaturation.BythedesignofPISD,nodeh5willjamthechannelbysendinganumberofpacketswithasmallcontentionwindowsize.Then,ow(h3;h4)willfeelthejammingthroughitsincreasingqueuelength.Notethatthereisnodierencebetweendetectingajamminganddetectingthechanneltobesaturated.Oncethejammingisdetectedbyh3,italsoperformsthesamething|jammingthechannel.Thejammingmadebyh3willeventuallycause(h1;h2)toexecuteamultiplicativedecrease,althoughatthistimeS1isnotreallysaturated. Arstintuitiononsolvingtheaboveproblemistosimplyreducethestrengthofeachjamming.Thisisreasonable,althoughitstillhasproblemsaswewillshowlater.Theideaisthat,whenacontentiongroupSkissaturated,anyow(i;j)inSktriestojamthechannelwithaproperstrengththatisjustenoughtomaketheowsinSkfeelthejamming,butwillnotmakeotherneighborsof(i;j)feelit.Thus,wechangethejammingschemeofPISDtothefollowing: 80

PAGE 81

whereandaresystemparameters,andwi;jistheweightofow(i;j).Thedefaultvalueofanywi;jis1.Itisworthofnotingthat,asasafeguardthatisalsointheoriginalPISD,multiplicativedecreaseshouldnotbeperformedfortwoconsecutivetimeunits. WenametheaboverevisedprotocolasPISD-RS(PISDwithReducedjammingStrength).ItseectivenesscanbeimmediatelyseenfromasimulationagainonthenetworkshowninFig. 4-25 .Inthesimulation,=10kBps(equivalentto10packetspersecond),=0:5,and=0:2.ResultsareshowninTable 4-4 .ComparedwithTable 4-3 ,therateattainedbyow(h1;h2)ismuchincreased,from74PPSto275.5PPS.Thecombinedratesoftheowsineachofthecontentiongroups(S1andS2)arealsoincreased,whichshowsthatthechannelbandwidthisnowbetterutilizedthentheoriginalPISD.Infact,aswewillprovelater,PISD-RSapproximatesproportionalfairness.Asetoftheoreticalvaluesgivenbyproportionalfairnessareshowninthethirdcolumnofthetable,assumingthecapacityofeachcontentiongrouptobe450PPS.Notethatthetheoreticalvaluesarejusttogiveusasenseonproportionalfairness.ForapurelydistributedalgorithmlikePISD-RSinwhichthereismerelynocommunicationorcollaborationsamongthehosts,itisveryhardtostrictlyachievethetheoreticalvalues.AlthoughtheratesattainedbyPISD-RSarelowerthenthetheoreticalvaluesforvariousextentsduetointerferencesinrealprotocolexecution,theystillshowaconsistencytotheproportionalfairness.Asabaseline,intheforthcolumnofthetable,weshowtheratesoftheowsattainedunderordinaryIEEE802.11DCF.Weseethatverysevereunfairnessexists.Infact,(h3;h4)canhardlytransmitanypackets. 81

PAGE 82

Westilluseasimulationtoshowtheproblem.Wetakeawayow(h1;h2)fromthenetworkshowninFig. 4-25 ,andhencealltheremainingowsareinacommoncontentiongroup,throughwhichwecanbetterseetheproblem.ThenewnetworkisshowninFig. 4-26 .Clearly,ifsynchronizedmultiplicativedecreaseisalwaysguaranteed,theveowsshouldgainequalrates.However,thesimulationresultshowsthattheratesoftheows(h3;h4),(h5;h6),(h7;h8),(h9;h10),and(h11;h12)attainedbyPISD-RSare47.3,86.6,84.2,89.3,and90.9packetspersecond,respectively.Theresultindicatesthat(h3;h4)sometimesmakesunsuccessfuljammingsandperformsmultiplicativedecreasesonitsown.Theproblemisresultedfromaninsucientjammingstrength. Canweremedytheproblemsimplybyincreasingthevalueofsothatthejammingstrengthisincreased?Unfortunately,theanswerisno.BecauseunderPISD-RS,duringasynchronizedmultiplicativedecreaseinacontentiongroup,everyowindeedperformsajamming.AssumingthatthecombinedrateoftheowsbeforethemultiplicativedecreaseisapproximatelyequaltothechannelcapacityC,thetotaljammingstrengthisaboutC.Increasingthevalueofdirectlyleadstoalargertotaljammingstrengthandwilllikelypropagatethejammingtoneighboringcontentiongroups,whichisjustthesameproblemthattheoriginalPISDsuersfrom. Ontheotherhand,aswehaveseen,keepingasmallvaluemayresultaninsucientstrengthforasinglejamming.BacktotheexampleofFig. 4-26 ,assumingthatthetargetrateofow(h3;h4)beforeperformingajammingis1 5C,thejammingstrengthis1 25C,if 82

PAGE 83

Thoughtheintuitionisstraightforward,theimplementationisnot.Supposeow(i;j)istherstonethatdetectsasaturatedchannelandwantstoinitiateasynchronizedmultiplicativedecreaseinasaturatedcontentiongroup.(i;j)mustbysomemeansnotifytheotherowsinthecontentiongroup.Forthoseowsbeingnotied,wewantthemtoonlyperformmultiplicativedecreasesbutnotanyadditionaljammings.Tomeetthisrequirement,weintroduceatwo-phasejammingscheme.Intherstphase,therstowthatdetectschannelsaturationandwantstojamthechannelclaimsitselfastheleadertoallitsneighboringows.Inthesecondphase,theleaderjamsthechannel,whileitsneighborstrytodetectthejammingbuttheywillnotjamthechannelagain.Detailedoperationsaredepictedbelow. Intherstphase,ifaow(i;j)detectsthatitsqueuelengthpassesaconstantthreshold1,itstartstoclaimitselfasthejamleaderinthecurrenttimeunit.Toclaimtobetheleader,(i;j)conductsanoisejamming,whichisdierentfromaregularjammingthatwehaveseen.Forregularjammings,nodesjustincreasetheirtargetratesanddecreasethecontentionwindowsize,andallpackettransmissionsstillfollowtheCSMA/CAprotocol.Inanoisejamming,nodeirstsignalsnodejwithanoisejammingbitcarriedinacontrolpacketthatistransmittedwithasmallcontentionwindow.Afterthesignalisreceivedbyj,bothiandjjamsthechannelwithnoiseforacertainperiod 83

PAGE 84

Inthesecondphase,theleaderjamsthechannelwithastrengthofC,bywhichwemeantheleadertemporarilyincreaseitstargetratefromri;j(t)tori;j(t)+C.Still,issetto0.2inoursimulations.Notethat,beingdierentfromPISD-RS,herethejamstrengthistimesthechannelcapacityCinsteadofthetargetrateoftheowitself.Afterjammingthechannel,theleadermakesamultiplicativedecreaseattheendofthetimeunit.FortheMDcandidates,theyusethesamejammingdetectionschemethatisusedinPISD-RS.IfthequeuelengthofaMDcandidatepassesaconstantthreshold2(normally,2issmallerthan1),itmakesamultiplicativedecreaseattheendofthetimeunit.Otherwise,thenodedoesnothing.Forallnodes,theirrolesofbeingleadersorMDcandidatesareclearedatthenexttimeunit. Wenametheaboveredesignedsolutionwithatwo-phasejammingschemePFS(ProportionalFairScheduling).Exceptthejammingscheme,theremainingAIMDrateadaptationpartofPFSisthesameasPISD-RS.WecanunderstandhowPFSworksthroughanexampleagainonthenetworkshowninFig. 4-25 .Supposeatthebeginningoftimeunitt,contentiongroupS2issaturated(thetotalrateoftheveowsinS2isalmostC),whileatthesametimecontentiongroupS1isnotsaturated.Alsosupposeow(h3;h4)rstdetectsthatitsqueuelengthpasses1.Notethat(h3;h4)ismorelikelytodetectchannelsaturationthanotherowsbecauseoftworeasons.First,(h3;h4)istheonlyowinthenetworkthatbelongstobothofthetwocontentiongroups.Second,duetolocationdependantcontentionsandthehiddenterminalproblem,allthefourowstotherightof(h3;h4)havepriorityonaccessingthechannelthan(h3;h4)[ 35 ; 19 ].Bythe 84

PAGE 85

4-5 .Inthesimulation,Tnoise=0.016seconds,=0.2,1=30packets,2=10packets.ComparedwithPISD-RS,therateoftheweakestow:(h3;h4)isincreased.Infact,theows'ratesattainedbyPFSbetterapproximatethetheoreticalvaluesofproportionalfairnessshowninTable 4-4 Next,weshowthattheproblemofPISD-RSisavoidedbyPFS.WerunasimulationonthenetworkofFig. 4-26 underPFS.Theresultedratesoftheows(h3;h4),(h5;h6),(h7;h8),(h9;h10),and(h11;h12)are72.4,74.8,87.5,88.3,and75.0packetspersecond,respectively.UnlikePISD-RS,alltheveowsgainsimilarratesnowunderPFS. WethentesttheweightedfairnessachievedbyPFS.Wechangetheweightofow(h3;h4)to2,andtheweightsofallotherowsstillremainthedefaultvalue:1.Simulationshowsthattheratesofthesixowsare122.8,61.8,61.7,61.4,and73.7packetspersecond.WeseethattheirratesattainedunderPFSareinproportiontotheirweights. WestressthatPFSissuitableformultihopwirelessnetworksormultipleWLANswithmultiplecontentiongroups,andtheratesattainedbyPFSapproximateproportionalfairness.Inthefollowingsections,weprovetheabovefactsthroughboththeoreticalanalysisandadditionalsimulations. 85

PAGE 86

54 ]. Weformulatetheproblemasfollows.AnetworkismodeledasGN=(VN;EN),whereVNrepresentsthesetofallnodesinthenetwork,andENrepresentsthesetofallsingle-hopMACows.ThecontentionrelationsamongtheowscanbesummarizedusingacontentiongraphGC=(VC;EC)basedonthenetworkgraphGN.Hereeveryowen2ENhasacorrespondingvertexvi2VC,hencejENj=jVCj.AnypairofverticesinGChasanedgeconnectingthemifandonlyifthecorrespondingtwoowsinGNcontendwitheachother.Asubgraphiscalledacliqueifeverypairofverticesinthesubgraphhasanedgebetweenthem.Amaximalcliqueisacliquethatisnotcontainedinanyotherclique.Forsimplicity,intherestofthissectionweusethetermcliquetodenotemaximalclique.WeconsiderthecontentiongraphGCandassumethattheresourceisdesignatedtoeveryclique. Wecanuseaconvexoptimizationmodeltodescribethenetwork.Theconstraintsaredeterminedbythenetwork'stopology,whiledierentfairnessobjectivescanbeachievedbydierentutilityfunctions.Accordingtotheoptimizationproblem,wehavearateupdatefunctionusingthegradientdecentmethod.Bycomparingtheupdatefunction 86

PAGE 87

SupposeaconcavefunctionUi(xi)istheutilityfunction,wherexi0denotestherateofaow.Theoptimizationproblemcanbeformalizedasfollows:maxxiXUi(xi) (4{4)subjecttoXi2KjxiC 4{5 )representsthejthclique.Theseconstraintsmeanthat,foreachclique,thesummationrateofalltheowsinthatcliquecannotexceedthecapacityC.Duetothecontinuousandtheconcavepropertiesoftheobjectivefunctionandthecompactnessoftheconstraints,anoptimalsolutionexists.Lettingx(t)tobetheprimalvariableforeachowandp(t)tobethedualvariableforeachclique,wehavetheLagrangiandual:L(x;p)=XUi(xi)+Xjpj(CXi2Kjxi): Sincetheobjectivefunctionisconcaveandtheconstrainsarelinear,strongdualityholds.Accordingtothestrongdualitytheory,wehavethefollowing:x=argmaxx0L(x;p) (4{8) Herexistheoptimalsolutionfortheprimalproblemandpistheoptimalsolutionforthedualproblem.Notethat,herethedualvariablepcanbeinterpretedastheshadowpricesoftheconstraints.Eachpjindicatesthetensionofeachclique.Whentheclique'scapacityhasbeenalmostusedout,thepricepjishigh;ontheopposite,whentheclique's 87

PAGE 88

Heresi(xi)isastepsize.Itcanbeaconstantorafunctionofxi.Usually,thisfunctionisalinearfunctionofxi. IntheAIMDalgorithmusedinPFS(andPISD-RS)wehavexi(t+1)=8><>:xi(t)+wiifPi2Kj
PAGE 89

(4{14) Equations( 4{13 )and( 4{14 )areforthepositiveandnegativepartsrespectively.TosatisfyEquation( 4{14 ),weshouldlet then or Hereisaconstantfactor.Theequalityofthenegativepartsshowsthat,whentheconstraintsforsomecliquesbecometighterandthedualpricesbecomehigher,theowsinthesecliquesperformmultiplicativedecreasesmorefrequently. ByapplyingEquation( 4{16 )intoEquation( 4{13 ),weobtainthat: whereDisaconstant.Sincetheutilityfunctionisalogarithmfunction[ 39 ],proportionalfairnessisachieved. 4.7 and 4.8 ,buttheyareinrathersmallnetworks.Inthissection,weconstructarelativelylargenetworkandperformsimulationsonit.Werevisedns2(v2.32)[ 36 ]toimplementPISD-RS,andPFSontopoftheIEEE802.11bDCF.TheparametersforDCFusethedefaultvaluessetbyns-2accordingtotheprotocolstandards.Thetransmissionrangeis250meters,andthecarrier 89

PAGE 90

ThenetworkthatwebuiltisshowninFig. 4-27 .Thereisanodeateachofthecrossingpointsofa9by9grid.Inthegrid,thedistancebetweenanytwoadjacent(horizontalorvertical)nodesis200meters.Thenetworkconsistsof23MACowsasshowninthegure.Contentionisdeterminedbythecarriersensingrange.Forexamples,owf15contendswithf9,f10,f13,f14,f16,f17,f18,andf19. WerunIEEE802.11DCF,PISD,PISD-RS,andPFSonebyoneonthenetwork,andthesimulationresultsareshowninFig. 4-28 .First,wesee802.11DCFhasseverefairnessproblem.Someowsgainextremelyhighrates,whilesomeowsarestarved.Inthe23ows,f8,f15,f16,f17,f20,andf21,areindeedgettingazero(ornearlyzero)throughput,whichshowsthat802.11DCFistotallyunacceptable.WethenruntheoriginalPISDprotocol,underwhichallowsgotsimilarratesaround60packetspersecond.ThisresultisconsistenttoourstudyontheproblemofPISDinSection 4.7 .BecauseunderPISDmultiplicativedecreasesaresynchronizedthroughoutthewholenetwork,manyowscannotfullyutilizetheirlocalchannelcapacity.TheresultsgivenbyPISD-RSandPFSdonothavethisproblem,wheresomeowsgainmuchhigherratesthanthoseinPISD,resultingamuchbetterchannelcapacityutilization.Basically,bothPISD-RSandPFSworkwellinthisnetworktopology.PFSslightlyoutperformsPISD-RS,sincesomeveryweakows,suchasf8,f11,f15,andf20,gainhigherratesunderPFS.Thereasonisthat,underPISD-RSowswithweakchannelaccessabilityarelikelytofailinchanneljammings. Finally,wewanttotakeabriefdiscussionontheissueoftotalthroughput.Thetotalrateofallthe23owsin802.11DCFishigherthanthoseinPISD-RSandPFS.Therearetworeasons.First,fairnessisindeedacontraryobjectivetototal 90

PAGE 91

91

PAGE 92

Flowratesachievedbydierentprotocols. owf1f2f3f4f5f6f7f8f9f10 PISD42.743.243.743.743.442.843.743.743.743.6PISD-b43.045.648.143.048.846.245.943.446.146.0DCF74.755.199.751.451.415.315.315.340.840.8H.-B.18.617.236.835.535.547.547.547.584.084.0 Table4-2. Underdierentweightassignments,owratesarealwaysproportionaltoowweights. owf1f2f3f4f5f6f7f8 weight11111111rate53.351.953.253.353.153.453.353.0 weight12111121rate43.482.141.643.441.741.782.543.4 weight12112121rate38.975.238.638.977.238.977.3138.7 weight11121211rate43.641.242.986.743.486.643.143.3 weight21311112rate72.435.4105.535.836.236.336.072.3 weight12141411rate28.455.530.9112.830.9112.630.827.8 Table4-3. FlowratesachievedbyPISD. FlowRate (h1;h2)74.0(h3;h4)73.9(h5;h6)73.5(h7;h8)73.6(h9;h10)73.9(h11;h12)74.0 92

PAGE 93

FlowratesachievedbyPISD-RS,comparedwithproportionalfairnessandbaseline802.11DCF. FlowPISD-RSP-Fair802.11 (h1;h2)275.5375.0436.7(h3;h4)32.975.01.2(h5;h6)81.793.75114.5(h7;h8)98.093.75118.1(h9;h10)96.593.75112.3(h11;h12)98.793.75115.8 Table4-5. FlowratesachievedbyPFS. FlowRate (h1;h2)223.5(h3;h4)49.7(h5;h6)90.3(h7;h8)92.0(h9;h10)86.2(h11;h12)92.3 93

PAGE 94

Networkoftwoows,(a;b)and(c;d). Figure4-2. Ratesoftwoowswithrespecttothedistancebetweenbandc.Thedistancefromatobandthatfromctodareboth150m. Figure4-3. Manycontendingnodescannotbeoverheardbyi. 94

PAGE 95

Ingeneral,Huang-Bensaouprotocoldoesnotworkifanyoneofthecontendinglinkscannotbeoverheard. Figure4-5. AdditiveIncreaseMultiplicativeDecrease:multiplicativedecreaseoccursupontransmissionfailure. Figure4-6. AdditiveIncreaseMultiplicativeDecrease:multiplicativedecreaseoccurswhenbueroccupancypassesacertainthreshold. Figure4-7. Idlesense:Thesamecontentionwindowsizedoesnotensurefairness. 95

PAGE 96

SynchronizedmultiplicativedecreaseinTCPachievesfairness. Figure4-9. UnsynchronizedmultiplicativedecreaseinCSMA/CAcannotachievefairness. Figure4-10. SynchronizedmultiplicativedecreaseequalizestheowratesforthenetworkinFig. 4-1 Figure4-11. RatesoftwocontendingowsunderAISDwithrespecttotime. 96

PAGE 97

Givenwa;b=3andwc;d=1,underPISD,therateofow(a;b)isaboutthreetimesthatofow(c;d). Figure4-13. RatesoftwocontendingowsunderPISDwithrespecttotime.wa;b=3,wc;d=1,andthedistancefrombtocis100m. Figure4-14. Backgroundtransmissionwillutilizesomeunusedchannelbandwidthforpackettransmission. 97

PAGE 98

Networktopology (a) (b) Figure4-16. (a):FlowratesunderIEEE802.11DCF.Thelowercurveshowstheratesofows(a;h1),(a;h2)and(a;h3),whicharethesame.Theuppercurveshowstheratesofows(b;h4),(b;h5)and(b;h6),whicharethesame;(b):FlowratesunderHuang-Bensaouprotocol.Similarly,thelowercurveshowstheratesofows(a;h1),(a;h2)and(a;h3).Theuppercurveshowstheratesoftheotherthreeows. 98

PAGE 99

ProportionalIncreaseSynchronizedMultiplicativeDecreaseachievesfairnessamongallows. Figure4-18. Flowratesareslightlyimprovedwithbackgroundtransmission.. Figure4-19. Increasingthevalueofreducesbothconvergencetimeandchannelcoverage.. Figure4-20. Increasingthevalueofreducesbothconvergencetimeandconvergenceaccuracy. Figure4-21. Hostsh2andh6arechangedtoservers. 99

PAGE 100

(b) Figure4-22. (a)Whentheserverseachhaveweight3andtheclientseachhaveweight1,therateofaserveristhreetimesthatofaclient;(b)Downwardspikesarereducedwhenisdecreased. Figure4-23. FourWLANsnetworktopology Figure4-24. Adhocnetworktopology 100

PAGE 101

Multihopnetworkwithsixows. Figure4-26. Multihopnetworkwithveows. Figure4-27. Multihopwirelessnetworkwith23MACows. 101

PAGE 102

Flowratesattainedunderdierentapproaches 102

PAGE 103

Twoimportantproblemsinwirelessnetworksarestudied.Theyareend-to-endservicedierentiationandrateassuranceformultihopows,andfairnessforsingle-hopMAC-layerlinks. Wepresentanewadaptiveratecontrolfunctionbasedontwonovelprotocols,dynamicweightadaptation(DWA)andproportionalpacketscheduling(PPS),whichtogetherenableprioritizedrateassuranceandsophisticatedbandwidthdierentiationamongend-to-endowsinmultihopwirelessnetworks.Thenewadaptivefunctionrepresentsaparadigmchangeinne-levelbandwidthmanagementwithoutresourcereservationandadmissioncontrol.Threeobjectivesareachieved:rateassuranceobjective,bandwidthdierentiationobjective,andno-starvationobjective. WedemonstratethatCSMA/CAnetworks,includingIEEE802.11networks,haveseverefairnessproblemintheirMAClayer.Mostexistingsolutionsrequirenodestooverheartransmissionsmadebytheirneighbors,andtheeectivenessofthesesolutionsmaybeverylimitedwhenthecarriersensingrangeandtheinterferencerangearemuchlargerthanthetransmissionrange.Weproposeanewratecontrolprotocol,PISD,whichisabletoprovidefairnesswithonlylocalizedoperationsandwithoutrelyingonoverhearing.Moreover,asenhancementstoPISD,PISD-RSandPFSareproposedtoachieveproportionalfairnessinnetworksconsistingofmultiplecontentiongroups. 103

PAGE 104

[1] A.BanchsandX.Perez,\ProvidingThroughputGuaranteesinIEEE802.11WirelessLANs,"Proc.IEEEWCNC,2002. [2] S.Shah,K.Chen,andK.Nahrstedt,\DynamicBandwidthManagementinSingle-HopAdHocWirelessNetworks,"Proc.IEEEPERCOM,2003. [3] C.-T.ChouandK.G.Shin,\AnalysisofAdaptiveBandwidthAllocationinWirelessNetworkswithMultilevelDegradableQualityofService,"IEEETransactionsonMobileCompuping,vol.3,no.1,2004. [4] P.Gupta,Y.Sankarasubramaniam,andA.L.Stolyar,\Random-AccessSchedulingwithServiceDierentiationinWirelessNetworks,"Proc.IEEEINFOCOM,2005. [5] S.SarkarandL.Tassiulas,\End-to-endBandwidthGuaranteesThroughFairLocalSpectrumShareinWirelessAd-hocNetworks,"Proc.IEEEConferenceonDecisionandControl,2003. [6] A.Veres,A.Campbell,M.Barry,andL.Sun,\SupportingServiceDierentiationinWirelessPacketNetworksUsingDistributedControl,"IEEEJSAC,Vol.19,No.10,2001. [7] S.KangandM.Mutka,\ProvisioningServiceDierentiationinAdHocNetworksbyTheModicationofBackoAlgorithm,"Proc.IEEEICCCN,2001. [8] G.Ahn,A.Campbell,A.Veres,andL.Sun,\SWAN:ServiceDierentiationinStatelessWirelessAdHocNetworks,"Proc.IEEEINFOCOM,2002. [9] K.Karenos,V.Kalogeraki,andS.V.Krishnamurthy,\ARateControlFrameworkforSupportingMultipleClassesofTracinSensorNetworks,"Proc.IEEEReal-TimeSystemsSymposium,2005. [10] \IEEEStandardforTelecommunicationsandInformationExchangeBetweenSystems-LAN/MANSpecicRequirements-Part11:WirelessLANMediumAccessControl(MAC)andPhysicalLayer(PHY)Specications:MediumAccessControl(MAC)QualityofServiceEnhancements.IEEEStd802.11e,"Nov2005. [11] Y.YangandR.Kravets,\Contention-AwareAdmissionControlforAdHocNetworks,"IEEETransactionsonMobileComputing,vol.4,no.4,2005. [12] M.Kazantzidis,M.Gerla,andS.-J.Lee,\PermissibleThroughputNetworkFeedbackforAdaptiveMultimediainAODVMANETs,"Proc.IEEEICC,2001. [13] G.Bianchi,\PerformanceAnalysisoftheIEEE802.11DistributedCoordinationFunction,"IEEEJSAC,vol.18,no.3,2000. 104

PAGE 105

F.Cali,M.Conti,andE.Gregori,\DynamicTuningoftheIEEE802.11ProtocoltoAchieveaTheoreticalThroughputLimit,"IEEE/ACMTransactionsonNetworking,vol.8,no.6,2000. [15] B.LiandR.Battiti,\PerformanceAnalysisofAnEnhancedIEEE802.11DistributedCoordinationFunctionSupportingServiceDierentiation,"Int'lWork-shoponQualityofFutureInternetService,2003. [16] Y.YangandR.Kravets,\ThroughputGuaranteesforMulti-priorityTracinAdHocNetworks,"ElsevierAdHocNetworksJournal,vol.5,no.2,2007. [17] Y.Jian,S.Chen,L.Zhang,andY.Fang,\NewAdaptiveProtocolsforFine-LevelEnd-to-EndRateControlinWirelessNetworks,"Proc.IEEEICNP,2008. [18] B.Bensaou,Y.Wang,andC.Ko,\FairMediumAccessin802.11BasedWirelessAdHocNetworks,"Proc.ACMMOBIHOC,2000. [19] X.HuangandB.Bensaou,\OnMax-MinFairnessandSchedulinginWirelessAdHocNetworks:AnalyticalFrameworkandImplementation,"Proc.ACMMOBIHOC,2001. [20] N.H.Vaidya,P.Bahl,andS.Gupta,\DistributedFairSchedulinginaWirelessLAN,"Proc.ACMMOBICOM,2000. [21] H.Luo,J.Cheng,andS.Lu,\Self-CoordinatingLocalizedFairQueueinginWirelessAdHocNetworks,"IEEETransactionsonMobileCompuping,vol.3,no.1,2004. [22] S.ChenandZ.Zhang,\LocalizedAlgorithmforAggregateFairnessinWirelessSensorNetworks,"Proc.ACMMobicom,2006. [23] A.RaoandI.Stoica,\AnOverlayMacLayerfor802.11Networks,"Proc.ACMMOBISYS,2005. [24] S.Cai,Y.Liu,andW.Gong,\AnalysisofanAIMDbasedCollisionAvoidanceProtocolinWirelessDataNetworks,"Proc.IEEECDC,2003. [25] M.Heusse,F.Rousseau,R.Guillier,andA.Duda,\IdleSense:AnOptimalAccessMethodforHighThroughputandFairnessinRateDiverseWirelessLANs,"Proc.ACMSIGCOMM,2005. [26] Q.Xue,W.Gong,andA.Ganz,\ProportionalServiceDierentiationinWirelessLANsUsingSpacing-basedChannelOccupancyRegulation,"MobileNetworksandApplications,vol.11,no.2,2006. [27] Y.Grunenberger,M.Heusse,F.Rousseau,andA.Duda,\ExperiencewithanImplementationoftheIdleSenseWirelessAccessMethod,"Proc.CoNEXT,2007. [28] Y.JianandS.Chen,\CanCSMA/CAnetworksbemadefair?"Proc.ACMMOBI-COM,2008. 105

PAGE 106

H.Luo,S.Lu,andV.Bhurghawn.,\ANewModelforPacketSchedulinginMultihopWirelessNetworks,"Proc.ACMMOBICOM,2000. [30] L.TassiulasandS.Sarkar,\MaxminFairSchedulinginWirelessNetworks,"Proc.IEEEINFOCOM,2002. [31] H.Luo,P.Medvedev,J.Cheng,andS.Lu.,\ASelf-CoordinatingApproachtoDistributedFairQueueinginAdHocWirelessNetworks,"Proc.IEEEINFOCOM,2001. [32] Y.WangandB.Bensaou,\AchievingFairnessinIEEE802.11DFWMACwithVariablePacketLengths,"Proc.IEEEGLOBECOM,2001. [33] B.Li,\End-to-EndFairBandwidthAllocationinMulti-HopWirelessAdHocNetworks,"Proc.IEEEICDCS,2005. [34] T.Nandagopal,T.Kim,X.Gao,andV.Bharghavan,\AchievingMACLayerFairnessinWirelessPacketNetworks,"Proc.ACMMOBICOM,2000. [35] V.Bharghavan,A.Demers,S.Shenker,andL.Zhang,\MACAW:AMediaAccessProtocolforWirelessLANs,"Proc.ACMSIGCOMM,1994. [36] TheNetworkSimulator-ns-2, http://www.isi.edu/nsnam/ns/ [37] M.Barry,A.Campbell,andA.Veres,\DistributedControlAlgorithmsforServiceDierentiationinWirelessPacketNetworks,"Proc.IEEEINFOCOM,2001. [38] C.R.LinandM.Gerla,\MACA/PR:AnAsynchronousMultimediaMulti-HopWirelessNetwork,"Proc.IEEEINFOCOM,1997. [39] F.Kelly,A.Maulloo,andD.Tan,\RateControlinCommunicationNetworks:ShadowPrices,ProportionalFairnessandStability,"JournaloftheOperationalResearch,vol.49,1998. [40] V.Jacobson,K.Nichols,andK.Paduri,\AnExpeditedForwardingPHB,"IETFNetworkWorkingGroupRFC2598,June1999. [41] J.Heinanen,F.Baker,W.Weiss,andJ.Wroclawski,\AssuredForwardingPHBGroup,"IETFNetworkWorkingGroupRFC2598,June1999. [42] L.Zhang,S.Deering,D.Estrin,S.Shenker,andD.Zappala,\RSVP:ANewResourceReSerVationProtocol,"IEEENetwork,Sep1993. [43] H.Zhang,\ServiceDisciplinesforGuaranteedPerformanceServiceInpacket-SwitchingNetworks,"Proc.IEEE,vol.83,no.10,1995. [44] S.ChenandN.Yang,\CongestionAvoidancebasedonLight-WeightBuerManagementinSensorNetworks,"IEEETransactionsonParallelandDistributedSystems,SpecialIssueonLocalizedCommunicationandTopologyProtocolsforAdHocNetworks,vol.17,no.9,Sep2006. 106

PAGE 107

S.SarkarandL.Tassiulas,\End-to-endBandwidthGuaranteesThroughFairLocalSpectrumShareinWirelessAdhocNetworks,"IEEETransactionsonAutomaticControl,vol.50,no.9,September2005. [46] D.M.ChiuandR.Jain,\AnalysisoftheIncreaseandDecreaseAlgorithmsforCongestionAvoidanceinComputerNetworks,"ComputerNetworksandISDNSystems,vol.17,pp.1{14,1989. [47] Y.R.YangandS.S.Lam,\GeneralAIMDCongestionControl,"Proc.IEEEICNP,2000. [48] S..Jin,L.Guo,I..Matta,andA.Bestavros,\TCP-friendlySIMDCongestionControlandItsConvergenceBehavior,"Proc.IEEEICNP,2001. [49] J.CrowcroftandP.Oechslin,\DierentiatedEnd-to-EndInternetServicesUsingaWeightedProportionalFairSharingTCP,"ACMSIGCOMMComputerCommunica-tionReview,vol.28,no.3,pp.53{69,1998. [50] C.-Y.Wan,S.B.Eisenman,andA.T.Campbell,\CODA:CongestionDetectionandAvoidanceinSensorNetworks,"Proc.ACMSenSys'03,November2003. [51] T.He,J.A.Stankovic,C.Lu,andT.F.Abdelzaher,\SPEED:AStatelessProtocolforReal-TimeCommunicationinSensorNetworks,"Proc.ICDCS,May2003. [52] K.Xu,M.Gerla,andS.Bae,\HoweectiveistheIEEE802.11RTS/CTShandshakeinadhocnetworks?"Proc.IEEEGLOBECOM,2002. [53] M.Heusse,F.Rousseau,G.Berger-Sabbatel,andA.Duda,\PerformanceAnomalyof802.11b,"Proc.IEEEINFOCOM,2003. [54] S.Low,\ADualityModelofTCPandQueueManagementAlgorithms,"IEEE/ACMTransactionsonNetworking,vol.11,no.4,pp.525{536,August2003. 107

PAGE 108

YingJianwasborninBeijing,China,in1979.HereceivedhisB.E.andM.E.degreesincomputerscienceandtechnologyfromTsinghuaUniversity,China,in2001and2004,respectively.In2004,hejoinedtheDepartmentofComputerandInformationScienceandEngineeringattheUniversityofFlorida,topursuehisPh.D.degree.HisadvisorisDr.ShigangChen.Hisresearchfocusedonqualityofserviceinwirelessnetworksandnetworksecurity. 108