Blockswitched Networks A New Paradigm for Wireless Transport Ming Li Devesh Agrawal Deepak - PDF document

vertsintoopportunities.First,end-to-endblockretrans-missionsareconsiderablymoreexpensivethanpacketretransmissions.Hopensuresend-to-endreliabilitythroughanovelretransmissionschemecalledvirtualre-transmissions.Hoprouterscachelargein-transitblocks.Uponanend-to-endtimeouttriggeredbyanoutstand-ingblock,aHopsendersendsatokencorrespondingtotheblockalongportionsoftheroutewheretheblockisalreadycached,andonlyphysicallyretransmitsblocksalongnon-overlappingportionsoftheroutewhereitisnotcached.Second,largeblocksastheunitoftransmis-sionexacerbateshiddenterminalsituations.Hopusesanovelackwithholdingmechanismthatsequencesblocktransferacrossmultiplesenderstransmittingtoasinglereceiver.Thislightweightschemereducescollisionsinhiddenterminalscenarioswhileincurringnoadditionalcontroloverhead.Insummary,ourmaincontributionistoshowthatreliableper-hopblocktransferisfundamentallybetterthanthetraditionalend-to-endpacketstreamabstractionthroughthedesign,implementation,andevaluationofHop.TheindividualcomponentsofHop'sdesignaresimpleandperhapsrightoutofanundergraduatenet-workingtextbook,buttheyprovidedramaticimprove-mentsincombination.Incomparisontothebestvariantof1)TCP,2)Hop-by-hopTCP,and3)DTN2.5,adelaytoleranttransportprotocol[8], I Hopachievesamediangoodputbenetof1.6and2.3oversingle-andmulti-hoppathsrespectively.Thecorrespondinglowerquartilegainsare28and2.7showingthatHopdegradesgracefully. I Underhighload,Hopachievesoveranorderofmagnitudebenetinmediangoodput(e.g.,90overTCPwith30concurrentlargeows),whileachievingcomparableorbetteraggregategoodputandtransferdelayforlargeaswellassmallles. I Hopisrobusttopartitions,andmaintainsitsperfor-mancegainsinwell-connectedWLANsandmeshnetworksaswellasdisruption-pronenetworks.Hopalsoco-existswellwithdelay-sensitiveVoIPtrafc. 2 Whyreliableper-hopblocktransfer?Inthissection,wegivesomeelementaryargumentsforwhyreliableper-hopblocktransferwithhop-by-hopowcontrolisbetterthanTCP'send-to-endpacketstreamwithend-to-endratecontrolinwirelessnetworks.Blockvs.packet:Amajorsourceofinefciencyistransportlayerper-packetoverheadfortimeouts,ac-knowledgementsandretransmissions.Theseoverheadsarelowinnetworkswithlowcontentionandlossbutin-creasesignicantlyaswirelesscontentionandlossratesincrease.Transferringdatainblocksasopposedtopack-etsprovidestwokeybenets.First,itamortizestheover-headofeachcontrolpacketoverlargernumberofdata packets.Thisallowsustouseadditionalcontrolpackets,forexample,toexploitin-networkcaching,whichwouldbeprohibitivelyexpensiveatthegranularityofapacket.Second,itenablestransporttoleveragelink-layertech-niquessuchas802.11bursttransfercapability[1],whosebenetsincreasewithlargeblocks.Transportvs.link-layerreliability:Wirelesschan-nelscanbelossywithextremelyhighrawchannellossratesinhighinterferenceconditions.Insuchnetworks,theend-to-enddeliveryratedecreasesexponentiallywiththenumberofhopsalongthepath,severelydegradingTCPthroughput.Thestate-of-the-artresponsetodayistouseasufcientlylargenumberof802.11link-layeracknowledgements(ARQ)toprovideareliablechannelabstractiontoTCP.However,802.11ARQ1)interactspoorlywithTCPend-to-endratecontrolasitincreasesRTTvariance,2)increasesper-packetoverheadduetomorecarriersensing,backoffs,andacknowledgments,especiallyunderhighcontentionandloss(inx5.1.1,weshowthat802.11bARQhas35%overhead).NotethatTCP'swoescannotbeaddressedbyjustsettingthe802.11ARQlimittoalargevalueasitwouldreducetheoverallthroughputbydisproportionatelyusingthechan-nelfortransmittingpacketsoverbadlinks.UnlikeTCP,Hopreliessolelyontransport-layerreliabilityandavoidslink-layerretransmissionsfordata,therebyavoidingneg-ativeinteractionsbetweenthelinkandtransportlayers.Hop-by-hopvs.end-to-endcongestioncontrol:RatecontrolinTCPoccursinresponsetoend-to-endlossanddelayfeedbackreportedbyeachpacket.However,end-to-endfeedbackiserror-proneandhashighvarianceinmulti-hopwirelessnetworksaseachpacketobservessig-nicantlydifferentwirelessinterferenceacrossdifferentcontentiondomainsalongtheroute.ThisvariancehurtsTCP'sutilizationas:1)itswindowsizeshrinksconserva-tivelyinresponsetoloss,and2)itexperiencesfrequentretransmissiontimeoutswhennodataissent.OurpositionisthatxingTCP'sratecontrolalgorithminenvironmentswithhighvariabilityisfundamentallydifcult.Instead,wecircumventend-to-endratecontrol,andreplaceitwithhop-by-hopbackpressure.Ourap-proachhastwokeybenets:1)hop-by-hopfeedbackismorerobustthanend-to-endfeedbackasitinvolvesonlyasinglecontentiondomain,and2)block-levelfeedbackprovidesanaggregatedlinkqualityestimatethathaslessvariabilitythanpacket-levelfeedback.In-networkcaching:Theuseofreliableper-hopblocktransferenablesustoexploitcachingatinterme-diatehopsfortwobenets.First,cachingobviatesre-dundantretransmissionsalongpreviouslytraversedseg-mentsofaroute.Second,cachingismorerobusttointer-mittentdisconnectionsasitenablesprogressevenwhenacontemporaneousend-to-endrouteisunavailable.Hop Figure3:Virtualretransmissionduetonodefailure. Figure4:Exampleshowingneedforbackpressure.Withoutbackpressure,NodeAwouldallocate1=5thofout-goingca-pacitytoeachow,resultinginqueuesincreasingunboundedatnodesBthroughE.Withbackpressure,mostdataissenttonodeF,therebyincreasingutilization. 3.3 BackpressurecongestioncontrolRatecontrolinresponsetocongestioniscriticalinTCPtopreventcongestioncollapseandimproveutilization.Inwirelessnetworks,congestioncollapsecanoccurbothduetoincreasedpacketlossduetocontention[11],andincreasedlossduetobufferdrops[9].Bothcasesresultinwastedwork,whereapackettraversesseveralhopsonlytobedroppedbeforereachingthedestination.Priorworkhasobservedthatend-to-endlossanddelayfeed-backhashighvarianceandisdifculttointerpretunam-biguouslyinwirelessnetworks,whichcomplicatesthedesignofcongestioncontrol[2,32].Hopreliesonlyonhop-by-hopbackpressuretoavoidcongestion.Foreachow,aHopnodemonitorsthedif-ferencebetweenthenumberofblocksreceivedandthenumberreliablytransmittedtoitsnext-hopasshowninFigure4.Hoplimitsthisdifferencetoasmallxedvalue,H,andimplementsitwithnoadditionalover-headtotheBSYN/BACKexchange.AfterreceivingHcompleteblocks,aHopnodedoesnotrespondtofur-therBSYNrequestsfromanupstreamnodeuntilithasmovedatleastonemoreblocktoitsdownstreamnode.ThedefaultvalueofHissetto1block.BackpressureinHopsignicantlyimprovesutiliza-tion.Toappreciatewhy,considerthefollowingscenariowhereows1;:::;kallsharetherstlinkwithalowlossrate.Assumethattherestofow1'sroutehasasimilarlowlossrate,whileows2;:::;(k�1)tra-verseapoorrouteorarepartitionedfromtheirdestina-tions.LetCbethelinkcapacity,p1betheend-to-end lossobservedbytherstow,andp2betheend-to-endlossrateobservedbyotherows(p1p2).With-outbackpressure,Hopwouldallocatea1=kfractionoflinkcapacitytoeachow,yieldingatotalgoodputofC((1�p1)+(1�p2)
(k�1)) k.Andthenumberofbufferedblocksatthenext-hopsofthelatterk�1owsgrowsunbounded.Ontheotherhand,limitingthenumberofbufferedblocksforeachowyieldsagoodputclosetoC
(1�p1)inthisexample.WhydoesHoplimitthenumberofbufferedblocks,H,toasmalldefaultvalue?Notethattheexampleabovecanbeaddressedsimplybychoosingtheblockcorrespond-ingtotheowwiththelargestdifferentialbacklog(alongA-F).Indeed,classicalbackpressurealgorithmsknowntoachieveoptimalthroughput[33]worksimilarly.Hoplimitsthenumberofbufferedblockstoasmallvalueinordertoensuresmalltransferdelayfornite-sizedles,aswellastolimitintra-pathcontention. 3.4 RobustnesstopartitionsAfundamentalbenetofHopisthatitcontinuestomakeprogressevenwhenthenetworkisintermittentlyparti-tioned.Hoptransfersablocksinahop-by-hopmannerwithoutwaitingforend-to-endfeedback.Thus,evenifanend-to-endrouteiscurrentlyunavailable,Hopcontin-uestomakeprogressalongotherhops.Theabilitytomakeprogressduringpartitionsreliesonknowingwhichnext-hoptouse.Unliketypicalmeshroutingprotocols[23,4],routingprotocolsdesignedfordisruption-toleranceexposenext-hopinformationevenifanend-to-endrouteisunavailable(e.g.RAPID[3],DTLSR[7]).Inconjunctionwithsuchadisruption-tolerantroutingprotocol,Hopcanaccomplishdatatrans-ferevenifacontemporaneousend-to-endrouteisneveravailable,i.e.,thenetworkisalwayspartitioned.Indisruption-pronenetworks,aHopnodemayneedtocacheblocksforalongerdurationinordertomakeprogressuponreconnection.Inthiscase,thebackpres-surelimitneedstobesettakingintoaccountthefractionoftimeanodeispartitionedandtheexpectedlengthofaconnectionopportunitywithanext-hopnodealongaroutetothedestination(seex5.7foranexample). 3.5 HandlinghiddenterminalsTheeliminationofcontroloverheadforblocktransferimprovesefciencybuthasanundesirableside-effect—itexacerbateslossinhiddenterminalsituations.Hoptransmitsblockswithoutratecontrolorlink-layerre-transmissions,whichcanresultinacontinuousstreamofcollisionsatareceiverifthesendersarehiddenfromeachother.WhilehiddenterminalsareaproblemevenforTCP,ratecontrolmitigatesitsimpactonoverallthroughput.Flowsthatcollideatareceiverobservein-creasedlossandthrottletheirrate.Sincedifferentows getdifferentperceptionsofloss,somereducetheirratemoreaggressivelythanothers,resultinginmostowsbeingcompletelyshutoutandbandwidthbeingdevotedtooneorfewows[37].Thus,TCPishighlyunfairbuthasgoodaggregatethroughput.Hopusesanovelackwithholdingtechniquetomit-igatetheimpactofhiddenterminals.Here,areceiveracknowledgesonlyoneBSYNpacketatanytime,andwithholdsacknowledgementtootherconcurrentBSYNpacketsuntiltheoutstandingblockhascompleted.Inthismanner,thereceiverensuresthatitisonlyreceivingoneblockfromanysenderatagiventime,andothersenderswaittheirturn.Oncetheblockhascompleted,there-ceivertransmitstheBACKtooneoftheothertransmit-ters,whichstartstransmittingitsblock.Althoughackwithholdingdoesnotaddresshiddenterminalscausedbyowstodifferentreceivers,itof-fersalightweightalternativetoexpensiveandconser-vativetechniqueslikeRTS/CTSforthecommonsingle-terminalhiddenterminalcase.ThehighoverheadofRTS/CTSarisesfromtheadditionalcontrolpackets,es-peciallysincethesearebroadcastpacketsthataretrans-mittedatthelowestbit-rate.TheuseofbroadcastalsomakesRTS/CTSmoreconservativesincealargercon-tentionregionisclearedthantypicallyrequired[39].Incontrast,ackwithholdingrequiresnoadditionalcontrolpackets(asBSYNsandBACKsarealreadyinplaceforblocktransfer). 3.6 PacketschedulingHop'sunitoflinklayertransmissionisatxop,whichisthemaximumdurationforwhichthenetworkinterfacecard(NIC)ispermittedtosendpacketsinaburstwithoutcontendingforaccess[1].Hop'sschedulerleveragestheburstmodeandsendsatxop'sworthofdatafromeachconcurrentowatatimeinaround-robinmanner.Hoptrafcisisolatedfromdelay-sensitivetrafcsuchasVoIPorvideobyusinglink-layerprioritiza-tion.802.11chipsetssupportfourpriorityqueues—voice,video,best-effort,andbackgroundindecreasingorderofpriority—withthehigherpriorityqueuesalsohavingsmallercontentionwindows[1].Hoptrafcissentusingthelowestprioritybackgroundqueuetomini-mizeimpactondelay-sensitivedatagrams.Thedesignchoicesthatwehavepresentedsofarcanbedetrimentaltodelayforsmallles(referredtoasmicro-blocks)inthreeways:1)theinitialBSYN/BACKexchangeincreasesdelayformicro-blocks,2)asendermaybeservicingmultipleows,inwhichcaseamicro-blockmayneedtowaitformultipletxops,and3)ack-withholdingcanresultinmicro-blocksbeingdelayedbyoneormorelargeblocksthatareacknowledgedbeforeitsturn.Hopemploysthreetechniquestooptimizedelayformicro-blocks.First,micro-blocksofsizelessthana xedBSYNbatchthreshold(fewtensofKB)aresentpiggybackedwiththeBSYNwithlink-layerARQviathevoicequeue.Thisoptimizationeliminatestheini-tialBSYN/BACKdelay,andavoidshavingtowaitforaBACKbeforeproceeding,therebycircumventingack-withholdingdelay.Second,thepacketscheduleratthesenderprioritizesmicro-blocksoverlargerblocks.Fi-nally,Hopuseablock-sizebasedack-withholdingpolicythatprioritizesmicro-blocksoverlargerblocks. 4 ImplementationWehaveimplementedaprototypeofHopwithallthefeaturesdescribedinx3.HopisimplementedinLinux2.6asanevent-baseduser-spacedaemoninroughly5100linesofCcode.HopiscurrentlyimplementedontopofUDP(i.e.,thereisaUDPheaderinbetweentheIPandHopheadersineachframeinFigure1).Below,wedescribeimportantaspectsofHop'simplementation. 4.1 MACparametersOurimplementationusestheAtheros-basedwirelesschipsetandtheMadwiopensource802.11devicedriver[18],apopularcommodityimplementation.Bydefault,theMadWidriver(aswellasothercommodityimplementations)supportsthe802.11eQoSextension.However,MadWiFisupportstheextensiononlyintheaccesspointmode,sowemodifythedrivertoenableitinthead-hocmodeaswell.Hopusesdefault802.11settings,exceptforthefollowing.Thetransmissionop-portunity(txop)forthebackgroundqueueissettothemaximumvaluepermittedbytheMadWidriver(8160sorroughly8KBofdata).Link-layerARQisdisabledforalldataframessentviaHopbutenabledforcontrolpackets(BSYN,BACK,etc). 4.2 Hopimplementation Parameters Alargeblocksizeincreasesbatchingben-ets[15],sowesetthedefaultmaximumblocksizeto1MB.NotethatthismeansthataHopblockisallowedtobeupto1MBinsize,butmaybeanysmallersize.Hopneverwaitsidlyinanticipationofmoreapplicationdatainordertoobtainbatchingbenets.TheBSYNbatchthresholdformicro-blocksissettoadefaultvalueof16KB,andthebackpressurelimit,H,issetto1.Thevirtualretransmissiontimeoutissettoaninitialvalueof60secondsandsimplyresettotheround-tripblockdelayreportedbythemostrecentblock.TheTTLlimitforavirtualretransmissionsissetto50hops.Inthecurrentimplementation,anintermediateHopnodekeepsalltheblocksthatithasreceivedinmemory. Headerformat: TheHopheaderconsistsofthefol-lowingelds.Allframescontainthemsg_typethatidentiesiftheframeisadata,BSYN,BACK,virtualretransmissionBSYN,oranend-to-endBACKframe; (a)SortedSingle-hopFlows (b)ImpactofLoss.Figure8:Gracefuldegradationtoadversechannelconditions.Firstplotshowsper-linkgoodputsfromone-hopexperimentsortedinTCPorder.Secondplotshowscontrolledexperimentsdemonstratingimpactofloss.Inbothcases,Hopismorero-bustanddegradesfarmoregracefullythanTCP. linksinthetestbedwhereTCP'sthroughputisnear-zero.Theaveragegoodputfortheworst20TCPowsis334Kbps,whereasHop'sgoodputforthesameowsis2.37Mbps,adifferenceof7.TounderstandthecauseofTCP'sfragilebehavior,weevaluatetheimpactoflossperceivedatthetrans-portlayerontheperformanceofHopandTCP.Westartwithaperfectlinkthathasanear-zerolossrateandin-troducelossbymodifyingtheMadWidevicedrivertorandomlydropaspeciedfractionofincomingpack-ets.Figure8(b)showsthat,unsurprisingly,TCPgoodputdropstonear-zerowhenlossrateisroughly20%.Hopshowsgracefulnear-lineardegradationandisoperationaluntilthelossrateisabout80%. 5.2 Multi-hopmicrobenchmarksHowdoesHopperformonmulti-hoppathscomparedtoexistingalternatives?Tostudythisquestion,wepickasequenceof100nodepairsrandomlywithrepetitionfromthetestbed.Staticroutesaresetupaprioribetweenallnodepairstoisolatetheimpactofrouteux(consid-eredinx5.3).Thestaticrouteswereobtainedbyrun-ningOLSRwiththedefaultETXmetricuntiltheroutingtopologystabilizedatthebeginningoftheexperiment.Amongthe100randomlychosenows,30%aretwo-hop,30%arethree-hop,10%arefour-hop,20%areve-hop,andtheremaining10%areseven-hopows.Wecomparethemulti-hopgoodputofHoptoTCP,Hop-by-HopTCP,DTN2.5,andUDP. Figure9:Experimentwithmulti-hopows.Hopimproveslowerquartilegoodputby2.7,mediangoodputby2.3,andmeangoodputby2. Figure10:Boxplotofmulti-hopsingle-owbenchmarks.Hophas2-3median,and2-4meanimprovementsoverotherreliabletransportprotocols.HopiscomparabletoUDP/no-ARQ/txopintermsofmedian/mean—thelatterisextremelyfastsinceithasnooverhead,butexperiencesmoreloss. Figure9showstheCDFofgoodputforjustHopandTCP,whileFigure10showsthesummarystatisticsforalltheprotocols.Hopconsistentlyoutperformsallotherprotocols.TheQ1,Q2,andQ3gainsoverTCPare2.7,2.3and1.9respectively.TheQ1gainoverTCPislowerthanforthesingle-hopexperimentbe-causeonlygoodlinksselectedbyOLSRareusedinthisexperiment(asevidencedbythebetterperformanceofUDP/no-ARQ/txopcomparedtoUDP/max-ARQ/txop).Overlossierpaths,Hop'sgainsaremuchhigher.Wealsondthatthegainsalsogrowwithincreasingnum-berofhops(refertechnicalreport[15]).Forexample,thelowerquartilegainsgrowfromabout2.7fortwohopstomorethan4forveandsixhops. 5.3 HopunderhighloadTheexperimentssofarconsideredoneowinisolation.Next,weevaluateHopinaheavilyloadednetworktoun-derstandtheeffectofincreasedcontentionandcollisionsonHop'sperformanceandfairness.WecompareHop,TCP,andHop-by-HopTCP.Theexperimentconsistsofthirtyconcurrentowsthattransferdatacontinuallybe-tweenrandomlychosennodepairsinthetestbed.Allprotocolsarerunoverastaticmeshtopologyidenticaltox5.2.Tofocusonmultihopbenets,wepicksrc-dstpairsthatarenotimmediateneighborsofeachother.Weruntheexperimentvetimes,andforeachrun,wemeasure batchingwithFECandbulkacknowledgmentsatthelinklayeroverlong-distanceunidirectional802.11links.Kimetal.[35]aggregateTCPframesusingthe802.11nburstmodetoamortizetheMACprotocoloverhead.Incomparison,Hopjointlyleveragesbatchingbothatthelinkandtransportlayers. 7 ConclusionsThelastdecadehasseenahugebodyofresearchonTCP'sproblemsoverwirelessnetworks,butTCPforgoodreasonscontinuestotobethedominantreal-worldalternativetoday.OnereasonmaybethatTCPisgoodenoughinthecommoncaseofwirelessLANs,andso-lutionsproposedformorechallengedenvironmentsdonotperformwellinthecommoncase.Anaturalques-tionisifwecanhaveonesimpletransportprotocolthatyieldsrobustperformanceacrossdiversenetworkssuchasWLANs,meshes,MANETs,sensornets,andDTNs.OurworkonHopsuggeststhatthisgoalisachievable.Hopachievessignicantthroughput,fairness,andde-laygainsbothinwell-connectedWLANsandmeshnet-worksaswellasdisruption-pronenetworks.References [1] http://standards.ieee.org/getieee802/download/802.11e-2005.pdf.802.11e:QualityofServiceenhancementsto802.11. [2] HariBalakrishnan,VenkataN.Padmanabhan,SrinivasanSeshan,andRandyH.Katz.Acomparisonofmechanismsforimprovingtcpperfor-manceoverwirelesslinks.InSIGCOMM,1996. [3] A.Balasubramanian,B.Levine,andA.Venkataramani.Dtnroutingasaresourceallocationproblem.SIGCOMM,2007. [4] JohnBicket,DanielAguayo,SanjitBiswas,andRobertMorris.Architec-tureandevaluationofanunplanned802.11bmeshnetwork.InMobiCom,2005. [5] L.Breslau,PeiCao,LiFan,G.Phillips,andS.Shenker.Webcachingandzipf-likedistributions:evidenceandimplications.INFOCOM,1999. [6] R.G.ColeandJ.H.Rosenbluth.Voiceoveripperformancemonitoring.SIGCOMMComput.Commun.Rev.,2001. [7] M.DemmerandK.Fall.Dtlsr:Delaytolerantroutingfordevelopingre-gions.NSDR,2007. [8] http://www.dtnrg.org/.DelayTolerantNetworking(DTN)Refer-enceGroup. [9] Z.Fu,P.Zerfos,H.Luo,S.Lu,L.Zhang,andM.Gerla.Theimpactofmultihopwirelesschannelontcpthroughputandloss.InINFOCOM'03,2003. [10] S.Ha,I.Rhee,andL.Xu.Cubic:anewtcp-friendlyhigh-speedtcpvariant.InSIGOPS,2008. [11] B.Hull,K.Jamieson,andH.Balakrishnan.Mitigatingcongestioninwire-lesssensornetworks.InSenSys,pages134–147,NewYork,NY,USA,2004.ACMPress. [12] http://www.ircache.net/.IRCache:TheNLANRWebCachingProject. [13] S.Jain,K.Fall,andR.Patra.Routinginadelaytolerantnetwork.InSIGCOMM,2004. [14] SukunKim,RodrigoFonseca,PrabalDutta,ArsalanTavakoli,DavidCuller,PhilipLevis,ScottShenker,andIonStoica.Flush:areliablebulktransportprotocolformultihopwirelessnetworks.InSenSys,2007. [15] M.Li,D.Agrawal,D.Ganesan,andA.Venkataramani.Block-switchednetworks:Anewparadigmforwirelesstransport.TechnicalReportTRUM-CS-2008-27,UMassAmherst,2008. [16] ChaegwonLim,HaiyunLuo,andChong-HoChoi.RAIN:Areliablewire-lessnetworkarchitecture.InProceedingsofIEEEICNP,2006. [17] S.Liu,J.;Singh.ATCP:Tcpformobileadhocnetworks.IEEEJSAC,2001. [18] http://www.madwifi.org/.MadwiDeviceDriver. [19] S.Mascolo,C.Casetti,M.Gerla,M.Y.Sanadidi,andR.Wang.TCPwest-wood:Bandwidthestimationforenhancedtransportoverwirelesslinks.InMobicom,2001. [20] ParthoP.MishraandHemantKanakia.Ahopbyhoprate-basedcongestioncontrolscheme.SIGCOMM,1992. [21] KitaeNahm,AhmedHelmy,andC.-C.JayKuo.Tcpovermultihop802.11networks:issuesandperformanceenhancement.InMobiHoc,2005. [22] AnthonyC.H.Ng,DavidMalone,andDouglasJ.Leith.Experimentalevaluationoftcpperformanceandfairnessinan802.11etest-bed.InE-WIND,2005. [23] http://www.olsr.org/.OptimizedLinkStateRoutingProtocol. [24] C¨uneyt¨Ozveren,RobertSimcoe,andGeorgeVarghese.Reliableandef-cienthop-by-hopowcontrol.SIGCOMM,1994. [25] SinhaP.,T.Nandagopal,N.Venkitaraman,R.Sivakumar,andV.Bhargha-van.Wtcp:areliabletransportprotocolforwirelesswide-areanetworks.WirelessNetworks,2002. [26] R.Patra,S.Nedevschi,S.Surana,A.Sheth,L.Subramanian,andE.Brewer.WiLDNet:DesignandImplementationofHighPerformanceWiFi-basedLongDistanceNetworks.InNSDI,2007. [27] B.Radunovic,C.Gkantsidis,D.Gunawardena,andP.Key.Horizon:Bal-ancingtcpovermultiplepathsinwirelessmeshnetwork.InMobiCom,2008. [28] GojkoBabicRajJain,ArjanDurresi.Throughputfairnessindex:Anex-planation,atmforum/99-0045,february1999. [29] S.Rangwala,R.Gummadi,R.Govindan,andK.Psounis.Interference-awarefairratecontrolinwirelesssensornetworks.SIGCOMM,2006. [30] S.Rangwala,A.Jindal,K.Jang,K.Psounis,andR.Govindan.Understand-ingcongestioncontrolinmulti-hopwirelessmeshnetworks.Mobicom,2008. [31] B.Scheuermann,C.Lochert,andM.Mauve.Implicithop-by-hopconges-tioncontrolinwirelessmultihopnetworks.AdHocNetw.,2008. [32] K.Sundaresan,V.Anantharaman,H.Hsieh,andR.Sivakumar.Atp:Areliabletransportprotocolforad-hocnetworks.InInProceedingsofMO-BIHOC2003,2003. [33] L.Tassiulas,andA.Ephremides.Stabilitypropertiesofconstrainedqueue-ingsystemsandschedulingpoliciesformaximumthroughputinmultihopradionetworks.InIEEETransactionsonAutomaticControl,37(12),1992. [34] L.Georgiadis,M.Neely,andL.Tassiulas.Resourceallocationandcross-layercontrolinwirelessnetworks.InFoundationsandTrendsinNetwork-ing,1(1):1-144,2006. [35] W.Kim,H.WrightandS.NettlesImprovingthePerformanceofMulti-hopWirelessNetworksusingFrameAggregationandBroadcastforTCPACKsInCoNext,2008 [36] M.Vutukuru,K.Jamieson,andH.Balakrishnan.Harnessingexposedter-minalsinwirelessnetworks.InNSDI,SanFrancisco,USA,April2008. [37] K.Xu,M.Gerla,L.Qi,andY.Shu.Enhancingtcpfairnessinadhocwirelessnetworksusingneighborhoodred.InMobiCom,2003. [38] X.Yu.Improvingtcpperformanceovermobileadhocnetworksbyexploit-ingcross-layerinformationawareness.InMobiCom,2004. [39] J.Kurose,K.RossComputernetworking:atopdownapproach.AddisonWesley,2007