Abstract—Packets in the Internet can experience large queueing delays during busy periods. Backbone routers are generally engineered to have large buffers, in which packets may wait as long as half a second (assuming FIFO service, longer otherwise). Durin
QueueManagementforShort-LivedTCPFlowsin
BackboneRouters
AnshulKantawalaandJonathanTurnerDepartmentofComputerScience
WashingtonUniversitySt.Louis,MO63130anshul,jst@arl.wustl.edu
Abstract—PacketsintheInternetcanexperiencelargequeueingdelaysduringbusyperiods.Backboneroutersaregenerallyengineeredtohavelargebuffers,inwhichpacketsmaywaitaslongashalfasecond(assumingFIFOservice,longerotherwise).Duringcongestionperiods,thesebuffersmaystayclosetofull,subjectingpacketstolongdelays,evenwhenthein-trinsiclatencyofthepathisrelativelysmall.Thispaperstudiestheper-formanceimprovementsthatcanbeobtainedforshort-livedTCP owsbyusingmoresophisticatedpacketschedulers,thanaretypicalofInter-netrouters.TheresultsshowthatthelargebuffersfoundinWANrouterscontributeonlymarginallytoimprovingrouterthroughput,andthehigherdelaysthatcomewithlargebuffersmakesthemadubiousinvestment.Theresultsalsoshowthatbetterpacketschedulingalgorithmscanproducedra-maticimprovementsinfairness.Usingns-2simulations,weshowthatalgo-rithmsusingmultiplequeuescansigni cantlyoutperformREDandBlue,especiallyatsmallerbuffersizes.Givenatraf cmixofshort-livedTCP owswithdifferentround-triptimes,longerround-triptime owsachieve
oftheirfair-shareusingmultiqueueschedulers,comparedtoun-derREDandBlue.Weobserveasimilarperformanceimprovementformulti-hoppaths.Wealsoshowthatperformanceresultscanbereliablyscaledacrossawiderangeofparametervalues,solongastheratioofthebuffersizetothelinkbandwidth-delayproductisheldinvariant.
I.INTRODUCTION
BackboneroutersintheInternetaretypicallycon guredwithbuffersthatareseveraltimestimeslargerthantheproductofthelinkbandwidthandthetypicalround-tripdelayonlongnetworkpaths.Suchbufferscandelaypacketsforasmuchashalfasec-ondduringcongestionperiods.WhensuchlargequeuescarryheavyTCPtraf cloads,andareservicedusingtheTailDroppolicy,thelargequeuesremainclosetofullmostofthetime.Thus,evenifeachTCP owisabletoobtainitsshareofthelinkbandwidth,theend-to-enddelayremainsveryhigh.Thisisexacerbatedfor owswithmultiplehops,sincepacketsmayexperiencehighqueueingdelaysateachhop.Thisphenomenoniswell-knownandhasbeendiscussedbyHashem[1]andMor-ris[2],amongothers.
Toaddressthisissue,researchershavedevelopedalternativequeueingalgorithmswhichtrytokeepaveragequeuesizeslow,whilestillprovidinghighthroughputandlinkutilization.ThemostpopularoftheseisRandomEarlyDiscardorRED[3].REDmaintainsanexponentially-weightedmovingaverageofthequeuelengthwhichisusedtodetectcongestion.Tomakeitoperateitrobustlyunderwidelyvaryingconditions,onemusteitherdynamicallyadjusttheparametersoroperateusingrel-ativelylargebuffersizes[4],[5].RecentlyanotherqueueingalgorithmcalledBlue[6],wasproposedtoimproveuponRED.Blueadjustsitsparametersautomaticallyinresponsetoqueue
ThisworkissupportedinpartbyNSFGrantANI-9714698
over owandunder owevents.AlthoughBluedoesimproveoverREDincertainscenarios,itsparametersarealsosensitivetodifferentcongestionconditionsandnetworktopologies.
Inthispaper,weinvestigatehowpacketschedulersusingmul-tiplequeuescanimproveperformanceoverexistingmethods.Ourgoalisto ndschedulersthatsatisfythefollowingobjec-tives:
Highthroughputwhenbuffersaresmall.Thisallowsqueue-ingdelaystobekeptlow.
Insensitivitytooperatingconditionsandtraf c.Thisreducestheneedtotuneparameters,orcompromiseonperformance.Fairtreatmentofdifferent ows.Thisshouldholdregardlessofdifferencesinround-tripdelayornumberofhopstraversed.TheresultspresentedhereshowthatbothREDandBluearede- cientintheserespects.Bothperformfairlypoorlywhenbufferspaceislimitedtoasmallfractionoftheround-tripdelay.
Inapreviousstudy[9],weinvestigatedtheperformanceofmultiplequeuesusinglong-livedTCP owtraf c.Thefocusofthispaperisstudyingtheperformanceofmultiplequeuesusingshort-livedTCP owtraf c.AmajorityoftheInternettraf ctodayisHTTP(web-traf c),whichconsistsofshort-livedTCP owstransferringwebpages.Evenwithp-HTTP(persistentHTTP)connections,theresultingTCPtraf cisbursty(down-loadingawebpage)withlongidleperiods(userpausebetweenwebpagedownloads)andcanbeemulatedbymultipleshort-livedTCP ows.
AnotherregularlyobservedphenomenonforqueueswithTailDropisbigswingsintheoccupancyofthebottlenecklinkqueue.OneofthemaincausesforthisisthesynchronizationofTCPsourcesgoingthroughthebottlenecklink.AlthoughREDandBluetrytoalleviatethesynchronizationproblembyusingarandomdroppolicy,theydonotperformwellwithbufferswhichareafractionofthebandwidth-delayproduct.Whenbuffersareverysmall,evenwitharandomdroppolicy,thereisahighprob-abilitythatall owssufferapacketloss.However,withper- owqueueing,wecanexplicitycontrolthenumberof owsthatsuf-ferapacketlossandthussigni cantlyreducesynchronizationamong ows.
Oneofthelimitationsofevaluatingalgorithmsusingsimu-lationsisthatwecannotreplicatethescaleofrealnetworks.Thus,acommonquestioniswhethertheresultsobtainedwillholdwhenthenumberofsourcesand/orbottleneckbandwidthisincreasedbyoneormoreordersofmagnitude.Inthispaper,weaddresstheissueastohowtheaveragegoodputofTCP owsisaffectedbychangesinnetworkparameters,namelynumber
& Meegan C.A., 1995, ARA&A, 33, 415 c 0000 RAS, MNRAS 000, 000–000 Evidence for anisotropy in the distribution of short-lived gamma-ray bursts...
Queue Management (AQM) routers serving a large ...Models for short-lived TCP ?ows and time-outs ...Most network links, especially in backbone networks...
queue management at routers have limited ability to assure even minimal ...In contrast to long-lived FTP flows, HTTP flows are short-lived with an...
queue management at routers have limited ability to assure even minimal ...In contrast to long-lived FTP flows, HTTP flows are short-lived with an...
热门文档