您好,欢迎访问三七文档
1.“Sizingrouterbuffers”,Appenzelleretal,Sigcomm20042.“Safeandeffectivefine-grainedTCP…”,Vasudevanetal,Sigcomm2009NickMcKeownCS244Spring2012Lecture4BufferSizing期刊杂志)(logO22)2()1(WnCTCT10,00020#packetsat10Gb/s1,000,000(1)Assume:Largenumberofdesynchronizedflows;100%utilization(2)Assume:Largenumberofdesynchronizedflows;100%utilizationIntuition&ProofSawtoothPeak-to-troughSmoothingofmanysawtoothsNon-burstyarrivalsEvidenceSimulation,SingleTCPFlowSimulations,MeasurementsSimulations,Measurements3TimeevolutionofasingleTCPflowthrougharouter.Bufferis2T*CTimeEvolutionofaSingleTCPFlowTimeevolutionofasingleTCPflowthrougharouter.Bufferis2T*C4TheStory)(logO22)2()1(WnCTCT10,00020#packetsat10Gb/s1,000,000EvidenceIntuition&Proof(1)Assume:Largenumberofdesynchronizedflows;100%utilization(2)Assume:Largenumberofdesynchronizedflows;100%utilizationSawtoothPeak-to-troughSmoothingofmanysawtoothsNon-burstyarrivalsSimulation,SingleTCPFlowSimulations,ExperimentsSimulations,Experiments5SynchronizedFlows•Aggregatewindowhassamedynamics•Thereforebufferoccupancyhassamedynamics•Rule-of-thumbstillholds.2maxWtmax2WmaxWmaxW6ProbabilityDistributionB0BufferSizeWManyTCPFlowsWhatyousaidIfounditquiteshockingthatbefore2004,nobodyhadbotheredputtingeffortintoresearchingsomethingthatwasclearlycausinghugeengineering(andthusfinancial)hurdlesinscalinguptheinternetinfrastructure.DavidSchneiderYouhavetowonderhowthebuffersizingruleofthumblastedadecade,andfurther,whethertheanswertothisquestionmightbethatlargebuffersdon'tposeasgraveathreatastheauthorswouldhaveusbelieve.”JustinCosta-Roberts78TheStory)(logO22)2()1(WnCTCT10,00020#packetsat10Gb/s1,000,000EvidenceIntuition&Proof(1)Assume:Largenumberofdesynchronizedflows;100%utilization(2)Assume:Largenumberofdesynchronizedflows;100%utilizationSawtoothPeak-to-troughSmoothingofmanysawtoothsNon-burstyarrivalsSimulation,SingleTCPFlowSimulations,ExperimentsSimulations,Experiments92TCnSimulationRequiredBufferSize-SimulationsWhatyousaidSincethispaperisalmosttenyearsoldIkeptwonderingwhethertheserecommendationswhereevertestedinthebackboneroutersandconfirmed.”JackDubieThispaper'stitleisSizingRouterBuffers,butitmightbemoreaccuratelytitledSizingCoreInternetRouterBuffers.”CarlCase1011InternetBackboneExperiment•Highlinkutilization•Longduration(abouttwoweeks)•Buffersizes190ms(250Kpackets),10ms(10Kpackets),1ms(1000packets)BufferSizinginInternetRouters12Dropvs.Load,Buffer=190ms,10ms00.040.080.120.160.2020406080100Load(%)PacketDropRate(%)BufferSizinginInternetRouters13Dropvs.Load,Buffer=1ms00.040.080.120.160.2020406080100Load(%)PacketDropRate(%)14TheStory)(logO22)2()1(WnCTCT10,00020#packetsat10Gb/s1,000,000EvidenceIntuition&Proof(1)Assume:Largenumberofdesynchronizedflows;100%utilization(2)Assume:Largenumberofdesynchronizedflows;100%utilizationSawtoothPeak-to-troughSmoothingofmanysawtoothsNon-burstyarrivalsSimulation,SingleTCPFlowSimulations,ExperimentsSimulations,Experiments15SmoothTraffic•Theory:Forsmoothtrafficverysmallbuffersareenough.•PoissonTraffic–Lossindependentoflinkrate,–RTT,numberofflows,etc.Canwemaketrafficlook“Poisson-enough”whenitarrivestotherouters…?M/D/1PoissonBDBloss%1losspkts2080%,..Bei16VerySmallBuffers-TheoryAssumption:linkover-provisionedby1/,andpacketsfromeachflowlaunchedsmoothlyTheorem:WeneedtWi)(log12log2max/1WOWBtWindowsizeRTTtWPoissoniTransmit17VerySmallBuffers-Theory•Assumptions:–Minimumdistancebetweenconsecutivepacketsofthesameflow;–desynchronizedflows–randomandindependentstarttimesforflows•Theorem:BuffersizeofO(logW)guaranteesverylowpacketdropprobability.•Note:Distancebetweenpackets–Slowaccesslinks–TCPPacingWithabuffersizeof10-20packetswecangainhighthroughput.18TheStory)(logO22)2()1(WnCTCT10,00020#packetsat10Gb/s1,000,000EvidenceIntuition&Proof(1)Assume:Largenumberofdesynchronizedflows;100%utilization(2)Assume:Largenumberofdesynchronizedflows;100%utilizationSawtoothPeak-to-troughSmoothingofmanysawtoothsNon-burstyarrivalsSimulation,SingleTCPFlowSimulations,ExperimentsSimulations,Experiments19O(logW)Buffers20ExperimentResultsSummary21RTT´CRTT´CNAbout50pktsAbout90%100%ThroughputBufferSize22Consequences10-50packetsonachipAnnouncementsCritiques:Wewillpostsomeselectedcommentspublicly.Ifyoudon'twantyournamelisted,markas“anon”andwewon'tuseyourname.PA#2:Youhave2weeks.StartsMonday,dueMondayApril30th.ItwillbepostedbymidnightFriday,notofficiallystartinguntilMonday.23AnnouncementsPA#3:Sayalittlebitaboutit.Whenreadingpapers,keepalogofwhatwouldmakeaninterestingexperimenttheywouldhavelikedtosee.24TCPIncast“Safeandeffectivefine-grainedTCPRetransmissionsforDatacenterCommunication”Vasudevanetal,Sigcomm200925Questions•Summaryofpaper•Whatarethemainresults?•Whywasthepaperwrittenin2009,notearlier?•Whatdoyouthinkoftheevaluationmethods?•Whatdoyouthinkofthewriting?26WhatyousaidAfterreadingthispaperandthelastIhavebecomeverysuspiciousofanyhard-codedconstantsinsoftwarethatwillinterfacewithmultiplenetworks.”AndrewHaven2728Fig
本文标题:怎样制作电子期刊
链接地址:https://www.777doc.com/doc-70883 .html