Figure1:Con uenceDiagram:In ectionPoint,MilieuChange,andTransformedIde - PDF document

Figure1:Con uenceDiagram:In ectionPoint,MilieuChange,andTransformedIdeastheseconceptsconcrete,considerrstanexamplefrompainting. Figure2:ImpressionismasaCon uenceFigure2depictsImpressionismasacon uencewhenmainstreampaintinginthe1800swasimpactedbytheemergingeldofpsychologytoformnewstreamsofpaintingsuchasimpressionism,andlaterexpression-ism.Thein ectionpointwasthearrivalofcheappho-tographywhichmadepaintersquestionthevalueofmerelyrealisticrendering.Whynot,painterssuchasMonetmayhavereasoned,paintthesubjectiveresponsetoalandscape,anim-pression,somethingacameracannotdo.Thiswasamilieuchangebecauseimpressionscapturedasconceptsinpsychologynowhadtobeincarnatedinpaint.Therewerealsotransformedideas:thin,precisebrushstrokesthatdelineatedbordersgavewaytoblurrythickstrokesthatmixintheeyeatacertaindistance.Whylookatexistingandnewworkthroughthelensofcon uence?Iwilldevelopthisthesisindetailelse-where.Fornow,mayIsuggestthatthecon uencelensallowsustoseparatetrendsfromfadsbylookingforthein ectionpoint;further,themilieuchange,onceidentied,providesathemeforresearchandaspringofspecicresearchideas.Thein ectionpointmakesitmorelikelythatthere-sultingresearchwillhaveimpact,andthemilieuchangeallowscreativefreedomtorethinkideasinsometimesbeautifulways,balancingthetwindesireswehaveascomputerscientistsforbothbeautyandimpact.Fi-nally,adiscernedcon uencecansometimessuggestaneweldinthemaking{agreeneldarea,especiallyaboonwhentheoriginalelds(thinkTCPpapers)havematured3.NETWORKALGORITHMICSHISTORYIwouldliketotakeawhirlwindtourofvariouscon-ceptsthathelpedmaketheInternetfast,butlookedatretrospectivelythoughthecon uencelens.FastServers:MyrstencounterwithalgorithmicswaswhenIjoinedDigitalEquipmentCorporationandfoundabeautifulcon uencebetweencomputerarchi-tectureandnetworking(Figure3)thatledtosomethingthattodayiscalledRDMA[3]butwaspartoftheVAX-ClustersysteminventedbyKronenberg,StreckerandLevy[21].Thein ectionpointwastherealizationthatonecouldcreatecheapclustersofminicomputers;themilieuchangewasgoingfromabusinasinglecomputertomanycomputersconnectedbyanetwork.Inparticular,theinventorsofVAXClusters[21]rea-sonedthatsinceDirectMemoryAccesswasastream-linedwayofsendinglargeamountsofdatadirectlyfromthedisktomemorywithoutbotheringtheCPU,theideacouldbeextendedtoRemoteDMAfromthemem-oryofcomputer1tothememoryofcomputer2withouttheinterventionofeitherCPU.Datawascopiedonlyonceoverthebus,andtheoverheadofsystemscallsandinterruptswasminimal.Ofcourse,todayRDMAisamajorforceinstoragenetworks,butpeopleforgetitwasinventedin1986. Figure3:RDMAasaCon uenceSoonafterVAXClusters,anewin ectionpointaroseastheInternetbegantoheatup.Serverswerefoundtobewoefullyslowbecausetheycopieddataacrosslay-ersofsoftware,andbecauseoftheoverheadofsystemcalls.WhileRDMAdidavoidtheseoverheads,itre-quiredprotocolchanges,andonlyworkedforlargedatatransfers.ThusbeganastreamofworkinSIGCOMMin uencedbytheOperatingSystemcommunity(theimpactingstream),tomatchthespeedgainsofRDMA usingonlylocalOperatingSystemchangeswithoutpro-tocolchanges,whileretainingstructureandprotection.Ipickthreerepresentativepapers.FbufsfromDr-uschelandPeterson[10]showeditwaspossibletoavoidcopiesbyleveragingpagetables.ApplicationDeviceChannelsfromDruschel,PetersonandDavy[11]showedhowtoavoidsystemcalls.BothideasarealiveandwellinwhatpeoplecallZeroCopyInterfaces[6]andVirtualInterfaceArchitecture[5].Finally,ActiveMessages[33]wasroughlyconcurrentwithApplicationLevelDeviceChannels,andagainavoidedcontroloverheadbypass-inginformationinpackets.Beyondwaystostream-linedatamovement,VanJacobsonandMikeKarelsshowedthatTCPperformancecouldbeoptimizedintheexpectedcaseusing\headerprediction"[4]whensegmentsarriveinorder.Nonewtransportprotocolswereneeded.Thestagehadbeensetforfastservers.FastRouters:Therstglimmerofarealcon uencebetweenalgorithmsandnetworksthatIencounteredarrivedbecauseofanewin ectionpoint(Figure4)around1996.IPv6wasrumoredtobeimminentandaddresseswerenowW=128bits.Simpletrie-basedschemeswerelinearinthenumberofaddressbitswhichwastooslow.Ofcourse,theoreticalcomputersciencehadsomefastprexalgorithmsbuttheyweremostlyO(logN)schemeswhereNisthenumberofprexes,andthemilieuwasdierent(Figure4)becausememoryaccessesandnotcomputationisthedominantmetric.Whilemosttheoreticalalgorithmswerecontentwithcomputingaprexlookupinmilliseconds,anarrivingpackethadlessthanamicrosecondtobeforwarded.ItwasinpuzzlingoverIPv6thatwediscoveredaprexlookupschemethattookO(logW)memoryaccesses,whichforIPv6was7memoryaccesses.Thistomeseemedtobeatransformedconcept.WhileO(loglogN)algorithmswereknownforlookups[29],theywereforexactlookupsandmorecomplicated. Figure4:BinarySearchonPrexLengthsasaCon u-ence.ThestoryofBinarySearchonprexlengths[34]isaromantictaleofanencounterwithtwooutsiders,andhowideasare\intheair"atacertaintimeperiod.ImetJonTurneronthestairsofBryanHallinWashing-tonUniversityandheaskedwhyonecouldntdobinarysearchonprexlengths.Prexeswouldrstbesegre-gatedbytheirlengthandthenateachlengthasearchforaprexrequiredonlyanexactmatchbyhashing.Theusualmethodstartswiththelongestlengthandworksbackward.Jon,however,wantedtostartwiththemiddleprexlengthandperformbinarysearch.Ithoughtabouthisideaforafewminutesandshowedhimasimplecounterexamplewithtwoprexes,oneatlength1andoneatlength3.Iaskedhimhowonecouldsearchinthemiddlelength(2)hashtablewhentherewasnoprexoflength2.Somedayslater,hemetmeonthesamestairsandtoldme\Easy:foreverylongerlengthprex,addanarticialprex(calledamarker)atalllengthtablesthatbinarysearchcanreach".ThusinFigure5,marker10isplacedintheLength2hashtable.Thiswaswonderfulasfaritwent,buttherewasa aw.Sometimesmarkerstakeyouonawildgoosechasetothesecondhalf.Forexample,inFigure5,whensearchingforthestring100,marker10takessearchtothesecondhalftowardstheentryfor101whentheanswer(1)insteadlurksinthersthalf.RatherthantellTurneraboutthebug,Idecidedtoxitmyselfbyprecomputingthebestmatchingprexofeverymarker.Forexample,thebestmatchingprexof10isprecom-putedtobe1.Ifthesearchprocessremembersthematchingprexofthelastmarkerencountered,thisbe-comestheanswerwhensearchfailswithouttheneedforbacktracking. Figure5:BinarySearchonPrexLengths.Thetwoprexesare1and101.10isanarticialentrycalledamarkerusedtoguidebinarysearch.Amazingly,onabusafewdayslater,IsatnexttoareallysmartSwissstudent,MarcelWaldvogel,whowasvisitingWashingtonUniversity.HehadallthesameideasasJonandendedwiththesamebug.Sowebeganworkingtogether.Ofcourse,Marceldidallthework, andaddedanumberofelegantideasofhisownsuchasRopeSearch[34].Next,everypacketarrivingonaninputlinkofarouterissubjecttoIPlookuptodetermineitsoutputportandthenmustbetransferredviathegutsoftherouter,oftencalledaswitch.Earlyswitchesinthe90s,suchastheCatalyst5KfromCiscodesignedbyTomEdsall,usedasimplebusakintotheolderPCIbusinaCPU.Butasspeedswentup,designersrealizedtheyhadtouseacrossbar,whichisasetofparallelbusses.Thesimplesttechniquetoscheduleacrossbarusesin-putqueuing,wherepacketswaitingforanoutputareplacedinasinglequeueattheinputlink.However,thatmeantthatapacketonaninputin-terfacedestinedtoaredoutputinterfacecouldwait(attheinput)behindapacketdestinedtoabusyblueout-putinterface,eventhoughtheredoutputinterfacewasfree.Thisistheso-calledHead-of-Lineblockingprob-lemwhichcanreducethroughputbynearlyhalf.Thisproblemresultedinresearchersproposingmorecomplexoutputqueuingdesigns.But,asfarasIknow,outputqueuingdesignsneverenteredthemainstreamroutermarketbecauseoftheircomplexity.Abreakthroughoccurred,orsoitseemstome,around1992whenTomAnderson,ChuckThacker,andothersfromDECSRC[8]introducedtwonewideas.First,theychangedtheFIFOinterfacetooneinterfaceateachinputforeachoutput,sometimesreferredtoasVOQs(virtualoutputqueues),asshowninFigure6.Theninformationaboutallnon-emptyVOQsissenttoascheduler.Theirsecondremarkableideawasamaximalmatch-ingalgorithmcalledPIM[8]thatcouldbedoneinhardwareinaround5iterations.OnecanthinkoftheirapproachasanEthernet-likeapproachperout-putport.Eachoutputportrandomlyselectsamongallinputportsthatwishtosendtoit;inputportsrejectedbecauseof\collisions"retryinthenextiteration. Figure6:VirtualOutputqueuesandMaximalMatch-ingeliminateHeadofLineBlockingAfewyearslaterNickMcKeownintroducediSLIP[25]whichcanberoughlythoughtofasatokenringequiv-alentoftheEthernet-likeapproachofPIM.Thesched-ulergrantsaccesstoeachoutputlinkbasedonarotat-ingprioritythatcyclesthroughtheinputports.WhileiSLIPcanstartbadly,itgenerallyconvergestoverygoodmatchesafterafewiterations.iSLIPwasusedintheCiscoGSR.Whatiscompellinghereisnotjusttheimpactintermsofswitchperformancebutthetransformedidea.Maximummatchisnormallyatleastlinearinthenum-berofinputsintheory.PIM,andtheniSLIP,intro-ducednewalgorithmsformaximalmatchthatcom-putedamatchinnanosecondsandworkedverywellinpractice.WhilePIMandiSLIPworkwellforsmallswitchesandunicasttrac,JonTurnershowedhowtobuildscalablebroadcastswitches[28].Betweenalgorithmicsforfastswitching,IPLookups,andpacketscheduling,therewasgenerallyasensebythe2000sthatpeopleknewhowtobuildfastrouters.TherewasCisco'sCatalyst6K(therstrouterwithacrossbar),Juniper'sM40(arguablytherstroutertouseASICsforforwarding),andCisco'sGSR(oneoftherstcommercialrouterstoavoidhead-of-lineblocking).Betterstill,thesolutionsscaledwithlinkspeeds.Sili-conValley,withalittlehelpfromacademia,hadguredoutrouters.Wastherenothingleftforrouteralgorith-mics?4.MEASUREMENTALGORITHMICSIntheearly2000,anewin ectionpointarrivedforrouters.TheperilsthataccompanysuccessbesettheInternetintermsoflargescaleattackssuchaswormsandDenial-of-Serviceattacks.Compoundingthesetac-ticalissueswasthestrategicdicultyofndingtraf-cestimatestoprovisionnetworks.Itturnsoutthatbothdetectingsecurityattacksandner-grainmea-surementrequiredetectingpatternsacrosspackets,amilieuchangefromstandardalgorithmicsthatonlyde-tectspatterns(suchasprexes)withinpackets.Naivemethodsrequirekeepingmassiveamountofstateacrosspackets.However,theoreticalboundsshowthatrandomizedalgorithmscandomuchbetter.Toil-lustratemeasurementalgorithmicsasacon uence(Fig-ure8)betweenrandomizedalgorithmsandalgorith-mics,considerthefollowingalgorithmcalledSample-and-Hold,whichdiersfromstandardsampling,asem-ployedinsayCisco'sNetFlow[1].Considertheproblemofestimatingthetracsentbytheheavy\elephant" owsuchasF1withoutkeepingtrackofpotentiallymillionsof\mice" owslikeF3.ThebasicideainSample-and-Hold[12]isthatordinarysamplingisusedtodecidewhethera owlikeF1issampled.ButonceF1issampled,itisplacedinahash Figure7:MeasurementAlgorithmicsasaCon uencetable(FlowTable)andallsubsequenttracsentbyF1iswatched.Unlikesampling,theuncertaintyinthemeasurementofF1occursonlyatthestartandthistranslatesintoareductionofstandarderrorfrom1=p Mto1=M,whereMistheamountofmemoryavailableforthe owtable.Mistypicallylimited,beinghighspeedon-chipmem-ory.Hence,areductionoferrorfrom1=100to1=10;000isappreciable.Ofcourse,therealedgeofSample-and-Holdoversamplingoccursbecauseofamilieuchange:theroutergetstoseeeverypacket.Bycontrast,theGallupPollndsitexpensivetosurveyindividuals,andhencemustresorttostandardsampling. Figure8:Sample-and-Holdsamples owsbutthenkeepstrackofallsubsequentpacketsforeachsampled owSample-and-HoldwasrstdescribedinaverynicepaperbyGibbonsandMathiasondatabases[15].We[12]did,however,addanewanalysisandmadeotherchangestotthenetworkingmileu.Measurementandsecurityalgorithmicshavebeenwelldevelopedbymanyresearchers.Thefollowingisasam-pleofthreepiecesofworkIlike.First,Elephanttraps[17]improvesample-and-holdbyperiodicallyremovingmice owsthathavedriftedintothe owtablebyrandomchance.Next,researchershavegonebeyondheavy-hittermeasurementstoobtainmorecomplexestimatesincluding owdistributions[22]usingsmallspace.Finally,super-spreaderalgorithms[32]detectmorecomplexsecuritypredicatesinstreamingfashionsuchassourcesthatsendpacketstoalargenumberofdes-tinations,sometimesindicativeofanintrudertryingtobreakintoanycompromisedmachine.Someaspectsoftheseideasareinchipsandsoftware;forexample,Ciscofabricatedachipbasedonautomatedwormdetectiontechnology[27].Despitethis,Idonotthinkthatmea-surementalgorithmicsismainstreamasyet.5.OTHERNETWORKINGCONFLUENCESCon uencesinnetworkingarenotnew.Examplesofpastcon uenceswiththeimpactingeldshowninparenthesesincludequeuingnetworks(QueuingThe-ory),PricingtheInternet(Economics),andNetworkSecurity(ComputerSecurity).Morecurrentcon u-encesincludeDataCenterNetworks(HighPerformanceComputing)andWirelessNetworkCoding(Informa-tionTheory).Onecanarguethateachcon uencewastriggeredbyanin ectionpointsuchastheneedtounderstandpacketdelaysintheearlyInternet(queuingnetworks),theshifttothecommercialInternet(InternetPricing),theadventoflargescaleattacksandcyber-crime(Net-workSecurity),theneedforlargescaledatacenterstosupportCloudServices(DataCenterNetworks),andtheadventofSoftwareDenedRadiostogetherwiththedearthofwirelessspectrum(WirelessNetworkCoding).Eachintroducesamilieuchangefromtheirimpact-ingelds:networksofqueues,thereducedimportanceofmarginalcostsinInterneteconomics,theabilitytorapidlyamplifyasecurityattack,theneedtoscaleclus-terstohundredsofthousandsofnodes,andthepossibil-ityofembracingwirelessinterferenceinsteadofshun-ningit.Eachcon uencehasproducednewideasin-cludingtheindependenceassumption[20],edgepric-ing[26],ecosystemanalysis[23],datacentertransportswithlowerlatencythanTCP[7],andZigZagcodingtorecoverinformationinthefaceofcollisions[16].Anareathatexcitesmepersonallyiswhatseveralresearcherscall\NetworkVerication".NickMcKe-own'sSIGCOMMKeynotetwoyearsago,describedtheopportunityinnetworkvericationbycomparingittohardwareandsoftwareverication[24].Networkver-icationcanalsobeviewedasacon uencebetweenProgrammingLanguagesandNetworkingasshowninFigure9.Thein ectionpointthatmakesNetworkVericationcompellingistheemergenceofcloudservices.Stud-ies[35]haveshownthatnetworkfailuresareasigni-cantanddebilitatingsourceofoperationalexpendituresthatcanimpacttheeconomicviabilityofsuchservices. Figure9:NetworkVericationasaCon uenceOntheotherhand,theeldofprogramminglanguagehasproducedavarietyoftoolsfromdebuggerstostaticcheckers.Networkvericationseekstondanalogoustoolsfornetworks.Themilieuchangesingoingfromprogramstonet-works.Networkscanberegardedasprogramsthattransformpackets,andsuchnetwork\programs"typi-callyhavesimplecontrol ow.However,thelargepos-siblespaceofpacketheaderscomplicatesthetaskcom-paredeventolarge-scalesoftwaresuchasoperatingsystems.Newideashaveemergedinthiscon uenceincludingnewformsofcompressiontocompactlyrep-resenttheheaderspace[19,18],theautomaticsynthe-sisofforwardingrulesatrouters[17],theextensionofthenotionoftestcoveragetocoveringlinksandrouterqueues[35],andtheuseofcausalityinnetworkdebug-ging[13].IstartedwithlifeinthefastlaneandIendedinthe2000swithmeasurementalgorithmicsandsecurityal-gorithmicsbothofwhichhavebeenwellstudied.Arethereanyunexploreddirectionsfornetworkalgorith-mics?OneIndpromising,acon uencebetweennet-workalgorithmicsandvirtualizationasdepictedinFig-ure10,wasbroughttomyattentionbyDanielFire-stone,RamanaKompella,andSylviaRatnasamy.Themilieuchangeistheplacementofnetworkfunc-tions[2]invirtualmachinesrunningonmultiplecoresinsteadofonpipelinedrouterhardware,onceagaincausedbythein ectionpointofcheapcloudservices.Examplesoftransformedideasinthisspaceincludere-thinkingTCPperformance[14]invirtualizedenviron-mentsandtheRouteBricksapproachtosoftwarerouterdesign[9].6.CONCLUSIONNetworkalgorithmicshasplayedoutfromafocusonspeedandscaleinthe90stoafocusonmeasure-mentandsecurityalgorithmicsinthe2000s.Acon u-encethatsuggestsnewproblemsinnetworkalgorith- Figure10:NetworkingusingVirtualMachinesasaCon uencemicsmayarisefromthein ectionpointcausedbynet-workprocessinginsoftwareonvirtualmachines.Be-sidescurrentcon uencesinnetworkingsuchasDataCenterNetworkingandWirelessNetworkCoding,Net-workVericationisapromisingnewcon uence.Con uencethinkingisusefulbecauseitallowsare-searchertodiscernanewdirection,ndaunifyingtheme,andproduceresearchthatbalancesbeauty(viatrans-formedideas)andimpact(viathein ectionpoint).Themilieuchangeallowsrethinkingideasinboththeexist-ingandimpactingeldstoproduceresearchideallyofinteresttobothcommunities.Ofcourse,thisbegsthequestion:aretheremoresystematictechniquestousecon uencethinkinginresearch?Theslidespublishedonline[30]havesomehintssuchas\embracecollisions"and\seekcoherence".Thisarticlebeganwithareviewofnetworkalgorith-micsbutgraduallyseguedtoaframeworkforthinkingaboutinterdisciplinaryresearch.Perhapstheultimateexcitementisnotmakingthingsfast,butthethrillofdiscerninganeweldwithideastoexploreandimpactthatpotentiallyawaits.Ihopeseekingcon uenceswillprovidereaderswiththatsamerush.7.REFERENCES[1]Net ow.http://en.wikipedia.org/wiki/NetFlow.[2]NetworkFunctionsVirtualizationhttp://en.wikipedia.org/wiki/Network_Functions_Virtualization.[3]Remotedirectmemoryaccess.http://en.wikipedia.org/wiki/Remote_direct_memory_access.[4]V.Jacobson'snotesonTCPheaderprediction.http://yangchi.me/v-jacobsons-notes-on-tcp-header-prediction.html.[5]VirtualInterfaceArchitecture.http://en.wikipedia.org/wiki/Virtual_Interface_Architecture.[6]Zero-copy.http://en.wikipedia.org/wiki/Zero-copy.[7]M.Alizadehandetal.DatacenterTCP(DCTCP).InProceedingsofACMSIGCOMM2010.[8]T.Anderson,S.Owicki,J.Saxe,andC.Thacker.High-speedswitchschedulingforlocal-areanetworks.ACMTrans.Comput.Syst.,11(4),Nov.1993.[9]M.Dobrescuandetal.Routebricks:Exploitingparallelismtoscalesoftwarerouters.ProceedingsofSOSP'09. [10]P.DruschelandL.Peterson.Fbufs:Ahigh-bandwidthcross-domaintransferfacility.SIGOPSOper.Syst.Rev.,27(5),Dec.1993.[11]P.Druschel,L.Peterson,andB.Davie.Experienceswithahigh-speednetworkadaptor:Asoftwareperspective.SIGCOMMComput.Commun.Rev.,24(4),Oct.1994.[12]C.EstanandG.Varghese.Newdirectionsintracmeasurementandaccounting.SIGCOMMComput.Commun.Rev.,32(4),Aug.2002.[13]R.Fonseca,G.Porter,R.Katz,S.Shenker,andI.Stoica.X-trace:Apervasivenetworktracingframework.InProceedingsofthe4thUSENIXNSDI,2007.[14]S.Gamage,R.Kompella,D.Xu,andA.Kangarlou.ProtocolresponsibilityooadingtoimproveTCPthroughputinvirtualizedenvironments.ACMTrans.Comput.Syst.,31(3),Aug.2013.[15]P.GibbonsandY.Matias.Newsampling-basedsummarystatisticsforimprovingapproximatequeryanswers.InProceedingsofSIGMOD1998.[16]S.GollakotaandD.Katabi.Zigzagdecoding:Combatinghiddenterminalsinwirelessnetworks.InProceedingsofACMSIGCOMM2008.[17]N.Kang,Z.Liu,J.Rexford,andD.Walker.Optimizingthe"onebigswitch"abstractioninsoftware-denednetworks.InProceedingsoftheNinthACMConferenceonEmergingNetworkingExperimentsandTechnologies,2013.[18]P.Kazemian,G.Varghese,andN.McKeown.Headerspaceanalysis:Staticcheckingfornetworks.InProceedingsofthe9thUSENIXConferenceonNetworkedSystemsDesignandImplementation,2012.[19]A.Khurshid,W.Zhou,M.Caesar,andN.Godfrey.Veri ow:Verifyingnetwork-wideinvariantsinrealtime.InProceedingsoftheFirstWorkshoponHotTopicsinSoftwareDenedNetworks,2012.[20]L.Kleinrock.Theory,Volume2,ComputerApplications.Wiley-Interscience,1975.[21]N.Kronenberg,H.Levy,andW.Strecker.VAXcluster:Aclosely-coupleddistributedsystem.ACMTrans.Comput.Syst.,4(2),May1986.[22]A.Kumar,M.Sung,J.Xu,andJ.Wang.Datastreamingalgorithmsforecientandaccurateestimationof owsizedistribution.SIGMETRICSPerform.Eval.Rev.[23]K.Levchenkoandetal.Clicktrajectories:End-to-endanalysisofthespamvaluechain.InProceedingsofthe2011IEEESymposiumonSecurityandPrivacy.[24]N.McKeown.Mindthegap.http://yuba.stanford.edu/~nickm/talks/Sigcomm%202012%20POSTED.ppt.[25]N.McKeown.iSLIP:Aschedulingalgorithmforinput-queuedswitches.IEEE/ACMTrans.Netw.,7(2),Apr.1999.[26]S.Shenker,D.Clark,D.Estrin,andS.Herzog.Pricingincomputernetworks:Reshapingtheresearchagenda.SIGCOMMComput.Commun.Rev.,26,1996.[27]S.Singh,C.Estan,G.Varghese,andS.Savage.Automatedwormngerprinting.InProceedingsofthe6thConferenceonSymposiumonOperatingSystemsDesign&Implementation,2004.[28]J.Turner.Designofabroadcastpacketswitchingnetwork.InProceedingsofInfocom1986.[29]P.vanEmdeBoas.Preservingorderinaforestinlessthanlogarithmictime.InProceedingsofthe16thAnnualSymposiumonFoundationsofComputerScience,SFCS'75,1975.[30]G.Varghese.Lifeinthefastlane.http://conferences.sigcomm.org/sigcomm/2014/doc/slides/2.pdf.[31]G.Varghese.NetworkAlgorithmics.Morgan-Kaufman,2004.[32]S.Venkataraman,D.Song,P.Gibbons,andA.Blum.Newstreamingalgorithmsforfastdetectionofsuperspreaders.IninProceedingsof15thIEEESymposiumonHighPerformanceInterconnects,2007.[33]T.vonEickenandetal.Activemessages:Amechanismforintegratedcommunicationandcomputation.SIGARCHComput.Archit.News,20(2),Apr.1992.[34]M.Waldvogel,G.Varghese,J.Turner,andB.Plattner.Scalablehighspeediproutinglookups.InProceedingsoftheACMSIGCOMM'97,1997.[35]H.Zeng,P.Kazemian,G.Varghese,andN.McKeown.Automatictestpacketgeneration.InProceedingsofthe8thInternationalConferenceonEmergingNetworkingExperimentsandTechnologies,2012.