opnet业务流自相似性

OntheSelf-SimilarNatureofEthernetTrafficWillE.Leland†MuradS.Taqqu§[e-mail]@bu-ma.bu.eduWalterWillinger†DanielV.Wilson†[e-mail]@bellcore.com†Bellcore§DepartmentofMathematics445SouthStreetBostonUniversityMorristown,NJ07960-6438Boston,MA02215AbstractWedemonstratethatEthernetlocalareanetwork(LAN)trafficisstatisticallyself-similar,thatnoneofthecommonlyusedtrafficmodelsisabletocapturethisfractalbehavior,andthatsuchbehaviorhasseriousimplicationsforthedesign,control,andanalysisofhigh-speed,cell-basednetworks.Intuitively,thecriticalcharacteristicofthisself-similartrafficisthatthereisnonaturallengthofaburst:ateverytimescalerangingfromafewmillisecondstominutesandhours,similar-lookingtrafficburstsareevident;wefindthataggregatingstreamsofsuchtraffictypicallyintensifiestheself-similarity(burstiness)insteadofsmoothingit.OurconclusionsaresupportedbyarigorousstatisticalanalysisofhundredsofmillionsofhighqualityEthernettrafficmeasurementscollectedbetween1989and1992,coupledwithadiscussionoftheunderlyingmathematicalandstatisticalpropertiesofself-similarityandtheirrelationshipwithactualnetworkbehavior.Wealsoconsidersomeimplicationsforcongestioncontrolinhigh-bandwidthnetworksandpresenttrafficmodelsbasedonself-similarstochasticprocessesthataresimple,accurate,andrealisticforaggregatetraffic.1.INTRODUCTIONThemainobjectivesofthispaperare(i)toestablishinastatisticallyrigorousmannertheself-similaritycharacteristicor,touseamorepopularnotion,thefractalnatureofthehightime-resolutionEthernettrafficmeasurementsofLelandandWilson(1991),(ii)toillustratesomeofthemoststrikingdifferencesbetweenself-similarmodelsandthestandardmodelsforpackettrafficcurrentlyconsideredintheliterature,and(iii)todemonstratesomeoftheseriousimplicationsofself-similarnetworktrafficforthedesign,control,andperformanceanalysisofhigh-speed,cell-basedcommunicationssystems.Intuitively,self-similarphenomenadisplaystructuralsimilaritiesacrossall(oratleastaverywiderangeof)timescales.InthecaseofEthernetLANtraffic,self-similarityismanifestedintheabsenceofanaturallengthofaburst;ateverytimescalerangingfromafewmillisecondstominutesandhours,burstsconsistofburstysubperiodsseparatedbylessburstysubperiods.Wealsoshowthatthedegreeofself-similarity(definedviatheHurstparameter)typicallydependsontheutilizationleveloftheEthernetandcanbeusedtomeasureburstinessofLANtraffic.Thetermself-similarwasformallydefinedbyMandelbrot.Forapplicationsandreferencesonthetheoryofself-similarprocesses,seeMandelbrot(1983)andtheextensivebibliographybyTaqqu(1985).Foranearlyapplicationoftheself-similarityconcepttocommunicationssystems,seetheseminalpaperbyMandelbrot(1965).Inthispaper,weuseveryhighquality,hightime-resolutionLANtrafficdatacollectedbetweenAugust1989andFebruary1992onseveralEthernetLANsattheBellcoreMorristownResearchandEngineeringCenter(MRE).LelandandWilson(1991)presentapreliminarystatisticalanalysisofthisuniquehigh-qualitydataandcommentindetailonthepresenceofburstinessacrossanextremelywiderangeoftimescales:trafficspikesrideonlonger-termripples,thatinturnrideonstilllongertermswells,etc.Thisself-similarorapparentlyfractal-likebehaviorofaggregateEthernetLANtrafficisverydifferentbothfromconventionaltelephonetrafficandfromcurrentlyconsideredformalmodelsforpackettraffic(e.g.,purePoissonorPoisson-relatedmodelssuchasPoisson-batchorMarkov-ModulatedPoissonprocesses(HeffesandLucantoni(1986)),packet-trainmodels(JainandRouthier(1986)),andfluidflowmodels(Anicketal.(1982)),etc.).Thesedifferencesrequireanewlookatmodelingthetrafficandperformanceofbroadbandnetworks.Forexample,ouranalysisoftheEthernetdatashowsthatthegenerallyacceptedargumentforthePoisson-likenatureofaggregatetraffic,namely,thataggregatetrafficbecomessmoother(lessbursty)asthenumberoftrafficsourcesincreases,hasverylittletodowithreality.Infact,theburstiness(degreeofself-similarity)ofLANtraffictypicallyintensifiesasthenumberofactivetrafficsourcesincreases,contrarytocommonlyheldviews.BecauseofthegrowingmarketforLANinterconnectionservices,LANtrafficisrapidlybecomingoneofthemajorpotentialtrafficcontributorsforhighspeednetworksofthefuturesuchasB-ISDN.Anotherexpectedmajorcontributorisvariable-bit-rate(VBR)videoservice.SinceVBRvideotraffichasrecentlybeenshowntodisplaythesameself-similaritypropertyasLANtraffic(seeBeranetal.(1992)),self-similarmodelsprovidesimple,accurate,andrealisticdescriptionsoftrafficscenariostobeencounteredduringhigh-bandwidthnetworkdeployment.Inlightofthisnewunderstandingofthenatureofbroadbandtraffic,wealsoaddresssomeoftheseriousimplicationsofself-similartrafficforissuesrelatedtothedesign,control,andperformanceanalysisofhigh-speed,cell-basednetworks.Asonespecificexample,weconsidertheareaofcongestionmanagementandshowthatthenatureofcongestionproducedbyself-similartrafficdiffersdrasticallyfromthatpredictedbytrafficmodelscurrentlyconsideredintheliteratureandisfarmorecomplexthanhasbeentypicallyassumedinthepast.Asaresult,proposedcongestioncontrolschemesthatworkwellassumingconventionaltrafficmodelstypicallyperformlessthansatisfactorilyinaself-similartrafficenvironment.Fi