一个在线拍卖系统的性能研究外文翻译可编辑.docx
《一个在线拍卖系统的性能研究外文翻译可编辑.docx》由会员分享,可在线阅读,更多相关《一个在线拍卖系统的性能研究外文翻译可编辑.docx(6页珍藏版)》请在冰豆网上搜索。
一个在线拍卖系统的性能研究外文翻译可编辑
一个在线拍卖系统的性能研究外文翻译(可编辑)
一个在线拍卖系统的性能研究外文翻译
外文翻译
PerformanceInvestigationofanOn-LineAuctionSystemMaterialSource:
////0>.
42.9402Author:
JaneHillston,LeilaKloulAbstract:
Thestandarddesignofon-lineauctionsystemsplacesmostofthecomputationalloadontheserveranditsadjacentlinks,resultinginabottleneckinthesystem.Inthispaper,weinvestigatetheimpact,intermsoftheperformanceoftheserveranditsadjacentlinks,ofintroducingactivenodesintothenetwork.TheperformancestudyofthesystemisdoneusingthestochasticprocessalgebraformalismPEPA.
1IntroductionInthispaperweinvestigatetheinterplayoftwoemergingtechnologies:
activenetworksandsoftwareagentstosupportelectroniccommerce.
Activenetworksareacompellingnewinitiativeinnetworking.Anactivenetworkextendsaconventionalonewiththeabilityfornetworkswitchestoprocessdataasitisbeingtransmitted.Theprocessingwhichistobeperformedcanbecustomisedbythenetworkuseronaper-applicationorevenper-messagebasis.Thisinnovationisadramaticdeparturefromtraditionalnetworkdesignwheretheemphasisisonthe
avoidanceofexaminationormodificationofdata.Activenetworksaresupportedbyavarietyofsoftwaretechnologies,executionenvironmentsandnodeoperatingsystemsSomerecentworkhasfocusedontheperformancegainsthatactivenetworktechnologymaybringtodistributedapplications.Examplesincludeactivereliable-multicastandcacherouting.Inon-lineauctionsaresuggestedasapplicationswhichmightbenefitfromprocessingatactivenodeswithinthenetworkTheInternetoffersexcitingnewprospectsforelectroniccommerce,butitcannotalwaysdelivertheperformancenecessarytomakethemviable.Theelectronictransactionstakeavarietyofformsbutincreasinglythereisamovetowardsagent-basedsystemsinwhichpersonalizedsemi-autonomoussoftwareagent’sactonbehalfofconsumersorbusinesses.Inmanycasessuchsystemsrelyontheexchangeofinformationandnegotiation.Iftheintegrityofsuchtransactionsistobemaintained,thereisaclearneedfortimelybehaviouroftheunderlyinginfrastructureSeveralon-lineauctionsystemshavebeendevelopedexperimentally,suchastheAuctionBotsystemfromtheUniversityofMichigan,theFishmarketProjectortheeAuctionHousewhichsupportscombinatorialauctionsandwhichisfromtheUniversityofWashington.Insuchsystems,competitivebehaviouronthepartofthebidderreliesonarapidresponsetosubmittedbids.Howeverthismaybejeopardizedbynetworklatencyand/orserveroverload.Assuggested,butnot
investigatedin,thein-networkprocessingcapabilitiesprovidedbyanactivenetworkappearstoprovideasolutiontothisproblem.AsfarasweareawarenothoroughperformanceanalysisofsuchascenariohasbeencarriedoutInthispaper,weinvestigatetheperformanceissuessurroundingsuchasituation.Theideawedevelopinvolvesreplacingstandard,basicintermediarynodesofthenetworkbyactivenodes,thegoalbeingtotransfersometasksfromtheservertothesenodes.Thisshouldresultinasignificantbenefitintermsofbothsystemthroughputandsystemlatency.TheresultingsystemisthenanalysesusingthestochasticprocessalgebramodelingformalismPEPAThepaperisorganizedasfollows.InSection2,wedescribetheon-lineauctionsystemweinvestigate,andthemotivationfortheapproachthatwetake.Then,inSection3,afterabriefintroductiontothemodelingformalismweuse,PEPA,wepresentthedetailsofourmodel.OursolutiontechniqueisoutlinedinSection4,togetherwiththeexperimentsweconductedandthenumericalresultswehaveobtained.Someconclusionsofthiswork,togetherwiththepossibleextensions,arediscussedinSection5.
2TheOn-lineAuctionSystemInanon-lineauctionsystem,aserverreceivesandprocessesbidsfromremotesoftwareagentsrepresentinginterestedconsumers.Thesesemi-autonomousagentssubmitbidsaccordingtoapredeterminedstrategytogetherwiththeinformationthattheycanascertainfromtheserver.Theserverprocessesbids,eitheraccepting
themorrejectingthem,dependingontheirvalue.InsomesystemsadditionalattributesmaybeconsideredwhencomparingbidsofthesameorclosevalueInadditiontobids,bidderagentsmayalsosubmitpricenotificationrequests,askingtheservertotellthemthelatestbiddingprice.Notethatthebidderagentscanneverbecertainthattheyhaveanaccuraterepresentationofthecurrentpriceduetonetworklatency.Theycan,however,becertainthattheircurrentrepresentationisoutofdatewhenasubmittedbid,whichishigherthantheirideaofthe'current'price,isrejectedTheeffectivenessofthebidderagentswilldependontheproportionoftimethattheirpriceinformationisaccurate.Maintainingsuchaccurateinformationplacesstringentperformancerequirementsontheunderlyinginfrastructure.Moreoverthescalabilityofsuchsystems,intermsofthenumberofbidderagentsthatcanbesatisfactorilyaccommodated,couldbeseverelylimitedbytheperformanceofthenetwork.Fromthepointofviewofaccessibilityitisimportantthatsuchauctionsystemsuseexistinginfrastructure,i.e.theInternet,andsotheabilitytodirectlyaddressperformanceproblemsmaybelimited.However,inthispaperweconsiderhowsuchperformancelimitationsmaybecircumvented,byincorporatingactivenodeswithinthenetworkThestandarddesignofanon-lineauctionsystemnecessarilyplacesmostofthecomputationalloadontheserveranditsadjacentlinks,formingabottleneckinthesystem.Weinvestigatetheadvantagesofintroducing
bidcachesatintermediarynodeswithinthenetworkbetweentheserverandsomebidders.Suchnodesare"active"inthesensethat,inadditiontorouting,theyexaminethecontentsofbidandpricerequestmessagesandformcaches,storingrecentbidandpriceinformation.Thesecachesdonothavetheabilitytoacceptbidsbutmayactasfilters,byrejectingbidswhichareknowntobetoolow.Thisreducestheloadonboththeserveranditsadjacentnetworkelements.Weassumethatthecache'spriceinformationisgleanedfromthebidswhichithandlestogetherwiththeirrespectiveresponses,butalsofromaperiodicupdatemessagesentfromtheserver.Forbidswhicharepassedtotheserverthelatencyduetoprocessingwithinthenetworkisincreased;howevertheintentionisthatthiswillbemorethancompensatedbythereducedtrafficreachingthebottleneckofthesystemInordertoquantifytheadvantagesofintroducinganactivenodetoactasacachefortheon-lineauctionsystem,wecomputeandcomparetheperformanceoftwosuchsystems:
onesystemdeployedonasimplenetworkwithoutactivenodesandanotheronthesamenetworkinwhichonenodebecomesactive.TheapproachwetaketodevelopperformancemeasuresisbasedonthestochasticprocessalgebraPEPAPerformanceEvaluationProcessAlgebra.PEPAservesasahigh-levelnotationforMarkovmodeling:
itispossibletoautomaticallygenerateacontinuous-timeMarkovprocessdirectlyfromthePEPAmodelwhichfaithfullyencodesthebehavioralandtemporalaspectsofthemodeled
system.Detailsofthismappingcanbefoundelsewhere.Otherhigh-levelnotationsforMarkovprocesses,suchasGSPNorSANcouldequallyhavebeenusedbutthecompositionalstructureofPEPAseemedwell-suitedtothestructureoftheauctionsystem.Usingaformalmodel,suchasprocessalgebra,allowsustoadditionallyverifythefunctionalcorrectnessoftheproposedsystem.Moreover,PEPAsupportsanautomaticaggregationtechniquewhichallowsthestatespaceofthemodeltobereducedwithoutlossofinformation,transparentlytotheuserOurstudyisdecomposedintotwoparts.Inthefirstpart,weconsiderthesystemwithonlytraditionalnodesasgivenbyFigure1.Wemodelthatsystemandcomputeitsperformanceintermsofthroughput.Inthesecondpartofthestudy,wereplaceoneofthetraditionalnodesbyanactiveone.Inthemodelsweadoptanumberofassumptionsandconventions.
Allbidderagentsinthesystemadoptthesamestrategy.Accordingtothisstrategy,wheneverabidisrejectedthebiddersubmitsapricenotificationrequesttogetanupdatedpriceestimatebeforesubmittinganymorebids
Intheserver,aservingagentisspawnedtocorrespondwitheachbidderagent.Thisagentisresponsibleformaintainingthecurrentstateofinteractionswiththecorrespondingbidder.
Inthemodelalldatadependentbehaviourisabstracted.Thismeansthatwedonotrepresentthecurrentprice,orthevalueofabid.Nor
dowerepresentdetailsofanybiddingstrategy,orselectionstrategyforchoosingbetweenbidsofcomparablevalue.Insteadweuseprobabilitiestorepresenttherelativefrequencywithwhichbidsaresuccessful.
Sincebidswhicharesubjecttolongerlatenciesaremorelikelytobeunsuccessful,weadjusttheacceptanceprobabilitiesaccordingtotheroutesbywhichbidswillarriveattheserver,orthecache.
Wemakeadistinctionbetweenthenodesofthesystem.Thisdistinctionreflectsthepositionofthenodeinthesystemandthustheelementstowhichitisconnected.However,notethattheessentialbehaviourandtimingcharacteristicsofthenodeswithoutcachearethesameineachcaseOurmodelisstochastic,meaningthatalltimesarerepresentedasrandomvariables.SincewewilluseMonroviananalysistocalculateperformancemeasures,allrandomvariablesareassumedtobeexponentiallydistributed.Performancemeasuresarederivedfromequilibrium,orsteadystate,behaviourInthefollowingsectionwebrieflyintroducethePEPAformalism,beforepresentingthemodels