最小化分布式计算系统中CPU可用性预测误差的自动预测方法(IJISA-V8-N9-2)

I.J.IntelligentSystemsandApplications,2016,9,8-21PublishedOnlineSeptember2016inMECS()DOI:10.5815/ijisa.2016.09.02Copyright©2016MECSI.J.IntelligentSystemsandApplications,2016,9,8-21AutomatedForecastingApproachMinimizingPredictionErrorsofCPUAvailabilityinDistributedComputingSystemsN.ChabbahSekmaNationalEngineeringSchoolofTunis,UniversityofTunisElManar,Tunis,1002,TunisiaE-mail:[e-mail],[e-mail]puterSciences,UniversityofManouba,Manouba,2010,TunisiaE-mail:[e-mail]—ForecastingCPUavailabilityinvolunteercomputingsystemsusingasinglepredictionalgorithmisinsufficientduetothediversityoftheworld-widedistributedresources.Inthispaper,wedraw-upthemainguidelinestodevelopanappropriateCPUavailabilitypredictionsystemforsuchcomputinginfrastructures.Toreducesolutiontimeandtoenhanceprecision,weusesimplepredictiontechniques,preciselyvectorautoregressivemodelsandatendency-basedtechnique.Weproposeapredictorconstructionprocesswhichautomaticallychecksassumptionsofvectorautoregressivemodelsintimeseries.Threedifferentpastanalysesareperformed.Foragivenvolunteerresource,theproposedpredictionsystemselectstheappropriatepredictorusingthemulti-statebasedpredictiontechnique.Then,itusestheselectedpredictortoforecastCPUavailabilityindicators.Weevalua[e-mail]edictionsystemimprovesthepredictionaccuracybyaround24%.IndexTerms—CPUavailabilityprediction,predictionsystem,multivariatetimeseries,multi-statebasedprediction,volunteercomputingsystem.I.INTRODUCTIONManyresourcesconnectedtotheInternetareidleformostofthetime.Theyrepresentaconsiderablereserveofcomputingpower.Volunteercomputing(VC)systemsaimtoharnessthisextensivenumberofunderusedcomputerresourcesandtoreachahighcomputingperformance.Whiletheseworld-widedistributedresourcesareheterogeneous,unreliableandbelongtoindependentadministrativedomains,appropriatemiddlewareisdeployedtoaggregate,on-demand,theunusedprocessingpower.Tasks,submittedtoaVCsystembyindependentusers,shouldbescheduledontheappropriatecomputingresources.Howevertheiravailability,forVCsystemusage,ishighlyvariabledependingondemand,owners’behavior,theirtimezonesandtheirlocation(athome,schoolorwork),etc.[1,2,38].Consequently,theschedulerhasnoavailabilityorspeedguarantees.TheschedulingoptimizationinsuchenvironmentsrequiresforecastingthefutureCPUresourceavailability.AreviewofrelatedworksshowsthatthereisnosinglepredictionmodelwhichisoptimalforalltheconsideredCPUtimeseries[3,4,5,7].Duetothediversityofworld-widedistributedresources,thepredictionaccuracyisnotalwaysensuredusingasinglepredictor.Forsuchcomputingresources,thepredictionsystemshouldbeabletoselectautomaticallytheappropriatepredictorforeachCPUresourceamongseveralintegratedpredictors.Besides,usualpredictionsystemsaretimeconsumingandconsequentlyinappropriateforlarge-scalecomputinginfrastructures[3,8].Inthiswork,weareparticularlyinterestedinpredictingCPUavailabilityofvolunteerresourcesinlarge-scaleVCsystems.Foreachcomputingresource,wepredictpreciselytwoCPUavailabilityindicators(i.e.variables)thatarethenumberandthemeandurationofCPUavailabilityintervalsoverthenexthour.Toreducethesolutiontime,welimitourstudytosimpleapproaches,whichmayoutperformthemostcomplexcompetitors[5,9,10]andensurereasonableaccuracies.Weextendtheapproachproposedin[7,9]inordertodrawupguidelinestoconceiveapredictionsystemofresourceavailabilityinVCinfrastructures.Aspointedin[7],avolunteerresourcemaybeinoneofthethreefollowingstates:totallyavailable,totallyunavailableorpartiallyavailableoverthewholehour.Multi-statebasedpredictorsareappropriatetoforecastdiscretevaluescorrespondingtothepossibleavailabilitystatesofaresource.Inthispaper,weanalyzetheperformanceofseveralmulti-statepredictiontechniquesinordertoretainthemostaccurateones.Wenoticethattheiraccuraciesdependonthemeandurationoftheavailabilityandunavailabilityintervalsofvolunteerresources.Consequently,weproposeanautomatedapproachtoidentifytheappropriatemulti-statepredictiontechniqueforeachvolunteerresourceregardingitsavailabilityandunavailabilityfrequencies.AutomatedForecastingApproachMinimizingPredictionErrorsofCPU9AvailabilityinDistributedComputingSystemsCopyright©2016MECSI.J.IntelligentSystemsandApplications,2016,9,8-21Forthetotallyavailableandunavailablestates,thevaluesofCPUavailabilityindicatorsareknown.However,inthecaseofthethirdavailabilitystate(partiallyavailable),thevolunteerresourceisunavailableduringsomeintervalsofthehour.So,CPUavailabilityvariablescorrespondtocontinuousvaluedata.Inordertopredicttheirvalues,werequirepredictorssuchastimeseriesmodels.Tendency-basedstrategyhasbeenconsideredasanautomated,simpleandimprovedpredictiontechniquereferencedinmanyrecentCPUloadpredictionresearches[11,4].Autoregressivemodelshavebeenshowntobeamongthesimplesttimeseriesmodelsusingbothautocorrelationandcross-correlationbetweenmultivariatetimeseries[12,7].Theyareasaccurateasthemostcompl