通过预测模型探索高效的建筑设计空间

EfficientlyExploringArchitecturalDesignSpacesviaPredictiveModelingEngin˙IpekSallyA.McKeeCornellUniversity{engin,sam}@csl.cornell.eduBronisR.deSupinskiMartinSchulzLawrenceLivermoreNationalLaboratory{bronis,schulzm}@[e-mail]cycle-by-cyclesimulationtoevaluatedesignchoicesandunderstandtradeoffsandinteractionsamongdesignparam-eters.Efficientlyexploringexponential-sizedesignspaceswithmanyinteractingparametersremainsanopenproblem:thesheernumberofexperimentsrendersdetailedsimulationintractable.Weattackthisproblemviaanautomatedapproachthatbuildsaccurate,confidentpredictivedesign-spacemodels.Wesimulatesampledpoints,usingtheresultstoteachourmodelsthefunctiondescribingrelationshipsamongdesignparameters.Themodelsproducehighlyaccurateperformanceestimatesforotherpointsinthespace,canbequeriedtopredictperformanceimpactsofarchitecturalchanges,andareveryfastcomparedtosimulation,enablingefficientdiscov-eryoftradeoffsamongparametersindifferentregions.WevalidateourapproachviasensitivitystudiesonmemoryhierarchyandCPUdesignspaces:ourmodelsgenerallypredictIPCwithonly1-2%errorandreducerequiredsimulationbytwoordersofmagnitude.Wealsoshowtheefficacyofourtechniqueforexploringchipmul-tiprocessor(CMP)designspaces:whentrainedona1%sampledrawnfromaCMPdesignspacewith250Kpointsandupto55×performanceswingsamongdifferentsystemconfigurations,ourmodelspredictperformancewithonly4-5%erroronaverage.Ourapproachcombineswithtechniquestoreducetimepersimulation,achievingnettimesavingsofthree-fourordersofmagnitude.CategoriesandSubjectDescriptorsI.6.5ComputingMethod-ologies[SimulationandModeling]:ModelDevelopment;B.8.2Hardware[PerformanceandReliability]:PerformanceAnalysisandDesignAidsGeneralTermsDesign,Experimentation,MeasurementKeywordsdesignspaceexploration,sensitivitystudies,artificialneuralnetworks,performanceprediction1.IntroductionArchitectsquantifytheimpactofdesignparametersonevaluationmetricstounderstandtradeoffsandinteractionsamongthosepa-rameters.Suchanalysesusuallyemploycycle-by-cyclesimulationCopyright2006AssociationforComputingMachinery.ACMacknowledgesthatthiscontributionwasauthoredorco-authoredbyacontractororaffiliateoftheU.S.Government.Assuch,theGovernmentretainsanonexclusive,royalty-freerighttopublishorreproducethisarticle,ortoallowotherstodoso,forGovernmentpurposesonly.ASPLOS’06October21–25,2006,SanJose,California,USA.Copyrightc2006ACM1-59593-451-0/06/0010...$5.00ofatargetmachineeithertopredictperformanceimpactsofar-chitecturalchanges,ortofindpromisingdesignsubspacessatisfy-ingdifferentperformance/cost/complexity/powerconstraints.Sev-eralfactorshaveunacceptablyincreasedthetimeandresourcesre-quiredforthelattertask,includingthedesiretomodelmorere-alisticworkloads,theincreasingcomplexityofmodeledarchitec-tures,andtheexponentialdesignspacesspannedbymanyindepen-dentparameters.Thoroughstudyofevenrelativelymodestdesignspacesbecomeschallenging,ifnotinfeasible[22,16,5].Nonetheless,sensitivitystudiesoflargedesignspacesareoftenessentialtomakinggoodchoices:forinstance,Kumaretal.[20]findthatdesigndecisionsnotaccountingforinteractionswiththeinterconnectinaCMPareoftenoppositetothoseindicatedwhensuchfactorsareconsidered.Researchonreducingtimeperexperi-mentoridentifyingthemostimportantsubspacestoexplorewithinafullparameterspacehassignificantlyimprovedourabilitytocon-ductmorethoroughstudies.Evenso,simulationtimesforthoroughdesignspaceexplorationremainintractableformostresearchers.Weattackthisproblembyusingartificialneuralnetworks(ANNs)topredictperformanceformostpointsinthedesignspace.WeviewthesimulatorasanonlinearfunctionofitsM-parameterconfiguration:SIM(p0,p1,...pM).Insteadofsamplingthisfunc-tionateverypoint(parametervector)ofinterest,weemploynon-linearregressiontoapproximateit.Werepeatedlysamplesmallnumbersofpointsinthedesignspace,simulatethem,andusetheresultstoteachtheANNstoapproximatethefunction.Ateachteaching(training)step,weobtainhighlyaccurateerrorestimatesofourapproximationforthefullspace.WecontinuerefiningtheapproximationbytrainingtheANNsonfurthersamplepointsuntilerrorestimatesdropsufficientlylow.BytrainingtheANNson1-2%ofadesignspace,wepredictresultsforotherdesignpointswith98-99%accuracy.TheANNsareextremelyfastcomparedtosimulation(trainingtypicallytakesfewminutes),andourapproachisfullyautomated.CombiningourmodelswithSimPoint[34]reducesrequiredCPUtimebythree-fourordersofmagnitude,enablingdetailedstudyofarchitecturaldesignspacespreviouslybeyondthereachofcurrentsimulationtechnology.Thisallowsthearchitecttopurgemostoftheunin-terestingdesignpointsquicklyandfocusdetailedsimulationonpromisingdesignregions.Mostimportantly,ourapproachfundamentallydiffersfromheuristicsearchalgorithmsinscopeanduse.Itcancertainlybeusedforoptimization(predictedoptimumwith1%samplingiswithin3%ofglobaloptimumperformanceforapplicationsinourprocessorandmemorysystemstudies),butweprovideasuper-setofthecapabilitiesofheuristicsthatintelligentlysearchdesignspacestooptimizeanobjectivefunction(e.g.,thosestudiedbyEyermanetal.[10]).Spec