Huang Dept of Electrical Computer Engineering Dept of Computer Science University of Rochester Rochester NY 14627 USA pariharece cdingcs michaelhuang rochesteredu Abstract A hot concept in program optimization is hotness For example program optimiz ID: 42415
Download Pdf The PPT/PDF document "A Coldness Metric for Cache Optimization..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Astheoptimizationlevelincreases,thecoldnessagainde-creases.Forthe4MBcache,wemustoptimizetheaccesstoatleast344KB,2.4MB,and5.4MBdatatoreducethemissratioby10%,50%,and90%respectively.Inthelastcase,thecoldnessmet-ricshowsthatitisnecessarytooptimizeforadatasizemorethanthecachesizetoobtaintheneededreduction.Nextwedescribetheexperimentsthatproducedthesecoldnessdata.OursimulationframeworkisbasedonSimpleScalarandwemodelaPOWER7likemicroarchitecture.Onanaverage,wefastforwardeachapplicationsabout5billioninstructionsbeforewebegincollectingthestatisticsfora200millioninstructionwindow.DatacachesinthisstudyarefullyassociativewithLRUreplace-mentpolicy.Thisensuresthatallthemissesarecapacityandcom-pulsorymissesofthedatawhosereusedistanceislargerthanthecachesizeandtherearenoconictmisses.InFigure2,wepresenttheminimalnumberofdistinctaddressesthataccountforagivenpercentageofcachemisses.Thecoldnessmetricisthenegationofthisnumber.Theaveragenumberofmostmissedaddressesin-creasebyabout100xfortop10%and50%missesasthecachesizeincreases. (a)Distinctaddressesaccountingfortop10%misses (b)Distinctaddressesaccountingfortop50%missesFigure2:Distinctaddressesaccountingfortop10%and50%missescorrespondingtovariousreusedistance.Basedontheindividualresults,weclassifyapplicationsintotwogroups.Applicationsthatareconsistentlycolderthantheme-dianareknownasbelowmediancoldandapplicationsthatarecon-sistentlynotascoldasthemedianareknownasabovemediancoldapplications.Thenexttableshowsthetwocoldnesscategories. Temperaturezones SPEC2006Applications abovemedian h264ref,sphinx3,astar,xalancbmk, (lesswidespreadmisses) gobmk,hmmer,dealII,namd belowmedian lbm,bwaves,libquantum,perlbench, (morewidespreadmisses) zeusmp,gromacs,mcf,soplex,sjeng 3.DiscussionandFutureWorkThenewmetricaugmentsthegrowingsetofdata-centricmetrics.Reusedistancehasbeenusedtoshowthetemporalandspatiallo-calityforindividualorcollectionofdata.Hotdatastreamsshowedtheregularityinconsecutivedataloads[5].Arecenttool,HPC-Toolkit,showsmostmisseddatatohelpprogramtuning[8].Unlikecoldness,theothermetricsdonotquantifytheminimalnumberofdistinctdataaddressesthatmustbetargetedbyanoptimization.Asfuturework,weplantocalibratethecoldnessmetricmorethoroughlyusingthefullprogramtraceandmeasuringtheeffectofprograminputandcacheassociativity.Wewillstudyhardwaresolutions.FromFigure2,itisevidentthattherearenumerousdistinctaddresseswhichaccountfortopmisses.Thesenumberofaddressesincreaserapidly(showninlog-scale)asthecachesizeincreases.Fromasimilarstudy,wealsoobservedthatthesemissesareincurredbyalargenumberofdis-tinctstaticinstructions,notjustafewdelinquentinstructions.Apossiblesolutionismoreeffectiveprefetching.Aspecicimple-mentationoflook-ahead,whichwecalldecoupledlook-ahead[4],isabletoreducetheprimarymissesby88xandsecondarymissesby38xforacachesizeof4MBthatincursabout100,000distinctmissaddressesintop90%misses.Wewillstudythenewmetricasaguidetoprogramoptimiza-tion.TheprogrammcfinFigure2showstwodistinguishingchar-acteristics.First,itisoneoftheprogramsthatarebelowmediancold,whichmeansthatitsmissesaremorewidespread.Second,for50%miss-ratioreduction,itscoldnessisamongtheleastvary-ingacrossdifferentcachesizes.Aneffectiveoptimizationhasbeenfoundformcf,upto35%improvementbystructuresplitting[7].Thecoldnesscharacteristicssuggestthatstructuresplittingisef-fectiveinremovingwidespreadmissesbutmaybeapplicableonlyforcertaintypeofprograms.Wewillalsousethecoldnessmetricprogramtuningtoestimatethedifcultyandsuggestdatatargets.Insummary,wehopetoidentifyeffectivetechniquestodealwithcolddata,andgeneralizeandimprovethesetechniques.Aswefaceincreasinglyseverememoryproblems,wemustunderstandandexpandthewaystooptimizeforextremelycolddata.References[1]D.Callahan,J.Cocke,andK.Kennedy.Estimatinginterlockandimprovingbalanceforpipelinedarchitectures.JPDC,5(4),1988.[2]D.CallahanandJ.Gray.Designconsiderationsforpar-allelprogramming,2008.http://msdn.microsoft.com/en-us/magazine/cc872852.aspx.[3]J.F.CantinandM.D.Hill.CacheperformanceforSPECCPU2000benchmarks.http://www.cs.wisc.edu/multifacet/misc/spec2000cache-data.[4]A.GargandM.Huang.APerformance-CorrectnessExplicitlyDecou-pledArchitectureInProc.Int'lSymp.onMicroarch.,pages306–317,November2008.[5]T.M.ChilimbiandM.Hirzel.Dynamichotdatastreamprefetchingforgeneral-purposeprograms.InPLDI,pages199–209,2002.[6]C.DingandK.Kennedy.Thememorybandwidthbottleneckanditsameliorationbyacompiler.InIPDPS,pages181–190,2000.[7]G.ChakrabartiandF.Chow.StructurelayoutoptimizationsintheOpen64compiler.InOpen64Workshop,2008.[8]X.LiuandJ.M.Mellor-Crummey.Pinpointingdatalocalityproblemsusingdata-centricanalysis.InCGO,pages171–180,2011.