Rice ELEC 525 - Thereshold Based Markov Prefetchers

Unformatted text preview:

 1Threshold-BasedMarkovPrefetchersCarlosMarchaniTamerMohamedLerzanCelikkanatGeorgeAbiNaderRiceUniversity,DepartmentofElectricalandComputerEngineeringELEC525,Spring2006AbstractInthisreportwepresentanoveltechniqueforaMarkovprefetcher designed to hide memory latency. This pre-fetchermakesuseoftheinformationaboutpreviouscachemisses in a memory trace and effectively predicts futuremisseswithhighaccuracyandcoveragerate.Theanalysisofthememory tracesoftheexecutedbenchmarksrevealsthat our threshold-based prefetcher poses an optimizedmemory bandwidth overhead while not hurting the pre-fetcher coverage and effectively increasing its accuracy.Ourdataindicatethattheaveragelossincoverageis1%while the gain in accuracy is 13% and the reduction inmemory bandwidth overhead is 39%. The design alsoachieves more than an order of magnitude reduction inmispredictionsascomparedtopreviousresults.KeywordsHidingMemoryLatency,MarkovTablesBACKGROUNDAsmodernsuperscalarprocessorsgethigherclockspeedsanddeeperpipelinesthecostofmemorylatencybecomesanincreasingburdenonthesystemperformance.Prefetch-ersareamechanismthataimstohidethememorylatencyinmodernprocessorsbyfetchingdatafrommemorybeforethe processor actually requests it. The data is typicallyfetched following a model pre-established by the systemdesigneranditisstoredindedicatedbuffers.Severalresearchershaveworkedondifferentmodelsforaprefetcher design, including compiler-dependant prefetch-ers,strideprefetchers,streambuffersandcorrelation-basedprefetchers. Joseph and Grunwald studied Markov pre-fetchersasamodelforcorrelationprefetchers[1].Intheirwork,theyconsideredandsimulatedtheaboveprefetchersinadditiontoaMarkovmodel.TheyfoundthattheMarkovprefetcherprovidedthebestperformancefortheSPEC95applicationsthatwereconsidered.Nonetheless,theyrecog-nizedtheneedtoreducethebandwidthusageandthepos-sibilitytoimprovetheprefetcheraccuracy.The main idea behind utilizing a Markov table in a pre-fetcheristoaccountfortheprobabilitiesoftransitionsfromonestatetoanotherstate.Inthiscontext,astatereferstoacertain missaddress following a givenmissaddress. Thefrequencies of transition are used to populate a table asshown in table 1, and these frequencies are utilized as ameasureof probability forthe prediction of futuretransi-tionsfromareachedstate.Thetableiscontinuouslybeingupdatedandastrideprefetcherisalsousedinordertoen-hancetheperformanceofbenchmarksthatexhibitalotofstridedaccesses.Table 1: A Markov table populated by the transitions andfrequencyoftransitionoccurrenceforthesequenceofstates(cachemisses):A,B,C,D,C,E,A,B,CMissAddress(currentstate)NextMiss(state)AB[2]  BC[1]  CD[1] E[1] DC[1]  EA[1]  HYPOTHESISIn[1],theauthorsimplementedaMarkovdesignbasedonthemissoccurrencesoftheL1-Cache;theybuiltaMarkovtreebasedon thenextreportedmisses(states)ofagivenmiss.Infacttheywouldalwaystrackandprefetchnextfourstatesofamiss.Weproposetocreateandsimulateamoredynamicmodelthatadaptstothefrequencyofoccurrenceofnextstates. Theprefetcherwouldonlyprefetchamongthetracked nextstatesthesesmissesthatexceeded inthepastagiventhresholdofappearance.Inouropinion,pre-fetchingmissesbasedonpastfrequencyofoccurrenceus-ingourdynamicthresholdwouldimproveprefetcheraccu-racyandreducethememorybandwidth.Oursolutionassumesthattheprobabilityofoccurrenceofamissiscorrelatedtothehistoryofoccurrenceofthatmiss.Inotherwords,if“missB”occurredafter“missA”halfthetimeinthepastthenithasa50%chanceofoccurringnowandthusisusefultoprefetch.Toconfirmourassumptionweexaminedthedata-L1missstreamofseveralSPEC2000benchmarksandfoundgoodevidencetoourintuition.Figure1showsasampleofthebehaviorthatprovestheviabilityofourhypothesis.Inthe 2VPR benchmark, the frequency of occurrence of missesgrows as theprogram progresses, i.e. the missesthat ap-pearedinthepastarelikelytoappearinthefuture.Theirrelativefrequencyofoccurrenceismorelikelytostaycon-stantorchangeslowlythroughouttheprogram’slifethantochangebehaviordrastically.0246810T6 T7 T8 T9 T107FFF734C7FFF73A0100091DC7FFF75907FFF73C07FFF73207FFF75281001D72C1001DF881001C4D0Figure1:VPRsamplenextmissesfromL1missstreamshowsthenumberofoccurrencesof“next-miss”atprogressivetimeperiodsoftheprogram.Early examination of the


View Full Document
Download Thereshold Based Markov Prefetchers
Our administrator received your request to download this document. We will send you the file to your email shortly.
Loading Unlocking...
Login

Join to view Thereshold Based Markov Prefetchers and access 3M+ class-specific study document.

or
We will never post anything without your permission.
Don't have an account?
Sign Up

Join to view Thereshold Based Markov Prefetchers 2 2 and access 3M+ class-specific study document.

or

By creating an account you agree to our Privacy Policy and Terms Of Use

Already a member?