简介:Whenqueryingonalarge-scaleknowledgebase,amajortechniqueofim-provingperformanceistopreloadknowledgetominimizethenumberofroundtripstotheknowledgebase.Inthispaper,anontology-basedsemanticcacheisproposedforanagentandontology-orientedknowledgebase(AOKB).InAOKB,anontologyisthecollectionofre-lationshipsbetweenagroupofknowledgeunits(agentsand/orothersub-ontologies).WhenloadingsomeagentA,itsrelationshipswithotherknowledgeunitsareexamined,andthosewhohaveatightsemantictiewithAwillbepreloadedatthesametime,includingagentsandsub-ontologiesinthesameontologywhereAis.Thepreloadedagentsandontologiesaresavedatasemanticcachelocatedinthememory.Testresultsshowthatupto50%reductioninrunningtimeisachieved.
简介:TheproblemofcontinuouslymonitoringmultipleK-nearestneighbor(K-NN)querieswithdynamicobjectandquerydatasetisvaluableformanylocation-basedapplications.Apracticalmethodistopartitionthedataspaceintogridcells,withbothobjectandquerytablebeingindexedbythisgridstructure,whilesolvingtheproblembyperiodicallyjoiningcellsofobjectswithquerieshavingtheirinfluenceregionsintersectingthecells.Intheworstcase,allcellsofobjectswillbeaccessedonce.ObjectandquerycachestrategiesareproposedtofurtherreducetheI/Ocost.Withobjectcachestrategy,queriesremainingstaticincurrentprocessingcycleseldomneedI/Ocost,theycanbereturnedquickly.ThemainI/Ocostcomesfrommovingqueries,thequerycachestrategyisusedtorestricttheirsearch-regions,whichusescurrentresultsofqueriesinthemainmemorybuffer.Thequeriescansharenotonlytheaccessingofobjectpages,butalsotheirinfluenceregions.TheoreticalanalysisoftheexpectedI/Ocostispresented,withtheI/Ocostbeingabout40%thatoftheSEA-CNNmethodintheexperimentresults.
简介:mc2llvmisaprocess-levelARM-to-x86binarytranslatordevelopedinourlabinthepastseveralyears.Currently,itisabletoemulatesinglethreadedprograms.Weextendmc2llvmtoemulatemulti-threadedprograms.Ourmaintaskistoreconstructitsarchitectureformulti-threadedprograms.Registermapping,codecachemanagement,andaddressmappinginmc2llvmhaveallbeenmodified.Inaddition,tofurtherspeeduptheemulation,wecollecthotpaths,aggressivelyoptimizeandgeneratecodeforthematruntime.Additionalthreadsareusedtoalleviatetheoverhead.Thus,whenthesamehotpathiswalkedthroughagain,thecorrespondingoptimizednativecodewillbeexecutedinstead.Inourexperiments,oursystemis8.8XfasterthanQEMU(quickemulator)onaveragewhenemulatingthespecifiedbenchmarkswith8guestthreads.
简介:Deduplication通常在两个企业存储系统和云存储被使用了。克服性能挑战为选择恢复deduplication系统的操作,solid-state-drive-based(即,基于SSD)读的缓存能为由缓冲加快被部署流行动态地恢复内容。不幸地,经常的数据更改由古典缓存计划导致了(例如,LRU和LFU)显著地弄短SSD一生当在SSD减慢I/O进程时。处理这个问题,我们建议新解决方案砍缓存极大地由扩大比例象I/O性能一样改进SSD的write耐久性长期流行(砍)在写进基于SSD的缓存的数据之中的数据。砍缓存保留很长时间在SSD缓存砍数据减少的时期缓存代替的数字。而且,它在deduplication集装箱阻止不得人心或不必要的数据被写进SSD缓存。我们在一个原型deduplication系统实现了砍缓存评估它的性能。我们的试验性的结果显示砍缓存弄短潜伏选择与仅仅deduplicated数据的5.56%能力以小基于SSD的缓存的成本由37.3%的一般水准恢复。重要地,砍缓存由9.77的一个因素改进SSD一生。砍缓存为一个成本效率的基于SSD的读的缓存解决方案提供到的证据表演增加性能选择为deduplication恢复系统。
简介:Shingledmagneticrecording(SMR)caneffectivelyincreasethecapacityofharddiskdrives(HDDs).HostawareSMR(HA-SMR)isexpectedtobemorepopularthanotherSMRmodelsbecauseofitsbackwardcompatibilityandnewSMR-specificAPIs.However,anHA-SMRdriveoftensuffersperformancedegradationunderwrite-intensiveworkloadsbecauseoffrequentnon-sequentialwritesbufferedinthediskcache.Thenon-sequentialwritesmainlycomefromupdatewrites,smallrandomwritesandout-of-orderwrites.Inthispaper,weproposeahybridstoragesystemcalledROCOwhichaimstouseasolidstatedrive(SSD)cachetoimprovetheperformanceofanHA-SMRdrive.ROCOreordersout-of-orderwritesbelongingtothesamezoneandusestheSSDcachetoabsorbupdatewritesandsmallrandomwrites.WealsodesignadatareplacementalgorithmcalledCREAfortheSSDcache.CREAfirstconductszone-orientedhot/colddataidentificationtoidentifycold-cachedzonesandhot-cachedzones,andthenevictsdatablocksbelongingtocolderzoneswithhigherprioritiesthatcanbesequentiallywrittenorwrittenthroughhost-sideread-modify-writeoperations.Itgivesthelowestprioritytodatablocksbelongingtothehottest-cachedzonethathavetobenon-sequentiallywritten.ExperimentalresultsshowthatROCOcaneffectivelyreducenon-sequentialwritestotheHA-SMRdriveandimprovetheperformanceoftheHA-SMRdrive.