The Essence of Caching

《The Essence of Caching》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、TheEssenceofCachingGregLuck,FounderandCTOEhcache,TerracottaJavaOne2011Session24241Tuesday,4October2011TheProblemTuesday,4October2011ApplicationApplicationApplicationApplication>100msDataStoreAverageResponseTimeSpeedCostScalabilityTuesday,4October2011TypesofScalingUPSc

2、aleCommodityServerApplicationScaleOUTTuesday,4October2011CapacityPlanning-PeakLoadWORKLOADSANDSOFTWAREINFRASTRUCTURE25FIGURE2.2:Exampleofdailytrafficfluctuationforasearchserviceinonedatacenter;Google’s Web Searches (1 Datacenter)x-axisisa24-h5periodandthey-axisistrafficmeas

3、uredinqueriespersecond.Tuesday,4October2011knowntoprovidegood-qualitysimilarityscores.Hereweconsideronesuchtypeofanalysis,calledco-citation.TheunderlyingideaistocounteveryarticlethatcitesarticlesAandBasavoteforthesimilaritybetweenAandB.Afterthatisdoneforallarticlesandapp

4、ropriatelynormalized,weob-tainanumericalscoreforthe(co-citation)similaritybetweenallpairsofarticlesandcreateadatastructurethatforeacharticlereturnsanorderedlist(byco-citationscore)ofsimilararticles.Thisdatastructureisperiodicallyupdated,andeachupdatethenbecomespartofthes

5、ervingstatefortheonlineservice.Thecomputationstartswithacitationgraphthatcreatesamappingfromeacharticleidenti-fiertoasetofarticlescitedbyit.Theinputdataaredividedintohundredsoffilesofapproximatelythesamesize(e.g.,thiscanbedonebytakingafingerprintofthearticleidentifier,dividi

6、ngitbythenumberofinputfiles,andusingtheremainderasthefileID)toenableefficientparallelexecution.WeuseasequenceofMapReducerunstotakeacitationgraphandproduceco-citationsimilarityscorevectorforallarticles.InthefirstMapphase,wetakeeachcitationlist(A1,A2,A3,...,An)andgenerateallpa

7、irsofdocumentsinthecitationlist,feedingthemtotheReducephase,whichcountsalloccurrencesofeachpair.Thisfirststepresultsinastructurethatassociatesallpairsofco-citeddocumentswithaco-citationcount.NotethatthisbecomesmuchlessthanaquadraticThe Elephant Curve6Tuesday,4October2011D

8、esirablePropertiesofaSolutionTuesday,4October20118Tuesday,4October20119Tuesday,4October2011APerformance