资源描述:
《Lookup Tables Fine-Grained Partitioning for Distributed Databases.pdf》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、LookupTables:Fine-GrainedPartitioningforDistributedDatabasesAubreyL.Tatarowicz#1,CarloCurino#2,EvanP.C.Jones#3,SamMadden#4#MassachusettsInstituteofTechnology,USA1altat@alum.mit.edu2krl@yahoo-inc.com3evanj@csail.mit.edu4madden@csail.mit.eduAbstract—ThestandardwaytoscaleadistributedO
2、LTPe.g.,orderprocessingapplicationswhereordersareissuedtoDBMSistohorizontallypartitiondataacrossseveralnodes.suppliersorbrokersthatservicemanycustomers(asmodeledIdeally,thisresultsineachquery/transactionbeingexecutedatbyTPC-E)ormessageboardswhereuserspostonmultiplejustonenode,toavo
3、idtheoverheadofdistributionandallowforums(withqueriestosearchforpostsbyforumoruser).thesystemtoscalebyaddingnodes.Forsomeapplications,simplestrategiessuchashashingonprimarykeyprovidethisOnesolutiontothisproblemistouseafine-grainedproperty.Unfortunately,formanyapplications,includings
4、ocialpartitioningstrategy,wheretuplesareallocatedtopartitionsnetworkingandorder-fulfillment,simplepartitioningschemesinawaythatexploitsrelationshipsbetweenrecords.Inourappliedtomany-to-manyrelationshipscreatealargefractionsocialnetworkingexample,auserandhisorherfriendscanofdistribut
5、edqueries/transactions.Whatisneededisafine-beco-locatedsuchthatsomequeriesgotojustonepartition.grainedpartitioning,whererelatedindividualtuples(e.g.,cliquesoffriends)areco-locatedtogetherinthesamepartition.Thus,acarefulassignmentoftuplestopartitionscanreduceMaintainingafine-grainedpa
6、rtitioningrequiresstoringtheoreliminatedistributedtransactions,allowingaworkloadtolocationofeachtuple.Wecallthismetadataalookuptable.Webeefficientlyscaledacrossmultiplemachines.presentadesignthatefficientlystoresverylargetablesandmain-Asecondproblemwithtraditionalpartitioningisthatwh
7、iletainsthemasthedatabaseismodified.Weshowtheyimprovequeriesonthepartitioningattributegotoasinglepartition,scalabilityforseveraldifficulttopartitiondatabaseworkloads,includingWikipedia,Twitter,andTPC-E.Ourimplementationqueriesonotherattributesmustbebroadcasttoallpartitions.provides40
8、%to300%betterthroughputont