RAMCloud Scalable High-Performance Storage Entirel

RAMCloud:ScalableDatacenterStorageEntirelyinDRAMJohnOusterhoutStanfordUniversityOctober27,2009RAMCloud/HPTSSlide2Introduction●NewresearchprojectatStanford●Createlarge-scalestoragesystemsentirelyinDRAM●Interestingcombination:scale,lowlatency●Thefutureofdatacenterstorage●LowlatencydisruptivetodatabasecommunityOctober27,2009RAMCloud/HPTSSlide3RAMCloudOverview●Storagefordatacenters●10-10000commodityservers●~64GBDRAM/server●AlldataalwaysinRAM●Durableandavailable●Highthroughput:1Mops/sec/server●Low-latencyaccess:5-10µsRPCApplicationServersStorageServersDatacenterOctober27,2009RAMCloud/HPTSSlide4ExampleConfigurations$4M$4MTotalservercost$4$60$/GB1PB64TBTotalcapacity64GB1000Today1024GBGB/server1000#servers5-10yearsOctober27,2009RAMCloud/HPTSSlide5RAMCloudMotivation●Relationaldatabasesdon’tscale●Everylarge-scaleWebapplicationhasproblems:Facebook:4000MySQLservers+2000memcachedservers●Newformsofstoragestartingtoappear:BigtableDynamoPNUTSH-storememcachedOctober27,2009RAMCloud/HPTSSlide6RAMCloudMotivation,cont’dDiskaccessratenotkeepingupwithcapacity:●Disksmustbecomemorearchival●Moreinformationmustmovetomemory8333x58days600sCapacity/bandwidth(1KBblocks)360x30hrs5minJimGray'srule5000s10ms100MB/s500GB2009333x2x50x16667xChange15sCapacity/bandwidth(largeblocks)20msLatency(seek&rotate)2MB/sMax.transferrate30MBDiskcapacityMid-1980’sOctober27,2009RAMCloud/HPTSSlide7ImpactofLatency●Large-scaleappsstrugglewithhighlatency●RAMCloudgoal:lowlatencyandlargescale●Enableanewbreedofinformation-intensiveapplicationsUIApp.LogicDataStructuresTraditionalApplicationUIBus.LogicApplicationServersStorageServersWebApplication1µslatency0.5-10mslatencySinglemachineDatacenterOctober27,2009RAMCloud/HPTSSlide8ResearchIssues●Achieving5-10µsRPC●Durabilityatlowlatency●Datamodel●Concurrency/consistencymodel●Datadistribution,scaling●Automatedmanagement●Multi-tenancy●NodearchitectureOctober27,2009RAMCloud/HPTSSlide9LowLatency:SQLisDead?●Relationalquerymodeltiedtohighlatency:DescribewhatyouneedupfronDBMSoptimizesretrieval●Withsufficientlylowlatency:Don'tneedoptimization;makeindividualrequestsasneededCan'taffordqueryprocessingoverheadTherelationalquerymodelwilldisappear●Question:whatsystemsofferverylowlatencyanduserelationalmodel?October27,2009RAMCloud/HPTSSlide10LowLatency:StrongerConsistency?●Costofconsistencyriseswithtransactionoverlap:O~R*DO=#overlappingtransactionsR=arrivalrateofnewtransactionsD=durationofeachtransaction●RincreaseswithsystemscaleEventually,scalemakesconsistencyunaffordable●But,DdecreaseswithlowerlatencyStrongerconsistencyaffordableatlargerscaleIsthisphenomenonstrongenoughtomatter?October27,2009RAMCloud/HPTSSlide11LowLatency:OneSizeFitsAllAgain?●One-size-fits-allisdead-MikeStonebraker●Specializeddatabasesproliferating:50xperformanceimprovementsinspecializeddomainsOptimizedisklayouttoeliminateseeks●Withlowlatency:Layoutdoesn'tmatterGeneral-purposeisfastOne-size-fits-allridesagain?October27,2009RAMCloud/HPTSSlide12Conclusions●AllonlinedataismovingtoRAM●RAMClouds=thefutureofdatacenterstorage●Lowlatencywillchangeeverything:NewapplicationsStrongerconsistencyatscaleOne-size-fits-allagainSQLisdead●1000-10000clientsaccessing100TB-1PB@5-10µslatencyOctober27,2009RAMCloud/HPTSSlide13Questions/Comments?FormoreonRAMCloudmotivation&researchissues:“TheCaseforRAMClouds:ScalableHigh-PerformanceStorageEntirelyinDRAM”ToappearinOperatingSystemsReview~ouster/cgi-bin/papers/ramcloud.pdfOr,google“RAMCloud”October27,2009RAMCloud/HPTSSlide14BackupSlidesOctober27,2009RAMCloud/HPTSSlide15WhyNotaCachingApproach?●Lostperformance:1%misses→10xperformancedegradation●Won’tsavemuchmoney:AlreadyhavetokeepinformationinmemoryExample:Facebookcaches~75%ofdatasize●Changesdiskmanagementissues:Optimizeforreads,vs.writes&recoveryOctober27,2009RAMCloud/HPTSSlide16WhynotFlashMemory?●ManycandidatetechnologiesbesidesDRAMFlash(NAND,NOR)PCRAM…●DRAMenableslowestlatency:5-10xfasterthanflash●MostRAMCloudtechniqueswillapplytoothertechnologies●Ultimately,choosestoragetechnologybasedoncost,performance,energy,notvolatilityOctober27,2009RAMCloud/HPTSSlide17IsRAMCloudCapacitySufficient?●Facebook:200TBof(non-image)datatoday●Amazon:Revenues/year:$16BOrders/year:400M?($40/order?)Bytes/order:1000-10000?Orderdata/year:0.4-4.0TB?RAMCloudcost:$24K-240K?●UnitedAirlines:Totalflights/day:4000?(30,000forallairlinesinU.S.)Passengerflights/year:200M?Bytes/passenger-flight:1000-10000?Orderdata/year:0.2-2.0TB?RAMCloudcost:$13K-130K?●Readytodayforallonlinedata;mediasoonOctober27,2009RAMCloud/HPTSSlide18DataDurability/Availability●DatamustbedurablewhenwriteRPCreturns●Unattractivepossibilities:Synchronousdiskwrite(100-1000xtooslow)Replicateinothermemories(tooexpensive)●Onepossibility:logtoRAM,thendiskDRAMdiskDRAMdiskStorageServerswritelogasync,batchDRAMdisklogOctober27,2009RAMCloud/HPTSSlide19Durability/Availability,cont’d●Bufferedloggingsupports~50Kwrites/sec./server(vs.1Mreads)●Needfastrecoveryaftercrashes:Read64GBfromdisk?10minutesShardbackupdata