big data systems - courses.cs.washington.edumapreduce input • where does input come from? •...

BigDataSystems

BigDataParallelism

• Hugedataset

• crawleddocuments,webrequestlogs,etc.

• Naturalparallelism:

• canworkondifferentpartsofdataindependently

• imageprocessing,grep,indexing,manymore

Challenges

• ParallelizeapplicaFon

• Wheretoplaceinputandoutputdata?

• WheretoplacecomputaFon?

• Howtocommunicatedata?Howtomanagethreads?HowtoavoidnetworkboJlenecks?

• BalancecomputaFons

• HandlefailuresofnodesduringcomputaFon

• SchedulingseveralapplicaFonswhowanttoshareinfrastructure

GoalofMapReduce

• TosolvethesedistribuFon/fault-toleranceissuesonceinareusablelibrary

• Toshieldtheprogrammerfromhavingtore-solvethemforeachprogram

• Toobtainadequatethroughputandscalability

• Toprovidetheprogrammerwithaconceptualframeworkfordesigningtheirparallelprogram

MapReduce

• Overview:

• ParFFonlargedatasetintoMsplits

• RunmaponeachparFFon,whichproducesRlocalparFFons;usingaparFFonfuncFonR

• Hiddenintermediateshufflephase

• RunreduceoneachintermediateparFFon,whichproducesRoutputfiles

Details

• Inputvalues:setofkey-valuepairs

• Jobwillreadchunksofkey-valuepairs

• “key-value”pairsagoodenoughabstracFon

• Map(key,value):

• SystemwillexecutethisfuncFononeachkey-valuepair

• Generateasetofintermediatekey-valuepairs

• Reduce(key,values):

• Intermediatekey-valuepairsaresorted

• ReducefuncFonisexecutedontheseintermediatekey-values

Countwordsinweb-pages

Map(key,value){//keyisurl//valueisthecontentoftheurlForeachwordWinthecontentGenerate(W,1);}

Reduce(key,values){//keyisword(W)//valuesarebasicallyall1sSum=Sumall1sinvalues

//generateword-countpairsGenerate(key,sum);}

Reverseweb-linkgraph

Gotogoogleadvancedsearch:"findpagesthatlinktothepage:"cnn.com

Map(key,value){//key=url//value=contentForeachurl,linkingtotargetGenerate(outputtarget,url);}

Reduce(key,values){//key=targeturl//values=allurlsthatpointtothetargeturlGenerate(key,listofvalues);}

• QuesFon:howdoweimplement“join”inMapReduce?

• ImagineyouhavealogtableLandsomeothertableRthatcontainssayuserinformaFon

• PerformJoin(L.uid==R.uid)

• SaysizeofL>>sizeofR

• Bonus:considerrealworldzipfdistribuFons

Comparisons

• Worthcomparingittootherprogrammingmodels:

• distributedsharedmemorysystems

• bulksynchronousparallelprograms

• key-valuestorageaccessedbygeneralprograms

• MoreconstrainedprogrammingmodelforMapReduce

• OthermodelsarelatencysensiFve,havepoorthroughputefficiency

• MapReduceprovidesforeasyfaultrecovery

ImplementaFon

• Dependsontheunderlyinghardware:sharedmemory,messagepassing,NUMAsharedmemory,etc.

• InsideGoogle:

• commodityworkstaFons

• commoditynetworkinghardware(1Gbps-10Gbpsnow-atnodelevelandmuchsmallerbisecFonbandwidth)

• cluster=100sor1000sofmachines

• storageisthroughGFS

MapReduceInput

• Wheredoesinputcomefrom?

• Inputisstriped+replicatedoverGFSin64MBchunks

• ButinfactMapalwaysreadsfromalocaldisk

• TheyruntheMapsontheGFSserverthatholdsthedata

• Tradeoff:

• Good:Mapreadsatdiskspeed(localaccess)

• Bad:onlytwoorthreechoicesofwhereagivenMapcanrun

• potenFalproblemforloadbalance,stragglers

IntermediateData

• WheredoesMapReducestoreintermediatedata?

• OnthelocaldiskoftheMapserver(notinGFS)

• Tradeoff:

• Good:localdiskwriteisfasterthanwriFngovernetworktoGFSserver

• Bad:onlyonecopy,potenFalproblemforfault-toleranceandload-balance

OutputStorage

• WheredoesMapReducestoreoutput?

• InGFS,replicated,separatefileperReducetask

• SooutputrequiresnetworkcommunicaFon--slow

• ItcanthenbeusedasinputforsubsequentMapReduce

QuesFon

• WhatarethescalabilityboJlenecksforMapReduce?

Scaling

• Mapcallsprobablyscale

• butinputmightnotbeinfinitelyparFFonable,andsmallinput/intermediatefilesincurhighoverheads

• Reducecallsprobablyscale

• butcan’thavemoreworkersthankeys,andsomekeyscouldhavemorevaluesthanothers

• Networkmaylimitscaling

• Stragglerscouldbeaproblem

FaultTolerance

• Themainidea:MapandReducearedeterminisFc,funcFonal,andindependent

• soMapReducecandealwithfailuresbyre-execuFng

• WhatifaworkerfailswhilerunningMap?

• CanwerestartjustthatMaponanothermachine?

• Yes:GFSkeepscopyofeachinputspliton3machines

• Masterknows,tellsReduceworkerswheretofindintermediatefiles

FaultTolerance

• IfaMapfinishes,thenthatworkerfails,doweneedtore-runthatMap?

• Intermediateoutputnowinaccessibleonworker'slocaldisk.

• Thusneedtore-runMapelsewhereunlessallReduceworkershavealreadyfetchedthatMap'soutput.

• WhatifMaphadstartedtoproduceoutput,thencrashed?

• NeedtoensurethatReducedoesnotconsumetheoutputtwice

• WhatifaworkerfailswhilerunningReduce?

RoleoftheMaster

• Keepsstateregardingthestateofeachworkermachine(pingseachmachine)

• Reschedulesworkcorrespondingtofailedmachines

• OrchestratesthepassingoflocaFonstoreducefuncFons

LoadBalance

• WhatifsomeMapmachinesarefasterthanothers?

• Orsomeinputsplitstakelongertoprocess?

• SoluFon:manymoreinputsplitsthanmachines

• MasterhandsoutmoreMaptasksasmachinesfinish

• Thusfastermachinesdobiggershareofwork

• Butthere'saconstraint:

• WanttorunMaptaskonmachinethatstoresinputdata

• GFSkeeps3replicasofeachinputdatasplit

• onlythreeefficientchoicesofwheretoruneachMaptask

Stragglers

• Oqenonemachineisslowatfinishingverylasttask

• badhardware,overloadedwithsomeotherwork

• Loadbalanceonlybalancesnewlyassignedtasks

• SoluFon:alwaysschedulemulFplecopiesofverylasttasks!

HowmanyMRtasks?

• PaperusesM=10xnumberofworkers,R=2x.

• More=>

• finergrainedloadbalance.

• lessredundantworkforstragglerreducFon.

• spreadtasksoffailedworkerovermoremachines

• overlapMapandshuffle,shuffleandReduce.

• Less=>bigintermediatefilesw/lessoverhead.

• MandRalsomaybeconstrainedbyhowdataisstripedinGFS(e.g.,64MBchunks)

Discussion

• whataretheconstraintsimposedonmapandreducefuncFons?

• howwouldyouliketoexpandthecapabilityofmapreduce?

MapReduceCriFcism

• “Giantstepbackwards”inprogrammingmodel

• Sub-opFmalimplementaFon

• “Notnovelatall”

• MissingmostoftheDBfeatures

• IncompaFblewithalloftheDBtools

ComparisontoDatabases

• Hugesourceofcontroversy;claims:

• paralleldatabaseshavemuchmoreadvanceddataprocessingsupportthatleadstomuchmoreefficiency

• supportanindex;selecFonisaccelerated

• providesqueryopFmizaFon

• paralleldatabasessupportamuchrichersemanFcmodel

• supportaschema;sharingacrossapps

• supportSQL,efficientjoins,etc.

WheredoesMRwin?

• Scaling

• Loadingdataintosystem

• Faulttolerance(parFalrestarts)

• Approachability

SparkMoFvaFon

• MRProblems

• cannotsupportcomplexapplicaFonsefficiently

• cannotsupportinteracFveapplicaFonsefficiently

• Rootcause

• Inefficientdatasharing

In MapReduce, the only way to share data across jobs is stable storage -> slow!

MoFvaFon

Goal:In-MemoryDataSharing

Challenge

• HowtodesignadistributedmemoryabstracFonthatisbothfaulttolerantandefficient?

OtheropFons

• ExisFngstorageabstracFonshaveinterfacesbasedonfine-grainedupdatestomutablestate

• E.g.,RAMCloud,databases,distributedmem,Piccolo

• RequiresreplicaFngdataorlogsacrossnodesforfaulttolerance

• Costlyfordata-intensiveapps

• 10-100xslowerthanmemorywrite

RDDAbstracFon

• Restrictedformofdistributedsharedmemory

• immutable,parFFonedcollecFonofrecords

• canonlybebuiltthroughcoarse-graineddeterminisFctransformaFons(map,filter,join…)

• Efficient fault-tolerance using lineage

• Log coarse-grained operations instead of fine-grained data updates

• An RDD has enough information about how it’s derived from other dataset

• Recompute lost partitions on failure

Fault-tolerance

DesignSpace

OperaFons

• TransformaFons(e.g.map,filter,groupBy,join)

• LazyoperaFonstobuildRDDsfromotherRDDs

• AcFons(e.g.count,collect,save)

• Returnaresultorwriteittostorage

lines=spark.textFile(“hdfs://...”)

errors=lines.filter(lambdas:s.startswith(“ERROR”))messages=errors.map(lambdas:s.split(‘\t’)[2])

messages.persist()

messages.filter(lambdas:“foo”ins).count()messages.filter(lambdas:“bar”ins).count()...

Base RDDTransformed RDD

Action

Result: full-text search of Wikipedia in <1 sec(vs 20 sec for on-disk data) Result: scaled to 1 TB data in 5-7 sec

(vs 170 sec for on-disk data)

Example: Mining Console Logs

Load error messages from a log into memory, then interactively search

RDD Fault ToleranceRDDs track the transformations used to build them (their lineage) to recompute lost data

E.g:

messages=textFile(...).filter(lambdas:s.contains(“ERROR”)).map(lambdas:s.split(‘\t’)[2])

HadoopRDD path = hdfs://…

FilteredRDD func = contains(...)

MappedRDD func = split(…)

Lineage

• Sparkusesthelineagetoschedulejobs

• TransformaFononthesameparFFonformastage

• Joins,forexample,areastageboundary

• Needtoreshuffledata

• Ajobrunsasinglestage

• pipelinetransformaFonwithinastage

• SchedulejobwheretheRDDparFFonis

Lineage&FaultTolerance

• Greatopportunityforefficientfaulttolerance

• Let'ssayonemachinefails

• Wanttorecomputeonlyitsstate

• Thelineagetellsuswhattorecompute

• FollowthelineagetoidenFfyallparFFonsneeded

• Recomputethem

• Forlastexample,idenFfyparFFonsoflinesmissing

• Alldependenciesare“narrow”;eachparFFonisdependentononeparentparFFon

• NeedtoreadthemissingparFFonoflines;recomputethetransformaFons

FaultRecovery

Example:PageRank

Optimizing Placement

• links&ranksrepeatedlyjoined

• Canco-parFFonthem(e.g.,hashbothonURL)

• Canalsouseappknowledge,e.g.,hashonDNSname

PageRankPerformance

TensorFlow:SystemforML

• OpenSource,lotsofdevelopers,externalcontributors

• Usedin:RankBrain(rankresults),Photos(imagerecogniFon),SmartReply(automaFcemailresponses)

ThreetypesofML

• Largescaletraining:hugedatasets,generatemodels

• Google’spreviousDistBelieffor100sofmachines

• Lowlatencyinference:runningmodelsindatacenters,phones,etc.

• Customengines

• TesFngnewideas

• Singlenodeflexiblesystems(Torch,Theano)

TensorFlow

• Commonwaytowriteprograms

• Dataflow+Tensors

• Mutablestate

• SimplemathemaFcaloperaFons

• AutomaFcdifferenFaFon

Background: NN Training

• Takeinputimage

• ComputelossfuncFon(forwardpass)

• Computeerrorgradients(backwardpass)

• Updateweights

• Repeat

ComputaFonisaDFG

ExampleCode

ExampleCode

Parameter Server Architecture

Statelessworkers,statefulparameterservers(DHT)CommutaFveupdatestoparameterserver

TensorFlow

• Flexiblearchitectureformappingoperatorsandparameterserverstodifferentdevices

• SupportsmulFpleconcurrentexecuFonsonoverlappingsubgraphsoftheoverallgraph

• IndividualverFcesmayhavemutablestatethatcanbesharedbetweendifferentexecuFonsofthegraph

TensorFlowhandlestheglue

Synchrony?

• AsynchronousexecuFonissomeFmeshelpful,addressesstragglers

• Asynchronycausesconsistencyproblems

• TensorFlow:pursuessynchronoustraining

• Butaddskbackupmachinestoreducethestragglerproblem

• UsesdomainspecificknowledgetoenablethisopFmizaFon

OpenResearchProblems

• AutomaFcplacement:dataflow-greatmechanism,butnotclearhowtouseitappropriately

• mutablestate-splitround-robinacrossparameterservernodes,statelesstasksreplicatedonGPUsasmuchasitfits,restonCPUs

• HowtotakedataflowrepresentaFontogeneratemoreefficientcode?