Oracle大数据一体机性能评测

Jean-PierreDijcksEricKaczmarek10/26/2015SpeakersJean-PierreDijcks,Oraclejean-pierre.dijcks@oracle.comEricKaczmarek,Inteleric.kaczmarek@intel.com•YouWeencouragequestionsanddiscussions2LegalDisclaimers•INFORMATIONINTHISDOCUMENTISPROVIDEDINCONNECTIONWITHINTELPRODUCTS.NOLICENSE,EXPRESSORIMPLIED,BYESTOPPELOROTHERWISE,TOANYINTELLECTUALPROPERTYRIGHTSISGRANTEDBYTHISDOCUMENT.EXCEPTASPROVIDEDININTEL'STERMSANDCONDITIONSOFSALEFORSUCHPRODUCTS,INTELASSUMESNOLIABILITYWHATSOEVERANDINTELDISCLAIMSANYEXPRESSORIMPLIEDWARRANTY,RELATINGTOSALEAND/ORUSEOFINTELPRODUCTSINCLUDINGLIABILITYORWARRANTIESRELATINGTOFITNESSFORAPARTICULARPURPOSE,MERCHANTABILITY,ORINFRINGEMENTOFANYPATENT,COPYRIGHTOROTHERINTELLECTUALPROPERTYRIGHT.A"MissionCriticalApplication"isanyapplicationinwhichfailureoftheIntelProductcouldresult,directlyorindirectly,inpersonalinjuryordeath.SHOULDYOUPURCHASEORUSEINTEL'SPRODUCTSFORANYSUCHMISSIONCRITICALAPPLICATION,YOUSHALLINDEMNIFYANDHOLDINTELANDITSSUBSIDIARIES,SUBCONTRACTORSANDAFFILIATES,ANDTHEDIRECTORS,OFFICERS,ANDEMPLOYEESOFEACH,HARMLESSAGAINSTALLCLAIMSCOSTS,DAMAGES,ANDEXPENSESANDREASONABLEATTORNEYS'FEESARISINGOUTOF,DIRECTLYORINDIRECTLY,ANYCLAIMOFPRODUCTLIABILITY,PERSONALINJURY,ORDEATHARISINGINANYWAYOUTOFSUCHMISSIONCRITICALAPPLICATION,WHETHERORNOTINTELORITSSUBCONTRACTORWASNEGLIGENTINTHEDESIGN,MANUFACTURE,ORWARNINGOFTHEINTELPRODUCTORANYOFITSPARTS.•Intelmaymakechangestospecificationsandproductdescriptionsatanytime,withoutnotice.Designersmustnotrelyontheabsenceorcharacteristicsofanyfeaturesorinstructionsmarked"reserved"or"undefined".Intelreservestheseforfuturedefinitionandshallhavenoresponsibilitywhatsoeverforconflictsorincompatibilitiesarisingfromfuturechangestothem.Theinformationhereissubjecttochangewithoutnotice.Donotfinalizeadesignwiththisinformation.•Theproductsdescribedinthisdocumentmaycontaindesigndefectsorerrorsknownaserratawhichmaycausetheproducttodeviatefrompublishedspecifications.Currentcharacterizederrataareavailableonrequest.•ContactyourlocalIntelsalesofficeoryourdistributortoobtainthelatestspecificationsandbeforeplacingyourproductorder.•Copiesofdocumentswhichhaveanordernumberandarereferencedinthisdocument,orotherIntelliterature,maybeobtainedbycalling1-800-548-4725,orgoto:http://www.intel.com/design/literature.htm•Intel,theIntellogo,IntelXeon,andXeonlogosaretrademarksofIntelCorporationintheU.S.and/orothercountries.•Intelprocessornumbersarenotameasureofperformance.Processornumbersdifferentiatefeatureswithineachprocessorfamily,notacrossdifferentprocessorfamilies:Goto:LearnAboutIntel®ProcessorNumbershttp://www.intel.com/products/processor_number*Othernamesandbrandsmaybeclaimedasthepropertyofothers.Copyright©2015IntelCorporation.Allrightsreserved.3LegalDisclaimers-Continued•SomeresultshavebeenestimatedbasedoninternalIntelanalysisandareprovidedforinformationalpurposesonly.Anydifferenceinsystemhardwareorsoftwaredesignorconfigurationmayaffectactualperformance.•SoftwareandworkloadsusedinperformancetestsmayhavebeenoptimizedforperformanceonlyonIntelmicroprocessors.Performancetests,suchasSYSmarkandMobileMark,aremeasuredusingspecificcomputersystems,components,software,operationsandfunctions.Anychangetoanyofthosefactorsmaycausetheresultstovary.Youshouldconsultotherinformationandperformanceteststoassistyouinfullyevaluatingyourcontemplatedpurchases,includingtheperformanceofthatproductwhencombinedwithotherproducts.•InteldoesnotcontrolorauditthedesignorimplementationofthirdpartybenchmarksorWebsitesreferencedinthisdocument.IntelencouragesallofitscustomerstovisitthereferencedWebsitesorotherswheresimilarperformancebenchmarksarereportedandconfirmwhetherthereferencedbenchmarksareaccurateandreflectperformanceofsystemsavailableforpurchase.•Relativeperformanceiscalculatedbyassigningabaselinevalueof1.0toonebenchmarkresult,andthendividingtheactualbenchmarkresultforthebaselineplatformintoeachofthespecificbenchmarkresultsofeachoftheotherplatforms,andassigningthemarelativeperformancenumberthatcorrelateswiththeperformanceimprovementsreported.•SPEC,SPECint,SPECfp,SPECrate,SPECpower,SPECjbb,SPECompG,SPECMPI,andSPECjEnterprise*aretrademarksoftheStandardPerformanceEvaluationCorporation.Seehttp://www.spec.orgformoreinformation.•TPCBenchmark,TPC-C,TPC-H,andTPC-EaretrademarksoftheTransactionProcessingCouncil.Seehttp://www.tpc.orgformoreinformation.•Intel®AdvancedVectorExtensions(Intel®AVX)*aredesignedtoachievehigherthroughputtocertainintegerandfloatingpointoperations.Duetovaryingprocessorpowercharacteristics,utilizingAVXinstructionsmaycausea)somepartstooperateatlessthantheratedfrequencyandb)somepartswithIntel®TurboBoostTechnology2.0tonotachieveanyormaximumturbofrequencies.Performancevariesdependingonhardware,software,andsystemconfigurationandyoushouldconsultyoursystemmanufacturerformoreinformation.•Intel®AdvancedVectorExtensionsreferstoIntel®AVX,Intel®AVX2orIntel®AVX-512.FormoreinformationonIntel®TurboBoostTechnology2.0,visithttp://www.intel.com/go/turbo4TheDo-It-Yourself(DIY)MythMythUseanyoldhardware&freesoftware(opensource)EasytoimplementGeneratesresultsimmediatelyReality2007commodityhardwareisonlygoodforbatchjobsTodayBigDataanalyticsmovingtowardsrealtimeanalysisApacheHadoopisfree,buthardtodeployoutofopen-source;requiresaHadoopEngineeron-site5OracleBigDataApplianceX5-2SunOracleX5-2LServerswithperserver:•2*18CoreIntelXeonE5-2699v3Processors•128GBMemoryexpandableto768GB•96TBDiskspaceSoftwareStack(v4.3):•OracleLinux6.6•OracleBigDataSQL2.0*•ClouderaDistributionofApacheHadoop5.4–EDHEdition•ClouderaManager5.4•OracleRDistribution•OracleNoSQLDatabaseCE6*OracleBigDataSQLisseparatelylicensedOracleBigDataApplianceX5-27OperationalSimplicity30%21%QuickertoDeployCheapertoBuy“OracleBigDataApplianceisanexcellentchoiceforcustomerslookingtoworkwiththefullsuiteofCloudera’sleadingHadoop-basedtechnology.It’smorecost-effectiveandquickertodeploythanaDIYcluster.”MikeOlson,Clouderafounder,ChiefStrategyOfficer,andChairmanoftheBoardSource:ESGWhitePaperOracleBigDataApplianceX5-2830%21%QuickertoDeployCheapertoBuy“OracleBigDataApplianceisanexcellentchoiceforcustomerslookingtoworkwiththefullsuiteofCloudera’sleadingHadoop-basedtechnology.It’smorecost-effectiveandquickertodeploythanaDIYcluster.”MikeOlson,Clouderafounder,ChiefStrategyOfficer,andChairmanoftheBoardSource:ESGWhitePaperWorkingtowardsanadditional2xfasterwhencomparedwithDIYBenchmarkTesting:OracleBigDataApplianceVs.Do-It-YourselfApacheHadoopCluster9TestConfiguration10OracleBigDataApplianceDIYClusterNumberofnodes66ConfigurationpernodeCPUIntel®Xeon®E5-2699v32.30GHzprocessorwithIntel®Hyper-ThreadingTechnologyenabledIntel®Xeon®E5-2699v32.30GHzprocessorwithIntel®Hyper-ThreadingTechnologyenabledMemoryDDR4128GBRAMDDR4128GBRAMStorageTwelve4TBHDDsOne64GBsolid-statedrivefortheoperatingsysteminstallationTwelve4TBHDDsNetworkInfiniband*network10GbnetworkOSVersionOracleLinuxEnterprise6CentOS6.6CDHversion5.4.4+Oracleconfiguration5.3.3withbasicbestknownpractices11BigDataPerformanceBenchmarkingMicroBenchmarks†BigBenchproposalacceptedbyTPC-TCforconsiderationTeragen;Teravalidate;Terasort;Wordcount;TestDFSIO;OSUHiBDSRLHeavydiskandnetworkusage(writes/reads)UsedtoillustrateIObehavior12BigDataPerformanceBenchmarkinghttp://www.odbms.org/wp-content/uploads/2014/12/bigbench-workload-tpctc_final.pdf*https://github.com/intel-hadoop/Big-Data-Benchmark-for-Big-BenchMicroBenchmarksEndtoEndBenchmark†BigBenchproposalacceptedbyTPC-TCforconsiderationTeragen;Teravalidate;Terasort;Wordcount;TestDFSIO;OSUHiBDSRLHeavydiskandnetworkusage(writes/reads)UsedtoillustrateIObehaviorBigBenchAdaptedfromTPC-DSSimulatesanonlineretailenvironmentConsideredbyTPC-TCasaBigDataworkloadFuturesupportforApacheSpark,Impala,etc.Supportfrom10+ecosystempartnersBigBenchQueries13QueryAlgorithm/MethodImplementationQuery1BasketAnalysisMapReduce(Java)Query2BasketAnalysisMapReduce(Java)Query3nPathAnalysisMapReduce(Python)Query4Sessionstream/patternmatchingMapReduce(Python)Query5LogisticRegressionMahoutQuery8nPathAnalysisMapReduce(Python)Query10SentimentAnalysisOpenNLPQuery11CorrelationHIVEQuery15LinearRegressionMapReduce(Java)Query18SentimentAnalysis&LinearRegressionMapReduce(Java)&NLPQuery19SentimentAnalysisNLPQuery20K-MeansClusteringMahoutQuery25K-MeansClusteringMahoutQuery26K-MeansClusteringMahoutQuery27NamedEntityRecognitionNLPQuery28BayesClassificationMahoutQuery29AffinityAnalysisMapReduce(Python)Query30AffinityAnalysisMapReduce(Python)Findlast5productsviewedbeforeagivenproductwaspurchasedonline&restrictcategoriesviewedwithin10dayspriortopurchasedateOutoftheBoxResults14SurprisingresultsgivencomparablehardwareGainsattributedtoWiderNetwork:23GbEvs.10GbEAdvancedLinux&Apache*Hadoop*configuration1.47x1.67x1.38x1.20x0x1x1x2x2xTerasortTeravalidateWordcountBigBenchNormalizedBDAClustervs.DIYDIYClusterBaselineBDACluster~1.4xfasterwhencomparedwithDIYGoalsAndMotivation15GoalsAndMotivation16Engineerabigdataappliancethatoffersa>2xperformanceadvantageoverasimilarlyconfiguredDIYclusterEncryptingDataonHDFSNonoticeableperformancedegradationafterencryptingallHDFSdata17http://www.intel.com/newsroom/kits/xeon/e7v3/pdfs/Xeon_E7v3_Cloudera-aes-ni.pdf96%100%99%99%100%96%102%0%20%40%60%80%100%120%dfsio-readdfsio-writeTerasortBigBenchQuery-2BasketAnalysisBigBenchQuery-16SentimentAnalysisBigBenchQuery-20K-MeansClusteringBigBenchQuery-30StreamingMap/ReduceRuntimepercentage(relativetounencrypted)EncryptedUnencryptedIntel®Xeon®ProcessorE5-2699v3+BDAx5+CDH5.4encryptionon/offperformance18File-SystemHadoopDistributedFileSystem(HDFS)isJava-basedfilesystembuiltforhighthroughputHDFSnot“baremetal”;reliesLinuxunderlyingfilesystemsuchasext3,ext4,etc.EXT3,EXT4;verystable;notbestsuitedfordemandinghighthroughputBigDataapplicationsXFSisahighperformancejournalingfilesystemexcelingatexecutingparallelinputs/outputsVerystablewithnewerLinuxkernels19SwitchingtoXFSReplacedEXT4infavorofXFSUpsideexceededexpectationsBiggestgainswithanincreasednumberofparallelinputs/outputs8%30%5%8%40%0%5%10%15%20%25%30%35%40%45%BigBenchTeragenDFSIO(read)DFSIO(write)OSUHiBDSRLPerformanceupsideforXFSoverExt420IncreasingNumberofLargePagesLinuxLargePagesspeed-upmemorylook-upsbyreducing“pagewalks”Increasedthenumberoflargepagesavailablefrom64to16,000Largepagespre-allocatedandunavailabletothefilesystemcacheUnevenimprovementsacrossworkloadsBigDataapplicationmakeheavyuseofthefilesystemcacheBestsuitedforcomputeintensiveworkloadsAvoidforrepetitivediskaccessConsideredaworkloadspecifictuning-4%-2%0%2%4%6%1357911131517192123252729PerformanceupsidefromlargepagesonBigBenchJavaVirtualMachine21YearsofIntel/OraclecollaborationonJavaVMoptimizationsEx:JDK8u60includesmanyoptimizationsforIntel®Xeon®E5-2699v3optimizations(AVX/AVX2,TSX,Crypto,Checksumacceleration)HW+SWHW+SWHW+SWHW+SWHW+SWHW+SWHWHWHWHWHWHW0x5x10x15x20x25x30x35xXeonX5160XeonX5535XeonX5470XeonX5570XeonX5680XeonE5-2690XeonE5-2697HW+SWHWHistoricalGen2GenHW+SWImprovementsSPECjbb200534xJavaVirtualMachine22MigratedCDHinstallationtoJDK8u60Allservices(DataNode,NameNode,YARN,etc.)configuredwithGarbageFirst(G1GC)collector(bestsuitedforlowlatency)Tunedmappersandreducersforparallelgarbagecollection(bestforhighthroughput)-15%-10%-5%0%5%10%15%1357911131517192123252729PerformanceupsidefromJDK8u60onBigBenchJavaVirtualMachine23MigratedCDHinstallationtoJDK8u60Allservices(DataNode,NameNode,YARN,etc.)configuredwithGarbageFirst(G1GC)collector(bestsuitedforlowlatency)Tunedmappersandreducersfortheparallelgarbagecollection(bestforhighthroughput)-15%-10%-5%0%5%10%15%1357911131517192123252729PerformanceupsidefromJDK8u60onBigBenchKeepyourJavauptodatetobenefitfromnewhardwarefeaturesApache*Hadoop*TuningApache*Hadoop*tuningcomplexandconvolutedwith100sofparametersAdvancedparameterstuningcanbefrustratingandworkloadspecificOracle’sBigDataAppliancewellconfiguredoutoftheboxforgeneralpurposeapplications(YARNcontainers,HDFS,shuffling,etc.)24Apache*Hadoop*TuningIncreasedYARNcontainersizefrom1GBto2GBLargerdatasetpermapper/reducer.Smallernumberofconcurrentmappers/reducers.LowerJVMstartupcostEnabledParallelExecution–runindependentphasesconcurrentlyReduceMapReducetailsEnabledVectorizationAllowstoprocessbatchofrowstogetherinsteadofprocessingonerowatatimeEnabledUberTask-singleJavaVMtoprocessonesmalljobBenefitssmalljobsOccupylessmapper/reducerslotsLowerJavaVMoverheadandstartupcost25Apache*Hadoop*TuningIncreasedYARNcontainersizefrom1GBto2GBLargerdatasetpermapper/reducer.Smallernumberofconcurrentmappers/reducers.LowerJVMstartupcostEnabledParallelExecution–runindependentphasesconcurrentlyReduceMapReducetailsEnabledVectorizationAllowstoprocessbatchofrowstogetherinsteadofprocessingonerowatatimeEnabledUberTask-singleJavaVMtoprocessonesmalljobBenefitssmalljobsOccupylessmapper/reducerslotsLowerJavaVMoverheadandstartupcost26AdvancedApache*Hadoop*tuningfurtherreducedworkloadsruntimebyanaverageof20%Muchmore…Noteverythingresultedinperformanceimprovements.Forexample:HivetablecompressionNetworkqueuesProcessorsturbomodeMemorytuning(NUMA)ManyHadoopparameters27Puttingitalltogether282.09x2.25x1.75x1.93x0x1x2x3xTerasortTeravalidateWordcountBigBenchNormalizedBDAClustervs.DIYDIYClusterBaselineBDACluster~2xfasterwhencomparedwithDIYhttp://www.oracle.com/ocom/groups/public/%40otn/documents/webcontent/2745435.pdfhttps://software.intel.com/en-us/articles/deploying-an-apache-hadoop-cluster-spend-your-time-on-bi-not-diyNextSteps–IncludeintoBDAReleases29TuningFeatureSoftwareVersionDoubleYARNcontainersizeBDA4.3EnableParallelExecutionBDA4.3EnableUbertaskBDA4.3EnableRSSforIBBDA4.4FileSystemConfiguration(XFS)FuturereleaseEnableHugePagesFuturereleaseJVMadditionaltuningFuturerelease