Transcription
SAS - When Big Memory ApplicationsMeet Big Memory MachinesMaureen Chew, OracleGary Granito, Oracle
Agenda§ Big Memory Utilization Opportunities for SAS§ Oracle Database 12c§ Conventional SAS Usage§ Transparent Large Memory Utilization§ Directed Large Memory Utilization§ SAS 9.4 High Performance Analytics2
Big Memory Convergence with SAS§ Large product portfolio & diverse platform support§ Setting the Big Picture for Big Memory, Big Data§ Separation of Data and Compute Platforms§ Blurring of In-Memory and In-Database3
COMPREHENSIVE ORACLE ARCHITECTURES FOR SAS PLATFORMSFOCUS ON IN-MEMORYIn-MemoryData PlatformExadata*12c In-MemoryShared Storagefor SAS GridComputingCopyright 2012, SAS Institute Inc. All rights reserved. 12c In-Memory Big DataAppliance (BDA)ExalogicOracle VirtualComputeAppliance (OVCA)ACCESS12c In-MemoryACCESSBig DataAppliance* (BDA) SAS HPA - HighPerformanceAnalytics (SMP)SAS LASR – InMemory AnalyticServerSuperCluster(SSC - SPARC)Big MemoryMachine (M5/M6)SAS 9.4DataStorageZFS StorageAppliance*SAS VA - VisualAnalyticsSAS HPA - HighPerformanceAnalytics (Distributed)SAS LASR – InMemory AnalyticServerIMSTAT LASROracle SolarisBig MemoryMachine (M5/M6) Oracle LinuxDatabase(SSC - SPARC)SAS/Hadoop SAS In-Database EmbeddedProcessing (EP) Parallel Data Feeders Scoring AcceleratorSuperCluster* Engineered Systems*ACCESSCompute Platform
Oracle Engineered Systems IExadata DatabaseMachineRDBMS storagecompression anddatabaseparallelization via“Exadata StorageServers”Copyright 2012, SAS Institute Inc. All rights reserved.Exalogic ElasticCloudExtreme -performanceI/O connecting largeamount of computepower and memoryHARDWARE AND SOFTWAREOracle VirtualizedComputeAppliance (OVCA)VM Servervirtualization – runsOracle Linux, OracleSolaris, Windows.Software DefinedNetworkingBig DataApplianceMassive disk storagearray with highbandwidth I/O forloading ‘big’ dataSPARCSuperClusterSPARC servers, highperformance I/O andExadata storageservers in one rack
SAS on Oracle Big Memory MachineA 32 Socket / 32 TB High-End Datacenter Server§ Compute§ Up to 32 x SPARC M5 6-core 3.6GHz CPUs§ Up to 1024x DDR3 DIMMs for max memory of up to 32TB§ I/O and storage§ 32 x 2.5” SAS-2 internal drives§ 64x PCIe Gen3 low profile internal slots§ Scalability and investment protection§ Upgradable with M6 processor6
Oracle Database 12c – In-Memory§ Tony Baer, Ovum StraightTalk, Principal AnalystHow Oracle Database 12c embraces in-memory architecture§ Operational and transactional data platforms – historically separated fromanalytical stores – Natural evolution to blend the two§ In-memory allows for reduction / elimination of analytical indexes whichgeometrically increase storage requirements§ Analytical indexes speed query performance, but multiple indexes createOLTP performance cost§ Differentiator: pairing with disk and in-memory based row store that willinstantly replicate data to columnar tables§ Oracle Big Memory Machine – a sweet spot for OracleDatabase 12c In-Memory7
Conventional SAS – Transparent Benefit ofLarge Memory§ On Concurrency & Scalability§ Multi-User§ Multi-Threaded§ Multi-User, Multi-Threaded§ MEMSIZE / SORTSIZE§ User directed (in general)§ Choice made w/o due consideration to overall impact8
Conventional SAS – Directed Use/Benefit ofLarge Memory§ SASFILE – Holding a File in Memory§ sasfile mylibname.census open;data test1;set mydata.census; ç data now in memoryrun;sasfile mydata.census close;§ SASWORK & TMPFS§ Performance increase can be dramatic§ Is this different than FLASH drives?§ Won’t work with Checkpoint/Restart; requires SASWORKpersistence9
SAS on Oracle Big Memory Machine!"# %&#'()*( & ,-./-0SAS Scale Up, Scale Out – proc LOGISTIC§ SAS Scalability Test§ Compute Intensive proc LOGISTIC§ Add identical SAS Jobs» Observe Job Turnaround§ Scale Up: Single 48 core PhysicalDomain(PDOM)§ Scale Out: Single 48job workload, runconcurrently on 4 PDOMs – 192 jobsTesting Outcome§ § § !"# %&#'?,@'(*)#' ,'-,*./#0,1'§ Excellent Scalability ResultsSingle PDOM: 48 Jobs run in 2X thetime of 1 job.» 24X Work Accomplished – ScaleUpScaling to 4 PDOMs – 192 jobs in 2xthe time of 1 job»2'3'4'5'23'26';)1&/#'89,*'; %/#' '-,1 #1 '?,@A'34'73'45'89,*':'89,*'2'89,*'3'89,*'7'B //'; %/#' ''45'-,1 #1 '?,@A'.# '89,*'(, %/',C'3D6'-,1 #1 '?,@A'96X Work Accomplished – ScaleOut10
§ Single 48 core PhysicalDomain (PDOM)§ § § !"# %&#'(!(')* ''(,#-'./0#'SASWORKonTMPFSSAS Scalability Testing with TMPFSSASWORK SAN vs. TMPFS(In Memory FileSystem) comparisonScale Up Workload – run up to 64concurrent jobsSystem is oversubscribed at 48 jobs§ Testing Outcome§ SAN performance is fine but throughputhits expected max§ Avg step time remains nearly flat(redline) for TMPFS; exponential increasefor SAN§ Reducing SASWORK I/O using TMPFSenables excellent scalability!"#" " "()"* ,-"#"./01-"#"),0232-"%"!#"1*234 #2,')* 5'%"()"* ,-" "./01-"#"),0232-"%"()"* ,-" "./01-" "),0232-"!&"'#"& "4,"567829"(:36-9-067"11
Full Scale Concurrency on Big Memory Machine§ SAS Global Trading Application – Scale Up & Scale Out§ Use TMPFS for SASWORK§ 256 CPU/memory intensive complex sas streams running concurrently.Each uses 750MB RAM 80 GB I/O to TMPFS4 Concurrent PDom test compared to Single RunEven results at all levels of ConcurrencyAvg Job Time00:28:48Single PDom gle PDom Run00:07:1200:00:0012481216# Parallel Jobs32486412
SAS High Performance Analytics§ SAS LASR Analytic Server§ In-Memory Analytic platform that allows for concurrent access todata loaded into memory§ SAS High Performance Procedures§ hpreg, hplogistic, hpreduce, hpsplit, etc§ Supports distributed (MPP) and non-distributed (SMP)13
#&02%,-" 8(0 )?@? @?B?@? @ A?@? @EJ?@? @G?@? @AI?@?E@GF HEGJEGBM)%%-# /%,L)'GB8(0 )**,0'4(53- ,*- HEGJEGBHGFH" (7-?" (7- " (7-E" (7-JSAS on Oracle Big Memory MachineSAS High Performance Analytics (HPA) Scalability Results?@ E@J?@ @AI?@ G@E?@ B@GB?@ B@GB?@ F@?B?@ B@A?@ B@ A"*( -!"#"1K6§ SAS HPA Testing§ hpslit scale up test§ § THREADs/job 4Oversubscribe 48core/2TB PDOM with up to96 concurrent jobs» Observe Avg JobTimeTesting Outcome)* ,-./""012.""345665778"§ § Excellent scalability results§ At 48 jobs, avg time is 2x single job» 24X scalability§ At 96 jobs, avg time is 3X single job»!"#" "%"&"!#"!%"9:;" : ?@@. A"B?2;.@":C"9:;-"# "'#" &"% "(%" 32X workload scalability14
Why SAS on Big Memory Machine - LargeMemory .§ Supports New & Existing Features§ Enables IT Agility§ Reduces Performance Risk§ Masks resource consumption chaos§ Application Memory§ Kernel Memory§ File system caching§ Provides the best performance “Insurance”15
Thank Youmaureen.chew@oracle.comgary.granito@oracle.com
SAS/Hadoop 12c In-Memory 12c In-Memory 12c In-Memory ACCESS ACCESS s * Oracle Linux Big Data Appliance* (BDA) . In-Memory Analytic platform that allows for concurrent access to data loaded into memory ! SAS High Performance Procedures ! hpreg, hplogistic, hpreduce, hpsplit, etc ! Supports distributed (MPP) and non-distributed (SMP)
