From the Archives: Future of Supercomputing at Altparty 2009
0 likes989 views
This document summarizes the state of high performance computing (HPC) and looks towards the future. It discusses how HPC systems are transitioning to use more commodity components like standard servers and processors. Reaching exascale computing power, or 1018 floating point operations per second, by 2020 will be very challenging and require major innovations in hardware, software, and programming models. Emerging technologies like accelerators, 3D chip stacking, and new memory types have potential to help overcome current barriers and improve the efficiency of future exascale systems.
From the Archives: Future of Supercomputing at Altparty 2009
- 1. CSC –Tieteen tietotekniikan keskus Oy CSC –IT Center for Science Ltd. The Futureof Supercomputing Olli-Pekka Lehto Systems Specialist
- 2. CSC –IT Center for Science •Center for ScientificComputing –Officeslocatedin Keilaniemi, Espoo –Allsharesownedbythe ministryof education –Founded in 1970 as a technical support unit for the Univac 1108 •Providesa varietyof servicesto the Finnishresearchcommunity –HighPerformanceComputing(HPC)resources –Consultingservicesfor scientificcomputing –Scientificsoftware development(Chipster, Elmer etc.) –IT infrastructureservices –ISP services(FUNET)
- 3. CSC in numbers •~180 employees •3000 researchersusethe computingcapacityactively –Around500 projectsat anygiventime •~320 000 FUNET end- usersin 85 organizations
- 4. Louhi.csc.fi Model CrayXT4 (single-socketnodes) CrayXT5 (dual-socketnodes) Processors 10864 AMD Opteron2,3GHz cores 2716 QuadCoreprocessors 1012 XT4 + 852 XT5 Theoreticalpeakperformance >100 TeraFlop/s (= 2.3 * 10^9Hz * 4 Flop/Hz* 10864) Memory ~10.3 TeraBytes Interconnectnetwork CraySeaStar2 3D torus: 6*5.6GByte/s linksper node Power consumption 520.8 kW (highload) ~300 kW (nominalload) Localfilesystem 67TB Lustrefilesystem OperatingSystem Service nodes: SuSELinux Computenodes: CrayComputeNodeLinux A”capability” system: Fewlarge(64-10000 core) jobs
- 5. Murska.csc.fi Model HPProliantBladecluster Processors 2176AMD Opteron2,6GHz cores 1088 DualCoreprocessors 544 Bladeservers Theoreticalpeakperformance ~11.3 TeraFlop/s (= 2.6 * 10^9Hz * 4 Flop/Hz* 2176) Memory ~5 TB Interconnectnetwork Voltaire 4xDDR InfiniBand (16Gbit/sfat-treenetwork) Power consumption ~75 kW (highload) Localfilesystem 98 TB Lustrefilesystem OperatingSystem HP XCClusterSuite(RHEL basedLinux) A”capacity” system: Manysmall(1-128 core) jobs
- 6. Whyusesupercomputers? •Constraints –Resultsareneededin a reasonabletime •Impatientusers •Time-criticalproblems(e.g. weatherforecasting) –Largeproblemsizes •The problemdoesnotfitinto the memoryof a single system •Manyproblemtypesrequireallthe processingpowercloseto eachother –Distributedcomputing(BOINC etc.) workwellonlyon certainproblemtypes
- 7. WhousesHPC? MILITARY SCIENTIFICCOMMERCIAL Weaponsmodelling Signalsintelligence Radar imageprocessing Nuclearphysics MathematicsQuantumchemistryFusionenergyNanotechnology Climatechange Weatherforecasting Electronic Design Automation(EDA) Genomics Tacticalsimulation Aerodynamics Crashsimulations MovieSFX Feature-lengthmovies Searchengines Oilreservoirdiscovery Stockmarketprediction Banking& Insurance databases1960s1970s1980s 1990s 2000s Strategicsimulation”Wargames” 2010s Materialsscience Drugdesign Organmodelling
- 8. Stateof HPC 2009 •Movetowardscommoditycomponents –Clustersbuiltfromoff-the-shelfservers –Linux –Opensourcetools(compilers, debuggers, clusteringmgmt, applications) –Standard x86 processors •Price-performanceefficientcomponents –Low-latency, high-bandwitdhinterconnects •Standard PCIcards •InfiniBand, 10GigEthernet, Myrinet –ParallelFilesystems •StripedRAID (0)withfileservers •Lustre, GPFS, PVFS2etc.
- 9. ModernHPC systems Commodityclusters •A largenumberof regularserversconnectedtogether –Usuallya standardLinux OS –Possibleto evenmix and matchcomponentsfromdifferentvendors •Mayincludesomespecialcomponents –High-performanceinterconnectnetwork –Parallelfilesystems •Low-endand midrangesystems •Vendors: IBM, HP, Sun etc. Proprietarysupercomputers •Designedfromthe groundupfor HPC –Custominterconnectnetwork –CustomizedOS & software –Vendor-specificcomponents •High-endsupercomputersand specialapplications •Examples:CrayXT-series, IBM BlueGene
- 10. The ThreeWalls Therearethree”walls” whichCPU design is hittingnow: •Memorywall –Processorclockrateshavegrownfasterthanmemoryclockrates •Power wall –Processorsconsumean increasingamountof power –The increaseis non-linear •+13% performance= +73% powerconsumption •Microarchitecturewall –Addingmorecomplexityto the CPUsis nothelpingthatmuch •Pipelining, branchpredictionetc.
- 11. A TypicalHPC System •Builtfromcommodityservers –1U orBladeformfactor –1-10 management nodes –1-10 loginnodes •Programdevelopment, compilation –10s of storagenodes •Hostingparallelfilesystem –100s of computenodes •2-4 CPU socketsper node(4-24 cores), AMD OpteronorIntel Xeon •Linux OS •ConnectedwithInfiniBandorGigabitEthernet •Programsin C/C++ orFortran and areparallelizedusingMPI (MessagePassingInterface) API
- 12. The Exaflopsystem •Target:2015-2018 –10^18 (milliontrillion) floating-pointoperationsper second –Currentsystem0.00165 Exaflops •Expectationswithcurrenttechnologyevolution –Power draw100-300 MW •15-40% of a nuclearreactor(OlkiluotoI)! •$1M/MW/year! •Needto bringitdownto 30-50 MW –500000 -5000 000processorcores –Memory30-100 PB –Storage1 Exabyte
- 13. ProgrammingLanguages •Currenttrend(C/C++/Fortran + MPI) –Difficultto programportableandefficientcode –MPI is notfaulttolerantbydefault(1 taskdiesand the wholesystemcrashes) •PGASlanguagesto the rescue? –PartitionedGlobalAddressSpace –Lookslikeglobalsharedmemory •Butpossibleto definetask-localregions •Compilergeneratescommunicationcode –Currentstandards •UPC -UnifiedParallelC •CAF -Co-ArrayFortran –Languagesunderdevelopment •Titanium, Fortress, X10, Chapel
- 14. Whatto dowithan exaflop? •Long termclimate-changemodelling •Highresolutionweatherforecasts –Predictionbycity block –Extremeweather •Largeproteinfolding –Alzheimer, cancer, Parkinson’setc. •Simulationof a humanbrain •Veryrealisticvirtualenvironments •Design of nanostructures –Carbonnanotubes, nanobots •Beata humanpro playerin a 19x19 Go
- 15. Accelerators: GPGPU •General PurposeComputingon Graphics ProcessingUnits •NvidiaTesla/Fermi, ATI FireStream, IBM Cell, Intel Larrabee •Advantages –Highvolumeproductionrates, lowprice –HighmemorybandwidthonGPU(>100GB/s vs. 10-30GB/s of RAM) –Highfloprate, for certainapplications •Disadvantages –Lowperformancein precise(64-bit)computation –Gettingdata to the GPU memoryis a bottleneck(8GB/s PCI Express) –Vendorshavedifferentprogramminglanguages •Now: NvidiaCUDA, ATI Stream, Intel Ct, Celletc. •Future: OpenCLon everything(hopefully!) –Doesnotworkfor alltypesof applications •Branching, randommemoryaccess, hugedatasetsetc.
- 16. Case: NvidiaFermi •Announcedlastmonth, availablein 2010 •New HPC-orientedfeatures –Error-correctingmemory –Highdoubleprecisionperformance •512 computecores, ~3 billiontransistors –750 GFlops(DoublePrecision) –1.5 Tflops(Single Precision) •2011: Fermi-basedCraysupercomputerin OakRidgeNational Laboratory –”10 timesfasterthanthe currentstateof the art”:~20 Petaflops
- 17. Case: Intel Larrabee •Intel’snew GPU architecture, availablein 2010 •Basedon Pentium x86 processorcores –Initiallytensof coresper GPU –Pentium coreswithvectorunits –Compatiblewithx86 programs •Coresconnectedwitha ringbus
- 18. Accelerators:FPGA •FieldProgrammableGate Arrays •Vendors: Clearspeed, Mitrionics, Convey, Nallatech •Chipwithprogrammablelogicunits –Unitsconnectedwitha programmablenetwork •Advantages –Verylowpowerconsumption –Arbitraryprecision –Veryefficientin searchalgorithms –Severalin-socketimplementations •FPGA sitsdirectlyin the CPU socket •Disadvantages –Difficultto program –Limited numberof logicblocks
- 19. PerformanceSP and DP GFlops050010001500200025003000350040004500 Nvidia Geforce GTX280Nvidia Tesla C1060Nvidia Tesla S1070ATI Radeon 4870ATI Radeon X2 4870ATI FireStream 9250ClearSpeed e710ClearSpeed CATS700IBM PowerXCell 8iAMD Opteron Barcelona GFLop/s SP Gflop/sDP Gflop/s
- 20. Power EfficiencyPower efficiency0123456789 Nvidia Geforce GTX280Nvidia Tesla C1060Nvidia Tesla S1070ATI Radeon 4870ATI Radeon X2 4870ATI FireStream 9250ClearSpeed e710ClearSpeed CATS700IBM PowerXCell 8iAMD Opteron Barcelona GFlop/s/Watt SPDP
- 21. 3D IntegratedCircuits •Wafersstackedon top of eachother •Layersconnectedwiththrough-silicon”vias” •Manybenefits –Highbandwidthand lowlatency –Savesspaceand power –Addedfreedomin circuitdesign –The stackmayconsistof differenttypesof wafers •Severalchallenges –Heatdissipation –Complexdesign and manufacturing •HPC killer-app: Memorystackedon top of a CPU
- 22. OtherTechnologies To Watch •SSD (SolidState Disk) –Fasttransactions, lowpower, improvingreliability –Fastcheckpointingand restartingof programs •Opticson silicon –Lightpathsbothon a chipand on the PCB •New memorytechnologies –Phase-changememoryetc. –Low-power, low-latency, highbandwidth •Green datacentertechnologies •DNA computing •Quantumcomputing
- 23. Conclusions •Differencesbetweenclustersand proprietarysupercomputersis diminishing •Acceleratortechnologyis promimsing –Simple, vendorindependentprogrammingmodelsareneeded •Lotsof programmingchallengesin parallelisation –Similarchallengesin mainstreamcomputingtoday •Goingto Exaflopwillbeverytough –Innovationneededin bothsoftware and hardware
- 24. Questions