KEMBAR78
High performance computing | PPTX
High-Performance ComputingDr. Guy Tel-Zurtel-zur@computer.orgAugust 5th, 2010
Talk OutlineThe need for High-Performance Computing (HPC)Trends, Architecture, Systems, ModelsHigh-Throughput ComputingMore on Distributed ComputingGrid ComputingCloud ComputingSummary and Q&As
High-Performance Computing (HPC)Trends, Architecture, Systems, ModelsHigh-Throughput ComputingMore on Distributed ComputingGrid ComputingCloud ComputingSummary and Q&As
HPC ->max(FLOPS) -> Parallel ComputingSpeed: The problem takes too much timeSize: The problem doesn’t fit on my machineThe Nature of the Problem is CPU Intensive (CFD, Weather Forecast, Bio Informatics, Signal Processing, 
)Cost: Can’t afford the                                              real experiment Risk: The real problem is                                        too dangerous Image source: http://www.symscape.com/node/261
!Give me a stronger computerFact #1: Until 2003 Stronger == Faster by Freq.	Since 2003 Stronger == ParallelFact #2: All present and future processors are and will be ParallelFact #3: CPU intensive computer codes won’t perform well on future architectures using the traditional “Sequential” programming styleFact #4: The Challenge is in the Software
The Free lunch is overHerb Sutter, C++ Architect at Microsoft (March 2005)http://www.gotw.ca/publications/concurrency-ddj.htm
Modern High-End Parallel ComputersCommodities (Intel+AMD ≈ 100% market share)Open Source (Unix/Linux ≈ 100% market share)High Speed Interconnect (Infiniband   )Mostly running MPI (Distributed Memory) and OpenMP (Shared Memory)A Growing trend: GPGPUs “Many-Many” cores: Multi-Threading
The Top500
Front view of Dawning TC3600 Blade Server.June 2010 Top 3224,162 cores
GPGPU (a demo on my laptop)Source: Fast N-Body Simulation with CUDA. ByLars NylandNVIDIA Corporation, Mark Harris NVIDIA Corporation, Jan Prins University of North Carolina at Chapel Hill.
High-Performance Computing (HPC)Trends, Architecture, Systems, ModelsHigh-Throughput Computing (HTC)More on Distributed ComputingGrid ComputingCloud ComputingSummary and Q&As
High-Throughput Computing (HTC)FLOPY ≠ 60 X 60 X 24 X 7 X 52 FLOPSCondorMay 2010 @ UW-Madison pool:
High-Performance Computing (HPC)Trends, Architecture, Systems, ModelsHigh-Throughput Computing (HTC)More on Distributed ComputingGrid ComputingCloud ComputingSummary and Q&As
Grid ComputingDefinition According to Ian Foster*Coordinates resources that are not subject to centralized controlUsing standard, open, general-purpose protocols and interfacesDelivers nontrivial qualities of service (QoS)* Source: “What is the Grid? A Three Point Checklist” byIan Foster, Argonne National Laboratory & University of Chicago, July 20, 2002
http://rtm.hep.ph.ic.ac.uk/webstart.php   Real time monitoring July 25th, 2010EGEE
The Production Service infrastructure is a large multi-science Grid infrastructure, federating some 250 resource centers world-wide, providing some 40.000 CPUs and several Petabytes of storage. This infrastructure is used on a daily basis by several thousands of scientists federated in over 200 Virtual Organizations on a daily basis.
High-Performance Computing (HPC)Trends, Architecture, Systems, ModelsHigh-Throughput Computing (HTC)More on Distributed ComputingGrid ComputingCloud ComputingSummary and Q&As
My Cloud Computing Group
X as a Service (Xaas)Where X can be any Computing resource: Platform, Software, Infrastructure
A major revolution in the ITVirtualization & OutsourcingPay Per Use (PPU)However, many challenges unsolved yetQoSSecurityLegal Issues
An Example: Amazon Web Services (AWS)
HPC Clouds by Amazon EC2 and SGI
High-Performance Computing (HPC)Trends, Architecture, Systems, ModelsHigh-Throughput Computing (HTC)More on Distributed ComputingGrid ComputingCloud ComputingSummary and Q&As
Summary and OutlookTera-FLOPs processor by 2015How about an Israeli Supercomputer?SMPExa-FLOPs Supercomputer by 2019HPCGrid ComputingClustersHTC8.5 cent/hour CPUOpen Source Software

High performance computing

  • 1.
    High-Performance ComputingDr. GuyTel-Zurtel-zur@computer.orgAugust 5th, 2010
  • 2.
    Talk OutlineThe needfor High-Performance Computing (HPC)Trends, Architecture, Systems, ModelsHigh-Throughput ComputingMore on Distributed ComputingGrid ComputingCloud ComputingSummary and Q&As
  • 3.
    High-Performance Computing (HPC)Trends,Architecture, Systems, ModelsHigh-Throughput ComputingMore on Distributed ComputingGrid ComputingCloud ComputingSummary and Q&As
  • 4.
    HPC ->max(FLOPS) ->Parallel ComputingSpeed: The problem takes too much timeSize: The problem doesn’t fit on my machineThe Nature of the Problem is CPU Intensive (CFD, Weather Forecast, Bio Informatics, Signal Processing, 
)Cost: Can’t afford the real experiment Risk: The real problem is too dangerous Image source: http://www.symscape.com/node/261
  • 5.
    !Give me astronger computerFact #1: Until 2003 Stronger == Faster by Freq. Since 2003 Stronger == ParallelFact #2: All present and future processors are and will be ParallelFact #3: CPU intensive computer codes won’t perform well on future architectures using the traditional “Sequential” programming styleFact #4: The Challenge is in the Software
  • 6.
    The Free lunchis overHerb Sutter, C++ Architect at Microsoft (March 2005)http://www.gotw.ca/publications/concurrency-ddj.htm
  • 7.
    Modern High-End ParallelComputersCommodities (Intel+AMD ≈ 100% market share)Open Source (Unix/Linux ≈ 100% market share)High Speed Interconnect (Infiniband )Mostly running MPI (Distributed Memory) and OpenMP (Shared Memory)A Growing trend: GPGPUs “Many-Many” cores: Multi-Threading
  • 8.
  • 9.
    Front view ofDawning TC3600 Blade Server.June 2010 Top 3224,162 cores
  • 10.
    GPGPU (a demoon my laptop)Source: Fast N-Body Simulation with CUDA. ByLars NylandNVIDIA Corporation, Mark Harris NVIDIA Corporation, Jan Prins University of North Carolina at Chapel Hill.
  • 11.
    High-Performance Computing (HPC)Trends,Architecture, Systems, ModelsHigh-Throughput Computing (HTC)More on Distributed ComputingGrid ComputingCloud ComputingSummary and Q&As
  • 12.
    High-Throughput Computing (HTC)FLOPY≠ 60 X 60 X 24 X 7 X 52 FLOPSCondorMay 2010 @ UW-Madison pool:
  • 13.
    High-Performance Computing (HPC)Trends,Architecture, Systems, ModelsHigh-Throughput Computing (HTC)More on Distributed ComputingGrid ComputingCloud ComputingSummary and Q&As
  • 14.
    Grid ComputingDefinition Accordingto Ian Foster*Coordinates resources that are not subject to centralized controlUsing standard, open, general-purpose protocols and interfacesDelivers nontrivial qualities of service (QoS)* Source: “What is the Grid? A Three Point Checklist” byIan Foster, Argonne National Laboratory & University of Chicago, July 20, 2002
  • 15.
    http://rtm.hep.ph.ic.ac.uk/webstart.php Real time monitoring July 25th, 2010EGEE
  • 16.
    The Production Serviceinfrastructure is a large multi-science Grid infrastructure, federating some 250 resource centers world-wide, providing some 40.000 CPUs and several Petabytes of storage. This infrastructure is used on a daily basis by several thousands of scientists federated in over 200 Virtual Organizations on a daily basis.
  • 17.
    High-Performance Computing (HPC)Trends,Architecture, Systems, ModelsHigh-Throughput Computing (HTC)More on Distributed ComputingGrid ComputingCloud ComputingSummary and Q&As
  • 18.
  • 19.
    X as aService (Xaas)Where X can be any Computing resource: Platform, Software, Infrastructure
A major revolution in the ITVirtualization & OutsourcingPay Per Use (PPU)However, many challenges unsolved yetQoSSecurityLegal Issues
  • 20.
    An Example: AmazonWeb Services (AWS)
  • 21.
    HPC Clouds byAmazon EC2 and SGI
  • 22.
    High-Performance Computing (HPC)Trends,Architecture, Systems, ModelsHigh-Throughput Computing (HTC)More on Distributed ComputingGrid ComputingCloud ComputingSummary and Q&As
  • 23.
    Summary and OutlookTera-FLOPsprocessor by 2015How about an Israeli Supercomputer?SMPExa-FLOPs Supercomputer by 2019HPCGrid ComputingClustersHTC8.5 cent/hour CPUOpen Source Software

Editor's Notes

  • #6Â Ś›Śœ ڐڗړ ŚšŚ•ŚŠŚ” ŚžŚ—Ś©Ś‘ ڗږڧ...
  • #8Â ŚžŚŚ€Ś™Ś™Ś Ś™ ŚžŚąŚšŚ›Ś•ŚȘ high-end
  • #24Â Ś’Ś‘Ś•ŚœŚ•ŚȘ Ś”Ś’Ś™Ś–ŚšŚ” کڜ Ś›Śœ ŚžŚ•Ś Ś— ŚŚ™Ś Ś Ś—Ś“Ś™Ś.