1. High Performance Computing Kyaw Zwa Soe (Director) Ministry of Science & Technology Centre of Advanced Science & Technology

2. Why we need high performance computing? Today, computing means not only to crunch the numbers Today, computing means not only to crunch the numbers Add too many graphical add-ons over the computing ( e.g. 3D games, video editing programs) Add too many graphical add-ons over the computing ( e.g. 3D games, video editing programs) Have to execute too many programs at the same time Have to execute too many programs at the same time Most of the programs have characteristic of parallel computing (such as VHDL ) Most of the programs have characteristic of parallel computing (such as VHDL ) Very popular in Nural- Network and Genetic Algorithms computing Very popular in Nural- Network and Genetic Algorithms computing

3. Ways of Improving Increasing the Clock Speed (has limitation on silicon fabrication technology) Increasing the Clock Speed (has limitation on silicon fabrication technology) Improvements by Architecture Improvements by Architecture Pipe line Architecture Pipe line Architecture Vector or Array Processor Architecture Vector or Array Processor Architecture Cluster and Grid Computing Cluster and Grid Computing CISC (Complex Instruction Set Computers) vs. RISC (Reduced Instruction Set Computers) CISC (Complex Instruction Set Computers) vs. RISC (Reduced Instruction Set Computers)

4. Pipe Line Like an assembly line in the factory Like an assembly line in the factory Need particular attention in dividing the task of each step in the pipe line Need particular attention in dividing the task of each step in the pipe line Not a problem oriented solution, so each instruction of the predecessor processor can be implemented Not a problem oriented solution, so each instruction of the predecessor processor can be implemented Minimum cycle time of the pipe line is limited by the longest time required to complete the most time consuming task that can not be divided Minimum cycle time of the pipe line is limited by the longest time required to complete the most time consuming task that can not be divided Faster, but in the scale of 10 or more is very difficult Faster, but in the scale of 10 or more is very difficult

5. Array or Vector Processors Single Control Unit Processor Memory.....

6. Why we choose Cluster Easy to implement Easy to implement Can be constructed by commodity equipments Can be constructed by commodity equipments Scalable and reliable ( because failure occurs, repair and maintenance is very easy) Scalable and reliable ( because failure occurs, repair and maintenance is very easy) Very cost effective Very cost effective Application sensitive but not by HW. Application sensitive but not by HW.

7. What is Cluster Computing All About? First, what is a cluster? First, what is a cluster? A cluster is a large group of computers networked together, all controlled by a “head node”.

8. What is Cluster Computing All About? Why do we want a cluster? Why do we want a cluster? To do complex calculations To do complex calculations The other option is a single super computer with a small number of extremely fast processors. The other option is a single super computer with a small number of extremely fast processors. Expensive! Expensive! Vs.

9. What is Cluster Computing All About? Limitations of cluster computing: Limitations of cluster computing: Work needs to be easily divisible: Work needs to be easily divisible: Work can be divided up into computationally independent segments Work can be divided up into computationally independent segments Work to be done Finished work Divided up Each computer does its own piece of the whole

10. Closed Cluster Configuration compute node compute node compute node compute node High Speed Network Service Network gateway node External Network compute node compute node compute node compute node High Speed Network gateway node External Network File Server node Front-end

11. High Performance Computing Cluster High Performance Computing Cluster Load Balancing Load Balancing High Availability Cluster High Availability Cluster Different kinds of PC cluster

12. High Performance Computing Cluster (HPCC) Start from 1994 Start from 1994 Donald Becker of NASA assemble the world ’ s first cluster with 16 sets of DX4 PCs and 10 Mb/s ethernet Donald Becker of NASA assemble the world ’ s first cluster with 16 sets of DX4 PCs and 10 Mb/s ethernet Also called Beowulf cluster Also called Beowulf cluster Built from commodity off-the-shelf hardware Built from commodity off-the-shelf hardware Applications like data mining, simulations, parallel processing, weather modelling, computer graphical rendering, etc. Applications like data mining, simulations, parallel processing, weather modelling, computer graphical rendering, etc.

13. Load Balancing Cluster PC cluster deliver load balancing performance PC cluster deliver load balancing performance Commonly used with busy ftp and web servers with large client base Commonly used with busy ftp and web servers with large client base Large number of nodes to share load Large number of nodes to share load

14. High Availability Cluster Avoid downtime of services Avoid downtime of services Avoid single point of failure Avoid single point of failure Always with redundancy Always with redundancy Almost all load balancing cluster are with HA capability Almost all load balancing cluster are with HA capability

15. Beowulf cluster Beowulf cluster

16. Message Passing Interface Message Passing Interface MPICH (http://www-unix.mcs.anl.gov/mpi/mpich/) MPICH (http://www-unix.mcs.anl.gov/mpi/mpich/)http://www-unix.mcs.anl.gov/mpi/mpich/ LAM/MPI (http://lam-mpi.org) LAM/MPI (http://lam-mpi.org)http://lam-mpi.org Mathematical Mathematical fftw (fast fourier transform) fftw (fast fourier transform) pblas (parallel basic linear algebra software) pblas (parallel basic linear algebra software) atlas (a collections of mathematical library) atlas (a collections of mathematical library) sprng (scalable parallel random number generator) sprng (scalable parallel random number generator) sprng MPITB -- MPI toolbox for MATLAB MPITB -- MPI toolbox for MATLAB MPITB Quantum Chemistry software Quantum Chemistry software gaussian, qchem gaussian, qchemqchem Molecular Dynamic solver Molecular Dynamic solver NAMD, gromacs, gamess NAMD, gromacs, gamess NAMDgromacsgamess NAMDgromacsgamess Weather modelling Weather modelling MM5 (http://www.mmm.ucar.edu/mm5/mm5-home.html) MM5 (http://www.mmm.ucar.edu/mm5/mm5-home.html)http://www.mmm.ucar.edu/mm5/mm5-home.html Cluster computing applications

17. Relative Interconnect Performance LANSustainedMb/sTheoreticalMb/sLatencyµSec Fast Ethernet 6496100-400 Gigabit Ethernet 68096030-120 Myrinet136019607 Dolphin ScaLI 159226644.5

18. Necessary Software Modules for Building & Running a Cluster System Clustering Clustering Installation Installation Administration Administration Monitoring Monitoring Tools/Libraries Tools/Libraries File system File system Job Management Job Management Scheduler Scheduler

19. RankSite Country/Year Computer / Processors Manufacturer R max R peak 1DOE/NNSA/LLNL DOE/NNSA/LLNL United States/2005 BlueGene/L eServer Blue Gene Solution / 65536 IBM eServer Blue Gene Solution 136800 183500 2IBM Thomas J. Watson Research Center IBM Thomas J. Watson Research Center United States/2005 BGW eServer Blue Gene Solution / 40960 IBM eServer Blue Gene Solution 91290 114688 3NASA/Ames Research Center/NAS NASA/Ames Research Center/NAS United States/2004 Columbia SGI Altix 1.5 GHz, Voltaire Infiniband / 10160 SGI SGI Altix 1.5 GHz, Voltaire Infiniband 51870 60960 4The Earth Simulator Center The Earth Simulator Center Japan/2002 Earth-SimulatorEarth-Simulator / 5120 NEC 35860 40960 5Barcelona Supercomputer Center Barcelona Supercomputer Center Spain/2005 MareNostrum JS20 Cluster, PPC 970, 2.2 GHz, Myrinet / 4800 IBM JS20 Cluster, PPC 970, 2.2 GHz, Myrinet 27910 42144 6ASTRON/University Groningen ASTRON/University Groningen Netherlands/2005 eServer Blue Gene SolutioneServer Blue Gene Solution / 12288 IBM 27450 34406.4 7Lawrence Livermore National Laboratory Lawrence Livermore National Laboratory United States/2004 Thunder Intel Itanium2 Tiger4 1.4GHz - Quadrics / 4096 California Digital Corporation Intel Itanium2 Tiger4 1.4GHz - Quadrics 19940 22938 8Computational Biology Research Center, AIST Computational Biology Research Center, AIST Japan/2005 Blue Protein eServer Blue Gene Solution / 8192 IBM eServer Blue Gene Solution 18200 22937.6

20. 9Ecole Polytechnique Federale de Lausanne Ecole Polytechnique Federale de Lausanne Switzerland/2005 eServer Blue Gene SolutioneServer Blue Gene Solution / 8192 IBM 18200 22937.6 10Sandia National Laboratories Sandia National Laboratories United States/2005 Red Storm, Cray XT3, 2.0 GHzRed Storm, Cray XT3, 2.0 GHz / 5000 Cray Inc. 15250 20000 11Oak Ridge National Laboratory Oak Ridge National Laboratory United States/2005 Cray XT3, 2.4 GHzCray XT3, 2.4 GHz / 3748 Cray Inc. 14170 17990 12Los Alamos National Laboratory Los Alamos National Laboratory United States/2002 ASCI Q ASCI Q - AlphaServer SC45, 1.25 GHz / 8192 Hewlett-Packard ASCI Q - AlphaServer SC45, 1.25 GHz 13880 20480 13Lawrence Livermore National Laboratory Lawrence Livermore National Laboratory United States/2005 eServer pSeries p5 575 1.9 GHzeServer pSeries p5 575 1.9 GHz / 2048 IBM 13090 15564.8 14Virginia Tech Virginia Tech United States/2004 System X 1100 Dual 2.3 GHz Apple XServe/Mellanox Infiniband 4X/Cisco GigE / 2200 Self-made 1100 Dual 2.3 GHz Apple XServe/Mellanox Infiniband 4X/Cisco GigE 12250 20240 15Japan Atomic Energy Research Institute Japan Atomic Energy Research Institute Japan/2005 SGI Altix 3700 Bx2, 1.6 GHz, NUMALinkSGI Altix 3700 Bx2, 1.6 GHz, NUMALink / 2048 SGI 11814 13107 http://www.top500.org/