Centro Svizzero di Calcolo Scientifico CSCS

High Performance Computing HPC at Swiss National Supercomputing Centre in Lugano.

Transcript

1. CSCS HPC CENTRO SVIZZERO DI CALCOLO SCIENTIFICO CSCS Swiss National Supercomputing Centre Bruggmann Roland Bern University of Applied Sciences BUAS Department of Computer Science and Information Technology Division of Computer Science Spring Term 2014
2. CSCS HPC Outline 1 Swiss National Supercomputing Centre Facilty Piz Daint 2 High Performance Computation Support by Hardware Support by Code Fields of Application
3. CSCS Facilty Piz Daint HPC Swiss National Supercomputing Centre Facilty Founded 1991 in Manno TI, located in Lugano TI since 2012 Operated by the Swiss Federal Institute of Technology ETH Zurich Staff: 50 people 3 Computer-Clusters, 6 Supercomputers
4. CSCS Facilty Piz Daint HPC Swiss National Supercomputing Centre Facilty A datacenter in general consist of units as follows Computation units Data communication connections Power supply (emergency power supply by batteries, current generators evtl.) Supporting technology (e.g. cooling system, fire suppression by CO2) Security devices (eye-scanner etc.) Administration facilties
5. CSCS Facilty Piz Daint HPC Swiss National Supercomputing Centre Facilty CSCS – Cooling by Water Suction-strainers lowered into Lake Lugano (2.8 km from Parco Ciani to CSCS) 700 litres per second, heated water used to heat campus of USI1 Efficiency of energy PUE > 1.25 Figure: Parco Ciani: Suction strainers and pumping station | CSCS: Incoming water pipe 1Università della Svizzera Italiana (University of Lugano)
6. CSCS Facilty Piz Daint HPC Swiss National Supercomputing Centre Facilty Figure: Distribution deck (ground floor)
7. CSCS Facilty Piz Daint HPC Swiss National Supercomputing Centre Piz Daint Piz Daint Cray XC30 Supercomputer, fully operational since April 2014 7.79 PetaFLOPS2 Range: #1 in Europe, #6 worldwide3 Power consumption: 2’325.00 kW (5 x Mont Soleil) 21 PetaFLOPS = 1015 FLOPS = 1’000’000 x 1’000’000’000 Floating Point Operations Per Second 3http://www.top500.org, November 2013
8. CSCS Facilty Piz Daint HPC Swiss National Supercomputing Centre Piz Daint 28 Cabinets in 3 rows of 8/10/10 Cabinets Dragonfly Network: global Bandwidth matches injection Bandwidth
9. CSCS Facilty Piz Daint HPC Swiss National Supercomputing Centre Piz Daint Groups of 2 Cabinets Inter-Group Network: Rank 3, fiber (green) Inner-Group Network: Rank 2, copper (black, color coded)
10. CSCS Facilty Piz Daint HPC Swiss National Supercomputing Centre Piz Daint 5’272 Nodes (4 Nodes per Blade) 2 Sockets per Node: CPU/GPU-hybrid (8 Core Intel Xeon E5, NVIDIA Tesla K20X) Inter-Node Network: Rank 1, Bandwidth 33 TB/s Total Memory 201 TB (38 GB per Node), Storage 2.5 PB
11. CSCS Facilty Piz Daint HPC Swiss National Supercomputing Centre Piz Daint Operating system: Cray Linux Environment CLE Job sheduler: Simple Linux Unified Resource Manager SLURM
12. CSCS HPC HW SW Application High Performance Computation System Engineering: IaaS/PaaS Virtual Server: Colocation / datacenter of your choice. Cloud Computing: IBM, Google, Amazon etc. High Performance Computation HPC: UserLab at CSCS.
13. CSCS HPC HW SW Application High Performance Computation A system has to support HPC by its architecture: Hardware: processing units, internode network, storage Operating system: Scheduler etc. Executables: Parallel implementation4 4see https://computing.llnl.gov/tutorials/parallel_comp/
14. CSCS HPC HW SW Application High Performance Computation Support by Hardware How parallel computation works: CPU/GPU hybrid
15. CSCS HPC HW SW Application High Performance Computation Support by Hardware How parallel computation works: CPU/GPU hybrid
16. CSCS HPC HW SW Application High Performance Computation Support by Code How parallel computation works: Code adaption
17. CSCS HPC HW SW Application High Performance Computation Support by Code How parallel computation works: Code adaption Standards and Libraries OpenCL (Khronos): Parallel heterogenous systems, C openAAC (Cray, Nvidia e.a.): Standard for parallel computing, C/C++ MPI: Message Passing Interface for parallel computation, C++ (e.g. Boost.MPI) openMP: Multi-processing / shared memory API, C MPTL: Multi Processor Template Library, C++ trilinos: Scientific numerical analysis, C++ Cuda (Nvidia): GPU processing, C/C++
18. CSCS HPC HW SW Application High Performance Computation Fields of Application Fields of Application Materials – Materials Simulations Chemistry – Quantum Simulation Physics – Plasma Physics, Astrophysics and Multiscale Fluid Dynamics Life Sciences – Life Sciences Across Scales Earth – Solid Earth Dynamics Climate – Climate and Atmospheric Modelling
19. CSCS HPC HW SW Application High Performance Computation Fields of Application Application: European regional weather and climate modeling COSMO5 User: National forecasting services in Switzerland MeteoSwiss6, Germany, Italy, Poland, Greece, Romania and Russia 5see http://www.cosmo-model.org/ 6see http://www.meteoswiss.admin.ch/web/en/weather/models/cosmo.html
20. CSCS HPC HW SW Application High Performance Computation Fields of Application Figure: CSCS – Usage by Field
21. CSCS HPC HW SW Application High Performance Computation Fields of Application Figure: CSCS – Usage by Institution
22. CSCS HPC Thanks for your time.