Accelerating CST MWS Performance with GPU and MPI Computing. CST workshop series

Transcription

1 Accelerating CST MWS Performance with GPU and MPI Computing CST workshop series

2 Hardware Based Acceleration Techniques - Overview - Multithreading GPU Computing Distributed Computing MPI Computing CST workshop series

3 Acceleration Dialog All acceleration features can be configured at a central position for all solvers. Multithreading Press the "Acceleration..." button in the solver dialog. 32 GPU Computing Distributed Computing MPI Computing Token Calculator CST workshop series

4 Multicore Computing CST workshop series

5 Multicore Processors - Hardware Overview Nehalem EP - Intel Xeon 5600 (Core i7 workstation) Series Key Facts: Quickpath Interconnect Non Uniform Memory Access (NUMA) Architecture Triple Channel RAM (DDR3) Integrated Memory Controller Per Processor 4/6 Processor Cores CST workshop series

6 GPU Computing CST workshop series

7 GPU Computing - Key Facts - Available for OS: Licensing: Token Scheme Solvers supported: Current GPU hardware supported by CST, NVIDIA Tesla 10 series: 1 GPU 2 GPU 4 GPU Tesla C1060 (no display) - $1,200 Quadro FX $3,000 Quadro Plex 2200 D2 $10,000 Tesla S $8,000 Some Technical Specs of GPU Hardware: 240 Cores per GPU 4 GB GDDR3 memory per GPU Memory Bandwidth 102 GB/s (normal DDR3 RAM: 25.6 GB/s) CST workshop series

8 Windows XP x64 Windows Vista x64 Windows 7 x64 Windows Server 2003 R2 Windows Server 2008 R2 RHEL 4, 5 CentOS 4,5 Supported Configurations Workstations: HPZ800, Dell T7500, etc Servers: 1U, 2U SuperMicro GP-GPU servers CST can assist with hardware recommendation. Further information in support area at (FAQs hardware) CST workshop series

9 GPU Computing - New Hardware (Codename "Fermi") - Supported in CST2011(NVIDIA Tesla 20 series) Tesla C2050: $2,500 Tesla C2070: $4,000 1 GPU 4 GPU Tesla S2050: $13,000 Tesla S2070: Some Technical Specs of GPU Hardware: 512 Cores per GPU C2050/C2070 cards include graphics display 3 GB GDDR5 (C2050,S2050), 6 GB GDDR5 (C2070,S2070) memory per GPU First GPU with Error Protecting Code (ECC) Floating Point Performance strongly improved by factor of about 8 C2050 supported in CST2011 release; C2070 support in 2011 service pack CST workshop series

10 GPU Computing - Performance - Typical Speedup of Solver Loop (Compared to 2 x Quad Core Intel Xeon X5550, 2.66 GHz) "Please note that the performance graph shows the behavior for a certain benchmark. The performance as well as the soft memory limit are problem dependent." CST workshop series

11 GPU Computing - Performance SAM head and Cell phone TIME X GPU 2 X GPU 4 X GPU solver loop 21M Mesh Cells CPU: 2x Intel Xeon E5530, 72 GB RAM GPU: NVIDIA Tesla S CST workshop series

12 MPI Computing CST workshop series

13 MPI Computing - Key Facts - Available for OS: Licensing: Token Scheme Solvers supported: ( supported in v2011) Technical Requirements: TCP/IP network connection between nodes. (IB supported in v2011). Homogeneous cluster strongly recommended. Further information: CST workshop series

14 MPI Computing - Working Principle - The simulation model (computational domain) is split into parts (subdomains). Those subdomains are sent to the compute nodes. All calculations necessary for such a subdomain is done locally on the compute node. Subdomain Boundary Simulation Model Node 1 Node 2 Node 3 Data exchange for subdomain boundaries is necessary during each time step. Interconnection Network CST workshop series

15 MPI Computing - MPI Cluster Update - Reference Installation Update Package After the update all nodes are consistent with the reference installation. CST workshop series

16 CST workshop series MPI Computing - Airbus A320 Benchmark-

17 MPI Computing - Airbus A320 Benchmark - 8 MPI nodes; 8 Blade cluster Each node: (8) E5520, 2.3GHz; 24GB DDR3 RAM 384M cells; also run with 1B cells! CST workshop series

18 MPI Computing - Airbus A320 Benchmark - Plane Wave Excitation, FarField output CST workshop series

19 MPI Computing + GPU Computing - Combined MPI Computing and GPU Computing - As the capacity of the GPU hardware is limited. Combined MPI and GPU Computing provides a possibility to combine GPU accelerated nodes to form an MPI cluster. This helps to increase the maximum model size. Each MPI node can be accelerated by GPU hardware. CST workshop series

20 Domain Decomposition Cluster Computing Matrix calculation and transient solution distributed Memory & CPU balancing Complex model split into sub-structures (domains) for solution on a cluster CST workshop series

21 MPI+GPU Computing Performance typical model - 100M cells Test Case # Nodes # Cores per Node # GPUs per Node Total # GPUs Pure CPU 1 2x4 0 0 MPI 8 2x4 0 0 GPU 1 2x4 4 4 MPI+GPU 4 2x Speedup Pure CPU MPI GPU MPI+GPU CST workshop series

22 Mixing Acceleration Features The following table gives you an overview of which combinations of acceleration features are possible for the solvers supporting the feature. Multithreading GPU Computing MPI Computing Distributed Computing Multithreading - GPU Computing - MPI Computing - Distributed Computing - CST workshop series

23 Acceleration Features - Which Acceleration Technique should I use? - Solver Transient Transient Transient Frequency Domain Model Size below memory limit of GPU hardware below memory limit of GPU hardware above memory limit of GPU hardware Number of Simulations low medium/high Acceleration Technique GPU Computing GPU Computing on a DC Cluster - MPI or Combined MPI+GPU Computing - medium/high Distributed Computing (DC) MPI (v2011) Non HPC simulation getting faster too - cooperation with chip manufacturers helps enhance multithreading, memory layout optimizations, workflow improvement etc. CST workshop series

24 Acceleration Token Concept - Unified License for all Acceleration Features - GPU Computing Distributed Computing MPI Computing CST workshop series

25 Acceleration Token Concept - Token Table / Example - Example: CST workshop series

26 External Job Queuing Linux Batch Computing CST workshop series

27 Job Queuing on Linux Clusters CST MWS T!, F!, I! solver batch computing supported on Linux clusters Separate Linux DVD LSF, PBS, Torque, OGE job schedulers CST workshop series