Difference between revisions of "NSC GPU and Accelerator Pilot"
|Line 3:||Line 3:|
Latest revision as of 12:11, 7 October 2013
|Name||NSC GPU and Accelerator Pilot|
|Description||GPU/Accelerator Pilot Project at NSC|
NSC recently started the pilot project for constructing accelerator systems. Currently two candidate systems are nVidia GPU cluster and IntelMIC. nVidia cluster system configuration will be clear when they provide their new Kepler card (Presumably at the end of Q2 2012). IntelMIC system is planned to be equipped at the end of 2012.
Along with hardware construction, we will testify different programming models for accelerators. That will include
- OpenCL or CUDA programming model
- OpenACC formulation
- IntelMIC pragmas
Here are a list of to-dos in this project.
- Construction of nVidia and IntelMIC clusters
- Cluster Configuration (Performance test of synthetic benchmarks: 1*4 VS 2*2 VS 4*1 node*card/node)
- Synthetic Benchmarks (HPL, SHOC)
- Installation/benchmark of Open-source Scientific Codes (Gromacs, LAMMPS, Luxrender, PetaChem, Gamess, OpenFOAM: List of installed softwares are subject to change)
- Compiler comparison for OpenACC (PGI, CAPS, Pathscale)
- GPU debugger, profiler installation
- OpenCL (and/or) CUDA programming practice
- OpenACC programming practice
- OpenACC Code Porting: Any existing open-source scientific codes which attracts our interest
- IntelMIC Code Porting: The same code as OpenACC. Also compare the performance between those two systems/programming models
- Hybrid (OpenCL + MPI) Modeling/Code Tuning for Best Performance
Full plans available here (accessible from NSC internal network).
|Soon-Heum Ko (NSC)||NSC||Application expert||Computational fluid dynamics|