William Putman, Kaushik Datta, Amidu Oloso, Thomas Clune
Graphical Processing Units (GPU's) offer the potential for significant performance improvements over conventional processors for broad class of algorithms relevant to climate modeling. Intel has recently announced a competing technology in the form of the MIC architecture for handling parallelizable, vectorizable workloads. However, since MIC is based on the x86 ISA, Intel claims that it will support most programming models with minimal code changes. We have investigated the performance of a 2D finite-volume advection kernel extracted from the GEOS-5 dynamical core on a first generation MIC card ("Knights Ferry") and an NVIDIA Fermi GPU. On the Fermi, the performance of both compiler directives and CUDA Fortran were examined. We will report initial performance numbers as well as some qualitative observations about developer productivity.
Thoughout Kaushik Datta's career, he has focused on extracting performance from the latest computer architecturs. As a Ph.D. student at U.C. Berkeley, he studied how to use auto-tuning to best exploit cache-based multicore processors. He then put that knowledge to use at Reservoir Labs, a small compiler company based in New York. He has since shifted to working at NASA (through Northrop Grumman), where Kaushik is now looking at how to best exploit more exotic architectures, including GPUs and Intel's MIC architecture, for simple climate kernels.
Attachment | Size |
---|---|
advection_mic_SEA.pptx | 223.01 KB |