Supercomputing and Computation

SHARE

OLCF CAAR project: acceleration of CAM-SE using GPU


Future success in climate simulation requires not only added spatial resolution but improvement in the fidelity of physical processes modeled, which in turn increases the number of processes to be modeled, such as chemical tracer quantities to be transported. Since transport is the dominant computational cost for added tracers, OLCF's CAAR effort has ported the tracer transport routines to utilize GPUs. This port has been performed using PGI's CUDA FORTRAN language. Memory optimizations and overlapping of PCI-express transfers, MPI transfers, and message packing and unpacking routines have led to over 2.5x reduction in full CAM-SE runtime compared to the CPU-only code run on an XE6 with a 14km science problem with 100 tracers. Currently, further changes to the codebase are being re-ported into the tracer routines on GPUs, and an effort is underway to translate the CUDA FORTRAN code into a more readable, portable, and maintainable OpenACC port.

ASK ORNL

We're always happy to get feedback from our users. Please use the Comments form to send us your comments, questions, and observations.