Accelerating lattice QCD simulations with 2 flavours of staggered fermions on multiple GPUs using OpenACC - a first attempt

Abstract

We present the results of an effort to accelerate a Rational Hybrid Monte Carlo (RHMC) program for lattice quantum chromodynamics (QCD) simulation for 2 flavours of staggered fermions on multiple Kepler K20X GPUs distributed on different nodes of a Cray XC30. We do not use CUDA but adopt a higher level directive based programming approach using the OpenACC platform. The lattice QCD algorithm is known to be bandwidth bound; our timing results illustrate this clearly, and we discuss how this limits the parallelization gains. We achieve more than a factor three speed-up compared to the CPU only MPI program.

Comments: 14 pages, 6 figures and 2 tables. Details about benchmarking hardware added along with one paragraph on GPU memory bandwidth measurements. Results unchanged. Replaced to match published version