The Fat-Link Computation on Large GPU Clusters for Lattice QCD

Graphics Processing Units (GPU) are becoming increasingly popular in high performance computing due to their high performance, high power efficiency and low cost. In this paper, we present results of an effort to implement the fatlink computation -- an important component of many lattice quantum chromo dynamics (LQCD) calculations -- on GPU clusters using the QUDA framework. Two implementations, one similar to the original CPU algorithm in the MILC code and one based on the idea of reduced communication by redundant computations, are presented and their relative advantages are discussed. In strong-scaling tests on up to 384GPUs on Longhorn and 256 GPUs on Keene land GPU clusters, where the CPU core to GPU ratio is 4:1 in both clusters, we achieved up to 11.4x and 8.7x node speedup when running on the two GPU clusters, respectively.