Porting Production Level Quantum Chromodynamics Code to Graphics Processing Units - A Case Study

摘要

We present our findings and results of a project to port an existing large lattice QCD codebase to run on GPUs and clusters of GPUs. Our design principles from the start were to strive for both productivity and performance, while tackling the problems presented by a large constantly moving codebase. The resulting simulator reproduces the original results while running up to 11 times faster than our highly optimized CPU-code and meeting productivity requirements. Multi-GPU support was implemented using MPI and scaling across nodes shows good weak scaling. We also contemplate the consequences of the dawning of the parallel computing era from a lattice QCD point of view and analyze where state-of-the art contemporary parallel computing architecture could be improved.