Effective Minimally-Invasive GPU Acceleration of Distributed Sparse Matrix Factorization

Sparse matrix factorization, a critical algorithm in many science and engineering applications, has had difficulty leveraging the additional computational power afforded by the infusion of heterogeneous accelerators in HPC clusters. We present a minimally invasive approach to the GPU acceleration of a hybrid multifrontal solver, the Watson Sparse Matrix Package, which is already highly optimized for the CPU and exhibits leading performance on distributed architectures. The novel aspect of this work is to demonstrate techniques for achieving substantial GPU acceleration, up to 3.5x, of the sparse factorization with strategic, but contained changes to the original, CPU-only, code. Strong scaling results show that performance benefits scale to as many as 512 nodes (4096 cores) of the Blue Waters supercomputer at NCSA. The techniques presented here suggest that detailed code reorganization may not be necessary to achieve substantial acceleration from GPUs, even for complex algorithms with highly irregular compute and data access patterns, like those used for distributed sparse factorization.