Using Hierarchical Parallelism to Accelerate the Solution of Many Small Partial Differential Equations

TL;DR

This paper explores hierarchical parallelism techniques to enhance GPU performance in solving many small PDEs, comparing NVIDIA Multi-Process Service and Kokkos-based methods, with the latter showing superior results.

Contribution

It introduces and compares two novel hierarchical parallelism strategies for GPU acceleration of small PDEs, demonstrating significant performance improvements.

Findings

Kokkos hierarchical parallelism outperforms NVIDIA Multi-Process Service

Both methods improve GPU parallel performance

The second method yields the greatest performance gains

Abstract

This paper presents efforts to improve the hierarchical parallelism of a two scale simulation code. Two methods to improve the GPU parallel performance were developed and compared. The first used the NVIDIA Multi-Process Service and the second moved the entire sub-problem loop into a single kernel using Kokkos hierarchical parallelism and a PackedView data structure. Both approaches improved parallel performance with the second method providing the greatest improvements.