Performance Comparison of HPX Versus Traditional Parallelization Strategies for the Discontinuous Galerkin Method

Abstract

As high performance computing moves towards the exascale computing regime, applications are required to expose increasingly fine grain parallelism to efficiently use next generation supercomputers. Intended as a solution to the programming challenges associated with these architectures, High Performance ParalleX (HPX) is a task-based C++ runtime, which emphasizes the use of lightweight threads and algorithm-dependent synchronization to maximize parallelism exposed by the application to the machine. The aim of this work is to explore the performance benefits of an HPX parallelization versus a MPI parallelization for the discontinuous Galerkin finite element method for the two-dimensional shallow water equations. We present strong and weak scaling results comparing the performance of HPX versus a MPI parallelization strategy on Knights Landing architectures. Our results indicate that for average task sizes of $$3.6\,{\mathrm {m s}}$$ , HPX’s runtime overhead is offset by more efficient execution of the application. Furthermore, we demonstrate that running with sufficiently large task granularity, HPX is able to outperform the MPI parallelization by a factor of approximately 1.2 for up to 128 nodes.