Performance optimization and analysis of the unstructured Discontinuous Galerkin solver on multi-core and many-core architectures

TL;DR

This paper evaluates the performance of a Discontinuous Galerkin solver on multi-core and many-core architectures, demonstrating significant speedups through various programming models and optimization techniques for CFD applications.

Contribution

It presents an in-depth comparison of CUDA, OpenACC, and OpenMP implementations of a DG solver, including optimization strategies and performance analysis on CPU and GPU systems.

Findings

CUDA achieved up to 42.9x speedup

OpenACC achieved up to 35.3x speedup

OpenMP achieved up to 8.1x speedup

Abstract

The discontinuous Galerkin (DG) algorithm is a representative high order method in Computational Fluid Dynamics (CFD) area which possesses considerable mathematical advantages such as high resolution, low dissipation, and dispersion. However, DG is rather computationally intensive to demonstrate practical engineering problems. This paper discusses the implementation of our in-house practical DG application in three different programming models, as well as some optimization techniques, including grid renumbering and mixed precision to maximize the performance improvements in a single node system. The experiment on CPU and GPU shows that our CUDA, OpenACC, and OpenMP-based code obtains a maximum speedup of 42.9x, 35.3x, and 8.1x compared with serial execution by the original application, respectively. Besides, we systematically compare the programming models in two aspects: performance and productivity. Our empirical conclusions facilitate the programmers to select the right platform with a suitable programming model according to their target applications.