A GPU-Accelerated Hybridizable Discontinuous Galerkin Method for Linear Elasticity

Authors

  • Maurice S. Fabien Department of Computational and Applied Mathematics, Rice University, 6100 Main MS-134, Houston, TX 77005, USA

DOI:

https://doi.org/10.4208/cicp.OA-2018-0235

Keywords:

GPU-acceleration, discontinuous Galerkin, hybridization, multigrid, performance analysis.

Abstract

We design and analyze an efficient GPU-accelerated hybridizable discontinuous Galerkin method for linear elasticity. Performance analysis of the method is done using the state-of-the-art Time-Accuracy-Size (TAS) spectrum. TAS is a new performance measure which takes into account the accuracy of the solution. Standard performance measures, like floating point operations or run-time, are not completely appropriate for gauging the performance of approximations of continuum mechanics problems, as they neglect the solutions accuracy. A standard roofline model demonstrates that our method is utilizing computational resources efficiently, and as such, significant speed ups over a serial implementation are obtained. By combining traditional performance measures and the novel time-accuracy measures [7] into our performance model, we are able to draw more complete conclusions about which discretizations are best suited for an application. Several numerical experiments validate and verify our numerical scheme.

Published

2019-12-07

Issue

Section

Articles