Challenges and Advancements for GPU-Accelerated Discontinuous Galerkin Solvers for Multi-Physics, Compressible Flow Simulations

Umesh Unnikrishnan, ALCF
Seminar Graphic

The emergence of GPU-accelerated hardware in high-performance computing (HPC) systems has significantly enhanced our capability to perform high-fidelity simulations, enabling the inclusion of increasingly complex physics models. This talk presents the development and optimization of a compressible Navier-Stokes solver based on the high-order spectral element discontinuous Galerkin (DG) method that is portable across GPU-accelerated HPC systems. 

We explore the integration of multi-physics modeling capabilities into the solver, including shock capturing techniques, multi-species transport, and chemically reacting flows. The presentation will cover the inherent challenges associated with these high-order numerical schemes and physics models, such as handling discontinuities and ensuring stability and accuracy in high-speed regimes. A key focus of the talk will be placed on identifying and addressing performance bottlenecks in the core computational kernels when implemented on GPUs. We will discuss various strategies and optimizations employed to enhance the performance and efficiency of the solver on GPU-accelerated platforms. The talk aims to provide insights into the practical aspects of developing performance portable DG solvers for advancing multi-physics flow simulations.