GPU-Accelerated Simulations of Turbulent Flows With the Immersed Boundary Method
William van Noordt, Joel A. McQuaid, Johan Larsson, Christoph Brehm
Abstract
In this work, a completely new, accelerated version of the Cartesian AMR/immersed-boundary code CHAMPS called CHAMPS+ is presented and validated for wall-modelled large-eddy simulations with the immersed boundary method. The GPU implementation strategy is presented for the performance-critical flux divergence algorithm, including data loading patterns and memory layout. Performance is evaluated for all GPU kernels for both a periodic flow test case (Taylor-Green vortex) and an external-flow case (A-airfoil at high Reynolds number). Considerable performance is shown to be achieved, with the mean floating-point operation throughput achieved over full simulations being 30-35% of the theoretical peak throughput of the device. Storage requirements of CHAMPS+ are shown to allow for 8.8 - 14.8 million cells per gigabyte of GPU memory.