Acceleration of CFD engineering software on GPU and MIC

Abstract

CartSolver is widely used three dimensional Euler solver software for Cartesian grids. In this paper, we use the latest many-core accelerators such as NVIDIA Fermi C2050, NVIDIA Kepler K20 and Intel MIC to do the acceleration, and achieve expected speedup over the serial solver. On the GPU platform, two versions of accelerated CartSolver are implemented and optimized. For MIC, we employ various optimization methods in order to achieve the best performance by an open source performance analysis tool. The differences in architecture and programming model between GPU and MIC are also discussed. In the experiments, the correctness and accuracy of the solvers is validated, and the great effect of optimization methods is also proved. Finally, a new criterion for measuring the workload is proposed, and several recommendations on selecting suitable accelerators for CFD engineering software are given on the base of the comparison of the criteria.