Some Experiences of Improving the Speed of Numerical Navier-Stokes Solver Using CUDA

Abstract

In this paper, the Graphics Processor Unit (GPU) is adopted to accelerate the codes of solving incompressible Navier-Stokes equations with two fully different numerical schemes: spectral method and finite difference method. Compared with the codes on CPU, the performance of the codes on GPU is much better, especially when the resolution increases. For the resolution of 2048 × 2048, the acceleration of spectral method reaches 14.45 times. 14.15 times acceleration is achieved when finite difference method in the resolution of 600 × 2400 is applied. We also try to combine MPI (Message Passing Interface) and CUDA (Compute Unified Device Architecture) in the spectral solver to accomplish larger simulations. Due to the inevitable frequent data transfers between Host and Device, the speedup is not so ideal compared with that of the single node, and only 1.82 times acceleration is obtained for the resolution of 4096 × 4096.