GPU Parallelization of a High Order Immersed Boundary Method Fluid Solver

A GPU parallelized high order immersed boundary method fluid solver is developed. Memory management, asynchronous, and algorithm optimization are required to have the highest GPU speed-up potential. Task parallelization must also be implemented through asynchronous and host parallelization (OpenMP). The Poisson solver is the speed-up bottle neck for high convergence iteration count. For small Poisson solver iteration count, the 5th order WENO scheme restricts speed-up. An overall speed-up of ∼4.9 is obtained for a single time step. Speed-up increases with grid size. Multi GPU parallelization requires OpenMP to decrease the GPUs’ idle time. With two GPUs, the increase in speed-up is ∼84.5%, with respect to single GPU, for the largest grid size currently examined.