A Hybrid CPU/GPU Parallel Algorithm for Coupled Eulerian and Vortex Particle Methods

Abstract

A combined CPU/GPU parallel algorithm for a hybrid Eulerian/Lagrangian CFD method is presented. Specifically, the optimized multi-GPU algorithm for direct Biot- Savart particle-particle interactions, associated with Vortex Particle Methods, is described in detail. The particle integration algorithm is implemented using multiple general purpose graphics processing units (GPU’s) via NVIDIA’s Compute Unified Device Architecture (CUDA). Details of the CUDA algorithm and the Eulerian/Lagrangian coupling method are described. Significant speed-up is reported when using the hybrid CPU-GPU parallel algorithm for the Biot-Savart induction operation for velocity and velocity-gradient. Further, multiple-levels of parallelism is demonstrated with linear scalability a distributed/shared-memory MPI + OpenMP + CUDA method.