An Improved GPU Acceleration Framework for Smoothed Particle Hydrodynamics

GPU has drawn much attention on accelerating SPH applications, which need high computational requirements. To eliminate the performance bottlenecks, this paper proposes an efficient GPU-accelerated framework for SPH computation on high-performance computing systems. To this end, several performance acceleration tools are developed to speed up the GPU implementation. The first one puts forward an efficient block size for GPU kernels, the second one determines the optimal workload for each CUDA thread and the third one uses reduction of global memory accesses to optimize data layout. Finally, an improved method makes use of simple GPU kernels instead of heavy ones when computing particle interactions. Comparison and analysis are made among the parallel results by the GPU implementations. As a result, with hundreds of thousands of particles run on advanced GPUs, the performance by the improved GPU implementation is at least 700 times higher than that by the serial CPU code. Compared to multi-thread CPU implementation, the performance increases by a factor over 200 on different GPUs.