基于FMM-PM方法的宇宙N体模拟在GPU上的实现和优化

【目的】本文在多GPU平台上,对基于快速多极子方法(FMM)和粒子网格方法(PM)的天文N体模拟软件PHoToNs的核心函数进行CUDA加速实现和性能优化。【方法】主要优化方法包括算法的参数优化、页锁定内存和CUDA流优化、混合精度和快速数学库优化等。【结果】优化后的短程力相互作用核心函数在Titan V的GPU平台上采用4张GPU卡的计算速度相对采用4个Intel Xeon CPU核提高了约410倍。【结论】本文的优化技术可为其它高性能GPU异构平台上的进一步算法研究和超大规模天文N体模拟提供支撑。

The Implementation and Optimization of Cosmological N-Body Simulation by FMM-PM Method on GPUs

[Objective]In this paper,the kernel functions of PhoToNs,which is an astronomical N-body simulation software based on the fast multipole method(FMM)and particle grid method(PM),are accelerated and optimized for CUDA on a multi-GPU platform.[Methods]The main optimization methods adopted in CUDA kernels include:algorithm parameter optimization,use of page-locked memory and CUDA streams,and use of mixed precision and fast math library.[Results]The kernel function of short range force interaction is deeply optimized,which achieves a speedup of about 410 times faster on four Titan V GPUs than the pure MPI code running on four Intel Xeon CPU cores.[Conclusions]Optimization methods in this paper can support further algorithm research and hyperscale N-body simulation on other high performance GPU-based heterogeneous platforms.