关键词: 激光等离子体相互作用, 粒子网格模拟, 统一计算设备架构（CUDA）, CUDA优化, GPUDirect RDMA

Abstract: The progress in generating intense ultra-short laser pulse demands more and more for kinetic descriptions of the interaction of such laser pulse with plasmas. Particle-in-cell (PIC) algorithm is a widely-used method in plasma physics to study the trajectories of charged particles under electromagnetic fields. Though there have been some implementations of PIC algorithm on GPU, some important issues still need to be clarified in detail, based on the characteristic of the laser-plasma-interaction simulation. This paper introduces a way to change the original CPU laser-plasma-interaction code into a parameterized adaptive GPU implementation with the whole algorithm ported. Then, this paper develops a series of methods to speed up the particle scatter phase: dynamic duplication algorithm, mix-precision computing and a parameterized particle sorting algorithm. Furthermore, this paper utilizes the GPUDirect RDMA (remote direct memory access) technique in a Kepler cluster and evaluates how it can benefit the MPI communication performance. The results from the numerical experiment show that these optimizations produce a 6.1x speed-up compared with the initial GPU version using the same number of GPUs for the key “Scatter” phase. The speed-up for the MPI communication part is 2.8x when the message size is over 3 KB. All the findings demonstrate that particular optimizations based on the features of the simulation and modern GPU cluster are essential for achieving significantly improved performance.

Key words: laser-plasma-interaction simulation, particle-in-cell (PIC), compute unified device architecture (CUDA), CUDA optimization, GPUDirect RDMA