PPMLRMHD数值模拟算法的并行性能优化.pdf

HPC China 2014,11,06-08  PPMLR-MHD 数值模拟算法的并行性能优化  1+ 1 王婷 , 张常有 1 ( 中国科学院软件研究所，并行软件与计算科学实验室，北京 100190) The Parallel Performance Optimization of PPMLR-MHD Numerical Simulation Algorithm WANG Ting1+, ZHANG Chang-You 1 1(Laboratory of Parallel Software and Computing Sciences, Institute of Software, CAS, Beijing 100190 , China) + Corresponding author: Phn +86 Fax +86 E-mail : wangting080@ , /jsHomepagePreview.do?deanId=30867 Abstract : For improving the computing speed of PPMLR-MHD (Lagrangian version of Piecewise Parabolic Method-Magnetohydrodynamics) numerical simulation algorithm , the parallel performance optimization is introduced. The piecewise parabolic method PPMLR numerical simulation algorithm on Lagrangian coordinates is used to solve the ideal MHD model, and OpenMP, MPI and an hybrid parallel techniques are used to improve the performance. Experiments are based on Tianhe- 1A and the fat node of DeepComp 7000. The results show that MPI and hybrid parallel algorithms are not as effective as shared memory OpenMP parallel optimization in this algorithm. Using OpenMP parallelization, the good performance result of the speedup of 29.3 at 256 threads on the fat node is showed. Key words: MHD simulation; OpenMP; MPI; Hybrid Parallel; Performance optimization 摘 要: 为提高PPMLR-MHD （Lagrangian version of Piecewise Parabolic Method-Magnetohydrodynamics） 数值模拟算法的计算速度，对其进行了并行化及性能调优。对于理想 MHD 模型，描述了基于拉格朗日坐标的 逐段抛物线方法PPMLR 数值模拟算法，并用OpenMP、MPI 及两者混合并行的方法进行了并行优化，在深腾7000 及天河 1A 系统进行了测试。对于此应用算法,MPI 及混合并行的优化效果不如共享内存的OpenMP 线程并行优 化，OpenMP 并行化在深腾7000 胖节点可扩展到256 线程，加速比为29.3，有较好的并行效果。 关键词: MHD 模拟；OpenMP；MPI；混合并行；性能优化 中图法分类号: TP316 文献标识码: A 1 前言 等离子体和磁场的相互作用的全球