Performance comparison of Eulerian kinetic Vlasov code between flat-MPI parallelism and hybrid parallelism on Fujitsu FX100 supercomputer
Performance comparison of Eulerian kinetic Vlasov code between flat-MPI parallelism and hybrid parallelism on Fujitsu FX100 supercomputer
复制标题
DOI：
10.1145/2966884.2966891
复制
发表时间：
2016-09
期刊：
影响因子：
--
通讯作者：
T. Umeda;K. Fukazawa
中科院分区：
文献类型：
--
作者：
T. Umeda;K. Fukazawa
本研究涉及弗拉索夫模拟程序，该程序求解空间等离子体的第一原理动力学方程称为弗拉索夫方程。本文采用平面- mpi和MPI-OpenMP混合并行两种方法对具有两个空间维度和三个速度维度的五维弗拉索夫码进行并行化。这两种并行的Vlasov代码在大规模并行超级计算机富士通FX100上进行了基准测试，该超级计算机是根据日本K计算机的第二代后架构开发的。在当前的性能比较中，我们将每个节点的线程数从1（平mpi并行度）更改为32。结果表明，对于任意数量的计算节点，OpenMP-MPI混合并行性优于扁平mpi。每个节点的最佳线程数取决于计算节点的数量。结果表明，计算节点越多，每个节点的最佳线程数越大。这是因为用于迭代方法收敛性检查的MPI集体通信子程序的通信时间可以通过使用OpenMP-MPI混合并行性减少进程总数来减少。
The present study deals with the Vlasov simulation code, which solves the first-principle kinetic equations called the Vlasov equation for space plasma. In the present study, a five-dimensional Vlasov code with two spatial dimension and three velocity dimensions is parallelized with two methods, the flat-MPI and the MPI-OpenMP hybrid parallelism. The two types of the parallel Vlasov code are benchmarked on massively-parallel supercomputer Fujitsu FX100, which has been developed with the second-generation post architecture of the K computer in Japan. In the present performance comparison, we vary the number of threads per nodes from 1 (the flat-MPI parallelism) to 32. The result shows that the OpenMP-MPI hybrid parallelism outperforms the flat-MPI for any number of compute nodes. There is an optimum number of threads per nodes depending on the number of compute nodes. It is shown that the optimum number of threads per node becomes larger on a larger number of compute nodes. This is because the communication time of an MPI collective communication subroutine, which is used for convergence check of iterative methods, can be reduced by decreasing the total number of processes with the OpenMP-MPI hybrid parallelism.