CAJ | 학술논문

本文分析了非结构网格多群粒子输运Sn方程求解的并行性,拟合多核机群系统的特点,设计了MPI/OpenMP混合程序,针对空间网格点采用区域分解划分,计算结点间基于消息传递MPI编程,每个MPI计算进程在计算过程中碰到关于能群的计算,就生成多个OpenMP线程,计算结点内针对能群进行多线程并行计算.数值测试结果表明,非结构网格上的粒子输运问题的混合并行计算能较好地匹配多核机群系统的硬件结构,具有良好的可扩展性,可以扩展到1 024个CPU核.
본문분석료비결구망격다군입자수운Sn방정구해적병행성,의합다핵궤군계통적특점,설계료MPI/OpenMP혼합정서,침대공간망격점채용구역분해화분,계산결점간기우소식전체MPI편정,매개MPI계산진정재계산과정중팽도관우능군적계산,취생성다개OpenMP선정,계산결점내침대능군진행다선정병행계산.수치측시결과표명,비결구망격상적입자수운문제적혼합병행계산능교호지필배다핵궤군계통적경건결구,구유량호적가확전성,가이확전도1 024개CPU핵.
The parallel performance of solving the multi-group particle transport equations on the unstructure meshes is analyzed Adapting to the characteristics of multi-core cluster systems, this paper desgins a MPI/OpenMP hybrid parallel code. For the meshes, the space domain decomposition is adopted, and MPI between the computations of multi-core CPU nodes is used. When each MPI process begin to compute the variables of the energy groups, several OpenMP threads will be forked, and the threads start to compute simultaneously in the same mutli-core CPU node. Using the MPI/OpenMP hybrid parallel code, we solve a 2D mutli-group particle transport equation on a cluster with mutli-core CPU nodes, and the results show that the code has good scalability and can be scaled to 1024 CPU cores.