三维中子/光子输运S_N程序的并行重构

Reconstruction and Parallelization of 3D SN Program for Neutron/Photon Transport

基于离散纵标方法(SN)的三维中子/光子输运模拟程序因计算量大、计算时间长,难以很好地满足实际工程应用需求。本文基于几何区域剖分方法,在并行自适应结构网格应用支撑软件框架(JASMIN)上通过重新设计底层数据结构、功能模块并组装计算流程,实现了遗产程序的并行重构,提高了程序效率,拓展了计算大规模物理问题的能力。针对外源和临界计算两类典型问题,利用遗产程序自带测试例题和国际基准题验证了重构后并行程序的正确性。在可扩展性测试中,512个处理器核上模拟数亿自由度临界问题时,并行效率可达37.07%。

The application of the neutron/photon transport simulation program based on 3D SN method is limited by the intensive computational requirement and prohibitive computational time. Based on the geometry domain partition method, this paper reports our efforts on the reconstruction and parallelization of the legacy SN program by redesigning the data structures, the functional module as well as the compute flow on J Adaptive Structured Meshes applications Infrastructure(JASMIN). The correctness of the reconstructed program is proved by modeling the fixed-source and k-eigenvalue benchmark problems. In the strong scalability test, the parallel efficiency of the reconstructed program can reach up to 37.07% on 512 CPU cores when modeling the k-eigenvalue problem with 100 million degree-of-freedom.