Additional axial decomposition parallelism for MOC solver in the neutron transport code PANDAS-MOC

Abstract

Method of Characteristics (MOC) is an extensively implemented technique for solving the Boltzmann neutron transport equation. In the PANDAS-MOC neutron transport code, MOC is used to determine the 2D radial solution. However, in the Whole-code OpenMP threading hybrid model (WCP) of PANDAS-MOC, the parallel performance of the MOC solver is still restricted by frequent synchronizations and spatial domain decomposition. This paper describes two new parallel algorithms: additional axial decomposition with barriers (AADb) and with status flags (AADsf). They further break down the axial layers based on the existing MPI domain decomposition and split the executed OpenMP threads into multiple groups to increase the workload size and decrease the number of synchronizations per thread. The AADb schedule synchronizes with #pragma omp barrier, but the AADsf schedule replaces those explicit barriers with status flags. C5G7 3D core is used to evaluate their parallel performance. The observed maximum efficiencies of the AAD algorithms are around 0.87, and their efficiencies were approximately 1.15x-1.32x that of the previous No-atomic schedule when using an identical amount of computing resources. Moreover, when launching many OpenMP threads, the AADsf schedule performs better than the AADb schedule.