Parallelization and optimization of RMC for criticality computing based on the heterogeneous architecture of the Sunway TaihuLight supercomputer

摘要

Heterogeneous parallel computing is being widely used for particle transport Monte Carlo codes on various platforms. However, there is currently no related research on the Sunway TaihuLight system, a heterogeneous supercomputer with a unique architecture. In this study, RMC, a key particle transport Monte Carlo code, was parallelized for use on Sunway TaihuLight. Some optimization methods including a thread-private k(eff) calculator mechanism based on Pthreads were introduced into RMC. The criticality computing module of the optimized RMC was then implemented on Sunway TaihuLight with an MPI-Athread parallel model and optimization methods oriented to the system's special architecture. The results were verified by comparison with the original RMC. The criticality calculation is 125.3 times faster with 512 core groups each with 65 cores than the original serial RMC on 1 core on an Intel Xeon X5670. Stable weak scalability of the parallel calculation is shown with over 1 million cores. (C) 2020 Elsevier Ltd. All rights reserved.