Optimization of heterogeneous parallel algorithm for Monte Carlo neutron transport simulation aiming at thread divergence Issues

• The heterogeneous parallel computing for Monte Carlo method is implemented. • Novel parallel optimization algorithms for thread divergence problems are proposed. • The AOS memory layout performs better than SOA in history-based tracking algorithm. • GPU-based parallelism provides significant acceleration for Monte Carlo simulations . The Monte Carlo simulation of large-scale neutron transport problems has always faced the problem of slow computation. In order to fully exploit the acceleration advantage of heterogeneous parallelism on the Monte Carlo neutron transport simulation, this paper carries out research around the history-based neutron tracking algorithm, deeply explores the adaptation of the Monte Carlo algorithm and heterogeneous parallelism. Aiming at the thread divergence problem, optimization strategies for particle tracking algorithm are proposed to ensure load balancing among parallel threads. In addition, to mitigate the impact of global memory access latency, the memory layout of the particle state data is reasonably arranged by comprehensively considering the random memory access of Monte Carlo algorithm and the hardware characteristics of GPU . The reliability and efficiency of heterogeneous parallel algorithm are validated in calculations of benchmarks, the computing performance on an NVIDIA A800 GPU is equivalent to the performance of 62–87 CPU cores.