Parallel optimization approaches for scattering source term and angle merging in MOC high-order anisotropic computation on GPUs

Abstract

The introduction of high-order anisotropic scattering and thousand-group calculations significantly increases the computational workload of the MOC method, posing difficulties for practical parallel computation on GPUs. In this paper, a GPU-based MOC algorithm is optimized, focusing on the performance optimization of the kernel function that merges high-order anisotropic angular flux density into scalar flux density moments, as well as the optimization of the scattering source term calculation method in thousand-group calculations. Two parallel algorithms based on azimuthal and polar angles are proposed for high-order anisotropic scattering. For the thousand-group calculation, a data structure is designed according to the physical characteristics of the anisotropic scattering matrix, and then two optimization algorithms are proposed from the perspectives of memory access and parallelism. Based on the KAIST-3A reference problem, the parallel expansion of the polar angle in the merging calculation is more advantageous. Based on the KAIST-3A reference problem, the parallel expansion of the polar angle in the merging calculation is found to be more advantageous. Based on a self-designed fast reactor fuel lattice, it is concluded that the parallelism-oriented calculation method for the thousand-group anisotropic scattering source term is more efficient. Finally, the Roofline model is used to analyze the advantages of the algorithm.