CPU-GPU concurrent computing algorithm of particle transport using discontinuous finite element discrete ordinates with unstructured grids

Abstract: Discontinuous finite element method (DFEM) based discrete ordinates (SN) for solving particle transport is a heavy burden that costs plenty of calculation time in multi-physics simulation. The emergence of the graphic processing unit (GPU) has triggered a change in computing capabilities, which provides a new way for DFEM SN calculation. However, the effect of the general GPU acceleration algorithm for particle transport with unstructured grids is limited due to the dependence of the SN spatial wavefront, and the computing capability of GPU is not fully utilized. This paper proposes a CPU-GPU concurrent computing algorithm, which is carried out as a large-scale linear system with a block Jacobi parallel strategy. In the algorithm, the fission and scattering source terms are calculated on the GPU, and the inflow terms are calculated on the CPU simultaneously. Moreover, the coarse-grained domain decomposition parallelism and fine-grained angular parallelism are adopted, where the data transmission and computation are performed simultaneously. This new algorithm not only takes full advantage of the huge number of threads on the GPU but also has no limits to the total GPU utilization. Numerical results of typical neutron transport benchmarks show that the CPU-GPU concurrent computing algorithm achieves a 24-60 times acceleration effect than the CPU algorithm and 6-11 times acceleration than the general GPU algorithm.