CPU-GPU协同计算在MOC中子输运异构并行计算中的应用研究

特征线方法（MOC）可以精确求解任意几何的中子输运方程，但该方法收敛慢、计算时间长。本研究基于空间区域分解和特征线并行技术，采用MPI+OpenMP/CUDA编程模型，实现了适用于中央处理器-图形处理器（CPU-GPU）异构系统的二维MOC异构并行算法。为充分利用异构系统中的CPU和GPU计算资源，实现CPU-GPU协同计算，提出动态任务分配模型，根据CPU和GPU的计算能力合理分配计算任务。数值验证结果表明:程序具有良好的计算精度；动态任务分配模型能根据硬件性能给出最佳任务分配方案；5异构节点（包含20块GPU）并行时，相对MPI+CUDA并行模式，采用CPU-GPU协同计算后，程序整体效率提升达到14%。 

Study on Heterogeneous Computing for MOC Neutron Transport Calculation with CPU-GPU Concurrent Calculation

The Method of Characteristics (MOC) is capable to accurately solve the neutron transport equation with arbitrary geometry. However, the MOC suffers from some drawbacks: slow convergence and time consuming. Based on the spatial domain decomposition and the ray parallelization, the parallel 2D MOC algorithm was implemented with MPI+OepnMP/CUDA programming model to leverage the computing power of Central Processing Unit-Graphics Processing Unit (CPU-GPU) heterogeneous high-performance computing systems. In addition, a dynamic workload partitioning scheme was proposed to efficiently take advantage of all the CPU and GPU resources. The workload is appropriately assigned to the CPU and GPU according to their computational capabilities, and all CPUs and GPUs perform the calculation concurrently. The numerical results demonstrate that the parallel algorithm maintains the desired accuracy. Meanwhile, the dynamic workload portioning scheme can provide the optimal workload partition based on the runtime performance. As a result, about 14% improvement is observed in the overall performance compared with the MPI+CUDA parallelization when the CPU-GPU heterogeneous computation is performed on 5 heterogeneous nodes (including 20 GPUs).