Solving Particle Transport Equation Based On CPU-GPU Heterogeneous Platform

Particle transport theory and particle transport equations have been widely used in many fields,which have promoted the development of large-scale scientific engineering and national economic construction.Sweep3D uses discrete ordinates method(Sn)to solve a time-independent,three-dimensional Cartesian geometry and single energy group particle transport equation,and the computational complexity is very huge.In recent years,GPU's floating-point performance and programmability are constantly improving,CPU-GPU heterogeneous architecture and collaborative parallel computing is an important trend in current high-performance computing systems,and their development has brought opportunities and challenges to the solution of particle transport equations.In this paper,we study the parallel implementation of Sweep3D based on GPU and the collaborative parallel implementation of Sweep3D based on CPU-GPU heterogeneous platform,and carry out the experiments on the "TianHe-1A" system in Changsha National Supercomputing center.The main works of this paper are as follows:(1)Research on GPU-based parallel algorithm of Sweep3D.This paper firstly explain the algorithm flow of Sweep3D on GPU,and then a three-diagonal sweep method is used to assign the grid cells to the GPU threads to deal in parallel,in the wavefront sweep process of Sweep3D.In order to improve the performance of GPU parallel algorithm,three optimization methods,including global memory access optimization,data dependency elimination and thread block size optimization,are used to optimize the GPU parallel algorithm.(2)Research on CPU-GPU-based collaborative parallel algorithm of Sweep3D.In this paper,the different computational processes of Sweep3D are assigned to the appropriate computing device,and the most time-consuming wavefront sweep process is assigned to both the CPU and the GPU for collaborative parallel computing,fully leveraging the computing resources of the CPU and the GPU.A two-diagonal sweep method is proposed to effectively map the grid cell computation to both the CPU and GPU.In order to ensure the load balancing between the CPU and the GPU,the static and dynamic task partitioning models are used to assign the grid cells.The problem of data transmission overhead in dynamic task partitioning model is effectively solved by combining discrete grid cell data and overlapping data transmission and computing.(3)Based on the MPI-OpenMP-CUDA parallel programming model,the task partitioning model,communication optimization and algorithm performance experiments of CPU-GPU collaborative parallel algorithm are carried out on the"Tianhe-1A" system.It is verified that the dynamic task partitioning model can ensure better load balancing between the CPU and the GPU than the static task partitioning model.After the data transmission optimization,data transmission time accounted for the total execution time significantly decreased on the GPU.Collaborative parallel algorithm of Sweep3D in a single-node and multi-node has achieved a better speedup relative to the CPU parallel algorithm.