Research On Large-Scale Particle Transport Heterogeneous Computing On Tianhe-2A Supercomputer

High performance computing(HPC)plays a vital role in the fields of scientific research and engineering technology,such as aerospace,astrophysics,biomedicine,weather,materials science,and nuclear engineering.Particle transport simulation is one of the most important applications among them.Since the discovery of neutrons in the1930 s,researchers have foucsed on various types of microscopic particles.Particle transport theory has been applied in many important fields such as astrophysics,nuclear physics,and medical radiotherapy.The particle transport equation(Boltzmann equation)is a mathematical physics equation describing the particle transport process,and its solution algorithm has always been the key to research.With the development of science and technology and the realistic need of practical application,the accuracy and real-time requirements for particle transport simulation are getting higher and higher.This paper is dedicated to using the Tianhe-2A high-performance computing platform to optimize the performance of particle transport simulation applications and improve the efficiency and utilization of high-performance computing platforms.The main contributions are as follows:1.We proposed a three-dimensional Cartesian geometry structured mesh particle transport data-level parallel algorithm,which fully exploits the finite parallelism in flux scanning and develops the parallelism of internal calculation of grid cells.In this algorithm,we effectively parallelized the solving of recursive discrete ordinate equations.We constructed a heterogeneous parallel algorithm by using BCL and ACL to transmit data between CPU and Matrix-2000.We use Open MP thread-level parallel optimization on the Matrix-2000 side to improve performance.Our optimization on thread-level parallelism includes iteration source calculation,I-line grid column calculation,and flux error calculation.In the single node test,this method achieves a maximum of 13.2 speedups on the Matrix-2000 superacceleration node.The extension test of the millioncore scale was completed on the Tianhe-2A supercomputer,the test efficiency was high,and the program has good scalability.2.Based on the existing particle transport Monte Carlo simulation algorithm MCNP program,we proposed a particle transport heterogeneous computing algorithm for CPU-MT2000 heterogeneous system.According to the architecture and access characteristics of the domestic accelerator Matrix-2000,we proposed an efficient data structure suitable for program parallelism and a task division method based on the number of particles and threads.Meanwhile,we optimized the original serial data collection communication mode,and proposed a new binary tree communication mode,which reduced the communication time greatly,with the speedup reaching to 17.7 at most.By optimizing the communication mode and based on the MPI-SCIF-Open MP programming model,the parallel program based on CPU-MT2000 heterogeneous computing can be scaled up to 450 k cores,and the parallel efficiency compared to 50 k cores is stabilized at 22.54%.