Research On Performance Modeling And Parallel Methods Of Monte Carlo Particle Transport Programs

Particle transport theory studies the migration laws of microscopic particles in media,and has been widely used in many scientific and engineering fields,such as nuclear physics and biomedicine.Monte Carlo(MC)method is widely used in solving particle transport problems.It makes up the shortcoming that deterministic methods are not suitable for solving problems with large deformation and complex three-dimensional geometry.However,MC particle transport program accesses memory randomly and hass many branches,resulting low single-node computation efficiency.Its randomness brings great difficulty in theoretical analysis and parallel method design for single-node performance modeling and research on parallel methods.In order to overcome these challenges,this paper analyzes and studies the performance modeling and parallel methods of the singlenode MC program.In order to further analyze the impact of factors,such as branch and memeory accesses,on the single-node performance,and improve the single-node performance,the main innovations of this paper are as follows:1.Research on the performance model of single-node MC particle transport program.Taking OpenMC as the target program,a single-core performance model based on sub-routine characteristic analysis is proposed.It divides the runtime of single-core runtime into base time,time to access memory and time introduced by branch perdiction failure.It analyze the program trace to obtain the features of subroutines,including instruction,memory access and branch.Based on these features,the single-core runtime of each subroutine is predicted.Based on the independence and similarity between threads of OpenMC multicore execution,a multicore performance model based on shared cache analysis is proposed.Based on the characteristic L3 cache shared by multicore,a prediction method of miss count of shared L3 cache and private L2 cache based on stack reuse distance is designed.The experimental results on two platforms show that the proposed model can accuratly predict the runtime proportion of each subroutine.Based on the model,this paper also analyzes the CPI composition of each subroutine and the effect of branch predictor and the size of L3 cache.This paper is the first research on performance modeling of single-node MC particle transport program.2.Research on parallel methods of single-node multigroup MC particle transport programs.Based on multigroup MC proxy program Quicksilver,an event-based method based on queue is proposed.This method improves the performance of event-based method on the GPU by replacing dead particles with new particles,hybrid history-based and event-based method and particle redistribution based on prefix sum.The experimental results show that the optimizations achieve more than 4x speedup than the original method.Besides,based on the implementation of different parallel methods,this paper compares the performance difference of hi story-based and event-based methods with different cross sections and analyzes the performance affecting factors in details,which provides a guidance for the further study of parallel methods of MC particle transport with different cross sections on different architectures.The experimental results show that memory access pattern and feature of cross sections are the main factors that affect the performance of MC program on the GPU.