| 注册
首页|期刊导航|现代应用物理|面向国产天河超算的粒子输运数值模拟程序性能优化研究

面向国产天河超算的粒子输运数值模拟程序性能优化研究

杨林 龚春叶 林鹏

现代应用物理2025,Vol.16Issue(6):20-30,11.
现代应用物理2025,Vol.16Issue(6):20-30,11.DOI:10.12061/j.issn.2095-6223.202508002

面向国产天河超算的粒子输运数值模拟程序性能优化研究

Performance Optimization of Numerical Simulation Programs for Deterministic Particle Transport on Domestic Tianhe Supercomputers

杨林 1龚春叶 2林鹏3

作者信息

  • 1. 国防科技大学计算机学院||国防科技大学高端装备数字化软件湖南省重点实验室||国防科技大学并行与分布式计算全国重点实验室:长沙 410073
  • 2. 国防科技大学计算机学院||国防科技大学高端装备数字化软件湖南省重点实验室||国防科技大学并行与分布式计算全国重点实验室:长沙 410073||国家超级计算天津中心,天津 300457
  • 3. 国防科技大学高端装备数字化软件湖南省重点实验室||国防科技大学并行与分布式计算全国重点实验室:长沙 410073||长沙理工大学计算机学院,长沙 410114
  • 折叠

摘要

Abstract

Numerical simulation of particle transport is one of the most important application areas of high-performance computing,and the strong data dependence in the solution of internal Sn recursive equations seriously restricts the improvement of parallel computing performance.The Tianhe supercomputer adopts a nationally produced processor architecture,which is a benchmark achievement based on the ARM technology route in the international high-performance computing field,demonstrating our country's major breakthrough in independent innovation of supercomputing architecture,and it is crucial to make full use of its scalable vector extension(SVE)acceleration components to improve the performance of applications such as numerical simulation of particle transport.Aiming at the problem of insufficient data-level parallelism and poor performance of Sweep3D on the domestic Tianhe supercomputing processor,this paper fully combines the SVE acceleration component to design a fine-grained data-level parallel algorithm for source item Pn calculation,flux Pn calculation and DSA face currents.The data dependence decoupling is realized through loop splitting and loop transformation technology,and the acceleration algorithm of flux modification and unmodified Sn recursive equations is realized.At the same time,an automatic performance optimization method based on polyhedral compilation technology is explored to further improve performance.The test results show that the full use of the SVE vector architecture can effectively improve the computational performance of the numerical simulation of particle transport,and show strong parallel scalability.

关键词

粒子输运并行算法/SVE向量化/Sweep3D/天河超级计算机/性能优化

Key words

particle transport parallel algorithm/SVE vectorization/Sweep3D/Tianhe Supercomputer/performance optimization

分类

能源科技

引用本文复制引用

杨林,龚春叶,林鹏..面向国产天河超算的粒子输运数值模拟程序性能优化研究[J].现代应用物理,2025,16(6):20-30,11.

基金项目

国家重点研发计划资金资助项目(2025YFB3003605) (2025YFB3003605)

国家自然科学基金资助项目(62032023) (62032023)

现代应用物理

2095-6223

访问量0
|
下载量0
段落导航相关论文