原子能科学技术 ›› 2021, Vol. 55 ›› Issue (9): 1581-1593.DOI: 10.7538/yzk.2021.youxian.0288

• 数值反应堆原型系统开发及示范应用 • 上一篇    下一篇

数值堆热工流体高精细并行模拟优化技术研究

董玲玉;周志锋;戴潮虎;赵珂;吴宗芸;刘天才;赵民富;杨文;胡长军   

  1. 北京科技大学 计算机与通信工程学院,北京100083;中国原子能科学研究院 反应堆工程技术研究所,北京102413
  • 出版日期:2021-09-20 发布日期:2021-09-20

High-precision Parallel Simulation and Optimization Technology of Thermal Fluid in Numerical Reactor

DONG Lingyu;ZHOU Zhifeng;DAI Chaohu;ZHAO Ke;WU Zongyun;LIU Tiancai;ZHAO Minfu;YANG Wen;HU Changjun   

  1. School of Computer and Communication Engineering, University of Science and Technology Beijing, Beijing 100083, China;Division of Reactor Engineering Technology Research, China Institute of Atomic Energy, Beijing 102413, China
  • Online:2021-09-20 Published:2021-09-20

摘要: 热工流体模拟是数值反应堆的重要组成部分,高精细、大规模数值计算是实现高保真数值模拟的基础。使用计算流体力学(CFD)软件进行高精细、大规模数值模拟对计算资源和存储资源提出了巨大挑战,需依赖超级计算机并行实现。本文以基于谱元法求解N-S(Navier-Stokes)方程的数值方法为研究对象,针对区域分解和基于典型混合架构国产超级计算机的并行优化两个核心问题,提出了一种面向海量精细网格的混合并行递归谱二分法实现的大规模区域分解方法,建立了一套以小矩阵乘为核心的申威(SW26010处理器)众核架构并行优化技术。混合并行大规模区域分解方法在天河二号超算上进行测试,相比开源CFD软件Nek5000的串行区域分解模块性能提升约95%;面向申威的小矩阵乘优化在神威·太湖之光超算上进行测试,当谱元阶数达到24时性能提高约51.9%。两种技术均在中国数值反应堆核心软件CVR-PACA中得以应用。

关键词: 计算流体力学, 谱元法, 并行区域分解, 矩阵乘并行优化技术, 数值反应堆

Abstract: Thermal fluid simulation is an important part of the numerical reactor. High-precision large-scale numerical calculation is the basis for realizing high-fidelity numerical simulation. The use of computational fluid dynamics (CFD) software for high-precision and large-scale numerical simulation poses a huge challenge to computing resources and storage resources, and requires parallel implementation of supercomputers. The numerical method based on the spectral element method was taken to solve the N-S (Navier-Stokes) equation as the research object in the paper. Aiming at the two core problems of domain decomposition and parallel optimization of Chinese supercomputers based on the typical hybrid architecture, a hybrid parallel recursive spectral bisection method for massive fine meshes was proposed to achieve large-scale domain decomposition, and a set of applications centered on small matrix-matrix multiplication was established. The hybrid parallel large-scale domain decomposition method was tested on the Tianhe-2 Supercomputer. Compared with the serial domain decomposition module of the open source CFD software Nek5000, the performance is improved by about 95%. The small matrix-matrix multiplication optimization for SW26010 was tested on Sunway TaihuLight Supercomputing. When the order of the spectral element reaches 24, the performance is improved by about 51.9%. Both technologies were applied in the core software CVR-PACA of China Numerical Reactor.

Key words: computational fluid dynamics, spectral element method, parallel domain decomposition, matrix multiplication parallel optimization technique, numerical reactor