数值堆结构力学高精细模拟的并行实现和优化技术

Parallel Implementation and Optimization Technology for High-precision Simulation of China Virtual Reactor Structural Mechanics

反应堆结构力学模拟是对堆芯组件的静动态形变位移和受力进行分析,以保证堆芯安全、稳定运行。不同于传统的结构力学模拟,本文结合异构超级计算机上的AMD GPU体系架构进行高精细大规模模拟。在数学物理模型方面使用有限元分析方法,通过充分利用反应堆组件的同构性,对单根组件或局部组件的网格数据进行坐标空间变换,自动生成全堆芯结构大规模网格文件,实现全堆芯百亿网格的剖分。采用共轭梯度法对具有正定特征的大规模稀疏刚度矩阵线性方程组进行高效数值求解,实现精细化分析获知全堆芯组件的应力应变空间分布状态。另外,采用数据预取和数据对齐的并行优化方法,使求解器计算性能提高了30%以上。模拟结果展示了该方法的可行性和正确性,并在计算核心数从9504扩展至615168时,对具有102亿网格规模的中国实验快堆(CEFR)全堆芯静力学问题进行计算,并行弱可扩展性达到53%。

China virtual reactor(CVR)structural mechanics is responsible for analyzing the static and dynamic deformation displacement and stress of the reactor core assembly to ensure the safety of the reactor core.Compared with the traditional simulation,the heterogeneous supercomputer based on AMD GPU architecture was used for high-precision large-scale simulation in this paper.The three-dimensional finite element analysis method was used in the mathematical and physical model.By using the isomorphism of reactor assembly,the mesh data of single or local assembly were transformed in coordinate space,and the large-scale mesh file of whole core structure was generated automatically.The whole core can be divided into 10 billion meshes.The conjugate gradient method was used to solve the large-scale sparse stiffness matrix linear equation with positive definite characteristics,and the stress-strain spatial distribution of the core assembly could be obtained.In addition,the parallel optimization method of data prefetching and data alignment were adopted to improve its performance by more than 30%.The simulation results verify the feasibility and correctness of the method.When the computing cores are extended from 9504 to 615168,China Experimental Fast Reactor(CEFR)whole core static with the meshes of 10.2 billion is calculated,and the parallel weak scalability reaches 53%.