| The virtual nuclear reactor is a complex software system built on supercomputers,and it conducts high-fidelity simulation for physical phenomena in reactors.The virtual reactor plays a significant role in many research domains including reactor design,safety,power uprate and life extension.CVR(China Virtual Reactor)1.0 is a software system developed in the Development of CVR Protype and Demonstration project,which is supported by the National Key Research and Development Program of China.CVR 1.0 aims at leveraging the Chinese supercomputers to conduct high-fidelity simulation for physical phenomena in reactors including structural mechanics,thermal-hydraulics,neutronics,and material radiation effects.This paper is based on the widely used structural mechanics methods,finite element method(FEM),according to the typical challenging problems of parallel large-scale mesh generation,physical parameter mapping and numerical calculation in CVR 1.0 high-fidelity structural mechanical simulation.On the supercomputing system of heterogeneous multi-core architecture of domestic CPU+AI accelerator,the research is carried out from three aspects:large-scale mesh generation technology of reactor core fuel assemblies,multi-file physical parameters mapping technology,and large-scale sparse stiffness matrix vector multiplication technology for heterogeneous multi-core architecture.The main work and achievements of the paper are as follows:(1)The fuel assembly amount is large and their structure are complex in fullcore and high-fidelity structural mechanical simulation.The widely-used mesh generation methods are time consuming for generating large-scale meshes.Based on the geometric structure characteristics of reactor calculation model,a large-scale mesh generation method for reactor core fuel assemblies is proposed in the paper.By using solid modeling method,the paper analyzes the modeling of full-core fuel assemblies,as well as the significance,key points and challenges of large-scale mesh generation.According to the data organization requirements of geometric modeling,the paper analyzes the structural features of geometric modeling,and builts a geometric model that can accurately describe the reactor core.An isomorphic modeling method based on key point coordinate space transformation is proposed,which realizes the unified modeling of isomorphic fuel assemblies and full-core fuel assemblies.Moreover,based on the isomorphic model,the efficient automatic generation of large-scale unstructured tetrahedral meshes from isomorphic fuel assemblies to full-core fuel assemblies is realized.The mothod improves user executing efficiency for preprocessing in the high-fidelity simulation for structural mechanics of virtual reactors.(2)In the case of the large-scale finite element mesh file,a multi-file physical parameter mapping technology is proposed to solve the problems that the physical parameter mapping requires a large amount of computing storage and the mapping time is too long.The method uses the minimum sufficient representation method to simplify the mesh data structure,and utilizes the logical organization of multi-file to efficiently map physical parameters with constant and non-constant properties to mesh nodes or elements in large-scale meshes.And according to the data organization form involved in physical parameter mapping,a method of multi-file organization and independent IO operation is proposed.The method has a hash table method and an index tree structure method,which correspond to different types of efficient physical parameter mappings,such as temperature and pressure fields.Based on this method,the general finite element preprocessing execution strategy is optimized to improve the preprocessing efficiency of high-fidelity simulation for structural mechanics of virtual reactors.(3)According to the architectures of Chinses home-grown supercomputers,the paper reports a large-scale sparse stiffness matrix vector multiplication method for reactor structural mechanics software using FEM.Based on the solution principle of structural mechanics,the paper analyzes the features of the solution core,the challenge of high-performance computing,the features of heterogeneous multi-core architecture of CPU+AI accelerator,and HIP(Heterogeneous-compute Interface for Portability)programming model.The technology proposes a CSRVector row algorithm strategy based on CSR storage format,and designs two adaptive reduction algorithms.In addition,the paper further improves the parallel performance by using overlap of access computation,continuous memory access,batch memory access and r-vector memory access.(4)Utilizing the methods proposed in this paper,the high-fidelity simulation of virtual reactor structural mechanics is realized.In addition,the feasibility and efficient parallel performance of the methods are verified on the heterogeneous multi-core supercomputing system of domestic CPU+AI accelerator.The functional verification passed the static simulation of IAEA standard example and the flow induced vibration simulation in single fuel assembly.The parallel performance tests were completed using the full-core calculation model with 712 fuel assemblies.The number of meshes expanded from 158.3 million to 10.25 billion,the computing cores are expanded from 9,804 to 615,330,and the number of multi-files is expanded from 776 to 4,656.The test results show that the parallel scalability is 53.27%,and the calculation error is only 2.7%. |
