Research On Large-scale Parallel Algorithm For Atomic Scale Material Irradiation Damage Simulation

Irradiation damages to nuclear materials in high temperature,high pressure and high radiation condition directly affect the safety and lifetime of the reactor.Numerical simulations are one of the major approaches for studying the irradiation damage of nuclear materials.Among them,in the atomic scale,Molecular Dynamics(MD)and kinetic Monte Carlo(KMC)are two important methods,which are usually used to study the primary damage production and the long-time evolution of defects in materials,respectively.In order to realize the fine simulation of MD and KMC and to predict the macroscopic properties of materials,it is crucial to extend the spatial and temporal scales of their simulations.Because of the large computational requirement and storage capacity of atomic-scale simulations,it can be an effective and inevitable approach to util the powerful computing power of supercomputers.However,this approach still suffers at least the following challenges:the "storage challenge" of efficient particle storage and organization,and the "scaling challenge" for large-scale simulation on supercomputer systems,and the "heterogeneity challenge" of the homegrown heterogeneous hybrid architecture,and the "coupling challenge" between atomic-scale methods.In this research,the key technologies for the massively parallel MD and KMC simulation on homegrown supercomputers are carried out.The main work and innovations are summarized as follows.1)To address the challenges of "storage","scaling","heterogeneity" and "coupling",the atomic-scale unified architecture was designed to unify the particle data structure,massively parallel and communication modes,and potential function computation operators among different atomic-scale methods.The unified architecture supports efficient simulation under tightly coupled mode and smooth inter-system migration for MD and KMC simulations.In this unified architecture,we designed an efficient particle data-structure for the homegrown supercomputers of heterogeneous hybrid architecture,which can effectively alleviate the challenges of "storage" and "heterogeneity".In our tests,the new-designed data-structure can reduce the memory consumption by 40%in MD simulation when compared with the famous software LAMMPS.It can also achieve the world record simulation of more than 3 trillion particles,and can break through the limitation of spatial scale of atomic scale material simulation.To address the challenge of large-scale scalability of MD and KMC simulation,the parallel algorithm and communication algorithm were designed and optimized for achieving material simulations at be millions of processor cores.2)To address the "heterogeneous" challenges,based on the atomic-scale unified architecture and heterogeneous hybrid architecture of homegrown supercomputers,the core algorithm library for potential computing was developed and optimized.After the optimization,the performance of the core algorithm library at the atomic scale has been accelerated by more than 100 times(compared to the CPU version).In particular,on Sunway-TaihuLight,the heterogeneous parallelism and optimization algorithms for the heterogeneous architecture have been designed for efficient potential function access mode,appropriate task partitioning algorithms,and computation-communication overlap.And finally,we achieved a speedup of nearly 56(64 CPE cluster cores versus a single MPE core).On the DCU architecture,the optimization of host-device communication,efficient computation strategy,and computation optimization(LDS cache,SoA data structure,etc.)are proposed and implemented,which finally achieved more than 130 times the performance acceleration of single DCU(compared to Hygon single CPU core).3)Targeting the architecture of homegrown supercomputer,based on the unified architecture and atomic-scale algorithm library,we developed the massively parallel Molecular Dynamics simulation software MISA-MD,and the parallel kinetic Monte Carlo materials simulation software MISA-AKMC,both of which have excellent performance and scalability.Among them,MSIA-MD achieved high scalability performance on both Sunway-Taihulight supercomputer and DCU computing platforms(strong scalability of 79%on the Sunway platform and 74%strong scalability on the DCU platform).Meanwhile,by comparing the memory utilization and computational performance of MISA-MD and LAMMPS,a well-known molecular dynamics simulation software in the world,the results show that on the CPU platform,MISA-MD can save about 60%of memory usage,and the overall computational performance is improved by about 10%to 20%.On the DCU heterogeneous platform,the memory consumption of MISA-MD on the device side is only 1/15 of LAMMPS,and the computational performance can be improved by 34%when compared to LAMMPS.The MISA-AKMC software was developed to support massively parallel off-lattice simulation,efficient coupling simulation with MISA-MD.The parallel efficiency of 60%is achieved for MISA-AKMC under strong scalability.Compared with the mainstream software of similar categories,MISA-AKMC has obvious advantages in model,performance,and coupling mode.The two software,as well as the algorithm library,are open-source for sharing,and the software has been installed and deployed to some supercomputer centers in China,enriching the ecology of scientific computing software applications on China supercomputers.The atomic-scale high performance simulation and material performance prediction have gained good benefits in supporting the irradiation effect of nuclear reactor materials.4)MISA-MD and MISA-AKMC software applications have been carried out for material irradiation damage application requirements.In particular,a super large-scale MD cascade collision simulation of 3 × 1013 particles was performed by using MISA-MD software on the Sunway platform(a new world record of MD simulation),and further breaking through the spatial scale of the material simulation.For the "coupling" challenge,based on the unified parallel mode and particle data structure under the atomic scale unified architecture,the MD-KMC tight-coupling model is designed and implemented to link different scales of material simulation and realize efficient coupling simulation.It breaks through the scale barrier of simulation,and can realize high-fine simulation analysis of materials at large spatial scales and long temporal scales.