| As a derivative of metallurgical,material,electronic information and computational science,metal solidification simulation technology not only has the basis of complete solidification theory,but also has been widely used in important research fields such as new materials and its preparation and processing.With the advancement of human society and the development of science and technology,metal solidification technology shows a deeper,higher,more refined,and even unconventional development trend.For the solidification of metal coagulation experiment,there are various kinds of interference factors,and the simulation environment is difficult to meet the experimental conditions.The molecular dynamics simulation of metal solidification process can be used to visualize the microstructure of the solidification process,save the capital investment and save the time cost.Especially with the rapid development of parallel computing technology,molecular dynamics simulation has more application opportunities.Nowadays,the research on the parallelization of metal solidification molecular dynamics simulation is mainly realized by multi-core CPU or GPU.However,on the one hand,in the face of large-scale molecular dynamics simulation problems,there is an urgent need to use high-performance computing system to achieve high-performance metal solidification parallel computing to enhance the efficiency of metal coagulation applications.On the other hand,multi-core CPU cluster and CPU-GPU heterogeneous system as a mainstream high-performance computing system,which can take into account the computational efficiency and programming versatility while greatly enhancing the overall performance of the system,but it is still necessary for the researchers to make full use of the optimized parallel algorithm and load balancing strategy to deepen the parallelism of the system in large-scale molecular dynamics simulation problems.This poses a great challenge to the application of the parallel algorithm of molecular dynamics.In this paper,we focus on the mainstream HPC architecture and typical metal solidification process simulation,the key technologies of large-scale molecular dynamics simulation,the multi-core CPU cluster and CPU-GPU heterogeneous parallel algorithm,CPUGPU load balancing optimization strategy for the metal solidification molecular dynamics simulation method,and describe the multi-potential parallel implementation based on the metal-atom interaction force.The main work includes:(1)Due to the characteristics of multi-core cluster system and molecular dynamics simulation method,a parallel algorithm for metal coagulation molecular dynamics simulation based on multi-core CPU cluster architecture is proposed.The parallel algorithm was designed for the calculation of the potential function with the highest proportion of calculation in the molecular dynamics simulation method.The parallel virtual machine(PVM)model program involved in molecular dynamics simulation was improved.The Parallel Algorithmic Program for Large Scale Molecular Dynamics Simulations is based on the combination of MPI and OpenMP.By using the method of region decomposition,the program mode of PVM model is improved,and the scale of sample space is greatly expanded which is physical estimation methods.The simulation results are accurate.(2)Based on the characteristics of CPU-GPU heterogeneous architecture,a parallel algorithm for metal-solidification dynamics simulation based on CPU-GPU heterogeneous platform was studied.The new parallel algorithm for molecular dynamics simulation is based on CUDA and OpenMP programming model.This algorithm can effectively utilize the powerful computational power and multi-core CPU of the GPU,which avoids the problem of GPU's memory shortage due to the solidification of molecular dynamics simulation.By running the algorithm on the Tianhe 1 A experimental platform,the parallel simulation of 10,000,000 metal atoms scale was realized.Compared to the calculation time of the multi-core CPU cluster system based on the mixed and heterogeneous system simulation system and the isomorphic system a speedup of 6 ื was obtained.(3)Due to the phenomenon that the heterogeneous platform is not efficient due to the heterogeneity of computational power and the uneven distribution of computing resources when the heterogeneous parallelism of CPU-GPU is used in the simulation of metal-solidification molecular dynamics simulation,load balancing strategy of molecular dynamics simulation based on CPU-GPU heterogeneous platform is proposed.Based on the ability of CPU-GPU heterogeneous computing resources,allocates the large-scale atoms by heuristic algorithm to achieve large-scale metal coagulation molecular dynamics simulations to minimize the total execution time.Based on this strategy,the optimization of CPU-GPU heterogeneous parallel algorithm for metal-solidification molecular dynamics simulation was carried out.Using the Tianhe 1 A supercomputer system as the experimental platform,the results show that this load balancing strategy can effectively enhance the performance of the simulation program for the whole solidification of molecular dynamics.(4)A parallel algorithm based on multi-body potential molecular dynamics simulation method is proposed.The algorithm also takes into account the potential energy function of heterogeneous parallel algorithm based on metal solidification molecular dynamics simulation.Firstly,the sub-term of the multi-body potential model is much more diverse than the potential model,and the calculation of the multi-body potential function is much more complicated than the potential functions,and the calculation of the intermolecular forces accounts for a high proportion in the whole algorithm.Therefore we consider the parallel analysis and optimization design of this part.Secondly,we make full use of the Tianhe 1A supercomputer system as the experimental platform to realize the parallel program simulation of multi-body metal solidification molecular dynamics in multi-core CPU cluster and CPU-GPU hybrid heterogeneous computing system.The experimental results show that not only the computational performance of parallel algorithm is enhanced,but also has good expansibility. |
PDF Full Text Request