Calculation And Optimization Of Multi-field Coupled Phase-field Model Based On Multi-GPU Parallel

With the continuous development of HPC(high performance computing)technology,all kinds of high performance computing methods are widely applied in the actual engineering fields,occupies a very important strategic position in the national economy,national defense construction and the development of science and technology.Phase-field method is one of the most effective methods for simulating dendritic microstructure.With the developing and perfecting of phase-field model,the external fields such as flow field,temperature field and solute field are coupled to the phase-field.The solution of multi-field coupled phase-field model presents problems such as large calculation amount,low computational efficiency and limited computing area.Moreover,in the field of practical application,there is an urgent need for the expansion of the phase-field model in simulation scale.Research on high performance computing method to solve phase field model has become a hot topic in the interdisciplinary field of computer science,engineering and materials science.Multi-node CPU+GPU heterogeneous high performance computing cluster provides new and strong support for improving the solution speed and expanding the solution scale of phase-field model.Based on the phase-field model and Lattice Boltzmann method,a multi-field coupled 3D PF-LBM model is established.The the PF-LBM model is realized on the CUDA+MPI programming platform,by implementing Multi-GPU parallel computing method on multi-node CPU+GPU heterogeneous cluster.Several factors that affect the performance of Multi-GPU parallel computing method are analyzed,and two corresponding optimization schemes are proposed for the problems of MPI communication delay,GPU calculation and communication overlap.In order to verify the reliability and validity of the proposed Multi-GPU parallel computing method,the classical theoretical values of PF-LBM model and the results of other scholars have been compared and analyzed with Multi-GPU method.At the same time,the simulation results before and after optimization are compared to verify the effectiveness of the optimization scheme.Comparison and analysis results show that in the high performance computing cluster,the solution result of the Multi-GPU parallel computing method is consistentwith the classical theoretical value,which proves that the Multi-GPU parallel computing is effective in solving the PF-LBM phase-field model.Compared with the previous simulation calculation methods,the Multi-GPU parallel computing method has been improved in both computational efficiency and simulation scale.Compared with the traditional CPU serial algorithm,the Multi-GPU parallel computing method has the maximum acceleration ratio of 58(21 GPU used).Compared with the MPI parallel algorithm with the same number of nodes,the Multi-GPU algorithm also has better acceleration solution effect.In phase-field simulation scale,the simulation scale of Multi-GPU parallel computing method is up to 819?819?819.Both optimization schemes M-GPU and CB-GPU can realize the optimization of Multi-GPU computing model to a certain extent,and the CB-GPU computing model has better optimization effect in the larger simulation scale.