Parallel Collaborative Algorithm For Large-Scale LBM Multiphase Flow On Heterogeneous Many-Core Platform

The Lattice Boltzmann method(LBM),a widely used method in Computational Fluid Dynamics(CFD),has natural parallelism and is especially suitable for large-scale parallel computation.In recent years,high-performance computer technology continues to evolve,heterogeneous many-core supercomputers using heterogeneous many-core processors as accelerating components have gradually become the mainstream.In this paper,on the typical heterogeneous many-core supercomputer Tianhe-2,launched a large-scale heterogeneous parallel computing research for open source 3D LBM multiphase flow software OpenLBMFlow.The main work and contributions are as follows:(1)The LBM multiphase flow algorithm based on lattice BGK and Shan-Chen model and OpenLBMFlow program are deeply analyzed.According to the characteristic of typical heterogeneous parallel architecture platform,a multi-level LBM flow field decomposition method is designed.The multilevel and multi granularity parallelism of LBM multiphase flow simulation is described from the task level,the heterogeneous collaboration layer,the data layer and the instruction layer.A large-scale heterogeneous cooperative many-core parallel algorithm is proposed,and the performance bottleneck of the algorithm is theoretically analyzed.On this basis,the corresponding optimization strategies are proposed from the aspects of communication,load balancing and LBM algorithm.(2)On the Tianhe-2 supercomputer,the OpenMP4.5 accelerator model is adopted to realize the CPU + MIC heterogeneous collaborative simulation of OpenLBMFlow based on MPI + OpenMP4.5 + SIMD.First through a series of serial code optimizations,the single-threaded efficiency of LBM code on CPUs and MICs was significantly improved,gaining 2.5 and 2.8 speedups,respectively,over the benchmark code.With SIMD optimizations enabled,CPU performance is further increased by 1.5X and MIC performance is more than 2x.Through load balancing optimization and asynchronous computing communication overlap,compared with pure CPU parallel simulation,heterogeneous cooperation has achieved good performance speedup.Using 128 node results as a benchmark,the program achieved over 80% efficiency at 2048 nodes,demonstrating good weak scalability.(3)Explored the large-scale high-performance computing and performance optimization methods of Python,and realized the world's first completely based on Python's large-scale parallel three-dimensional LBM multiphase flow simulation open source code PyLBMFlow.According to the characteristics of the Python language,a series of performance optimization methods are proposed and the LBM boundary algorithm is reconstructed,which greatly improves the computational efficiency of Python.The optimized serial performance is improved by two orders of magnitude compared with the benchmark implementation.On this basis,MPI + OpenMP hybrid parallelism is realized based on Mpi4 py and Cython,LBM gas-liquid two-phase flow is successfully simulated on Tianhe-2 supercomputer,and the parallel scale is 1024 nodes,parallel efficiency is more than 80%.