Research On Parallel Computation And Optimization Of Lattice Boltzmann Method On Many-core Platform

In recent years,high-performance computing has gotten a high-speed development,and numerical calculation method has become more and more mature as well.Parallel computing has become a major solution to large-scale scientific and engineering problems.The Tianhe-2 which belongs to the National University of Defense and Technology has won six championships consecutively on TOP 500 Conferences.In 2016,the "Sunway TaihuLight" turned out and won the first for two consecutive times.China has been in the front of the world on supercomputers.In the fields of aviation,meteorology,shipbuilding,civil engineering,computational fluid dynamics has been widely used as a powerful computational analysis tool.The demand for computing and storage is increasing because of the demand of high-precision,large-scale and real-time requirements.Parallel computing is necessary.The Computational Fluid Dynamics numerical simulation based on the Lattice Boltzmann Method(LBM)is particularly suitable for parallelization on multi-core high performance computing platforms.In this paper,openLBMflow,an open source lattice Boltzmann implementation available at web site www.lbmflow.com,is employed.We have achieved a parallel and optimization strategy on Xeon CPU and KNL(Knights Landing).The main achievements of this paper are as follows:We realize the serial optimization of the Lattice Boltzmann Method program and achieve a 3x performance optimization on the Tianhe-2 supercomputer platform.Based on the serial optimization,the parallel optimization of OpenMP and MPI is achieved.When 64 computational cores are used,the speed-up of nearly 57 is obtained.And we get the conclusion that the program has good strong scalability and weak scalability.By improving algorithm of the data migration module of the code,we have saved a lot of space and provide a viewpoint for researchers in related fields.And in the LBM program,grid compression technology is used and achieved a certain optimization.We test and run the optimized LBM program on KNL.We find that serial program cannot fully make use of the advantages of KNL,the calculation efficiency become lower when a serial program is running on it.The parallelization of the program is the key to making good use of the many-core platform.Using the up-to-512-bit bandwidth on KNL,we rebuild the code to achieve vectorization.Comparing with Xeon,we obtained twice the performance improvement.It indicates that KNL is very suitable for large-scale computing problems.