Research And Implementation Of Parallel Algorithm For Fluid-Solid Interaction Simulation(Lattice-Boltzmann&Immersed-Boundary) On Heterogeneous System Architecture

Over the past 10 years,computer technology and integrated circuit technology have been still developing rapidly.GPU particularly has evolved into a high specific CPU coprocessor for compute-intensive request,served in large-scale and highly parallelized computing applications.In a heterogeneous system architecture(HSA)constructed by CPU and GPU,CPU and GPU synergetic parallel computing has become a focus in high-performance computing research,mainly related to HSA being rational allocation of computing resources to the CPU and GPU according to the characteristics of the computing tasks,which can help to enhance the utilization of computing resources and to ensure high computing performance with low overheads.In this paper,the parallel algorithms of numerical simulation of fluid-solid interaction problem on CPU-GPU and CPU-MIC HSA are explored.Lattice Boltzmann method(LBM)is the main method of numerical simulation in complex flow field.Immersed boundary method(IBM)is one way to tackle the problem of microscopic interaction of solid structure immersed in the flow field.The LBM s the fluid into a large number of microscopic particles which will collide and stream on the lattice grid,and then,the macroscopic variables of fluid flow will be computed by the time-space evolution of the statistical function which represents the particle distribution.The boundary conditions of the solids are transformed into volme forces by IBM so that the fluid field can be perceptional.We proposed a numerical approach based on IB-LBM(Immersed Boundary-Lattice-Boltzmann method)to simulate and analyze the problem of interaction of solids with an incompressible fluid flow.However,the IB-LBM is compute-intensive,the lager simulation scale we computing,the higher computing performance we requiring.Thus,in this article the parallel algorithm based on HSA for IB-LBM simulation of fluid-solid interaction is designed via CUDA(Computer Unified Device Architecture)compiler framework.First of all,we establish complete fluid-solid interaction numerical framework by useing the LBM D2Q9 lattice model combing with IBM,and then verify the test data with the existing data.By given a certain distribution(fine-grained or coares-grained)acrosing cores,we present the parallel algorithms upon LBM update and IBM correction,and propose three data structure types to optimize implementation.Fuerher more,the implementation of the IB-LBM simulation of fluid-solid interactionare is transplanted to the CPU-GPU and CPU-MIC two different HSA.We design new computing resource allocation strategy on the two heterogeneous system to further optimize the algorithm performance.Finally,based on MFLUPS(Millions of Fluid Lattice Updates Per Second)ideal value standard,we completed the performance evaluation and analyses,included:the three types of data structure,scalability of thread number,two volume fractions and lattice grid size.