Parallel Optimization Of LBM Algorithm Based On CCPU＿GPU Heterogeneous System

In recent years,with the development of high-performance computing,numerical simulation on supercomputers has gradually become one of the important means to solve large-scale science and engineering problems.In addition to its advantages in processing graphics computing,GPU has been gradually applied to intensive numerical computing.At the same time,with the development of numerical technology,parallel technology,and supercomputing,computational fluid mechanics has gradually become one of the pillars of the study of fluid mechanics.The Lattice Boltzmann method is a computational fluid dynamics method based on a mesoscopic simulation scale.It obtains the motion law of fluid in a macroscopic state through statistical analysis of a large number of discrete particles and has been widely used in the fluid dynamics calculation of energy,chemical industry,and other fields.Due to the limitation of computing power,parallel computing design is very necessary.GPU-based parallel research has shown great potential in the field of computational fluid dynamics.In this thesis,the formula evolution process of the LBM algorithm and related discrete mathematical models are analyzed in detail at first,and then the hardware architecture of the GPU and CUDA programming model adapted to GPU are introduced.On this basis,the parallel optimization strategy of the LBM algorithm is designed.The main contents of this thesis are as follows:(1)Hot spot analysis of the LBM algorithm.This thesis tests the hot spot calculation of the three-dimensional square cavity flow model,and the hot spot calculation process and calculation task are analyzed in detail.The test results show that there are two main hotspots in the iterative process of the LBM algorithm: collision calculation and streaming calculation.(2)GPU optimization of LBM algorithm.Based on the characteristics of GPU architecture,the collision,and streaming are optimized in parallel.In the optimization of the collision computation part,the computation task is allocated to the computing unit of GPU by the method of address mapping and shared memory allocation.Streaming computation cannot be directly parallel due to data dependence.In this thesis,the causes of streaming calculation data dependence are analyzed in detail,and then the methods of model dimension reduction,data location,and region division are used to rearrange data reading methods,which successfully eliminates the influence of data dependence.After testing,the proposed algorithm has good parallel computing efficiency and can reach 1.92 times of acceleration ratio under 130 million grids.By changing the grid size,the performance of the algorithm under different computing scales is tested.The results show that the algorithm has scalability,which proves that the algorithm can obtain better performance under different computing scales and has good scalability.(3)Application of GPU optimization of LBM algorithm in electric convection model.Electrodynamics,as a form of non-contact-driven fluid motion,has attracted wide attention in the field of computational fluid dynamics.As a classical research model,the parallel design of square cavity convection can provide a case for the calculation of electrodynamics.The NS,NP,and Poisson iteration equations that need to be solved in the model can be solved by the LBM algorithm.This thesis designs parallel algorithms for the three core equations in the model and tests the computational efficiency of the program,which can reach 1.52 times the acceleration ratio in 5123 grids.In this thesis,the parallel optimization strategy can improve the algorithm efficiency of the source program at different computing scales,indicating that the algorithm can be applied in related fluid computing models,and has research significance for the parallel computation of other LBM algorithm fluid models.