Key Technologies Research Of Incompressible Pipe Flow Parallel Simulation Based On GPU

Incompressible pipe flow as an important object of fluid mechanics study, hasbeen widely used in plasma physics, magnetic fluid mechanics, astrophysics,controlled thermonuclear fusion reactors dual cooled lithium lead blanket and relatedfields of scientific research and industrial technology. With the development ofcomputer science and technology, incompressible pipe flow problem has beenresolved from field trials to computer simulation through mathematical models.However, computer simulation of incompressible pipe flow with accurate modelsrequires enormous computing resources and large time cost. Therefore, the design ofincompressible pipe flow simulation efficiently is always research difficulty andhotspot. This thesis studies two important steps, which are sparse matrix vectormultiplication calculation and finite difference Stencil algorithm, and adopts paralleloptimization techniques and data locality optimization techniques to improvenumerical simulation efficiency of incompressible pipe flow.Matrix vector multiplication problem in the numerical simulation ofincompressible pipe flow, the matrix shows of generally sparse but locally densecharacteristic. The traditional sparse matrix storage structures cannot make good useof this feature, and this thesis proposes QCSR sparse matrix storage structure for thisproblem. QCSR storage structure combines the advantages of quadtree structure andCSR storage structure, and stores sparse matrices by sparse matrices recursivedecomposition and rearrangement to improve data locality for sparse matrix vectormultiplication. Compared with the CSR storage structure, experiments in this thesisshow that sparse matrix vector multiplication using QCSR storage structure canimprove overall performance and reduce the distribution factor of nonzero elementswithin the matrices for computing processes. This thesis analyzes the programmingmodel CUDA for the CPU-GPU heterogeneous parallel system and studies theoptimization strategies of GPU memory access. By using four strategies includingthread mapping optimization, data access optimization, data transmission optimizationand data reuse optimization, this thesis implements sparse matrix vectormultiplication using QCSR storage structure based on GPU, which achieves betterresults for calculation acceleration and overall acceleration. Due to the interaction of pipe flow and magnetic field and the complexity ofheat transmission of incompressible pipe flow issue, the semi-implicit method forpressure and velocity field coupled equations requires grained mesh method in thesolution process to accurately find out the details of the internal fluid. But this willresult in larger scale of solving equations and higher time complexity. For finitedifference Stencil algorithm based on linear domain decomposition method exists datalocality poor and scalability problems, this thesis proposes finite difference Stencilparallel iterative algorithm based on symmetric staggered bars of multi-grid spacealgorithm. This finite difference Stencil parallel iterative algorithm adopts regionaldivision strategy based on staggered grid bars which divides iteration space intostaggered bars along the direction of added time axis by time-skewing technique,which can improve the data locality of internal bars. Further, it introduces multi-gridsymmetric operation strategy to improve the algorithm parallelism and accelerate theiterative convergence speed. And it uses staggered bar reordering strategy, which caneffectively reduce communication and synchronization overhead in iterative process.This thesis further implements multi-dimensional finite difference Stencil iterativealgorithm based on GPU. Experiments in this thesis show that finite difference Stenciliterative algorithm using corresponding optimization strategies can reduce thecomputing time.In this thesis, sparse matrix vector multiplication and finite difference Stenciliterative algorithm are optimized through parallel optimization techniques andCPU-GPU heterogeneous parallel system, which can improve the numericalsimulation effect of incompressible pipe flow. The results of this study can beextended to the corresponding numerical computing and engineering fields.