Research On CPU/GPU Parallel Computing For Multi-block Structural Grid CFD

At present, the CPU/GPU collaborative and parallel computing is one of the hot research areasin high performance computing applications. This article is based on a practical CFD application ofmulti-block structural grid’s flow field aerodynamics simulation, and the WCNS method, which isdesigned and developed independently by our own country. This paper studies and implements theGPU parallel computing and CPU/GPU collaborative and parallel computing of typical CFD solverswithin a computing node of “TH-1A” in Changsha National Supercomputing center. The majorachievements of this study are as follows.1) This paper analyzes the basic theories, algorithm procedures, and the data dependences oftypical CFD solving methods in depth, tests the time occupancy of major computationprocedures of relevant solvers. This paper classifies out these typical computation proceduresinto different categories, the data independent computation procedure, the week datadependence computation procedure, the strong data dependence computation procedure and thebranch intensive procedure, and proposes theirs GPU parallel programming methods, the gridpoint level parallel by3D GPU threads block and the grid line level parallel by2D GPU threadsblock. And then this paper implements and optimizes the GPU parallel computing ofRunge-Kutta solver and Jacobi Iteration solver, and tests their performances by real cases ofstructural grid in different scales. The test result shows that the speedups of Runge-Kutta solverand Jacobi Iteration solver increase along with the grid scale, in the2million single blockstructural grid they gain the speedups of4.62and8.85respectively, the Jacobi Iteration solveris proven to be much more suited for GPU parallel computing.2) This paper considers the heterogeneous architecture feature in current high performancecomputers and the multi-block characteristic of most practical CFD applications, studies theCPU/GPU collaborative and parallel computing for multi-block structural grid’s typical CFDsolvers. There are two collaborative and parallel programming models, one is based on nestedOpenMP threads, and the other is based on the GPU asynchronous communication. This paperfurther proposes a novel solver collaborate method, which uses solvers suited for large scalefine grain parallelization on GPU and solvers with better computing efficiency on CPU, andcompares it to the plain sole-solver collaborate method with specific examples. This paperdesigns and implements the collaborative and parallel computing of CPU/GPU by sole-solver(Jacobi Iteration solver) method and multi-solver (LU-SGS solver and Jacobi Iteration solver)method, using the nested OpenMP threads based collaborative and parallel programming modeland other methods mentioned in the preceding chapters, then tests and analyzes theirperformances with a two million structural grid of four blocks. These two implementations gainthe speedups of7.49and8.24respectively.