Research And Parallel Implementation For Incompressible Pipe Flow Numerical Simulation

With the development of parallel computer system and the increasing scale of CFD, parallel algorithm and parallel models in numerical simulation of steady flows have become an advanced and necessary research topic. Research on numerical simulation of incompressible pipe flow is one of hot topics in the study of dual coolant lead lithium blankets of a fusion reactor. In application of a fusion reactor blanket, the Hartmann number is104～105. The Hartmann layer is perpendicular to the external magnetic field which thickness is scaled with Ha-1. There must be several grids in the Hartmann layer in order to obtain the correct numerical simulation results. Duration of serial pipe flow simulations may take hours even days, because of the large-scale grids and slow speed of convergence of the iteration process caused by the interaction of magnetic field and liquid metal fluid.In this paper we investigate the issues of three-dimensional steady incompressible pipe flow based on the SIMPLE algorithm. Firstly we propose a parallel algorithm based on domain decomposition method, and then apply an iteration space parallel two-way finite difference stencil algorithm on the basis of domain decomposition. Compared with traditional domain decomposition method, the locality of serial stencil algorithm is improved through the introduction of the execution order of grid tiles. By decomposing the iteration space and reordering the grid tiles, we realized the parallelization of iteration. We reduce the cost of communication and synchronization with alternate tiling technique by partitioning the iteration space along the direction of time steps. Finally numerical simulations of the two parallel algorithms on DeepComp7000cluster confirm the effectiveness of the two parallel algorithm and locality tests for serial finite difference stencil algorithm is performed on a single node, which shows that the parallel two-way finite difference stencil algorithm based on iteration space reduces cache misses, have a good data locality, parallel efficiency, and improves the efficiency in numerical simulations of three-dimensional steady incompressible pipe flow.For the irregular problem, this paper presents an Inspector/Executor model based on distributed clusters, and improves parallel efficiency utilizing alternate tiling technology, on the basis of Strout’s work. The Inspector analyzes the memory reference pattern at run-time, reorders the data space and iteration space according to the tiling generating strategy, and creates new schedules and communication strategy. The Executor improves the data locality and parallelism by performing the new schedules and communication strategy. Moreover, we construct an auto-tuner for the algorithm to get the optimal efficiency of specific architecture by detecting the right combination of parameters and fixed it so that it can be effectively used later.Finally we give a summary of our work, which has specified the achievements and contributions, and the future work. Maybe we can try to improve the performance of parallel algorithms in incompressible pipe flow simulation on multi-core hybrid architecture, and expanded the irregular computing to other areas to improve the locality and parallelism.