Research On Auto-tuning Techniques For Parallel CFD Programs

Computational fluid dynamics (CFD) have been widely used in the areas ofindustrial design and manufacture. With the development of CFD, the physicalprocesses and geometric models needed to be simulated are becoming morecomplicated, and the grids scale involved are becoming tremendous; meanwhile theapplication requirement of CFD has transferred from analyzing the flow field tosearching for the optimal solution of the large scale involving solutions. This has posedgreat challenge on how to efficiently use large scale processing and super computationalcapability of high performance computers to improve the performance of large scaleparallel CFD programs. Auto-tuning technique has been proved a success that it caneffectively contain variety of optimization transformations, with better suitability forcomputational programs that are hard to modeling and portability for platforms withdifferent system constructions.Aimed at improving the whole program’s performance, this thesis carries outresearch on auto-tuning technique for parallel CFD program performance optimization.The main works and contributions are as follows:1. On mapping stratages of parallel CFD programs, we do researching intoauto-tuning techniques, extract the performance critical parameters from CFD programs,form the optimal space, search for the optimal parameter values and then feed themback to run the original program.2. Considering the performance impact of mapping optimization, we propose aheuristic mapping method based on the computation to communication ratio. Thismethod corporately restricts the load balance and communication cost, and also restrictsthe search space to some extent. It can obtain the optimal mapping that maximizes thetotal performance. Compared with block mapping strategy and cyclic mapping strategy,our method can significantly reduce the communication cost.3. We transform the mapping optimization problem of parallel CFD programs intograph partition problem, and solve this problem with graph partition and grid coarseningtechniques. Firstly, the grid topology file is transformed into corresponding graphformat, CSR format. Then the input parameters are defined. The graph format can betransformed back into grid topology and can be feed to the CFD program, with the inputinterface defined. This method is commonly applicable for parallel CFD programs. It can significantly reduce the communication cost. Meanwhile, it improves the loadbalance of the parallel CFD programs.4. We implement the two auto-tuning optimization strategies and apply them to areal parallel CFD program with a real application case. The results show theauto-tuning approach using the heuristic based on load balance and communicationimproves the performance for4.9%-14.5%, and the auto-tuning approach based ongraph partitioning and grids coarsening improves the performance for12.1%-27.2%.