Research On Large-scale Parallel Computing Of Unstructured Grid DSMC Applications

Direct Simulation Monte Carlo(DSMC)method is a powerful tool for solving real gas flow problems in the field of rarefied gases at present.As an important branch of the DSMC method,unstructured grids have been widely adopted in DSMC engineering calculations because of their advantages of high automation,short generation period and flexible distribution control in the generation of computational grids for complex geometric shapes.However,with the continuous expansion of applications,the problems solved by unstructured DSMC are becoming more and more complex,and the demand for computation and storage is increasing,it has become an inevitable trend to carry out efficient unstructured DSMC massively parallel computation.In the context of a DSMC application for solving multi-block unstructured flow fields,the characteristics of data structure,numerical methods and computational flow in the DSMC solver are analyzed in depth,and a study on the massively parallel computation of unstructured grid DSMC methods adapted to the architectural features and algorithmic characteristics is carried out.In terms of adapting architectural features: 1)for homogeneous computing systems,an efficient MPI communication method based on peer-to-peer mode and a two-level hybrid parallel algorithm of MPI+OpenMP are designed,and performance optimization strategies such as thread scheduling and cores binding are discussed.2)for heterogeneous computing systems,a hybrid parallel algorithm of MPI+OpenACC based on unified memory technology is proposed to realize the DSMC method on GPU for random number generation algorithm,which reduces the cumbersome GPU memory management and improves the localization of data.In terms of adaptation algorithm features,a dynamic load balancing strategy based on molecular number weighting is designed and implemented to improve the computational load imbalance of the DSMC method during the computation.The correctness of the parallel algorithm is verified by the hypersonic flat plate bypass calculation example.Performance tests were conducted using a large-scale class X38 vehicle.On the homogeneous system,the MPI+OpenMP hybrid parallel strategy achieved a performance improvement of 6.18 times compared to the MPI strategy at a computational scale of eight thousand cores,while the acceleration ratio of the 60,000-core hybrid parallel computation was 5.54 with 1024 cores as the benchmark.on the heterogeneous system,compared to the Intel Xeon Gold 6244 CPU single-core,eight NVIDIA A100 GPU achieve a 32.04 x acceleration ratio,and the parallel efficiency of eight cards reaches54% with good scalability based on a single GPU.Finally,with the application of dynamic load balancing strategy,the DSMC computation time is reduced by 50%,further improving the parallel efficiency.