Parallel Research On GPU-based DSMC Method Using Cartesian Grid

The direct simulation Monte Carlo(DSMC)method is a powerful computational tool for fluid mechanics simulation in the regime of rarefied gas flow.However,there are two main shortages in DSMC method: one is complex gridding processing,the other is large time consumption.For the first shortage,we use Cartesian grids for the flow field with high computational efficiency and unstructured triangular cells on body geometry surfaces with precise description.The moving stage and collision stage are decoupled in the DSMC method,therefor GPU is an eligible candidate as a result of the huge calculation and innate parallelism of the DSMC method,which overcomes the last shortage.Based on the CUDA Fortran programming platform,we execute parallel DSMC simulation on a GPU.Before the particle collision,on the moving stage,we investigate particle hashing techniques to record the particle's cell ID and offset in that cell,so that we can complete particle hashing by singlestep.We process the particle in/out separately by a data-parallel algorithms,then replace the front part of outer particles with the posterior part of inner particles.In addition,we studied the memory usage and random number generator on GPU.We also have tested and verified the accuracy of our implementation.Our results show substantial performance improvements over serial implementation without loss of accuracy.