Concurrent DSMC Method Combining Variable Time-Step and Adaptive Unstructured Mesh

A parallel direct simulation Monte Carlo method is developed, which uses dynamic domain decomposition to solve the problem of load unbalancing among processors. In addition, variable time-step method and unstructured adaptive mesh are implemented, respectively, for reducing the computational time and increasing the accuracy of solution. Associated flow chart of the proposed DSMC method is illustrated in Fig. 1. Major difference between the current study and the previous related parallel DSMC studies [2,3] is the utilization of the state-of-the-art graph partitioning technique developed in the community of computer science. The current DSMC method is implemented on unstructured adaptive mesh using particle ray tracing technique by taking the advantages of the cell connectivity information [5]. Standard Message Passage Interface (MPI) is used to communicate data between processors. In addition, different strategies of applying Stop at Rise (SAR) [4] scheme are utilized to decide when to adjust the workload among processors. Briefly speaking, SAR method defines a degradation function, which represents the average idle time for each processor including the cost of repartition. Decision to repartition is made based on the detection of a minimum value of degradation function during the simulation. Corresponding parallel performance is analyzed using the results of a high-speed driven cavity flow on IBM-SP2 (memory-distributed, CPU 160 MHz, RAM 256 MB each) and IBM-SMP (memory-distributed, CPU 375 MHz, RAM 1 GB each) parallel machines up to 64 and 128 processors, respectively. In general, the implementation of dynamic domain decomposition reduces the runtime in the range of 30-60% as compared with that of static domain decomposition.