MPI+X:Massive Parallelization and Dynamic Load Balance of a Production-level Unstructured DSMC Solver

This paper presents implementing and balancing a production-level unstructured DSMC solver using two variants of the hybrid programming model. Both approaches can be implemented with minimal modifications to the source code. Specifically, we port and optimize RariHV, a production-level unstructured DSMC application, to scalable homogeneous and heterogeneous systems using two variants of the hybrid MPI+X programming model: MPI+OpenMP and MPI+OpenACC. Additionally, a physics-awared mesh repartition during the execution is proposed to alleviate the load imbalance issues. Results demonstrate that the load balancing strategy has an improvement in performance of up to 50%, and the hybrid MPI+X parallelism method can provide considerable scalability benefits on multi-core and many-core platforms.