Parallelization and Optimization of a Combustion Simulation Application on GPU Platform

TURF sim (Target Unsteady Reacting Flow simulation) is a CFD application that solves engine combustion problems on structured grids. In this paper, PGI CUDA Fortran is used to implement the CPU + GPU heterogeneous parallelization. To reduce the data transfer overhead between the CPU and GPU and the MPI communication overhead, we design a packing/unpacking based method to decrease the volume of data transferred and the number of data transfer, and use the Pinned memory to improve data transfer rate. To overlap the CPU computation, GPU computation and CPU-GPU data transfer, we reorganize the computing process, and utilize the asynchronous data transfer and the stream processing technology. We also employ loop transformations to optimize the register utilization and task partitions on the GPU. Performance evaluation is performed on two GPU-based platforms. The experimental results show that our optimization techniques for GPU are effective and the resulting GPU version significantly outperforms the original pure CPU version. The GPU version, when using a NVIDIA GTX1060 GPU, achieves a 2.96X performance acceleration over the pure CPU version runs on an Intel i7-8700 CPU (6-cores). When using a NVIDIA V100 GPU, the GPU version achieves a 1.88X performance acceleration over the pure CPU version runs on two Intel Xeon Skylake Gold CPUs (24-cores).