Heterogeneous CPU-GPU parallelization for modeling supersonic reacting flows with detailed chemical kinetics

Abstract

Accurate simulations of complex combustion phenomena associated with supersonic reacting flows require the use of detailed chemical kinetic mechanisms. However, detailed chemical mechanisms may consist of a large number of species and reactions resulting in extremely high computation cost. In order to accelerate the simulations of supersonic reacting flows with detailed chemical mechanisms, a heterogeneous CPU-GPU parallel algorithm and its portable software implementation are presented. The parallel algorithm was broken down into two parts: the CPU handled the computation of fluid dynamics while the GPU evaluated the chemical source terms and gas physical properties. The use of overlapping computations of chemical source terms on GPU and calculations of viscous flux on CPU is also presented. A study of performance tests was conducted. The performance results show that evaluating chemical source terms and gas physical properties on 2 GPUs are about 157.8× and 78.5× faster than running on the 16-core CPU when using the most complex mechanism on a grid of 3.3 million cells, respectively, resulting in an excellent speedup of the whole iteration up to 47. The significant performance improvement provided by the parallel algorithm can provide a significant perspective for designing heterogeneous CPU-GPU algorithms for applications in simulating supersonic reacting flows with detailed chemical kinetics. Program summaryProgram Title: OpenHurricaneCPC Library link to program files:https://doi.org/10.17632/m9pphg9cjj.1Licensing provisions: GPLv3Programming language: C++, CUDANature of problem: The computation cost in evaluating chemical source terms and theirs diagonal Jacobian matrices, and in calculating gas physical properties dominates the simulations of supersonic reacting flows with detailed chemical kinetic models. And the parallelization of these computations can yield great benefits from GPU acceleration.Solution method: The parallel implementation of evaluating the chemical source terms and theirs diagonal Jacobian matrices, and calculating gas physical properties were conducted on GPU to speed up the computation.