Parallelizing and optimizing a detonation combustion simulation application on heterogeneous platform

Initition3d is a detonation combustion simulation application developed based on the AMROC (Adaptive Mesh Refinement Object-oriented C++) framework, originally parallelized in pure MPI and runs only on CPUs. To port this application to the heterogeneous platform, we add OpenMP parallelization to the code, implement the heterogeneous algorithm using the OpenMP 4.0 programming interface, and perform a series of performance optimizations to optimize the parallelism and to reduce the overhead. We also solve some issues related to the framework based application development. Performance evaluation shows that the application exhibits good OpenMP parallel efficiency, and the hybrid MPI/OpenMP version is at best 1.79 ×  faster than the original pure MPI version on one server node. The heterogeneous version successfully utilized the MIC (Many Integrated Core) coprocessors to accelerate the performance by a factor of 2.69-2.92 on the Tianhe-2 supercomputer. Good parallel scalability across multiple nodes is also achieved.