MIC-THPCM: MIC-Based Heterogeneous Parallel Optimization for Axial Compressor Rotor
Yu Lei, Xingjun Zhang, Han Li, Xiaoshe Dong, Jingbo Li · 2019
The computational scale of many computational fluid dynamics (CFD) numerical simulations is very large, such as fluid mechanics numerical simulation. The continuous development of high performance computers and parallel programming models provides a new way for CFD numerical simulation. How to adjust large CFD applications to adapt to heterogeneous system architecture and make full use of computing resources is still an urgent problem to be solved. This paper proposes a MIC-THPCM (MIC-based Three-layers Heterogeneous Parallel Computing Model) model for the application characteristics of the axial compressor rotor and the structural characteristics of the CPU + MIC (Intel Many Integrated Core Architecture) heterogeneous system. The MIC-THPCM model is realized by using the MPI (Message Passing Interface) + OpenMP (Open Multi-Processing) + Offload three-layer programming model. By analyzing data dependencies and parallel features of hotspots, the hotspots are parallelized from the aspects of multi-threading, vectorization, data transmission, and memory access based on MIC hardware features. The experimental results show that the MIC-THPCM achieves better system performance under the premise of ensuring the correctness of fluid mechanical simulation. Compared with CPU multi-threaded calculation, the MIC-THPCM has a 33.9% improvement in single-channel performance. The speedup of the hotspot optimized on the MIC can be up to 24. The results show that the parallel optimization of coprocessor computationally intensive tasks has a huge benefit for the overall performance improvement of the program.