Massively parallel simulations of multi-stage compressors on Sunway TaihuLight

ASPAC is an in-house computational fluid dynamics (CFD) software for the simulation of flow in turbomachinery. In this paper, with a dual-level hybrid and heterogeneous programming method, we optimized the ASPAC software and ran it on the Sunway TaihuLight supercomputer. Then the unsteady implicit simulation of the 13-stage twin-spool compressor with 6.116 billion grid cells was realized using 611,000 cores. This is an important step closer to solving one of the four challenges mentioned in the “CFD Vision 2030 Study” proposed by NASA. A general solution method for the twin-spool compressor was first proposed. In addition, an efficient large-scale parallel processing method of row interface data exchange was proposed. After that, a series of optimization methods, such as DMA transmission, data packaging, and linear equation system solving process reconstruction, were explored. The test results show that the proposed parallel processing method of row interface data exchange greatly reduces the memory occupation, and the speed of establishing the row interface exchange relationship is increased by two orders of magnitude. Strong scalability testing in the twin-spool 13-stage compressor showed that for 0.765 billion grid cells, ASPAC can achieve 85.0% parallel efficiency when the cores are increased from 10,400 to 624,000; for 6.161 billion grid cells, ASPAC can achieve nearly 100% parallel efficiency when the cores are increased from 208,000 to 611,000. The test shows that the heterogeneous parallel modification of ASPAC on Sunway does not affect the calculation accuracy, and has high scalability, which can play a major role in the large-scale simulation of compressors.