Affiliation:
1. China Ship Scientific Research Center Wuxi China
2. Zhejiang Lab Hangzhou China
3. Department of Computer Science and Technology Tsinghua University Beijing China
4. Jiangnan Institute of Computing Technology Wuxi China
Abstract
AbstractDecomposition and solver are the main performance bottlenecks of multi‐block structured CFD simulation involving complex industrial configurations such as aero‐engine, shock‐boundary layer interactions, turbulence modeling and so on. In this article, we proposed several optimization strategies to improve the computing efficiency of multi‐block structured CFD simulation based on Sunway TaihuLight super computing system, including: (1) a load balancing decomposition approach combined with recursive segmentation of undirected graphs and block mapping for multi‐structured blocks, (2) two‐level parallelism that utilizes MPI + OpenACC2.0* hybrid parallel paradigms with various performance optimizations such as data preprocessing, reducing unnecessary loops of subroutine calls, collapse, and tile syntax, memory access optimization between the main memory and local data memory (LDM), and (3) a carefully orchestrated pipeline and register communication strategy between computing processor elements (CPEs) to tackle the dependence of LU‐SGS (Lower‐Upper Symmetric Gauss–Seidel). Numerical simulations were conducted to evaluate the proposed optimization strategies. The results showed that our parallel implementation provides high load balance and efficiency, achieving a speedup of 8× + for one loop step, and a speed up of 2× + for strong correlation kernels.
Subject
General Engineering,General Computer Science