A Multi-level Parallel Tie-dye Algorithm for Auto-CFD
Junjie Gao, Yeping Zheng, Jun Liu, Jie Chen, Xiaoqi Nie · 2023
With the rapid growth of high-performance computing and automatic mesh generation, large-scale Automatic Computational Fluid Dynamics (Auto-CFD) not only improves efficiency and reduces manual labor but also reduces industry barriers and expands application scope, which has urgent engineering needs. The preliminary work of the tie-dye algorithm has completed the qualitative analysis of its suitability for large-scale Auto-CFD applications, and there is an urgent need for algorithm breakthroughs and quantitative research in engineering. Aiming at the uniqueness of the tie-dye algorithm based on the unstructured finite difference method, after systematically analyzing and designing multiple aspects such as algorithm framework, data structure, partitioning strategy, topology mapping, and communication optimization, a multi-level parallel tie-dye algorithm (MPTD) for Auto-CFD is proposed in this paper. The correctness and performance tests of MP-TD for NACA0024 airfoil and car exterior shape were carried out on the Sugon supercomputer. The test data show that the MP-TD algorithm for Auto-CFD can input 2D CAD models or images, set relevant parameters, automatically generate CFD grids, and quickly complete flow field calculations without manual participation in geometry cleaning. The algorithm has good parallel efficiency and scalability, with a maximum parallel computing scale of 100 million grids and 7.84 million surface nodes, and up to 256 computing nodes.