Multi-DCU Heterogeneous Parallel Tie-Dye Algorithm for Auto-CFD

On the one hand, with the development of high-performance computing technology, heterogeneous parallel computing has attracted widespread attention for its ability to match different parts of algorithms with computing resources, thereby providing more powerful computational performance. On the other hand, automation in grid generation and high-performance computing are of utmost concern in the Auto-CFD field; previous studies on the tie-dye algorithm have proven its suitability for Auto-CFD from theoretical and homogeneous parallel computing perspectives. Applying the tie-dye algorithm in actual engineering fields necessitates research on heterogeneous parallel tie-dye algorithms. Given the uniqueness of the tie-dye algorithm, comprehensive development from the ground up is required. By parallelizing the data structure, designing data area division and data transmission modes, optimizing kernel functions, and integrating advanced programming frameworks such as Compute Unified Device Architecture (CUDA) and Message Passing Interface (MPI), the Multi-DCU Heterogeneous Parallel Tie-Dye (MHPT) algorithm is proposed to achieve enhanced computational efficiency. Simulation results demonstrate the superiority of the MHPT algorithm. Compared to traditional methods, the MHPT algorithm achieves a speedup of over 80% when handling tens of millions of grid generation and flow field computation tasks. The weak scalability of the multi-DCU system under tens of millions of grids exceeds 82%. Compared to homogeneous parallel computing, its average computation time is reduced by more than 50%.