Implementation of Parallel I/O for Multi-Partition Structured Grids CFD Program Based on HDF5
Tao He, Xiaohang Liu, Chunye Gong · 2023
Computational fluid dynamics (CFD) is one of the important application fields of high performance computing. Parallelization of CFD programs can significantly increase the speed of computation, but the traditional serial I/O mode is no longer sufficient for efficient parallel computing and requires the development of corresponding parallel I/O libraries. Aiming at the I/O bottleneck in the calculation process of multi-zone grid CFD program, this paper achieves the parallelized I/O interface library of CFD program files, which based on hierarchical data format V5 (HDF5) and the message passing interface (MPI). By reconstructing, converting and encapsulating the underlying interface of HDF5, we optimized the read and write function of the program. On the strength of the parallel I/O library, we have developed a post-processing program HDF2Tec in order to facilitate the data format conversion and graphical display. We compared the single-core, two-core, and four-core write-out speed based on the RAE2822 structured grid model and tested the accuracy of writing files in parallel using post-processing software Tecplot. The results show that under the same computing conditions, the two-core and four-core write efficiency are 15–20 times higher than that of single-core with the 10000 steps of iteration, and the Mach number of distribution cloud chart obtained from tow-core, four-core computing and single-core computing is basically the same. I/O performance of the program is significantly improved.