Abstract:

[Background] Traditional Central Processing Unit (CPU) computing models are increasingly inadequate to meet the massive computational demands of large-scale OpenFOAM simulations. Existing Graphics Processing Unit (GPU) acceleration methods, while faster, suffer from performance bottlenecks due to frequent data transfers between the host and the device. Addressing this gap is critical for advancing computational fluid dynamics (CFD) research and industrial applications. [Objectives] This study aims to develop a novel heterogeneous computing framework for OpenFOAM specifically targeting domestic Chinese accelerator cards. The core objective is to significantly boost computational efficiency by achieving full-process device-side computation for the entire process, encompassing equation discretization and solution. This approach seeks to minimize data movement and maximize the utilization of the accelerator’s parallel processing capabilities. [Methods] The developed framework is built on OpenFOAM-v2112, introducing a set of independent data structures and computational logic designed to reside entirely on the device. By systematically overloading OpenFOAM’s discretization functions, the entire computational workflow is seamlessly ported to the domestic accelerator card. The architecture utilizes a composite design pattern, ensuring excellent version compatibility and maintainability while preserving the familiar structure and conventions of native OpenFOAM solver development. [Results] The framework’s performance was rigorously validated using two representative benchmark cases: DTCHULL and Dambreak3D. The results demonstrate that the heterogeneous framework successfully maintains simulation accuracy while achieving a substantial speedup of 2.64 to 72.43 times compared to the native CPU version. [Conclusions] This heterogeneous framework significantly improves the computational performance in OpenFOAM simulations by enabling efficient device-side processing. It establishes a robust foundation for future research into the heterogeneous acceleration of complex physical models and demonstrates immense potential for improving the efficiency of large-scale CFD simulations. [Limitations] The current framework only covers the InterFOAM solver and is only compatible with simple mesh structures.

Key words: OpenFOAM, heterogeneous computing, domestic accelerator card, full process, multi-phase flow