Multi-GPU parallelization of shallow water modelling on unstructured meshes

Abstract

Floods are one of the most devastating natural hazards globally, causing significant loss of life and extensive economic damage. Shallow water equation (SWE) models, due to their clear physical mechanism and good accuracy, can provide detailed predictions of flood behaviour, which are essential for flood risk evaluation and mitigation. However, traditional SWE models face significant limitations in supporting large-scale, long-duration, and high-resolution numerical simulations, which are increasingly demanded by modern applications such as flood forecasting and the establishment of warning systems. In response to the increasing demand for rapid and accurate flood modelling, this study presents a multi-GPU accelerated unstructured mesh SWE model. The proposed model employs MPI-OpenACC method to facilitate multi-GPU parallel computing for hydrodynamic simulations and incorporates a novel asynchronous communication strategy aimed at minimizing the overhead associated with parallel communication. Three representative flood cases were employed to assess the accuracy and efficiency of the proposed model. The results indicated that the speedup of the proposed model reached more than 800 when using eight GPUs in parallel, and the model could simulate a 30 h extreme flood in a 1,300 km² watershed within 0.35 h. Multi-GPU parallel computing holds great promise for applications in rapid flood simulation and real-time risk assessment.