Computational Performance of an OpenMP-enabled, MPI-enabled and GPU-accelerated Two-dimensional Flood Model

Floods are one of the devastating natural disasters that can occur irrespective of country, region or border. Computationally intensive models have been developed to simulate flood regime of an area to understand, predict and study its impact on the community. Recently, flood models with improved computational speed have garnered interest for many modelers due to the necessity of faster prediction to enable quicker response and to reduce loss of life and property. The main objective of this study is to compare the computational performance of a two-dimensional numerical flood model in shared memory multicore (OpenMP), GPU (CUDA), and distributed computing (MPI) environments. The model is based on shallow water equations (SWE), and uses upwind-finite difference numerical formulations to simulate inundation extents along with flood depth and velocity fields. Three study areas with different spatial extents are used to compare the performance of the models. The case study events include Taum-Sauk dam break event (Missouri); New Madrid Spillway breach (Illinois); and a hypothetical flooding of Carrollton city (Georgia) under PMF. Three comparison metrics are used to compare the increase in the performance in using the CPU or GPU MPI versions. Further discussions are provided on the computational approaches and accuracy of the models for the three study areas.