Benchmark Exercise for Comparing the Computational Performance of Two-Dimensional Flood Models in CPU, Multi-CPU, and GPU Frameworks

The objective of this study is to investigate the computational performance and accuracy of three different implementations of a 2D flood model: sequential (Flood2D-CPP), parallel (Flood2DPTH & Flood2D-OMP), and General Purpose Graphics Processing Unit (Flood2D-GPU). The model is based on shallow water equations (SWE) and uses an upwind-finite difference numerical formulation to simulate flood events. Two parallel versions of the model are implemented, one based on pthread and the other based on OpenMP. The GPU version has been developed using NVIDIA's CUDA library. For this study, these implementations are being applied to simulate a dam break event at the Taum Sauk pump-storage hydro-electric power plant in Missouri, which occurred on December 14, 2005. The GPU implementation provided a significant speed up, up to two orders of magnitude compared to the CPU model. As predicted, the sequential model (Flood2D-CPP) reported with the lowest performance compared to the parallel and GPGPU versions, because it would take longer for a single CPU thread to perform all the calculations as opposed to multiple threads or through multiple GPU cores . Results indicate that the computational performance of both Flood2D-PTH and Flood2D-OMP improves with increase in the number CPU threads. The speedups of Flood2D-PTH and Flood2D-OMP are maximized at 8 threads, but much less than the theoretical maximum. In general, even though Flood2D-GPU had significance performance the comparison indicated the potential for optimizing Flood2D-PTH and Flood2D-OMP models to simulate larger computational domains.