The multi-GPU Wetland DEM Ponding Model

Abstract

The Wetland DEM (Digital Elevation Model) Ponding Model (WDPM) is software that simulates how runoff water is distributed across the Canadian Prairies. Previous versions of the WDPM are able to run in parallel with a single CPU or GPU. Now that multi-device parallel computing has become an established method to increase computational throughput and efficiency, this study extends WDPM to a multi-GPU parallel algorithm with efficient data transmission methods via overlapping communication with computation. The new implementation is evaluated from several perspectives. First, the output summary and system are compared with the previous implementation to verify correctness and demonstrate convergence. Second, the multi-GPU code is profiled, showing that the algorithm carries out efficient data synchronization through optimized techniques. Finally, the new implementation was tested experimentally and showed improved performance and good scaling. Specifically, a speedup of 2.39 was achieved when using four GPUs compared to using one GPU.