Towards the Development of An Operational Hydrodynamic Flood Simulation Capability – A Multi-GPU TRITON Framework

Abstract:

Given the increasing frequency and intensity of extreme weather events in a changing environment, faster and more accurate flood inundation models are ever more important. When translating the forecasted rainfall depth to inundation, coupled hydrologic-hydraulic models are needed, in which the inundation model is the final step that simulates surface ponding based on simulated runoff and streamflow from hydrologic models. Despite the great benefits, mainly given the high computational cost, large-scale operational inundation forecasting based on hydrodynamic models is yet to be achieved.

To enable operational hydrodynamic inundation modeling, the use of modern graphics processing unit (GPU)-based high performance computing (HPC) system is necessary. Here we introduce a new open source multi-GPU 2D flood model named TRITON (Two-dimensional Runoff Inundation Toolkit for Operational Needs, https://code.ornl.gov/hydro/triton). TRITON solves the full 2D shallow water equations with source terms using an explicit first order upwind finite volume scheme. TRITON can take both in-stream hydrographs and surface runoff (i.e., excess rainfall) as model inputs and hence can simulate both pluvial and fluvial flood events. To be adaptable to heterogeneous computing platforms, TRITON has been designed for a multi-architecture paradigm, and can operate on several configurations including single or multiple central processing units (CPUs) and GPUs using a combination of Open Multi-Processing (OpenMP), Compute Unified Device Architecture (CUDA), and Message Passing Interface (MPI).

To demonstrate the performance of TRITON, we present case studies of the Hurricane Harvey and Tropical Storm Imelda flood events in the Houston and Beaumont, TX areas. We show that TRITON may provide reasonable inundation depth estimates, and can also achieve significant computing time reduction through GPU acceleration. In particular, under the CUDA-Aware MPI implementation, TRITON can scale to 384 GPUs for a 272 million grid cell application. We further discuss the unresolved challenges and required next steps to eventually achieve the goal of operational hydrodynamic flood forecasting.