Advanced GPU Parallelization for Two-Dimensional operational river flood forecasting.

---

Reinaldo Garcia, Pedro Restrepo, Mike DeWeese, Mark Ziemer, Justin Palmer, Jonathon Thornburg, Javier Murillo, Pilar Garcia-Navarro, Asier Lacasta, Mario Morales

Thursday 2 july 2015

11:30 - 11:45h at Oceania (level 0)

Themes: (T) Water engineering, (ST) Computational methods

Parallel session: 11E. Engineering - Computational

---

One-dimensional (1D) models have traditionally been used for river flood prediction in large river reaches, mainly because they run fast due to the simplified flow equations used. However, in complex floodplains or where the flow is unconfined, 1D models may not provide an adequate solution due to the limitations of uniform water velocity and constant water surface elevation on each cross section. This warrants the use of more accurate two-dimensional (2D) models. However, the numerical solution of the 2D dynamic equations and the requirement of flexible meshes to resolve complex terrain characteristics, had made until recently 2D models considerably much more demanding in computer times than 1D models. Recent advances in massive parallelization techniques for 2D hydraulic models are able to reduce computer times by orders of magnitude making 2D applications competitive and practical for operational flood prediction in large river reaches. Moreover, high performance code development can take advantage of general purpose and inexpensive Graphical Processing Units (GPU), allowing to run 2D simulations more than 100 times faster than old generation 2D codes, in some cases. This work describes the application of the RiverFlow2D GPU model to a 670-Km reach of the Red River in North Dakota and Minnesota, USA. This project is a collaborative effort between NOAA, Hydronia and University of Zaragoza to assess the accuracy and performance the RiverFlow2D model at NOAA North Central River Forecast Center as a 2D operational model for river flood forecasting. Test runs indicate that routing a 40-day hydrograph using between 600,000 and one-million cell meshes can run in as little 30 minutes on a NVIDIA GTX Titan Black 2,880 GPU-core hardware. This performance is even better than the existing 1D model currently used as a forecast tool in the same river reach. Ongoing efforts include tests on other advanced GPU hardware such as the NVIDIA TESLA K20 and K40. Results suggest that 2D GPU models such as RiverFlow2D are be able to achieve the performance required in hydrological forecasting such as that of NOAA River Forecast Centers.