Efficient Routing Computations with a Graph-Based Routing Framework

Abstract:

To resolve non-uniform and unsteady flows in the National Water Model (NWM), the Office of Water Prediction is developing additional routing engines to power simulations with the dynamic and diffusive approximations of the St Venant equations. This gives rise to two major computational challenges. First, the presence of both upstream and downstream boundary conditions requires tracking topological connectivity of the entire network within the computation. Second, all solution methods, whether explicit or implicit, become computationally expensive when scaled to continental domains. To be viable for operational modeling as an element of the National Water Model, the computational framework for dynamic routing must address these challenges.

We present a continental-scale flow routing framework that represents the flow network as a collection of directed acyclic graphs where edges point in the direction of downstream flow. We use information from this graph representation to efficiently drive a parallelized computation of flow from headwaters downstream to the tailwaters. This approach has achieved modest performance gains in terms of overall compute time and resources for the routing cases we have tested. The framework is publicly developed and we encourage interested community members to use our approach and provide feedback.

Initial results show that we can simulate 5 days of continental scale flow routing below all existing national weather service forecast points in approximately 10 minutes using only 4 processors. Also, the new framework permits computation using upstream dependencies in all timesteps, which is not possible in the present NWM routing framework. We will continue our work with the goal of significantly reducing barriers to efficient application of higher order routing solutions in the National Water Model, enabling more useful forecasts that help communities prepare for hydrologic hazards.