Examination of the Weather Research and Forecast (WRF) Model on Heterogeneous Distributed Systems

The Weather Research and Forecast model (WRF) is a numerical weather prediction model suitable for research and operations. WRF is designed to run on a variety of platforms, either serially or in parallel, with or without multi-threading. Currently, a model forecast can be run on only one system even though the choice of system is highly flexible. In this paper, we describe our implementation of WRFv1.3 using TCP/IP that allows division of a model domain between two heterogeneous systems distributed across a network. This option can help current and potential WRF users utilize all the computing power available to them; it allows users to run larger and/or higher resolution domains than would be feasible on only one of their systems. Forecast results are equivalent between distributed and non-distributed runs on the same domain with the same initial conditions. We examine the communication costs and their impact on performance, explore how to reduce those costs, and discuss next steps in development