3.3
Runtime Optimization for the High-Resolution Rapid Refresh Forecast Model

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Thursday, 8 January 2015: 2:00 PM
128AB (Phoenix Convention Center - West and North Buildings)
Curtis R. Alexander, NOAA/ESRL/Global Systems Division and CIRES/Univ. of Colorado, Boulder, CO; and J. Michalakes, T. G. Smirnova, M. Hu, and G. Manikin

The 3-km convective-allowing High-Resolution Rapid Refresh (HRRR) is an hourly updating weather forecast model that use a specially configured version of the Advanced Research WRF (ARW) model and assimilates many novel and most conventional observation types on an hourly basis using Gridpoint Statistical Interpolation (GSI). The HRRR is run hourly out to fifteen forecast hours over a domain covering the entire conterminous United States and is available in real-time to operational forecasters in both the private and public sectors. The HRRR is scheduled for operational implementation at NCEP in 2014.

To facilitate more advanced data assimilation and modeling capabilities in the HRRR development process while maintaining a timely delivery of forecast products that can be used in real-time, high-performance computer (HPC) code optimization has become increasingly important to achieve this balance. Additional constraints for runtime have been established with the transition to an operational HPC system where computer resource usage must be minimized to fit within a pre-determined production schedule and allocation.

This presentation will describe both the resource constraints for the operational HRRR and efforts to reduce the runtime of the model code including code profiling with message passing interface (MPI) traces, the deployment of openMP with threading, and use of dedicated output nodes using quilting with task geometry. Both the full model domain (~95 million grid points) and reduced domain testing will be demonstrated with estimates for runtime savings across all code optimization efforts.