Improving the Scalability of the Basin Scale HWRF Model

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Thursday, 8 January 2015: 2:15 PM
128AB (Phoenix Convention Center - West and North Buildings)
Javier Delgado, NOAA/AOML, Miami, FL; and T. Quirino, X. Zhang, and S. Gopalakrishnan

The operational Hurricane Weather Research and Forecasting (HWRF) model currently generates forecasts by running simulations consisting of a relatively coarse resolution parent domain and a pair of progressively higher resolution, storm-following nest domains. One of the latest developments for hurricane modeling being researched at AOML's hurricane research division is the basin scale HWRF, whose parent domain encompasses a much larger region spanning the North Atlantic and East Pacific basins. This development includes the ability to integrate multiple storm-following pairs of nests in the same simulation to capture storm to storm interactions. It has shown promising results in terms of hurricane prediction, but the increased computational requirements of having multiple storm-following nests prevents it from satisfying the operational time constraints. Furthermore, due to the sizes of the inner nests, simply increasing the number of processors is not a solution to the problem.

We present our solution towards making the basin scale HWRF worthy of consideration for operational use by reducing its execution time. Our solution takes advantage of the fact that moving nests can be integrated independent of one another. Hence, instead of sequentially integrating each pair of storm-following nests, they are integrated in parallel. While using different sets of workers for each nest is not trivial within the HWRF framework, we minimized the amount of changes needed by devising a solution that takes advantage of the fact that modern compute nodes contain multiple compute cores. Our modified code uses a subset of cores from each machine to operate on each of the storm-following pairs of nests.

Our results show that with our approach, computation time remains constant regardless of the number of storms, rather than linearly increasing as more storms are simulated. Furthermore, we confirmed that forecasts made using our parallel integration approach are statistically equivalent to those made using sequential integration. The system is currently running as part of the operational demonstration for the 2014 hurricane season. We conclude that the performance and robustness of this approach is a significant milestone towards making the basin scale HWRF meet the operational time constraints.