Optimizing Performance and Scalability of HWRF

The basin scale implementation of the Hurricane Weather Research and Forecast (HWRF) model has the ability to integrate multiple storm-following high resolution nests; however, meeting operational time constraints requires a new more scalable nest parallelization approach, as well as one that is supportable by NCAR as part of the community version of WRF. In the original nesting scheme in WRF and HWRF, every nested domain needed to be decomposed over all the MPI-compute tasks allocated to the HWRF job and then integrated serially, one-after-the-other,. Now, patterned after the nest-parallelization scheme developed for NCEP's NMM-B, this new implementation of nest parallelism in HWRF allows each nested domain to be run on a separate subset of the total number of MPI tasks assigned to an HWRF job. This approach takes advantage of the fact that nests over different storms may run concurrently. And since each nest is decomposed over fewer tasks, the multi-storm scenario runs more efficiently overall. We also report on improved efficiency when there is only one storm and thus no inter-storm concurrency. In the case of a single storm, there are still opportunities for concurrency between nest levels, provided that there are no sequential forcing/feedback dependencies between a nest and its parent.