Addressing the Computational Challenges of RRFS: a 3 km North America Ensemble Prediction System

The Rapid Refresh Forecast System (RRFS), an hourly updating 3 km grid spacing ensemble forecast system, is being developed to run over a large North American (NA) domain. This large domain has presented a number of computational challenges to overcome, as this NA domain has roughly 7x as many grid cells as the CONUS-sized High Resolution Rapid Refresh (HRRR) domain for which certain RRFS codes were originally developed.

This talk will review various RRFS computational bottlenecks, many tied to the extreme I/O and memory requirements of the system. These bottlenecks needed to be overcome to allow the system to run an hourly updating, ensemble data assimilation cycle over the NA domain in a timely fashion. The talk will discuss the impacts of high-performance flash storage, data compression, 32-bit physics, and other optimizations. We also will comment on how we designed the RRFS system to fit within the constraints of a future NWS operational modeling suite by developing a flexible Python tool to simulate the computational load relative to capacity for various configurations.