Computing capacity is crucial for advances in data processing in weather and climate models, but as newer model versions have increased data resolution, on-site capacity may not be sufficient. On September 21st, 2022, the NOAA National Weather Service and NOA Global Systems Laboratory development teams were able to demonstrate that NOAA's Rapid Refresh Forecast System (RRFS) could be run on Amazon Web Services (AWS) cloud. The RRFS is the next-generation, ensemble-based assimilation, and forecast system. Ensemble forecasts are a group of forecasts valid over the same time period that either start from different initial states, use different models, or both. The cloud-based experiments were successful and taught important lessons for future cloud use with high-resolution modeling.
The team built and tested the performance of RRFS on a non-cloud, on-site supercomputer. They then rebuilt and tested the system on the AWS cloud. AWS is web-based, with the ability to create and manage parallel data clusters, and report cost data. Using the cloud allows access to hundreds of readily-available convective case instances as well as the latest generation of hardware, which speeds up the processing time.
The experiments showed that using the AWS cloud is successful for data management and operations with RRFS. The tests helped develop a methodology to utilize on-site systems for cloud development work, leveraging the expertise of the on-site community to resolve cloud performance issues. The process used to port and rebuild the code to the cloud can be applied to future model developments. The team also realized the need to benchmark cloud costs and work with the High-Performance Computing team to refine cloud architecture. Ultimately, the AWS cloud provides a very viable option for data management and operations as research continues to develop advanced and high-resolution rapid weather and climate modeling.
Caption: Ensemble forecasts are a group of forecasts valid over the same time period that either start from different initial states, use different models, or both.