Addressing HPC Challenges in the Development of Global, Cloud-Resolving Weather Prediction Models

Continued advances in weather prediction models depend on (1) increasing spatial resolution, (2) increasing the number of ensemble members, (3) including of more physical processes, and (4) increasing the scale and accuracy of data assimilation to incorporate billions of observations collected by next-generation, satellites, radars, and a myriad of in situ sensors worldwide. It will also require next generation HPC systems with at least 1000 times more computing power to run global, cloud resolving prediction models.

However, massive increases in computing will not be sufficient to overcome models not designed for fine-grain processor technologies. Effective utilization of processors with thousands of compute cores, and systems with millions of compute cores will require adapting and rewriting applications to prepare them for the exascale computing era. Such investigations have already begun with investment by the European Union supporting efforts by ECMWF, UK Met Office and other centers worldwide. To date, investigations have focused on three main topic areas: new scientific algorithms that offer more parallelism, code rewriting to improve computational efficiencies, and improvements to systems design including more efficient I/O and workflows.

Recognizing these computational and scientific challenges for the U.S. led to the creation of the NOAA Exascale Computing Project in early 2018. This presentation will describe activities and results to date in dwarf model development, advanced data assimilation, machine learning, and I/O. References to other presentations at AMS, publications, and future plans will also be given.