Evolution of the ICON Model to Exploit Emerging Technology: An Overview
DescriptionThe Icosahedral Non-hydrostatic (ICON) model has been developed jointly in the last 20 years by the Max Planck Institute for Meteorology and the German Weather Service (Deutscher Wetterdienst) to support both climate simulations and numerical weather prediction (NWP). Due to the expense of such simulations, it was recognized already in 2010 -- well before its wide-spread use -- that ICON's performance could benefit from Graphics Processing Units (GPUs) and potentially other emerging technologies. The Swiss National Supercomputing Centre (CSCS) embarked at that point on a multi-year program to help prepare ICON for the inevitable emergence of GPUs as the primary technology for high-performance computing. This effort started with code rewrites of ICON's dynamical core (solver of the fully compressible Euler equations) in OpenCL and CUDAFortran. Favorably performance results led to a ten year crusade to port all of ICON to GPUs, first using OpenACC compiler directives, and more recently using a Python-enabled domain specific language called Gridtools-for-Python (GT4Py), involving software development partners from MeteoSuisse and the Center for Climate System Modeling (C2SM), among others. This presentation contains an overview of the evolution of ICON for HPC over the last ten years, highlighting both the successes and missteps.
TimeWednesday, June 2815:30 - 16:00 CEST
Climate, Weather and Earth Sciences