Performance Modeling and Scalability for the ICON Model

Global kilometre-scale resolving weather and climate models come with extreme compute requirements. Profiling, understanding and finally predicting the performance and scalability of these models is hence utterly important to efficiently exploit today’s supercomputers on the one hand and to achieve optimal time-to-solution in weather and climate predictions on the other hand. The Centre of Excellence in Simulation of Weather and Climate in Europe (ESiWACE) focuses on the development of global high-resolution models, which have fed into the international intercomparison project DYAMOND. A key challenge in these developments lies - besides the scientific case of DYAMOND - in the prediction of the performance of models used in the communities on upcoming exascale systems. I will present recent considerations on measuring and predicting performance of weather and climate models at the example of the ICOsahedral Non-hydrostatic (ICON) model. After introducing ESiWACE and DYAMOND, I will discuss scalability of high-resolution ICON runs and present a semi-analytical performance modeling approach to predict ICON scalability, given hardware properties and some additional knowledge on the model run configuration. I will close with the presentation of the sparse grid regression technique to predict performance from measured scalability data.