ICON Performance Analysis and Simulations for the Project “High Definition Clouds and Precipitation for advancing Climate Prediction – HD(CP)2”

Clouds and precipitation are an essential element of the climate system, but they are difficult to account for in theories and models. Clouds significantly influence the atmospheric radiative forcing and modulate many feedback processes of the Earth system. Interactions with the terrestrial biosphere, sea-ice interactions, and bio-geochemical interactions are critically dependent on the cloud-precipitation representation. Atmospheric clouds and precipitation are small- and meso-scale phenomena, which cannot be modeled explicitly on the global scale with current computer architectures and hence are parametrized in global climate models. Global climate simulations today show significant biases in precipitation patterns, which are related to insufficient parametrization of these processes.HD(CP)2 is a coordinated initiative to improve our understanding of cloud and precipitation events, to increase the quality and number of available observations, and to evaluate and improve modeling capabilities. One group of project partners develops a highly scalable regional weather model which will be used to provide a series of ultrahigh resolution summer-season hind-casts over Germany. Aiming on a horizontal grid spacing of approximately 100 m (and vertical grid spacings considerably finer) over domains of 1000 km, the new HD(CP)2 model will help to understand current limits on cloud-precipitation modeling and parametrization skills. The scale of the simulation allows the explicit representation and analysis of entirely new phenomena. Due to the fine horizontal grid resolution and simulation time steps the ICONLES model must be highly parallelized and optimized for Peta-scale systems. Numerous performance analyses of the ICON code with different MPI/OpenMP task/thread-combinations on JUQUEEN were done. The aim of the performance analyses was to find the best model-configuration, e.g. in terms of MPI/OpenMP splitting and to identify bottlenecks and scaling problems specifically related to massive-parallel systems. The performance analyses of the ICON code on JUQUEEN were done for different horizontal resolutions of 420 m, 240 m and 120 m. Our efforts lead to substantial performance improvements of the ICON code on JUQUEEN and will enable large-scale simulations of HD(CP)2 in phase 2 of the project