Software Engineering
Climate modelling is an essential part of CLICCS. The increasing complexity of climate or even earth system models and the arrival of new hardware concepts are posing great challenges to scientists. Therefore, the aim of this sub-project is to support CLICCS scientists in the area of climate model software engineering in order to enable the models to fully exploit advances in high performance computing technologies. One of our aims is to serve as a link between CliCCS and other research projects for future weather and climate modelling at the national and European level (e.g., ESCAPE, ESiWACE).
One starting point was to work out the computational requirements of CLICCS experiments together with the different research programs in the first phase of CLICCS HPC-DIS Software Engineering. Based on this, the suitability of different programming paradigms for ICON has been analyzed and tested by prototype. However, the existing approaches for high-level domain specific language (DSL) have either been developed for a specific application on a specific hardware, or are not past a prototype stage. Our main conclusion is that DSLs are a promising concept but none exist that would be ready for general use with ICON. A large community effort is needed to get towards a usable solution. This has also been recognized in the ICON community and will be addressed by the ICON-consolidated initiative to which we contribute and in the BMBF-funded projects preWarmWorld and WarmWorld, with which we will closely cooperate.
Furthermore the porting of ICON to GPU systems as an additional option to classical CPU technique is encouraged in the CLICCS HPC-DIS Software Engineering . The atmospheric component of ICON has already been mostly ported to Nvidia GPU systems using OpenACC directives and successfully runs on different high performance supercomputers. We are analyzing the performance of this solution on the JUWELS Booster system at the Jülich Forschungszentrum and contributing with this analysis also to a collaboration between DKRZ and the hardware vendor Atos to study the suitability of GPU systems for climate models and in particular ICON. We could show that the OpenACC version of ICON-A performs well on GPUs for high-resolution experiment setups. At low resolutions the runtime is better on CPUs at the scaling limit. In a next step, we are transferring the experience with OpenACC from ICON-A and the work on the advection component to the rest of the ICON ocean model.
Team: Dominik Zobel, Panagiotis Adamidis, Joachim Biercamp, Claudia Frauen (Alumni)
Contact: zobel@dkrz.de
