03.02.2016
About 120 Members of 18 different scientific instituts from all over Germany came together to discuss the steps that need to be taken into account to accomplish a simulation of 130000 years using a fully coupled atmosphere-ocean-ice model.
Besides the improvement of existing models towards more capable physical and biogeochemical systems, optimization with regard to computer wall clock time plays an important role. Therefore, DKRZ is part of the project and will work together with the other project partners on improving the simulations running on DKRZ's HPC system Mistral and exploring new algorithms and other computer architectures. Joachim Biercamp, Head of the Applications department at DKRZ, is one of the coordinators of PalMod working group 4 aiming at optimization of the Earth system models utilized.
Here, optimization has a twofold meaning: optimization of quality and optimization of performance. Both aspects are crucial for the intended initiative as long-term Earth system simulations only make sense if the quality of the results can be assured and the results can be obtained in an acceptable amount of wall clock time.
For example Work Package 4.4 "Algorithmic and Implementation Performance Optimization", led by Hendryk Bockelmann (DKRZ), utilizes several options to optimize the models. First, model specific developments and improvements will be undertaken to increase the capability of the biogeochemistry model HAMOCC, where grid coarsening strategies and an improved sediment acceleration tool will be developed. As a second task, new transport methods for tracer constituents based on semi-Lagrangian advection schemes will be analyzed. These two tasks aim at a significant reduction of the possible runtime of model years per day while keeping the model code still maintainable and understandable for researchers.
Further optimizations shall be provided by two more technically motivated optimization tasks. After evaluation of existing Earth System Model configurations with detailed performance analysis, the potential for performance gain that might be achieved by the use of specialized hardware or by new algorithms in selected components has to be quantified. Based on that, suitable implementations for actual and upcoming hardware will be promoted, ultimately leading to faster simulations.
Further information is available on the PalMod website: www.palmod.de