Performance optimisation and reproducibility analysis of distributed simulations of the Earth climate system

Abstract

Comprehensive Earth System Models (ESMs) are powerful scientific tools for the investigation of the processes and feedback mechanisms in the climate system and for assessing the system response to natural and anthropogenic perturbations. The ECHAM/MESSy Atmospheric Chemistry (EMAC) model system is specifically designed to allow a wide variety of model configurations with tailor made complexity for different scientific tasks. The requirements on the high-performance computing (HPC) system vary with the degree of complexity and the spatial and temporal resolution. In consequence, it is expected that the optimal (i.e., most efficient) HPC system differs for different model configurations. For an efficient exploitation of computational resources, in particular within a heterogeneous HPC grid, it is therefore desirable to select the best suited HPC system for each simulation setup. Further advantages for ESM simulations from a HPC grid arise for distributed ensemble simulations, distributed sensitivity simulations, and the distributed piecewise simulation of long time scales. The prerequisite is, however, the independence of the simulated climatology from the HPC system – a reproducibility characteristic, which is a-priori not guaranteed, due to the (deterministic) chaotic nature of the ESM and the peculiarities of the different HPC systems. The DEISA grid computing infrastructure provides an ideal frame for the efficient analysis (and potential optimisation) of the run-time performance of different EMAC setups on different HPC systems, and further to test the statistical independence (i.e., their reproducibility) of the results on the HPC system.