| Summary: | Simulations supporting the scientific consensus that human activity is changing the Earth’s climate have been derived from models run at O(100 km) resolutions. The impact of unresolved scales on these predictions is not precisely known: indeed it has been hypothesized that noise in the climate system (fluctuations on short spatial and temporal scales) could be "reddened," thereby influencing the low-frequency components of the climate signal. To test this hypothesis, we need to run high-resolution, century-scale simulations of the Earth System: this is the primary goal of our use of the Cray XT4 and XT5 systems at the National Institute of Computational Science (NICS) and the National Energy Research Scientific Computing Center (NERSC). These large-scale, resource-intensive climate experiments require careful tuning at scale to achieve a reasonable compromise between integration rate and efficiency. This paper presents preliminary performance and scaling data from a variety of Cray XT systems for a high-resolution (0.5° atmosphere and land surface coupled to 0.1° ocean/sea ice) development version of the Community Climate System Model (CCSM) in configurations capable of running efficiently on up to 5,844 processors. We have achieved integration rates of 2.3 simulated years/day for CCSM4_alpha on the Franklin XT4 and approximately 2.0 years/day on the Kraken XT5 in benchmarks with I/O turned off. An 18-fold variability in output times when CCSM4_alpha writes monthly history and restart files to Kraken’s multi-petabyte Lustre file system during the first seven simulated years of production is also presented and discussed.
|
|---|