CYBERINFRASTRUCTURE AND CLIMATE MODELING Bob Oglesby Department of Geosciences* *As of 2 Jan 2006 Presented at the UNL Cyberinfrastructure 2005 Workshop 15 August 2005
Climate Models, especially the big, high resolution GCMs, are among the most computationally-demanding programs that exist They require extensive CPU power BUT also produce vast amounts of output. (Reminding of the old adage that a ‘supercomputer is defined as a machine that takes a CPU-bound problem and makes it an I/O bound problem’) The vast amount of output requires both huge data-storage requirements, AND considerable computational power for processing and visualization.
At present I typically run global and regional climate models at the National Center for Atmospheric Research (NCAR) and at Oak Ridge National Laboratory (ORNL) on IBM platforms. The regional model ‘MM5’ with a typical high-resolution domain (12 km) requires about 40 wall clock hours per month of model simulation on a node of the IBM ‘Cheetah’ at ORNL, a 27-node IBM p690 system, where each node has thirty-two 1.3 GHz Power4 processors The global model ‘CCSM3’ at a T85 resolution (about 150km) requires 4.5 years per wall clock day on 192 processors on the NCAR IBM-SP Power 4 bluesky machine
Results from new CCSM3 IPCC runs currently being analyzed
Mesoamerican Deforestation Scenarios – Implications for Wet and Dry Seasons January (dry season) Temperature (left) – warms everywhere but only by 1-2deg C Rainfall (right) - little impact as it’s the dry season! July (wet season) Temperature (left) warms everywhere, but now as much as 2-5 deg C. Rainfall (right) shows a decrease of 20-30% over much of the Mayan region
Typically I do basic processing of the output where the model was run and then bring required data to my local machines for final analysis and processing Depending on length of run and frequency of writeouts, anywhere from a few gigabytes to a terabyte of output can be produced per run.
The key question as I see it for UNL is if the university wants to develop the capability of doing production runs with the climate models ‘in-house’ OR Maintain sufficient computational resources to do final processing and visualization of output from runs made elsewhere I have worked with Prof. Clint Rowe from Geosciences to port CCSM3 to the UNL ‘prairiefire’ cluster. It appears the model runs OK on the cluster, but I doubt that production runs would be feasible without considerable enhancement to the system and its storage/data handling capabilties