The Interactive Ensemble Coupled Modeling Strategy Ben Kirtman

The Interactive Ensemble Coupled Modeling Strategy Ben Kirtman Center for Ocean-Land-Atmosphere Studies And George Mason University

Science Questions • How Does Internal Atmospheric Dynamics Impact Climate Variability? – Is ENSO Stochastically Driven or Self. Sustained? – How Do Errors in Internal Atmospheric Dynamics Impact Climate Predictability? • Is it Possible to Combine Multiple Imperfect Climate Models to Improve Climate Simulations and Better Quantify Uncertainty?

Standard Approach • A Posteriori Combinations of Ensemble Realizations (Coupled and Uncoupled) – Does Not Isolate How Internal Dynamics Impacts Evolution of the Coupled System • Want to Employ Ensemble Averaging As the Coupled System Evolves

SST CAM_MOM 3 Heat Flux Momentum Flux MOM

SST COLA_MOM 3 COLA Heat Flux Momentum Flux MOM

SST CAM COLA Heat Flux Momentum Flux MOM CAM_COLA_MOM 3

CAM SST COLA Heat Flux Momentum Flux MOM COLA_CAM_MOM 3

Computational Issues (I) • Load Balance Between Oceanic and Atmospheric Components – More Processors • Communication of Coupling Information – Frequency Can be a Problem • Data Management – Dataportal (little MSS)

Computational Issues (II) • Use (Prospect) Blackforest, Bluesky – Conversion from Compaq to IBM Difficult • Communication: Did Not Know the Right Questions to Ask

Computational Issues (II) • Use (Prospect) Blackforest, Bluesky – Conversion from Compaq to IBM Difficult • Communication: Did Not Know the Right Questions to Ask • Documentation Incomplete (e. g. , loadleveler geometry) • Consulting Support Excellent • Usage (Typically 1 Node for each Component) – Blackforest: 4 x 75 min/simulation month (for each atmosphere) and 4 x 75 (for each ocean) – Bluesky: 8 x 20 min /simulation month (for each atmosphere) and 8 x 20 (for each ocean)

Example of CPU (GAU) Usage • Bluesky 12 Atmospheres and 1 Ocean 100 year simulation 13 * 8 * 20 min/month * 1200 month = 41, 600 CPU hours In Regular Queue Simulation Takes about 30 -50 days In Practice We Get About 2 -3 Experiments Completed Per Year This Should Double Every Year Over the Next 5 Years

Data Management • Some Reliance on MSS – Primarily Job Recovery • Initially ftp’d Data Back to COLA • Access to dataportal Grads-Dods-Server (GDS) Has Changed My Life • Heavy usage of /ptmp

Concluding Remarks • NCAR Facility Ideal for Testing Interactive Ensemble Coupling Strategy • Demand on Computing to Steadily Increase Over the Next 5 Years • Need for Local Visualization of Remote Data Large and Increasing