Ensemble Data Assimilation for Climate and Weather Models: A Fascinating Challenge for Scalable Computing
Nancy Collins
NCAR
Boulder, Colorado
Abstract
Data Assimilation (DA) combines information from observations of a system with model forecasts to produce improved initial conditions for subsequent forecasts. Modern models of the Earth's climate system and high resolution weather forecasting systems are large, parallel programs of high complexity that are challenging to run efficiently on today's high performance computing platforms. Ensemble DA methods involve running multiple copies of these large models and statistically combining the forecasts with millions of observations from a wide variety of observing instruments. The communication patterns of these large models are complex; ensemble DA adds significant additional complexity to the communication matrix.
The Data Assimilation Research Testbed (DART) is a community ensemble data assimilation software facility that must work efficiently with a variety of large geophysical models including coupled climate models like CESM and regional weather prediction models like WRF. We discuss the design of DART on current hardware and explore possible implementations on expected future hardware architectures.