Data Assimilation Research Testbed software

Data assimilation is providing rapid advances in geophysical studies. The Data Assimilation Research Section (DAReS) of IMAGe develops and maintains the Data Assimilation Research Testbed (DART), a software facility for ensemble data assimilation. DAReS also provides support to a growing community of NCAR, university, and government laboratory partners who apply ensemble data assimilation methods with DART.

CISL’s data assimilation research supports CISL’s computing imperative for software cyberinfrastructure. Further, developing and supporting the capabilities of the DART facility is specified as a strategic action item in the CISL Strategic Plan.

RMA version memory management
The remote memory access (RMA) version of DART uses MPI2 one-sided communication to allow any MPI task access to any part of the model state vector that is placed in a memory window by the task that “owns” it. This means that DART can now compute the expected value of observations while the state vector is distributed across many MPI tasks. This removes a previous limitation that the entire model state must fit in the memory of a single MPI task and allows DART to run with much larger prediction models.

A new version of DART that uses the remote memory access (RMA) capability of MPI-2 has been developed and is in production use by several groups producing real-time forecasts with WRF. The DART RMA version eliminates the need to store a complete model state vector in the memory space of a single process and allows much larger models to work with DART. Work has begun on merging the previous DART software that computed all forward observations locally with the RMA version into a single source that will work efficiently for models of all sizes.

An initial implementation of a cross-component assimilation system for fully coupled CESM has been completed. This implementation has revealed a number of new software engineering challenges, in particular, allowing simultaneous use of observations like temperature that exist in more than one model component (e.g., atmosphere, ocean, land).

The interfaces between DART and the CAM family of models (CAM-FV, CAM-SE, WACCM, CAM/Chem) have been updated to work efficiently with the latest versions of the models and DART.

A fundamentally new data assimilation algorithm, the fourth-order-accurate quadratic ensemble filter developed by Dan Hodyss at NRL, has been implemented in DART. The quad filter was tested with both low-order models and GCMs and will work with any model for which a DART interface exists.

DART interfaces to a number of new models were developed in collaboration with model developers and users. New models include: the WRFHydro hydrological modeling system, the CM1 non-hydrostatic convective model in collaboration with the University of Washington, the Joint UK Land Environment Simulator (JULES) land surface model in collaboration with the University of Bristol, the Laboratoire de Meteorologie Zoom (LMDZ) atmospheric GCM in collaboration with the Indian Institute of Technology, and the Unified Curvilinear Ocean Atmosphere Model / General Curvilinear Ocean Model (UCOAMS/GCOM) in collaboration with San Diego State University.