SIParCS 2015- Feng

Delilah Feng, University of Wyoming

Evaluating Coprocessor Effectiveness for the Data Assimilation Research Testbed

(Slides) (Recorded Talk)

The Data Assimilation Research Testbed (DART) provides modelers, observational scientists, and geophysicists with powerful, flexible data assimilation (DA) tools that are easy to implement and use and can be customized to support efficient operational DA applications.  The most computationally intensive routine in the DART algorithm has been successfully accelerated in a previous project.  The possibility of achieving further performance improvement by applying GPU programing techniques to other DART routines is investigated in this work. Results from TAU and Allinea MAP were used to select the routine ‘update_from_obs_inc’ as the target for acceleration.  This function does linear regression of a state variable onto an observation and computes state variable increments from observation increments.  Algorithm modification, device memory optimization, and memory transfer techniques were explored in GPU versions of the simplified stand-alone target code. All code was developed with CUDA FORTRAN on an NVIDIA Tesla K20x GPGPU.

Mentors: Nancy Collins and Helen Kershaw, CISL IMAGe