Abstract

Data Assimilation Research Testbed (DART) is a software environment, which makes it easy for modelers, observational scientists, and geophysicists to explore a variety of data assimilation methods and observations with different numerical models. As we all know, to fulfill the increasing demands ... Show moreData Assimilation Research Testbed (DART) is a software environment, which makes it easy for modelers, observational scientists, and geophysicists to explore a variety of data assimilation methods and observations with different numerical models. As we all know, to fulfill the increasing demands on computing efficiency, general-purpose graphics processing units (GPGPUs) are widely used as a less expensive, more efficient solution to accelerate scientific applications. The DART function get_close_obs was identified as a suitable target for coprocessor acceleration as it is both computationally intensive and called many times during a typical DART run. The purpose of this function is to find all observations close to a given base observation, where close is defined by the user. A performance improvement of this segment of the code would be beneficial across a variety of models. In this project, we are focusing on implementing a parallel version of the exhaustive search method in get_close_obs with CUDA Fortran on NVDIA GPUs, and evaluating the coprocessor feasibility and effectiveness for DART. Show less