Parallelization of the FV3 Model for GPU and MIC Processors

The NWS recently selected the FV3 model to become its next operational global weather prediction model.  To take advantage of next generation fine-grain computing, NOAA’s Earth System Research Laboratory has been parallelizing FV3 for both GPU and MIC processors. Initial parallelization of the dynamical core for the MIC was trivial with results for the latest generation KNL processor showing significant performance benefit versus the CPU. 

Code modifications were needed to expose sufficient parallelism in FV3 routines to run the FV3 efficiently on the CPU. Initial work focused on a single shallow water routine called “c_sw”, that served as a template for the parallelization of other routines in the dynamics portion of the FV3 model.  This talk will describe parallelization efforts to date, show some performance results and give a status of the work.