An Implementation of MPAS-Atmosphere Running on GPUs

Over the past several years, graphics processing units (GPUs) have been used to accelerate
high performance parallel computing workloads. In fact, many of today’s fastest
supercomputers are being built by coupling GPUs and CPUs within nodes. In these systems,
GPUs often account for an order of magnitude more computing power than CPUs. Today, many
scientific workloads, including numerical weather prediction (NWP) systems, are being ported
and in some cases built from the outset, to take advantage of the capabilities of GPUs.
The atmospheric component of the Model for Prediction Across Scales (MPAS), developed by
the National Center for Atmospheric Research (NCAR), is a general circulation model and NWP
system that is currently being ported to nVidia GPUs using OpenACC directives. Currently, the
dynamical core is fully ported, and runs several times faster on one nVidia Volta v100 GPU as
compared to a dual socket socket server with Intel Broadwell CPUs. As of this writing, a port of
a full suite of physics necessary to drive NWP is nearly completed. Once complete,
MPAS-Atmosphere will be fully capable of running running full numerical weather prediction on a
GPU-accelerated supercomputers.
This presentation will provide results that show the performance gained by porting the system to
GPUs, examples of the necessary code restructuring that is required to gain optimal
performance on GPUs, and how MPAS is scaling on systems with up to several hundred GPUs.
This port of MPAS-Atmosphere will run operationally at The Weather Company on an IBM
supercomputer, and results from this implementation will also be presented