IBM jumped into the weather forecasting business roughly a year and a half ago by purchasing The Weather Company. This week at ISC 2017, Big Blue rolled out plans to push deeper into climate science and develop more granular forecasting models and to port those models onto the Power platform including forthcoming Power9-based systems. Power9 is IBM’s next generation Power processor and will be used in the CORAL supercomputer project.
A key aspect of the new effort is leveraging GPUs on the Power platform. “If you know the weather space, the utilization of GPUs has been pretty limited. There’s the COSMO (Consortium for Small Scale Modeling) model used in Switzerland,” said Dave Turek, IBM vice president of exascale system. While not completely new, use of GPUs has been limited in weather forecasting (see HPCwire articles, Swiss Weather Forecasting Achieves 1.1km Resolution on ‘Piz Kesch’ and Today’s Outlook: GPU-accelerated Weather Forecasting).
GPUs have shown promise helping to tame one of weather forecasting’s thorniest problems which is being able to scale the models to both small areas and large areas and integrating those into forecast spanning local, regional, and global weather. Turek says GPUs will become increasingly important for achieving the necessary near real-time, multi-scale simulations required for accurate forecasting.
“The early indications are we should see significant performance gains and it is really going to be an enabling technology, if you will, to let us go down to that level of scale,” he said. “You should see some results [from this effort] around the end of the year.”
The collaboration announced at ISC includes IBM, The Weather Company, University Corporation for Atmospheric Research (UCAR) and the National Center for Atmospheric Research (NCAR), a federally funded research and development center for the atmosphere and Earth’s geospace systems.
Specific work will be around further developing MPAS – the Model for Prediction Across Scales (MPAS) – and porting it to the IBM system. MPAS itself is a collaborative project for developing atmosphere, ocean and other earth-system simulation components for use in climate, regional climate and weather studies. To date, the primary development partners working on MPAS have been the climate modeling group at Los Alamos National Laboratory (COSIM) and the NCAR.
As noted, the basic idea is to be able drill down on small scale events. Turek noted today’s operational global forecast models predict weather patterns down to regional-scale weather events, such as snowstorms and hurricanes. The new model could improve weather and climate forecasting by better accounting for the small-scale phenomena, such as thunderstorms, that can impact the weather.
Here’s a bit more detail on MPAS from GitHub: “The defining features of MPAS are unstructured Voronoi meshes and C-grid discretization used as the basis for many of the model components. The unstructuctured Voronoi meshes, formally Spherical Centroidal Voronoi Tesselations (SCVTs), allow for both quasi-uniform discretization of the sphere and local refinement. The C-grid discretization, where the normal component of velocity on cell edges is prognosed, is especially well-suited for higher-resolution, mesoscale atmosphere and oceanic simulations. The land ice model takes advantage of the SCVT-dual mesh, which is a triangular Deluanay tessellation appropriate for use with finite element-based discretization.”
A key goal is to adapt MPAS to run more efficiently on next-generation computers. While regional models have been run at scales that predict thunderstorms for over a decade, the enormity of global models have made this challenging on a global scale. Enabling “convection-allowing models” on a global scale will not only enable short-term thunderstorm forecasts, but also lead to more accurate long-range forecasts days, weeks and months in advance, according to IBM.
The UCAR/NCAR teams involved include the Computational Information System Laboratory and the Mesoscale & Microscale Meteorology Laboratory.
“This is a major public-private collaboration that aims to advance weather prediction and generate significant benefits for businesses making critical decisions based on weather forecasts,” said UCAR President Antonio Busalacchi in the official release. “We are gratified that taxpayer investments in the development of weather models are now helping U.S. industries compete in the global marketplace.”
Currently, Power9 silicon is not available although Turek said IBM is doing some early testing with “partners” not customers. Very likely that includes GPU specialist Nvidia, which along with IBM, is an OpenPOWER founder. The Power9 chip, said Turek, will be the first in the line to have cache coherency such that the GPU will be seen as a coprocessor which should make programming easier. Power9, of course, is intended to support several types of accelerators using different optimized interconnect schemes.