Computational science has benefitted tremendously from Moore's Law, by which computer performance grew exponentially over nearly five decades at constant cost. This is especially true for weather and climate modelling, where very complex models could be implemented, maintained and systematically improved over decades. During the past 10 years, however, there has been a number of challenges, such as exploding parallelism, power consumption and hardware cost, and most recently the tapering of scalability of the underlying CMOS technology, that have raised severe concerns about established approaches to implement weather and climate models on modern supercomputers. At the same time, supercomputing initiatives in the USA, China, Japan and Europe seem to be determined to deliver exaflops supercomputing capabilities early next decade, at almost any cost. It is understood that such performance will require substantial investments in refactoring of software to exploit the performance of emerging exascale computing architectures. We will discuss the results of a collaboration between CSCS and MeteoSwiss that could serve as a template for such developments. In order to meet the performance needs of MeteoSwiss’ new operational model within a constant budget envelop, it was necessary to substantially refactor the implementation of the COSMO code and build a system with GPU accelerators. The resulting software infrastructure is applicable to global models as well, and could serve as a basis for performance-portable implementations of weather and climate models that run on all emerging exascale computing architectures. Thomas Schulthess is Professor of Computational Physics at ETH Zurich and Director of the Swiss National Supercomputing Center (CSCS) since 2008. He studied physics at ETH Zurich and earned his doctorate in 1994 with a thesis on surface physics, in which he combined experiment and supercomputing-based simulations. He subsequently continued his research in the USA, working on the DARPA-funded spintronics project and spending more than 10 years at Oak Ridge National Laboratory (ORNL), from where he returned to Switzerland in 2008. Thomas led the teams that won the ACM Gordon Bell Prizes in 2008 and 2009 with the first production-level applications sustaining a petaflops on ORNL’s first peta-scale supercomputers. In the past 10 years, as director of CSCS that hosts the operational weather forecasting systems of MeteoSwiss, Thomas took interest in developing software systems for weather and climate simulations, as wells a design of computing systems for numerical weather prediction. Under his leadership, CSCS was the first center in Europe that deployed a productive GPU-accelerated supercomputing system in 2013, and co-designed with Cray, NVIDIA and MeteoSwiss the first GPU-based weather forecasting system that has been deployed in 2015, and since spring 2016 is running MeteoSwiss’ new COSMO-NEXT model at 1km horizontal resolution. For this work CSCS and MeteoSwiss received the 2016 Swiss ICT Award for Outstanding IT-Based Projects and Services.
GFDL Informal Seminar
Mon, Sep 17, 2018, 10:30 am to 11:30 am
Smagorinsky Seminar Room 209