Europe's proven and fastest algorithms for numerical weather and climate prediction, including both ocean and atmosphere components, represent substantial financial and intellectual investments, often in the form of up to millions of lines of code. Substantial progress has been made in improving the scalability of these codes which has enabled Europe's first ever prototype global kilometre-scale simulations. This capability is expected to eliminate climate uncertainty associated with parametrised processes such as clouds and convection (Schneider et al. 2017).
But fundamental changes are required to enable optimisation and adaptation for emerging HPC architectures as well as to obtaining further increases in parallelism towards exascale.
The ESCAPE project is addressing this issue on so-called weather and climate dwarfs that allow to focus adaptation and optimisation on very specific compute and data movement patterns. Dwarfs represent the wide range of different domain specific algorithms that are key components of weather and climate models, and they contribute significantly to the overall workload.
MeteoSwiss develops with partners of the project a novel high-level DSL toolchain that will be demonstrated on the dwarfs of the project.