Oliver Fuhrer from the Federal Office of Meteorology and Climatology Meteo Swiss is guest of Bjorn Stevens (MPI-M).
Time: 3:15 PM, Room 22/23 Bundestraße 53 (ground floor).
High performance computing for weather and climate: Learning how to forget
The weather and climate community has set ambitious goals to reach global km-scale modeling capability on future exascale high-performance computing (HPC) systems. Reducing the horizontal resolution to the kilometer-scale holds the promise of reducing some of the long-standing biases and uncertainties. But how far are we from achieving this goal?
The weather and climate community is struggling to keep up with the fast-paced developments as HPC system designs strive towards the Exascale. Only few production-ready community codes are able to leverage emerging hardware architectures. This has become known as the software productivity gap. While novel programming models - such as OpenACC or CUDA - are available to target these hardware architectures, they further increase the complexity of the code and thus exacerbate the problem.
In this talk, we give an introduction to some of the challenges we are facing. Further, we discuss the experience of a re-design effort for a regional weather and climate model to illustrate a possible alternative pathway. Significant parts of the code have been rewritten based on a domain-specific language (DSL) specifically targeted to the key algorithmic motifs in weather and climate models. The DSL allows a high-level, hardware agnostic formulation of the model which then can be compiled to different hardware targets. Finally, we will demonstrate the capability of the refactored code on the leadership-class heterogeneous HPC system Piz Daint and discuss the implications for global km-scale climate simulations.