Weather and Climate Simulations in Python using GT4Py and DaCe
(Presentation - presented in Virtual, Jan. 2022)
Abstract
We will discuss how Python, a productive high-level language, can be used to develop and accelerate real weather models. To achieve that, we first describe the principles of data-centric Python – a subset of Python and NumPy that can be compiled and transformed via dataflow analysis. We then extend the model to weather and climate modeling with GT4Py: a high-productivity embedded DSL that allows meteorologists and practitioners to write stencils with ease. GT4Py optimizes each stencil to its full extent, and the Data-Centric Python framework (DaCe) can optimize full weather models with multiple GT4Py stencils and NumPy primitives in tandem. We demonstrate GT4Py and DaCe on the FV3 dynamical core, showing that it increases productivity in authoring climate models, as well as portability and performance, running on CPUs and GPUs up to tenfold faster than the optimized FORTRAN implementation.
Documents
Recorded talk (best effort)
BibTeX
@misc{, author={Torsten Hoefler and Linus Groner and Tal Ben-Nun and Tobias Wicky}, title={{Weather and Climate Simulations in Python using GT4Py and DaCe}}, year={2022}, month={Jan.}, location={Virtual}, source={http://www.unixer.de/~htor/publications/}, }
