Abstract
We propose a software implementation strategy for complex weather and climate models that produces performance portable, architecture agnostic codes. It relies on domain and data structure specific tools that are usable within common model development frameworks - Fortran today and possibly high-level programming environments like Python in the future. We present the strategy in terms of a refactoring project of the atmospheric model COSMO, where we have rewritten the dynamical core and refactored the remaining Fortran code. The dynamical core is built on top of the domain specific Stencil Loop Language for stencil computations on structured grids, a generic framework for halo exchange and boundary conditions, as well as a generic communication library that handles data exchange on a distributed memory system. All these tools are implemented in C++ making extensive use of generic programming and template metaprogramming. The refactored code is shown to outperform the current production code and is performance portable to various hybrid CPU-GPU node architectures.
Highlights
The socio-economic value of numerical weather prediction and climate modelling is driving the continuous development and improvement of atmospheric models on current and emerging supercomputing architectures [22, 33]
We have discussed a reimplementation of the Consortium for Small-Scale Modeling (COSMO) numerical weather prediction and regional climate model that is performance portable across multiple hardware architectures such as distributed multi-core and hybrid CPU-GPU systems
Stencil Library (STELLA) is implemented using template metaprogramming and supports a multi-threaded model with fine grained parallelism on a variety of node architectures supported through architecture specific backends to STELLA The dynamics is built on top of a generic halo exchange and boundary condition framework that provides the necessary functionality for domain decomposition on distributed memory systems
Summary
The socio-economic value of numerical weather prediction and climate modelling is driving the continuous development and improvement of atmospheric models on current and emerging supercomputing architectures [22, 33]. Attaining high scalability and efficiency with atmospheric models is a challenging task for several reasons. The typical codebase of an atmospheric model has several hundred thousand to million lines of Fortran code. Atmospheric models are community codes with large developer communities, and even larger user communities. Due to the non-linear nature of the Earth’s atmosphere, the large range of spatial and temporal scales involved and the complexity of the underlying physical processes, assessing the correctness and predictive capabilities of an atmospheric model is in itself a scientifically and computationally arduous task. Many atmospheric models are characterized by a low arithmetic density and limited scalability. The low arithmetic density results from the algorithmic motifs in the most time-consuming parts of the codes, typically finite difference, finite element or finite volume discretizations on a structured or unstructured grid. As a consequence of these motifs, scalability is limited by the total number of horizontal
Sign-in/Register to view highlights & summary Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
