June 16-20, 2013

Leipzig, Germany

Presentation Details

Name: Architectures
(9) A GPU Capable Version of the COSMO Weather Prediction Model
Time: Monday, June 17, 2013
2:05 PM - 2:10 PM
Room:   Multi-Purpose Area 4 (MPA 4)
CCL - Congress Center Leipzig
Speakers:   Xavier Lapillonne, ETH Zurich
Abstract:   As a part of a larger effort to prepare the numerical weather prediction code COSMO (Consortium for Small-scale Modeling) to future high performance computing architectures, a prototype version of the model capable of running on Graphical Processor Units (GPU) is being developed. In order to avoid the high cost of transferring data between the GPU and the host CPU, a full port of the software to GPU is considered. More precisely, only the initialization phase and I/O remain on the CPU while most other computations, i.e. those occurring during the time integration, are done on the GPU. In view of achieving high performance and yet maintain a portable solution, different approaches were chosen for the dynamical core (solving the Euler equations), physics packages (treating sub-grid scale phenomena) and data assimilation (nudging the model towards observational data). The dynamical core was rewritten using a domain specific embedded language (DSEL) supporting different back-ends for multi-core CPUs, GPUs and potentially other hardware architectures. For the physics packages and data assimilation on the other hand a complete rewrite would neither be feasible nor acceptable by the user community. For this reason a compiler directives approach, based on the OpenACC standard is used. Details of the implementation strategy are presented together with some specific optimizations introduced to target the GPUs. Performance results for the different parts and for the integrated code, using a representative weather prediction benchmark, are shown.

Poster authors: Mauro Bianco, CSCS; Tiziano Diamanti, Meteoswiss; Oliver Fuhrer, Meteoswiss; Tobias Gysi, Supercomputing Systems; Xavier Lapillonne, ETH Zurich; Carlos Osuna, ETH Zurich; Thomas Schulthess, CSCS
Program may be subject to changes.