In this thesis, we develop and demonstrate a robust and scalable workflow
for building, deploying, and testing containerized weather models on
the HPC clusters of the EuroHPC Joint Undertaking. The workflow is
designed to support different cluster architectures, including x86-based
systems (LEONARDO, GALILEO100, MELUXINA) and Cray-based
systems (LUMI). Containerization is achieved using the Singularity
platform, with container images built via Singularity definition files and
optimized through multi-stage builds for lightweight and efficient images.
The model-specific software stack is managed and deployed within
the container using Spack, which enables the creation of customized
environments tailored to the model’s needs. We also highlight the
necessary compatibility conditions required for the containerized model
to fully leverage the underlying hardware. As a proof of concept, we
successfully containerize the RAPS bundle of the IFS global weather
model, developed by the European Centre for Medium-Range Weather
Forecasts (ECMWF), which utilizes hybrid parallelism (MPI+OpenMP)
and GPU acceleration. The containerized model is deployed and
benchmarked across multiple clusters, with performance comparisons
against native host installations. The results demonstrate that the
containerized model achieves performance on par with a host-based
execution, with no significant degradation, showcasing the scalability and
efficiency of the containerized approach for HPC environments.
