A workflow for cloud-based and HPC simulations with the NEMO ocean model using containers

In this work, we explored the deployment and execution of the NEMO ocean model using Singularity containers within the EDITO Model Lab, implementing the European Digital Twin of the Ocean. The Auto-NEMO workflow, a fork of Auto-EC-Earth used to run NEMO workflows using the NEMO Community reference code, was adapted to run simulations using containers. The use of a Singularity container ensures consistent execution by packaging all dependencies, making it easier to deploy the model across various HPC systems.

The containerized approach was tested on multiple HPC platforms, including MareNostrum5 and LUMI, to evaluate scaling performance. Our tests compared the use of mpich and openmp libraries, providing insights into how communication strategies impact the computational performance of the model in containerized setups. In addition, the runs are orchestrated by a content workflow manager, in this case Autosubmit, deployed in a cloud infrastructure in EDITO-Infra, making the entire solution (workflow manager and workflow itself) portable end-to-end. The benefits of portability and reproducibility make containers an attractive solution for streamlining workflows in diverse computational environments.

A comparison between containerized and non-containerized runs highlights the trade-offs involved. Direct execution may provide slightly better performance in some cases, but the containerized approach greatly reduces setup complexity. These findings demonstrate the potential of containerization to enhance efficiency and accessibility in large-scale ocean modeling efforts.