Performance analysis of TSMP2 on heterogeneous high-performance computing architectures
Regional Earth system modelling at kilometre-scale resolution increasingly depends on heterogeneous high-performance computing (HPC) systems to simulate tightly coupled interactions between the atmosphere, land surface, and subsurface processes. TSMP2, the second generation of the Terrestrial Systems Modelling Platform, addresses this challenge by coupling its component models through a Multiple Program Multiple Data (MPMD) approach that is explicitly designed for modern multi-core, many-core, and accelerator-based architectures. In our work, we assess the performance behaviour of TSMP2 across a broad spectrum of hardware configurations, ranging from homogeneous CPU-only systems to heterogeneous environments at the intra-chip, intra-node, and inter-node levels. The evaluation is conducted on several contemporary and emerging HPC platforms, including JURECA-DC, and the exascale JUPITER system. To ensure a systematic analysis, we employ an idealised experiment with homogeneous soil properties. Our analysis allows us to establish realistic performance baselines for each subsystem, identify hardware configurations that maximise efficiency for coupled simulations, and expose critical bottlenecks that arise in heterogeneous deployments. Overall, the experience and results underline the importance of emerging HPC architectures for regional Earth system modelling and provide practical guidance for the continued development of TSMP2 and similar coupled modelling frameworks. They also show that operating such models in the evolving HPC landscape is a non-trivial exercise and approaches such as this exercise will likely become more common.
Authors
- Jörg Benke (ORCID: https://orcid.org/0000-0003-2707-1629)
- Ana González‐Nicolás (ORCID: https://orcid.org/0000-0003-2869-8255)
- Daniel Caviedes‐Voullième (ORCID: https://orcid.org/0000-0001-7871-7544)
- Stefan Poll (ORCID: https://orcid.org/0000-0003-2527-0747)
- Paul Rigor (ORCID: https://orcid.org/0009-0003-5256-9215)
Institutions
- Forschungszentrum Jülich (DE)
- Jülich Supercomputing Centre (DE)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-16
- DOI
- https://doi.org/10.5281/zenodo.24274329
- Primary Topic
- Parallel Computing and Optimization Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00