Accelerating Marine Biogeochemistry Modelling on GPUs: The OGSTM-BFM Case Study

Marine Biogeochemical models developed in the context of environmental monitoring are becoming increasingly complex, as they are required to represent a growing number of ecosystem indicators. Modern modelling frameworks aim not only to quantify biogeochemical cycles and primary productivity, but also to investigate ecosystem biodiversity, interactions with higher trophic levels, and the impacts of anthropogenic pressures such as fishing. This continuous increase in model complexity poses significant computational challenges and calls for substantial upgrades of existing codes to fully exploit modern high-performance computing (HPC) platforms.

The MedBFM model system is based on the coupled physical–biogeochemical OGSTM-BFM framework. The physical transport component, the OGS Tracer Model (OGSTM), is based on the OPA 8.1 system and is developed at the National Institute of Oceanography and Applied Geophysics (OGS). The biogeochemical component is maintained by the Biogeochemical Flux Model (BFM) Consortium.

MedBFM is operated daily within the Copernicus Marine Service, providing essential information on Mediterranean plankton dynamics to support monitoring of biomass productivity, plankton diversity, carbon sequestration, and ocean acidification.

The OGSTM-BFM model is also employed within the EU Horizon project New Copernicus Capabilities for Trophic Ocean Networks (NECCTON), which unites major European institutes involved in marine forecasting in support of the Copernicus Marine Service. 

In addition, the same modelling framework has been used for long-term simulations, including centennial-scale scenario simulations extending up to the year 2100.

In this work, we present performance improvements to OGSTM-BFM achieved through a two-year collaboration within the ESiWACE initiative. ESiWACE3 supports exascale readiness for the European weather and climate modelling community by providing short- and long-term services aimed at improving model performance and facilitating knowledge transfer.

The main activities carried out during the collaboration focused on:

• completing the porting of the OGSTM horizontal and vertical diffusion schemes to NVIDIA GPUs;
  
  
• porting the complex carbonate system solver of the BFM to GPUs;
  
  
• assessing and optimizing the overall application performance, including the porting of critical code sections, kernels tuning, and improvements to data locality.

The current computational burden of the implementation corresponds to a problem size of approximately 2.6 billion computational elements, accounting for both spatial resolution and biogeochemical complexity. Performance tests conducted on the Leonardo system show a speedup of 7.41 using eight NVIDIA A100 GPUs compared to eight Intel Sapphire Rapids CPUs (112 cores each). This substantial acceleration makes long-term simulations more feasible, while leaving adequate time for data analysis and sensitivity studies. Moreover, enabling GPU support opens the way for efficient deployment of the model on current and future exascale computing platforms, and on-demand simulations within the Digital Twin of the Ocean.