Parallel implementation of a data assimilation scheme for operational oceanography: The case of the MedBFM model system

Abstract

The MedBFM model system provides forecasts and reanalysis of the Mediterranean Sea biogeochemistry for the European Copernicus Services. The system integrates model and observations through a 3D variational assimilation scheme, whose performance capitalizes on present HPC systems and ensures compliance with the service requirements. Domain decomposition with message passing paradigm was implemented to parallelize the assimilation code and maximize performance and scalability. In particular, the parallelization of the horizontal recursive filtering algorithm was implemented using a dynamic sliced domain decomposition. Moreover, the efficient parallel solver of the PETSc/TAO library was adopted for optimizing the cost function minimization on which the variational assimilation is based. Considering the complex domain decomposition of the Mediterranean Sea and the high variability of the data input, a modern and scalable software application for variational assimilation schemes that exploits the performance of modern HPC architectures on the whole node was developed.