Scalable sequential data assimilation with the Parallel Data Assimilation Framework PDAF

Data assimilation applications with high-dimensional numerical modelsshow extreme requirements on computational resources. Thus, goodscalability of the assimilation system is necessary to make theseapplications feasible. Sequential data assimilation methods based onensemble forecasts, like ensemble-based Kalman filters, provide suchgood scalability, because the forecast of each ensemble member can beperformed independently. However, this parallelism has to be combinedwith the parallelization of both the numerical model and the data assimilation algorithm. In order to simplify the implementation ofscalable data assimilation systems based on existing numerical models,the Parallel Data Assimilation Framework PDAF has been developed. Itprovides support for parallel ensemble forecasts and parallelnumerical models. Further, it includes several optimized parallel filteralgorithms, like the ensemble transform Kalman filter. We will discussthe features and scalability of data assimilation systems based onPDAF on the example of data assimilation with the finite element oceanmodel FEOM.