Speeding up the ensemble data assimilation system of the limited‐area model of Météo‐France using a block Krylov algorithm

Abstract

Ensembles of data assimilations (EDA) in numerical weather prediction (NWP) are frequently used for both initialization of ensemble prediction systems and provision of background‐error statistics to a deterministic variational data assimilation scheme. The EDA consists of running multiple data assimilation schemes in parallel with perturbed observations and backgrounds. This kind of ensemble is computationally expensive, in particular because it requires the solution of as many linear systems as there are members in the ensemble.Recently, we proposed the use of block Krylov methods to take advantage of the EDA structure to solve all the linear systems simultaneously, allowing us to reduce the number of iterations needed for convergence. This approach is studied further in this article. We develop advanced parallelization strategies for the block Krylov method formulated in observation space using the Object‐Oriented Prediction System (OOPS) jointly developed by the European Centre for Medium‐Range Weather Forecasts, Météo‐France, and their partners.The algorithm is fully tested with the EDA of the limited‐area model AROME of Météo‐France, for different weather situations and configurations representative of the current system, namely assimilating around 104–105 observations together with 25 members. This new block Krylov approach allows us to gain from 13–45% of the computational time of the minimizations. Experiments with extended configurations that assimilate more observations and use more ensemble members confirm the potential of the algorithm in the future, with gains of up to 65%.