Calibrating a hydrologic model against ET and TWS observation at the global scaling using a state-of-the-art parallelization scheme

Abstract:

Parameter estimation for global-scale, high-resolution hydrological models is and will be an important challenge. Within this topic, multiple observational data sets are available for different components of the water cycle. These cover spatially continuous (e.g. evapotranspiration (ET) or total water storage (TWS)) and discrete (e.g. river discharge) time series data. Multi-objective functions aggregate the residuals of modelled variables against those reference data. Efficient algorithms with hybrid parallelization schemes are required for an improved description of the global hydrology.

The parameters of the mesoscale Hydrologic Model (mHM) are calibrated against FLUXNET products for ET, GRACE data for TWS anomalies and a subset of 5500 river discharge data from the GRDC database. As there are no horizontal exchanges needed for the simulation of ET or TWS, a partitioning of the global domain into 258 independent, almost equally sized subdomains is employed. The parallelization of the routing of river discharge for large basins is done with the MPI parallelized decomposition of forests (MDF) algorithm (1). The partial objective functions for each variable are combined to a global multi-objective function. We run the model on the high-performance supercomputer JUWELS at Jülich Supercomputing Center (JSC) where the code efficiently scales on more than ~8 000 CPUs.

We show the dependency of the scaling efficiency on different parallelization components and parameters. Furthermore, we provide results on the convergence of the global calibration with regard to different regions and partial objective functions. We also investigate the sensitivity of the results influenced by the number and type of sampling of river domains.