Key points are not available for this paper at this time.
When developing large-scale numerical earth system models traditionally knowledge of a broad range of programming technologies is required to support hardware from laptops up to supercomputers. A knowledge that scientists specialized in a particular geosciences domain mostly do not have and often do not want to acquire. Their emphasis is on describing and implementing the processes rather than for instance dealing with parallelization of model equations. Moreover, when model characteristics or domain extents change their chosen parallelisation technique may already be obsolete or require significant refactoring to adapt to the new situation. We develop the open-source LUE modelling framework, a software environment allowing domain scientists who may not be familiar with the development of high-performance applications to develop numerical simulation models that seamlessly scale when adding additional hardware resources. LUE comprises of a data model for the storage of field-based and agent-based data, and provides a broad range of map algebra operations as building blocks for model construction. Each spatial operation is implemented in C++ using HPX, a library and runtime environment providing asynchronous execution of interdependent tasks on both shared-memory and distributed computing systems. LUE provides a Python module and therefore a single high-level API whether models are run on laptops or HPC systems. In our presentation we demonstrate two capabilities of LUE. First, using the built-in operations we implemented a spatially distributed hydrological model including surface water routing. The model runs for the African continent at 100 metres spatial and hourly temporal resolution. Secondly, to demonstrate the extensibility we utilise LUEs focal operation framework to implement an individual kernel calculating greenness visibility exposure. Our PICO presentation will also include future extensions of the framework in particular for agent-based modelling and integration of machine learning model components.
Karssenberg et al. (Sat,) studied this question.