Preparing a weather prediction and regional climate model
for current and emerging hardware architectures
Oliver Fuhrer
Being able to effiiciently leverage
the available compute power is the most important constraint limiting
the horizontal resolution, the complexity of the model system, and the
number of ensemble members of numerical weather prediction and regional
climate models. In order to leverage future supercomputers, which will
tend to have an increasing number of compute cores with reduced memory
access speed per core, current climate and weather prediction codes
will have to be adapted. In this presentation we will discuss general
characteristics of codes using finite difference or finite volume
methods on structured grids and their implications on the efficiency of
the algorithms on emerging computing hardware. Using the concrete
example of the dynamical core of the COSMO (Consortium for Small Scale
Modeling) model, an effort where the application code, numerical
libraries, and the computing system are co-designed will be discussed.