| Abstract: |
|
To efficiently use the emerging massively parallel computer architectures, many aspects of atmsopheric or coupled atmosphere-ocean models have to be optimized or even re-develolped on the basis of new conceptual ideas. The most obvious requirements to our model codes are (a) the minimziation of communication, particularly global communication, (b) memory scalability, implying that the allocation of global fields needs to be avoided as far as possible, and where indispensable, restricted to 2D horizontal slices, and (c) algorithmic optimization attempting to minimize the data traffic between memory and the computing cores. Because climate applications typically have smaller grids but much longer integration periods than numerical weather prediction (NWP) applications, efficiency in the strong scaling limit is even more important for climate modeling than for NWP, whereas memory scaling is more critical for NWP applications. For reason (a), there is a general trend to move from spectral models towards grid-point models with horizontally explicit time-stepping schemes. Moreover, there is a trend towards horizontally unstructured grids particularly for global models becuase of their larger flexibility. From an HPC point of view, unstructured grids have the disadvantage of needing indirect addressing for horizontal stencil operations, which usually is computationally more expensive than direct addressing, but the advantage that the halo points for MPI communication can easily be collected at the end of the index vector, which greatly improves cache efficiency in the strong scaling limit. In this presentation, all these aspects will be discussed in the light of the development of the ICON (ICOsahedral Nonhydrostatic Model), which is a joint project of Deutscher Wetterdienst (DWD) and Max-Planck-Institute for Meteorology (MPI-M) to develop a global nonhydrostatic coupled atmosphere-ocean model for numerical weather prediction and climate modeling. |
|
