Wetenschappelijke publicatie

24 More Years of Numerical Weather Prediction: A Model Performance Model

Left panels: Achievable maximum equivalent number of grid points for the sss and eEg models; the lower panel (b1) is an enlargement of the period 2020 to 2032. Right panels: distribution of elapsed time over computations (physics and dynamics) and communications (local and global); the dashed lines split communication times into start-up time (lower area) and bandwidth-related (upper area); top panel for the sss, bottom panel for the eEg model. The dotted lines in panels (b1) and (a2) show the development in case the number of processors is kept constant, at the level of 2024, 2026, 2028 and 2030, resp., for the sss model; in panel (b1) they show the development of the dynamics line, hence the separation of computation and communication. In panel (a1) the historical ECMWF models and an ECMWF projection (T1279 in 2010) are also shown; the ECMWF model in 2006 was used to calibrate the sss line.
GJ Cats

For two formulations of currently usual numerical weather prediction models the evolution of maximum achievable resolution is predicted for the next 24 years. One of those formulations is the semi-implicit, semi-Lagrangian and spectral (sss) method as used eg by ECMWF. This formulation requires global communications. The other one is the explicit, Eulerian, grid point (eEg) formulation, requiring local communications only. The evolution prediction is based on an analysis of computation and communication requirements and on extrapolation of computation and communication performance of future computer systems. The main result is that model performance will continue to grow, approximately exponentially, for another two decades. After that, however, the finite speed of light will make global communications very inefficient. As a result, towards the end of our investigation period, sss models lose their current advantage, which derives from their lower compute operations count, and eEg models start to perform better, because they do not need global communications. The extrapolation of computer performance into the future requires quite a few assumptions on future configurations. Our results are not very sensitive to those assumptions, except if the configuration parameters are chosen to model a Grid computer. On a Grid computer, even models without global communication requirements show some performance loss due to the finiteness of the speed of light towards the end of the investigation eriod.

Bibliografische gegevens

GJ Cats. 24 More Years of Numerical Weather Prediction: A Model Performance Model
26--2008, pp0

Download volledige publicatie

download PDF (346 KB)
Niet gevonden wat u zocht? Zoek meer wetenschappelijke publicaties