Scalability Study of Two Weather Prediction Models

In this paper we perform an analysis of the performance and strong scalability of the HRM and COSMO mesoscale weather prediction models in order to emphasize the close liaison between the applications and the hardware used to run these models. Moreover, the paper tries to provide ways to tune the applications to achieve the best possible speedup and best utilization of the processing power involved. The study was conducted through the following steps: porting the applications on available cluster systems, several runs to establish applications scaling, profiling, interpretation of profiling results, tuning communication type and domain distribution over the computing nodes, and a comparison between the results obtained for different parameters. The study revealed that both the HRM and COSMO models scale very good to 16 to 64 processing elements on cluster systems, depending on the communication overhead involved in each model. Several improvements are also reported with the use of dedicated I/O processors or by adapting the MPI communication mode to particular problem specifications. The results are relevant to actual users of both these models and to other interested parties because it offers useful insights on the configuration of both hardware and software systems used in the production of real meteorological forecasts. Increased performance is thus essential for delivering the best possible forecast data within a given timeframe.