Performance modeling for scientific production codes is of interest both for program tuning and for the selection of new machines. An empirical method is used for developing a model to predict the runtime of large parallel weather prediction programs. The different steps of the method are outlined giving examples from two different programs of the DWD (German Weather Service). The first one is the new local model (LM) of DWD. The second one is the old Deutschland Model (DM) which is characterized by more complicated algorithmic steps and parallelization strategies.