This paper introduces two Mixed Integer Linear Programming (MILP) models for railway traffic planning using a cumulative scheduling constraint and associated pre-processing filters. We compare standard solver performance for these models on three sets of problems from the railway domain and for two of them, where tasks have unitary resource consumption, we also compare them with two more conventional models. In the experiments, the solver performance of one of the cumulative models is clearly the best and is also shown to scale very well for a large scale practical railway scheduling problem.