Industrial environments that make use of Model-Driven Engineering (MDE) are starting to see the appearance of very large models, made by millions of elements. Such models are produced automatically (e.g., by reverse engineering complex systems) or manually by a large number of users (e.g., from social networks). The success of MDE in these application scenarios strongly depends on the scalability of model manipulation tools. While parallelization is one of the traditional ways of making computation systems scalable, developing parallel model transformations in a general-purpose language is a complex and error-prone task. In this paper we show that rule-based languages like ATL have strong parallelization properties. Transformations can be developed without taking into account concurrency concerns, and a transformation engine can automatically parallelize execution. We describe the implementation of a parallel transformation engine for the current version of the ATL language and experimentally evaluate the consequent gain in scalability.