Using Genetic Programming (GP) for classifying data streams is problematic as GP is slow compared with traditional single solution techniques. However, the availability of cheaper and better-performing distributed and parallel architectures make it possible to deal with complex problems previously hardly solved owing to the large amount of time necessary. This work presents a general framework based on a distributed GP ensemble algorithm for coping with different types of concept drift for the task of classification of large data streams. The framework is able to detect changes in a very efficient way using only a detection function based on the incoming unclassified data. Thus, only if a change is detected a distributed GP algorithm is performed in order to improve classification accuracy and this limits the overhead associated with the use of a population-based method. Real world data streams may present drifts of different types. The introduced detection function, based on the self-similarity fractal dimension, permits to cope in a very short time with the main types of different drifts, as demonstrated by the first experiments performed on some artificial datasets. Furthermore, having an adequate number of resources, distributed GP can handle very frequent concept drifts.