Inferring software behavioral models with MapReduce

This paper proposes a scalable, general approach to the inference of behavior models that can handle large execution logs via parallel and distributed algorithms implemented using the MapReduce programming model and executed on a cluster of interconnected execution nodes. The approach consists of two distributed phases that perform trace slicing and model synthesis. For each phase, a distributed algorithm using MapReduce is developed. With the parallel data processing capacity of MapReduce, the problem of inferring behavior models from large logs can be efficiently solved. The technique is implemented on top of Hadoop. Experiments on Amazon clusters show efficiency and scalability of our approach.