Differencing labeled transition systems

Abstract

Concurrent programs often use Labeled Transition Systems (LTSs) as their operational semantic models, which provide the basis for automatic system analysis and verification. System behaviors (generated from the operational semantics) evolve as programs evolve for fixing bugs or implementing new user requirements. Even when a program remains unchanged, its LTS models explored by a model checker or analyzer may be different due to the application of different exploration methods. In this paper, we introduce a novel approach (named SpecDiff) to computing the differences between two LTSs, representing the evolving behaviors of a concurrent program. SpecDiff considers LTSs as Typed Attributed Graphs (TAGs), in which states and transitions are encoded in finite dimensional vector spaces. It then computes a maximum common subgraph of two TAGs, which represents an optimal matching of states and transitions between two evolving LTSs of the concurrent program. SpecDiff has been implemented in our home grown model checker framework PAT. Our evaluation demonstrates that SpecDiff can assist in debugging system faults, understanding the impacts of state reduction techniques, and revealing system change patterns.