Statistical fault localization in decision support system based on probability distribution criterion

Finding the location of a fault in code is an important research and practical problem, which often requires much time and manual effort. To automate this time consuming task, a class of predicate-based statistical fault localization techniques have been proposed, which test the similarity of dynamic predicate spectra between non-failed runs and failed runs and suggest suspicious predicates to the programmers to facilitate the identification of faults. However, with the existence of coincidental correctness, how to efficiently and effectively compare the difference of predicate spectra distribution has become a crucial problem to be solved. In this paper, we make use of probability distribution criterion in developing a new statistical fault localization algorithm. Instead of using geometry distance, it calculates the overlapping of dynamic predicate spectra in two communities (non-failed runs and failed runs) to evaluate the difference. Empirical results show that our technique outperforms some representative predicate-based fault localization techniques for localizing faults in most subject programs of the Siemens suite and space program. To facilitate the debugging process and provide visual help to the debugger, we also designed a system software prototype, which integrates many recent fault localization algorithms, including the one proposed in this paper.