Wielding Statistical Fault Localization Statistically

Program debugging is a laborious but necessary phase of software development. It generally consists of fault localization, bug fix, and regression testing. Statistical software fault localization automates the manual and error-prone first task. It predicts fault locations by analyzing dynamic program spectrum captured in program runs. Previous studies mostly focused on how to provide reliable input data to such a technique and how to process the data accurately, but inadequately studied how to wield the output result of such a technique. In this work, we raise the assumption of symmetric distribution on the effectiveness of such a technique in locating faults, based on empirical results. We use maximum likelihood estimate and linear programming to develop a tuning method to enhance the result of a statistical fault localization technique. Experiments with two representative such techniques on two realistic UNIX utility programs validate our assumption and show our method effective.