PAFL: Fault Localization via Noise Reduction on Coverage Vector

Coverage-based fault localization techniques assess the extent of how much a program entity relates to faults by contrasting the execution spectra of passed executions and failed executions. However, previous studies show that different test cases may generate similar or identical coverage information in program execution, which makes the execution spectra of program entities indistinguishable to one another, thus involves noise and decreases the effectiveness of existing techniques. In this paper, we use the concept of coverage vector to model program coverage in execution, compare coverage vectors to capture the similarity among test cases, reduce noise by removing similar coverage vector to refine the execution spectra, and based on them assess the suspicious basic blocks being related to fault. We thus narrow down the search region and facilitate fault localization. The empirical evaluation using Siemens programs and realistic UNIX utilities shows that our technique effectively addresses the problem caused by similar test cases and outperforms existing representative techniques. Keywords-fault localization; execution path; noise reduction