On the uniformity of error propagation in software

Abstract

This paper presents an empirical study of an important aspect of software defect behavior: the propagation of data-state errors. A data-state error occurs when a fault is executed and affects a program's data-state, and it is said to propagate if it affects the outcome of the execution. Our results show that data-state errors appear to have a property that is quite useful when simulating faulty code: for a given input, it appears that either all data state errors injected at a given location tend to propagate to the output, or else none of them do. These results are interesting because of what they indicate about the behavior of data-state errors in software. They suggest that data state errors behave in an orderly way, and that the behavior of software may not be as unpredictable as it could theoretically be. Additionally, if all faults behave the same for a given input and a given location, then one can use simulation to get a good picture of how faults behave, regardless of whether the faults one has simulated are representative of real faults.