An investigation of the applicability of design of experiments to software testing

Abstract

Approaches to software testing based on methods from the field of design of experiments have been advocated as a means of providing high coverage at relatively low cost. Tools to generate all pairs, or higher n-degree combinations, of input values have been developed and demonstrated in a few applications, but little empirical evidence is available to aid developers in evaluating the effectiveness of these tools for particular problems. We investigate error reports from two large open-source software projects, a browser and Web server, to provide preliminary answers to three questions: Is there a point of diminishing returns at which generating all n-degree combinations is nearly as effective as all n+1-degree combinations? What is the appropriate value of n for particular classes of software? Does this value differ for different types of software, and by how much? Our findings suggest that more than 95% of errors in the software studied would be detected by test cases that cover all 4-way combinations of values, and that the browser and server software were similar in the percentage of errors detectable by combinations of degree 2 through 6.