Evaluating Interaction Patterns in Configurable Software Systems

Many modern software systems are designed to be highly configurable, which makes testing them a challenge. One popular approach is combinatorial configuration testing, which, given an interaction strength $t$, computes a set of configurations to test such that all $t $-way combinations of option settings appear at least once. Basically, this approach assumes that interactions are complete in the sense that any combination of $t$ options can interact and therefore must be tested. We conjecture, however, that in practical systems interactions are limited. If our conjecture is true, then new techniques might be developed to identify or approximate infeasible interactions, greatly reducing the number of configurations that must be tested. We evaluated this conjecture with an initial empirical study of several configurable software systems. In this study we used symbolic evaluation to analyze how the settings of run-time configuration options affected a test suite's line coverage. Our results strongly suggest that for these subject programs, test suites and configuration options, at least at the level of line coverage, interactions between configuration options are not complete