Evaluating Machine-Independent Metrics for State-Space Exploration

Many recent advancements in testing concurrent programs can be described as novel optimization and heuristic techniques for exploring the tests of such programs. To empirically evaluate these techniques, researchers apply them on subject programs and capture a set of metrics that characterize the techniques' effectiveness. From a user's perspective, the most important metric is often the amount of real time required to find a error (if one exists), but using real time for comparison can be misleading because it is necessarily dependent on the machine configuration used for the experiments. On the other hand, using machine-independent metrics can be meaningless if they do not correlate highly with real time. As a result, it can be difficult to select metrics for valid comparisons among exploration techniques. This paper presents a study of the commonly used machine-independent metrics for two different exploration frameworks for Java (JPF and ReEx) by revisiting and extending a previous study (Parallel Randomized State-Space Search) and evaluating the correlation of the metrics with real time both on a single machine and on a high-performance cluster of machines. Our study provides new evidence for selecting metrics in future evaluations of exploration techniques by showing that several machine-independent metrics are a good substitute for real time, and that reporting real time results even from clusters can provide useful information.