Addressing high dimensionality and lack of feature models in testing of software product lines

Abstract

Software Product Lines (SPLs) allow engineers to systematically build families of software products, defined by a unique combination of features—increments in functionality, improving both the efficiency of the software development process and the quality of the software developed. However, testing these kinds of systems is challenging, as it may require running each test against a combinatorial number of products. We call this problem the High Dimensionality Problem. Another obstacle to product line testing is the absence of Feature Models (FMs), making it difficult to discover the real causes for test failures. We call this problem the Lack of Feature Model Problem. The High Dimensionality Problem is associated to the large space of possible configurations that an SPL can reach. If an SPL has n boolean features, for example, there are 2n possible feature combinations. Therefore, systematically testing this kind of system may require running each test against all those combinations, in the worst case. The Lack of Feature Model Problem is related to the absence of feature models. The FM enables accurate categorization of failing tests as failures of programs or the tests themselves, not as failures due to inconsistent combinations of features. For this reason, the lack of FM presents a huge challenge to discover the true causes for test failures. Aiming to solve these problems, we propose two lightweight techniques: SPLat and SPLif. SPLat is a new approach to dynamically prune irrelevant configurations: the configurations to run for a test can be determined during test execution by monitoring accesses to configuration variables. As a result, SPLat reduces the number of configurations. Consequently, SPLat is lightweight compared to prior works that used static analysis and heavyweight dynamic execution. SPLif is a technique for testing SPLs that does not require a priori availability of feature models. Our insight is to use a profile of passing and failing test runs to quickly identify test failures that are indicative of a problem (in test or code) as opposed to a manifestation of execution against an inconsistent combination of features. Experimental results show that SPLat effectively identifies relevant configurations with a low overhead. We also apply SPLat on two large configurable systems (Groupon and GCC), and it scaled without much engineering effort. Experimental results demonstrate that SPLif is useful and effective to quickly find tests that fail on consistent configurations, regardless of how complete the FMs are. Furthermore, we evaluated SPLif on one large extensively tested configurable system, GCC, where it helped to reveal 5 new bugs, 3 of which have been fixed after our bug reports.