Improved prediction of non-functional properties in software product lines with domain context

Abstract

Software Product Lines (SPLs) enable software reuse by systematically managing commonalities and variability. Usually, commonalities and variability are expressed by features. Functional requirements of a software product are met by selecting appropriate features. However, selecting features also influences non-functional properties. To satisfy non-functional requirements of a software product, as well, the effect of a feature selection on non-functional properties has to be known. Often an SPL allows a vast number of valid products, which renders a test of non-functional properties on the basis of all valid products impractical. Recent research offers a solution to this problem: the effect of features on non-functional properties of software products is predicted by measuring in advance. A sample of feature configurations is generated, executed with a predefined benchmark, and then non-functional properties are measured. Based on these data a model is created that allows to predict non-functional properties of a software product before actually building it. However, in some domains contextual influences, such as input data, can heavily affect nonfunctional properties. We argue that the measurement of the effect of features on non-functional properties can be drastically improved by considering contextual influences of a domain. We study this assumption on input data as an example for a contextual influence and using an artificial but intuitive case study from the domain of compression algorithms. Our study shows that prediction accuracy of non-functional properties can be significantly improved.