Reasoning on Feature Models: Compilation-Based vs. Direct Approaches

Abstract

Analyzing a Feature Model (FM) and reasoning on the corresponding configuration space is a central task in Software Product Line (SPL) engineering. Problems such as deciding the satisfiability of the FM and eliminating inconsistent parts of the FM have been well resolved by translating the FM into a conjunctive normal form (CNF) formula, and then feeding the CNF to a SAT solver. However, this approach has some limits for other important reasoning issues about the FM, such as counting or enumerating configurations. Two mainstream approaches have been investigated in this direction: (i) direct approaches, using tools based on the CNF representation of the FM at hand, or (ii) compilation-based approaches, where the CNF representation of the FM has first been translated into another representation for which the reasoning queries are easier to address. Our contribution is twofold. First, we evaluate how both approaches compare when dealing with common reasoning operations on FM, namely counting configurations, pointing out one or several configurations, sampling configurations, and finding optimal configurations regarding a utility function. Our experimental results show that the compilation-based is efficient enough to possibly compete with the direct approaches and that the cost of translation (i.e., the compilation time) can be balanced when addressing sufficiently many complex reasoning operations on large configuration spaces. Second, we provide a Java-based automated reasoner that supports these operations for both approaches, thus eliminating the burden of selecting the appropriate tool and approach depending on the operation one wants to perform.

On the benefits of knowledge compilation for feature-model analyses

“…However, their work is each limited to a single feature-model analysis and target language while we consider a large variety of target languages and analyses. Bourhis et al [40] perform several analyses on feature models using d-DNNFs. However, they do not consider any other target language.…”

“…However, many capabilities of well-known knowledge-compilation target languages [25] have not been utilized yet for feature-model analyses. For instance, while BDDs have been employed repeatedly [21,[35][36][37] and d-DNNFs found some adoption in recent literature [22,[38][39][40], target languages that are popular in other domains, such as SDDs [41], have not been employed at all. Finally, regarding compiler scalability, knowledge compilation can only be sensibly applied for feature-model analyses if the compilation of feature models to the respective artifact scales appropriately.…”

“…If the compilation takes too long or does not scale at all, the benefits (i.e., efficient queries) of the knowledge-compilation artifact vanish. However, besides preliminary work on some target languages [21,22,39,40,42], it is still largely unknown which artifacts can be compiled from feature models in reasonable time.…”