The compilation of feature models into binary decision diagrams (BDDs) is a major challenge in the area of configurable systems analysis. Many large-scale feature models have been reported to exceed state-of-the-art compilation capabilities, e.g., for variants of the Linux kernel product line. However, experiments have been mainly conducted on standard settings of the BDD compilers themselves, not taking advanced configurations into account.
In this paper, we investigate the impact of various BDD compilation techniques for compiling feature models in conjunctive normal form. Specifically, we evaluate preprocessing techniques from satisfiability (SAT) solving, variable and clause ordering heuristics, non-incremental construction schemes, as well as parallelization of BDD construction. Our experiments on current feature models show that BDD compilation of feature models greatly benefits from these techniques, enabling to construct many previously not constructible large-scale feature models within seconds.