A BDD for Linux?

Abstract

What is the number of valid configurations for Linux? How to generate uniform random samples for Linux? Can we create a binary decision diagram for Linux? It seems that the product-line community tries hard to answer such questions for Linux and other configurable systems. However, attempts are often not published due to the publication bias (i.e., unsuccessful attempts are not published). As a consequence, researchers keep trying by potentially spending redundant effort. The goal of this challenge is to guide research on these computationally complex problems and to foster the exchange between researchers and practitioners.

A Monte Carlo tree search conceptual framework for feature model analyses

“…SAT solvers could face scalability problems for large-scale models (Liang et al, 2015;Mendonca et al, 2009). Some statistical analyses require the construction of BDDs (e.g., determining the distribution of the number of features among all valid configurations or testing the uniformity of a random sampler on complex models with thousands of features) (Heradio et al, 2019(Heradio et al, , 2022a, which can be intractable (Thüm, 2020). Other approaches like genetic algorithms (Lopez-Herrejon et al, 2015b,a) and metaheuristics (Yadav et al, 2020) require to incorporate specific domain knowledge, and analyzing and inferring results from the final solutions which is not straightforward.…”

“…Given a partial configuration, the BDD solver returns a sample of complete configurations that includes the features of the provided partial configuration. However, the BDD solver also presents scalability issues regarding memory when dealing with large-scale feature models (Thüm, 2020), and therefore, it limits the applicability of our MCTS framework to those feature models whose associated BDD can be built. OC7: Providing uniform random sampling for large-scale feature models is actually one of the open challenges in the SPL community (Pett et al, 2019).…”

“…In highly configurable systems, AAFM is a challenging task because it requires coping with a variety of problems which involve inter-related features and colossal search spaces. For example, we find multiple operations, such as counting the number of products (Fernández-Amorós et al, 2014;Heradio et al, 2016;Thüm, 2020) and optimizing configurations (Guo et al, 2019;Nair et al, 2017;Karimpour and Ruhe, 2017), which are performed on the entire configuration space. Other operations are performed on feature models (e.g., evolution (Marques et al, 2019) and reverse engineering (Lopez-Herrejon et al, 2015b; ✩ Editor: Raffaela Mirandola.…”

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

“…[44,3,6,45]. Unfortunately, OBDD has proved not to scale well when dealing with large configuration spaces such as the one of Linux [5] or the ones related to highlyconfigurable systems [6]. Despite these unappealing results, the compilation-based approach looks as well-suited for performing reasoning queries on large FM.…”

“…More precisely, a given representation may be more suitable than other ones regarding a certain analysis. For instance, checking the satisfiability of an Ordered Binary Decision Diagram (OBDD) [4] is an operation with constant effort [5], while there is no polynomial-time algorithm for addressing the same query when the FM is represented as a CNF formula (and it is likely that no such algorithm may exist, since it would show that P = NP). Yet, OBDD has proved not to scale well when dealing with large configuration spaces such as the one of Linux [5] or the ones related to highly-configurable systems [6].…”