Dominik Schreiber, Christoph Jabs, Jeremias Berg

From Scalable SAT to MaxSAT: Massively Parallel Solution Improving Search

Introduction

From scalable sat to maxsat: massively parallel solution improving search. Explore a new massively parallel MaxSAT solver. It scales to hundreds of cores, outperforming state-of-the-art SIS-based solvers and setting a new benchmark for scalable MaxSAT.

Abstract

Maximum Satisfiability (MaxSAT) is an essential framework for combinatorial optimization at the core of automated reasoning. However, to date, no notable parallelizations with convincing scaling behaviour exist. We suggest to exploit and transfer recent advances in massively parallel SAT solving to perform scalable solution improving search (SIS) for MaxSAT solving. Building upon the distributed job scheduling and SAT solving platform Mallob, we present the first MaxSAT solver that scales to hundreds of cores through a careful combination of parallel and distributed incremental SAT solving, task parallelism and flexible load balancing, and clause sharing within and across SAT solving tasks. Experiments on up to 768 cores (16 nodes) show that our approach clearly outscales state-of-the-art SIS-based MaxSAT solvers, marking a new baseline for parallel MaxSAT solving.

Review

The paper "From Scalable SAT to MaxSAT: Massively Parallel Solution Improving Search" addresses a critical gap in automated reasoning and combinatorial optimization: the absence of a truly scalable parallel Maximum Satisfiability (MaxSAT) solver. MaxSAT, while a foundational framework, has historically lacked parallelizations that exhibit convincing scaling behavior. The authors propose an innovative approach by leveraging recent advancements in massively parallel SAT solving and adapting them for solution improving search (SIS) within the MaxSAT domain. Their core contribution is the development of a novel MaxSAT solver, built upon the distributed job scheduling and SAT platform Mallob, which reportedly achieves unprecedented scalability. The technical strength of this work lies in its "careful combination" of several sophisticated parallelization techniques. Specifically, the abstract highlights the integration of parallel and distributed incremental SAT solving, task parallelism, flexible load balancing, and both internal and external clause sharing across SAT solving tasks. This multifaceted strategy appears to be the key to overcoming the inherent challenges of parallelizing MaxSAT SIS. The experimental evaluation, conducted on a significant scale of up to 768 cores across 16 nodes, provides strong evidence for the solver's superior performance, demonstrating that it "clearly outscales state-of-the-art SIS-based MaxSAT solvers." This marks a significant milestone, establishing a new baseline for parallel MaxSAT solving. This research represents a substantial advancement in the field of automated reasoning and optimization, with the potential to unlock the solution of previously intractable MaxSAT problems. By demonstrating the feasibility of massively parallel MaxSAT, the authors have not only set a new performance benchmark but also opened new avenues for research into the parallelization of other complex combinatorial problems. The established "new baseline" implies a paradigm shift, encouraging future work to build upon these foundational techniques and explore their applicability to an even broader range of MaxSAT instances and real-world applications. The approach's robustness and scalability could significantly accelerate progress in areas reliant on MaxSAT, from AI planning to circuit design and scheduling.

Full Text

Comments

You need to be logged in to post a comment.