Parallelism in Constraint Programming

TL;DR: This thesis study and present the first constraint framework that automatically parallelizes both the consistency and the search of the solving process in constraint programming (CP), and shows that automatic parallelism in CP can provide very good performance.

Abstract: Writing efficient parallel programs is the biggest challenge of the software industry for the foreseeable future. We are currently in a time when parallel computers are the norm, not the exception. Soon, parallel processors will be standard even in cell phones. Without drastic changes in hardware development, all software must be parallelized to its fullest extent. Parallelism can increase performance and reduce power consumption at the same time. Many programs will execute faster on a dual-core processor than a single core processor running at twice the speed. Halving the speed of a processor can reduce the power consumption up to four times. Hence, parallelism gives more performance per unit of power to efficient programs. In order to make use of parallel hardware, we need to overcome the difficulties of parallel programming. To many programmers, it is easier to learn a handful of small domain-specific programming languages than to learn efficient parallel programming. The frameworks for these languages can then automatically parallelize the program. Automatically parallelizing traditional programs is usually much more difficult. In this thesis, we study and present parallelism in constraint programming (CP). We have developed the first constraint framework that automatically parallelizes both the consistency and the search of the solving process. This allows programmers to avoid the difficult issues of parallel programming. We also study distributed CP with independent agents and propose solutions to this problem. Our results show that automatic parallelism in CP can provide very good performance. Our parallel consistency scales very well for problems with many large constraints. We also manage to combine parallel consistency and parallel search with a performance increase. The communication and load-balancing schemes we developed increase the scalability of parallel search. Our model for distributed CP is orders of magnitude faster than traditional approaches. As far as we know, it is the first to solve standard benchmark scheduling problems.