On the Scalability of Constraint Programming on Hierarchical Multiprocessor Systems

Abstract

Abstract-Recent developments in computer architecture progress towards systems with large core count, which expose more parallelism to applications, creating a hierarchical setup at the node and cluster levels. To take advantage of all this parallelism, applications must carefully exploit the different levels of the system which, if ignored, may yield surprising results. This aggravates the already difficult task of parallel programming.Declarative approaches such as those based on constraints are attractive to parallel programming because they concentrate on the logic of the problem. They have been successfully applied to hard problems, which usually involve searching through large problem spaces, a computationally intensive task but with potential for parallelization. Tree search algorithms play an important role in research areas such as constraint satisfaction or optimisation, and artificial intelligence. Tree search lends itself naturally to parallelization by exploiting different branches of the tree but scalability may be harder to achieve due to the high dynamic load balancing requirements.In this paper we present a high-level declarative approach based on constraints and show how it benefits from an efficient dynamic load balancing based on work stealing targeted at largescale. We focus on the implementation of a hierarchical work stealing scheme using a different programming model, GPI. Experimentation brought encouraging results on up to 512 cores on large instances of satisfaction and optimisation problems.

A review of literature on parallel constraint solving

“…This allows each solver to use a different search strategy, but no experiments were carried out doing this. Machado et al (2013) describe a distributed work-stealing approach built on top of MPI and pthreads, using a mix of local and global work pools. Results on the N-Queens enumeration problem with n = 17 show roughly linear speedups up to 512 processors, whilst results on a single quadratic assignment problem optimisation instance are similar.…”

Improving Constraint Solving on Parallel Hybrid Systems

“…Over the past couple of decades, several constraint solvers have been developed, and techniques to improve their performance have proliferated. Some of these are designed to make use of parallel architectures, such as multicore CPUs, and some are even capable of effectively running in distributed environments, as the Parallel Complete Constraint Solver (PaCCS) [15], [21]. PaCCS is a complete constraint solver which uses work stealing techniques to distribute the work among multi-threaded CPUs on a distributed environment.…”