Massively parallel constraint programming for supercomputers: Challenges and initial results

Abstract

In this paper we present initial results for implementing a constraint programming solver on a massively parallel supercomputer where coordination between processing elements is achieved through message passing. Previous work on message passing based constraint programming has been targeted towards clusters of computers (see [1,2] for some examples). Our target hardware platform is the IBM Blue Gene supercomputer. Blue Gene is designed to use a large number of relatively slow (800MHz) processors in order to achieve lower power consumption, compared to other supercomputing platforms. Blue Gene/P, the second generation of Blue Gene, can run continuously at 1 PFLOPS and can be scaled to 884,736-processors to achieve 3 PFLOPS performance. We present a dynamic scheme for allocating sub-problems to processors in a parallel, limited discrepancy tree search [3]. We evaluate this parallelization scheme on resource constrained project scheduling problems from PSPLIB [4]. © 2010 Springer-Verlag.