Details

Autor(en) / Beteiligte

• Xie, Feng
• Davenport, Andrew

Titel

Massively Parallel Constraint Programming for Supercomputers: Challenges and Initial Results

Ist Teil von

• Integration of AI and OR Techniques in Constraint Programming for Combinatorial Optimization Problems, p.334-338

Ort / Verlag

Berlin, Heidelberg: Springer Berlin Heidelberg

Link zum Volltext

Quelle

Alma/SFX Local Collection

Beschreibungen/Notizen

• 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].