Implementing YewPar: A Framework for Parallel Tree Search

Abstract

Combinatorial search is central to many applications yet hard to parallelise. We argue for improving the reuse of parallel searches, and present the design and implementation of a new parallel search framework. YewPar generalises search by abstracting search tree generation, and by providing algorithmic skeletons that support three search types, together with a set of search coordination strategies. The evaluation shows that the cost of YewPar generality is low (6.1%); global knowledge is inexpensively shared between workers; irregular tasks are effectively distributed; and YewPar delivers good runtimes, speedups and efficiency with up to 255 workers on 17 localities.