Anticipated Dynamic Load Balancing Strategy to Parallelize Constraint Programming Search
Tarek Menouer, Bertrand Le Cun · 2013
This paper presents a parallelization of the Constraint Programming solver OR-Tools, using the parallel framework Bobpp. The principle of the parallelization is to assign one sequential OR-Tools solver per core and to communicate work or nodes using the Bobpp global priority queue. As usual the communication or migration of work between solvers is a kind of Work Stealing. But here we have to deal with a specific over cost. By using OR-Tools, it is possible to record the path from the root of the tree to a node so as to stop the search at a precise node. However, to start the search on a sub-tree, the entire path from the root of the main tree to the root of the sub-tree is to be replayed. This leads to additional costs to the parallel search. This paper presents new strategies using a Work Stealing technique to choose good nodes of the search-space and assigns each sub-tree during the execution of the algorithm in order to reduce the extra costs due to OR-Tools solver. These strategies are tested with problem modeled in OR-Tools for imbalanced search tree, despite the additional costs add by OR-Tools a speed-up of 8.04 must to be reached with N-Queens problem using 12 cores and speed-up of 37.73 with Golomb Ruler problem using 48 cores.