Simple approaches to parallel Branch and Bound

Abstract

The Branch and Bound algorithm has been efficiently parallelized by several researchers, but usually on the expense of producing rather complicated algorithms with respect to communication management. This paper describes two new approaches to parallel Branch and Bound. The first approach employs static distribution of subproblems, and does not allow processors to communicate during the search. A work function is devised to enable initial construction of subproblem pools of equal difficulty. The algorithm performs surprisingly well, exhibiting an average processor utilization of about 85%. The second approach is an ordinary parallel Branch and Bound algorithm using dynamic distribution of subproblems, but communication is restricted to be synchronized. Processor utilization up to 98% have been measured for this algorithm.