Reusable Object-Oriented Parallelization of Branch-and-Bound Algorithms

In this paper we present a framework for branch-and-bound algorithms to separate problem specific logic and branch-and-bound implementation from the core branch-and-bound algorithm. By isolating the problem-specific logic and paradigm-specific branch-and-bound implementation via two instances of the Template Method design pattern, we simplify parallelizing branch-and-bound algorithms and allow multiple execution paradigms. Specifically, our framework includes multiple branch-and-bound algorithm implementations that can execute problem specific logic sequentially, in parallel via distribution of each node in the search tree (e.g., a single ’task’), or in parallel via execution of multiple ’tasks’ before distribution. Each implementation paradigm can employ the same problem-specific operations. Problematically, branch-and-bound algorithms are difficult to parallelize due to both the fine-grained nature of each ’task,’ and the uneven distribution in the search tree requiring re-balancing of workloads. Our approach allows low-effort changes between parallelization strategies that can operate on the same problem-specific implementations enabling significant speedups on a variety of problems. Our proof-of-concept framework is assessed across a variety of branch-and-bound algorithms for different problems to demonstrate its effectiveness.