Branch and bound based load balancing for parallel applications

Abstract. Many parallel applications are highly dynamic in nature. In some, computation and communication patterns change gradually dur-ing the run; in others those characteristics change abruptly. Such dy-namic applications require an adaptive load balancing strategy. We are exploring an adaptive approach based on multi-partition object-based decomposition, supported by object migration. For many applications, relatively infrequent load balancing is needed. In these cases it becomes economical to spend considerable computation time toward arriving at a nearly optimal mapping of objects to processors. We present an optimal-seeking branch and bound based strategy that finds nearly optimal so-lutions to such load balancing problems quickly, and can continuously improve such solutions as time permits.