Distributed reconfiguration of multiprocessor systems

Abstract

We propose distributed algorithms for assigning tasks to processors in a distributed system, as well as algorithms for reassigning tasks from a failed processor to the remaining fault free processors in the system. The assignment and reassignment of the tasks should 1) make an efficient use of the capacity of the processors in the system, 2) distribute the load in a balanced way among the processors in the system, 3) assign tasks to processors for which they have a large preference (affinity), and 4) minimize the disturbance that is introduced by the migration of some tasks from one processor to another. We show that the distributed algorithms perform very closely to the centralized ones, even though they hold and maintain information only about processors in their local neighborhood.<>