Dynamic Algorithms for Energy Minimization on Parallel Machines

Abstract

Static DVS (dynamic voltage scaling) algorithms for DAG (directed acyclic graph) execution use the estimated execution time which is, in practice, an upper bound on the actual execution time to guarantee that an application completes in a given deadline. Therefore, many tasks may complete earlier than expected during the actual execution. This allows that the extra available slack can be allocated to tasks that have not yet begun execution with the goal of reducing the total energy requirements while still meeting the deadline constraints. In this paper, we present novel dynamic algorithms for reallocating the slack to future tasks. Experimental results show that our algorithms are comparable to static algorithms applied at runtime in terms of energy minimization, but require considerably smaller computational time.