Agile Traffic Merging for DCNs

Abstract

Abstract. Data center networks (DCNs) have been growing in size and their power consumption is becoming a matter of concern. Many recent papers, including ElasticTree and CARPO, propose new nearenergy-proportional DCNs, aiming at reducing the power consumption by dynamically powering off idle network switches and links. In this paper, we examine the power optimization model for DCNs, and present a scalable heuristic algorithm that finds a near-optimal subset of network switches and links that satisfies a given traffic load and consumes minimal power. Furthermore, we apply merge networks to each switch in order to power off the idle interfaces of the active switches, thus further reducing the energy consumption of active switches and achieving greater energy savings than ElasticTree. We finish by simulating large-scale fattree DCNs and comparing the energy cost of our techniques versus the ElasticTree method. The results demonstrate that our solution is more energy-efficient.