Energy-Efficient Data Center Networks

Abstract

Data center networks (DCNs) are designed to be scalable, resilient, and tolerant to failures. This is achieved through redundancy of network devices and links. All devices and links are always operational, consuming full energy even when underutilized. Energy-aware routing (EAR) protocols leverage network-wide information to dynamically scale up or down network energy consumption according to utilization. However, many EARs are designed for networks of devices and links that are not energy-proportional, such as ElasticTree and VMPlanner. The advent of Energy Efficient Ethernet (EEE) brings new challenges into how EAR protocols are designed. In this paper, we propose Greener, an extension of ElasticTree to leverage the energy proportionality characteristics of EEE. We integrate the EEE energy model with the ElasticTree solution. Greener is designed for K-ary Fat-Tree multi-rooted topologies, to steer flows along the most energy-efficient paths and put into sleep mode unused switches and links. Simulation experiments of Greener in DCNs of different sizes and under varying traffic loads show that it brings significant improvements in energy savings. Moreover, Greener outperforms the benchmark ElasticTree executed on EEE-based DCNs by up to 10 percentage points of energy savings.