Reducing the Power Consumption of Internet Datacenters
Ganesh Ananthanarayanan and Randy H. Katz
Traditionally, the problem of power consumption in datacenters has not been looked at with a great deal of seriousness. This was mainly because power consumption assumed relevance only when machines had a "battery" life, and since these were wall-powered, infinite power availability was assumed. Of late, startling observations have been made regarding the fraction of power consumed by the datacenters with respect to the total energy consumption of the Internet. This fraction is disproportionately large compared to the fraction of machines these datacenters constitute in the whole Internet. We investigate the power consumption of the various protocols and mechanisms used in a datacenter and aim to devise techniques that lower their power consumption. Specifically, we concentrate on the networking component of the datacenters.
A traffic-prediction (TP) module would be used to analyze the traffic patterns and predict the "low-activity" periods. This would then trigger a suite of power-saving techniques in the datacenters. These techniques primarily concentrate on reducing the number of idle networking components or switching to lower-powered (and low-bandwidth) alternatives during the "low-activity" periods.
We aim to collect data about the power consumption of the individual networking components in various modes under varying network loads, and develop a simple and effective power model for the datacenter depending on the number and type of the active components. Using typical workload traces, we evaluate the power savings achieved by our techniques.