Bruce Nordman

Reducing the Energy Consumption of Ethernet with Adaptive Link Rate (ALR)

The rapidly increasing energy consumption by computing and communications equipment is a significant economic and environmental problem that needs to be addressed. Ethernet network interface controllers (NICs) in the US alone consume hundreds of millions of US dollars in electricity per year. Most Ethernet links are underutilized and link energy consumption can be reduced by operating at a lower data rate. In this paper, we investigate adaptive link rate (ALR) as a means of reducing the energy consumption of a typical Ethernet link by adaptively varying the link data rate in response to utilization. Policies to determine when to change the link data rate are studied. Simple policies that use output buffer queue length thresholds and fine-grain utilization monitoring are shown to be effective. A Markov model of a state-dependent service rate queue with rate transitions only at service completion is used to evaluate the performance of ALR with respect to the mean packet delay, the time spent in an energy-saving low link data rate, and the oscillation of link data rates. Simulation experiments using actual and synthetic traffic traces show that an Ethernet link with ALR can operate at a lower data rate for over 80 percent of the time, yielding significant energy savings with only a very small increase in packet delay.

IEEE 802.3az: the road to energy efficient ethernet

Ethernet is the dominant wireline communications technology for LANs with over 1 billion interfaces installed in the U.S. and over 3 billion worldwide. In 2006 the IEEE 802.3 Working Group started an effort to improve the energy efficiency of Ethernet. This effort became IEEE P802.3az Energy Efficient Ethernet (EEE) resulting in IEEE Std 802.3az-2010, which was approved September 30, 2010. EEE uses a Low Power Idle mode to reduce the energy consumption of a link when no packets are being sent. In this article, we describe the development of the EEE standard and how energy savings resulting from the adoption of EEE may exceed

Electricity used by office equipment and network equipment in the US

400 million per year in the U.S. alone (and over

Measuring building occupancy using existing network infrastructure

1 billion worldwide). We also present results from a simulation-based performance evaluation showing how packet coalescing can be used to improve the energy efficiency of EEE. Our results show that packet coalescing can significantly improve energy efficiency while keeping absolute packet delays to tolerable bounds. We are aware that coalescing may cause packet loss in downstream buffers, especially when using TCP/IP. We explore the effects of coalescing on TCP/IP flows with an ns-2 simulation, note that coalescing is already used to reduce packet processing load on the system CPU, and suggest open questions for future work. This article will help clarify what can be expected when EEE is deployed.

Data center power requirements: measurements from Silicon Valley

In spite of the recent explosive growth in the use of office and network equipment, there has been no recent study (until this one) that estimates in detail how much electricity is consumed by that equipment in the United States.

A Network Connection Proxy to Enable Hosts to Sleep and Save Energy

The primary focus of Green IT has been on reducing energy use of the IT infrastructure itself. Additional significant energy savings can be achieved by using the IT infrastructure to enable energy savings in both the IT and non-IT infrastructure. Our premise is that energy can be saved by driving building operation on information gleaned from existing IT infrastructure already installed for non-energy purposes. We call our idea implicit occupancy sensing where existing IT infrastructure can be used to replace and/or supplement traditional dedicated sensors to determine building occupancy. Our implicit sensing methods are largely based on monitoring MAC and IP addresses in routers and wireless access points, and then correlating these addresses to the occupancy of a building, zone, and/or room. Occupancy data can be used to control lighting, HVAC, and other building functions to improve building functionality and reduce energy use. We experimentally evaluate the feasibility of this dual-use of IT infrastructure and assess the accuracy of implicit sensing. Our findings, based on data collected from two facilities, show that there is significant promise in implicit sensing using the existing IT infrastructure present in most modern non-residential buildings.

Electricity used by office equipment and network equipment in the U.S.: Detailed report and appendices

Current estimates of data center power requirements are greatly overstated because they are based on criteria that incorporate oversized, redundant systems, and several safety factors. Furthermore, most estimates assume that data centers are filled to capacity. For the most part, these numbers are unsubstantiated. Although there are many estimates of the amount of electricity consumed by data centers, until this study, there were no publicly available measurements of power use. This paper examines some of the reasons why power requirements at data centers are overstated and adds actual measurements and the analysis of real-world data to the public policy debate over how much energy these facilities use.

Estimating the Energy Use and Efficiency Potential of U.S. Data Centers

Billions of dollars of electricity are being used to keep idle or unused network hosts fully powered-on only to maintain their network presence. We investigate how a network connectivity proxy (NCP) could enable significant energy savings by allowing idle hosts to enter a low-power sleep state and still maintain full network presence. An NCP must handle ARP, ICMP, DHCP, and other low-level network presence tasks for a network host. An NCP must also be able to maintain TCP connections and UDP data flows and to respond to application messages. The focus of this paper is on how TCP connections can be kept alive during periods of host sleep by using a SOCKS-based approach called green SOCKS (gSOCKS) as part of an NCP. The gSOCKS includes awareness of the power state of a host. A prototype implementation of gSOCKS in a Linksys router shows that TCP connections can be preserved.

Power management in networked devices

In spite of the recent explosive growth in the use of office and network equipment, there has been no recent study that estimates in detail how much electricity is consumed by that equipment in the United States. In this study, we examined energy use by office equipment and network equipment at the end of 1999. We classified office equipment into 11 types; for each type we estimated annual energy consumption for residential, commercial, and industrial use by combining estimates of stock, power requirements, usage, and saturation of power management. We also classified network equipment into six types and estimated the annual energy consumption for each type. We found that total direct power use by office and network equipment is about 74 TWh per year, which is about 2% of total electricity use in the U.S. When electricity used by telecommunications equipment and electronics manufacturing is included, that figure rises to 3% of all electricity use (Koomey 2000). More than 70% of the 74 TWh/year is dedicated to office equipment for commercial use. We also found that power management currently saves 23 TWh/year, and complete saturation and proper functioning of power management would achieve additional savings of 17 TWh/year. Furthermore, complete saturation of night shutdown for equipment not required to operate at night would reduce power use by an additional 7 TWh/year. Finally, we compared our current estimater with our 1995 forecast for 1999. We found that the total difference between our current estimate and the previous forecast is less than 15% and identified the factors that led to inaccuracies in the previous forecast. We also conducted a sensitivity analysis of the uncertainties in our current forecast and identified the data sets that have the largest impact on our current estimate of energy use.

Field surveys of office equipment operating patterns

Data centers are a significant and growing component of electricity demand in the United States. This paper presents a bottom-up model that can be used to estimate total data center electricity demand within a region as well as the potential electricity savings associated with energy efficiency improvements. The model is applied to estimate 2008 U.S. data center electricity demand and the technical potential for electricity savings associated with major measures for IT devices and infrastructure equipment. Results suggest that 2008 demand was approximately 69 billion kilowatt hours (1.8% of 2008 total U.S. electricity sales) and that it may be technically feasible to reduce this demand by up to 80% (to 13 billion kilowatt hours) through aggressive pursuit of energy efficiency measures. Measure-level savings estimates are provided, which shed light on the relative importance of different measures at the national level. Measures applied to servers are found to have the greatest contribution to potential savings.