Energy Impacts of Wired and Wireless Networks

Transcription

1 Energy mpacts of Wired and Wireless Networks H. Scott Matthews, Chris T. Hendrickson, Hui Min Chong, and Woon Sien Loh Green Design nitiative, Camegie Mellon University Pittsburgh, PA USA Contact: Abstract Many commercial and residential buildings now have wired telecommunications networks. An emerging trend is to substitute wired with wireless networks, which have generally lower bandwidth but are easier to setup and manage. n order to compare the relative impacts of these two types of computer networks, we consider a case stu& of Carnegie Mellon University s campus network, which includes ubiquitous wired and wireless networks. We Jind that the network inji-astructure alone consumes 6% of the campus electricity load. Further, while there is some diflerence in network performance between the two types of networks (and are thus not completely equivalent), the wireless network consumes considerably less energy. Since a college campus (especially a highly computer intensive one like the one studied) is not representative of all commercial, industrial, or residential buildings, these results are not generalizable, but are still useful in understanding some of the components of the 3% of all electrical load that is attributable to information technology in the US.. However, there are still important points to be made associated with the relative energy eficiency of wireless networking that can aid future deployment and policy issues.. Keywords Electricity, network, wireless, infrastructure NTRODUCTON Wired computer networks first began to appear in the late 1960s in the United States through the development of ARPANET and related projects [l]. By the 1990s, the industrialized world was heavily networked via copper and fiber wires. Now, wireless computer networking is becoming popular, with the wireless protocols leading to a proliferation of TCP/P-based business and home solutions [2]. Wired networks in use today operate at a variety of speeds or bandwidths. The most common type of wired network is via Ethernet. Ethernet networks can operate at 10, 100 or even 1000 Megabits per second for typical business applications. Wireless computer networks generally operate at much lower speeds. Some of the more popular wireless networking options are Bluetooth (radio), rda (infrared), HomeRF, and WiFi (EEE b). These standards generally operate in the less than 11 Megabits per second range, although nextgeneration la devices will be about 5 times faster. Several existing studies have estimated the total electrical load of information technology-based products at about 3% of total U.S. demand [3,4,5]. One of the key uncertainties in these assessments is the electricity used by network equipment, and this study provides real data on such electricity use for one highly computerized campus in Pittsburgh, PA. Electricity generation is one of the largest national sources of many pollutants, especially sulfur dioxide, nitrogen oxides, carbon dioxide, and particulates. Further, the manufacture of computers and related equipment consumes substantial amounts of energy and involves significant amounts of toxic materials, although improvements in design, energy efficiency, and manufacturing processes are being made. MOTVATON This research project was motivated by two related initiatives on our campus. First, a contract was signed in 2001 to purchase wind power equal to 5% of the main campus electricity demand. This wind power is produced in close proximity to campus and is one of several green initiatives on campus. This amount of generated wind power, compared to the predominantly fossil-fuel fired electricity of the northeastern United States, was estimated to prevent the release of 13 tons of nitrogen oxides (a precursor to ozone), 35 tons of sulfur dioxide, and 5,100 tons of carbon dioxide. Given the current economics of renewable energy production, there is cost premium for buying wind power compared to average grid electricity. n this case, the wind power extra cost was roughly $80,000. The second related initiative is the Carnegie Mellon University Challenge. This challenge is a coordinated effort to reduce - via conservation and efficiency projects - campus electricity demand by the added cost of the wind energy. At an average cost of about 5 cents per kilowatt-hour (kwh), the University Challenge target is 1,600,000 kwh per year. n order to meet this target, an inventory of major end-use electricity was needed. This inventory sought to estimate electricity use from areas like lighting, computers, office machines, etc. The goal of this research project was to focus on a specific site and estimate the energy consumption of the network and computing infrastructure (NC) needed to X/02/$ EEE 44

2 Core Building Wiring Office Network Network/ 4- External Closets (BWC) Equipment (ONE) & Wired Server Devices ~nternet +t) nfra- Wired Switches Connection structure Wireless * 1 (CNS) Transceivers Wireless Wireless (WT) Devices Cwi trhpf System Boundary Figure 1: Summary of Network and Computing nfrastructure (NC) on Campus use the network. The campus of Carnegie Mellon University was selected because of its proximity to the researchers, its concern for energy-related issues, and the existence of ubiquitous wired and wireless networks across campus. This is an interesting issue because NC electricity use is continuous, i.e hours per year. As shown in Figure 1, the network and computing infrastructure on campus was defined to include the core network and server infrastructure (CNS) in the main campus computing center building Cyert Hall, all building wiring closets (BWC), all office-level networking equipment (ONE), e.g. 4-port Ethernet hubs, and wireless networking equipment. Note that this estimate will not include any use of personal computers, printers, or other peripheral devices that make use of network services. While consuming considerable amounts of electricity, these items are assumed to be non-infrastructure related devices. Electricity inventory for this equipment was done separately as part of a broader study [6]. n this project, the campus has two particular types of network devices: hubs and Switches. While similar, they have a critical functionality difference. Both connect devices to a network. However the difference is how they pass on the traffic received to items connected to the hub or switch. A hub passes on all incoming network traffic to all connected devices. A switch will only pass on network traffic to the desired destination device connected to it. For example, if 4 computers are connected to a hub, any incoming data stream would be sent to all 4 computers. A switch would only send the data stream to the computer it is being sent to. This has network management benefits (unnecessary traffic does not clutter networks) but switches generally use more electricity than hubs. To complete the task of estimating total NC electricity use, gathering a large amount of end-use data was necessary. Fortunately, the campus has recently completed the first stage of installing industrial power meters in all buildings on campus. Data from these meters shows the aggregate building-by-building use of electricity on campus. For determining NC electricity use, these values were generally not helpful because the meters only track total building consumption of electricity. Further methods would need to be used to estimate electricity use of pieces of each building s NCJ electricity use. However, these aggregate building meters were useful in estimating the CNS electricity use located in Cyert Hall. Two electric meters were installed in this building, one for only the network and computerlserver rooms, and one for the rest of the building The annual consumption of electricity through these meters is 4.2 for CNS and 1.3 million kwh otherwise respectively. This centralized use of electricity represents a significant share of total campus electricity demand. The meter attached to the core networklserver rooms also includes the dedicated cooling system for these rooms. Since this cooling system (used to dissipate heat generated by the large amount of network and computer devices installed) is required to keep the CNS operating, we included it as part of the NC estimate. Note that in general no other cooling energy is included in this analysis (e.g. cooling of wiring closets or in-office air conditioning needed to dissipate heat). Thus we are underestimating total electricity use (although not by much given the low power dissipation of BWC and ONE in the study as discussed below). Estimating the remaining NC categories - BWC, ONE, and WT - was more complicated and required significant additional effort. With the help of the campus Computing Services group, a complete inventory of building wiring closet equipment was obtained. nstead of metering each individual switch in each wiring closet, a sample of each type of switch in use was metered, and these samples were used to extrapolate across all wiring closet switches. 45

3 These measurements were found by using a Brand Electronics Model /C power meter. This device features an LCD display showing instantaneous and cumulative power consumption as well as several days worth of memory for minute-by-minute logging that can be output via a serial port to a computer. Several observations were noted when metering network equipment. First, the power consumption of multi-port hubs and switches was fairly independent of the number of connected ports. For example, a Cisco switch might consume 40 Watts with no devices connected and 42 Watts with 16 devices connected. Thus the metered value for the switches was used as the power consumption of the switch with no adjustment for how many devices were connected. The sensitivity of final results to this difference is small. Another observation is the large difference in nameplate power versus metered power. For example, a personal computer might have a metal tag on the back of the case noting it contains a 250-Watt power supply but when in use and metered only consumes 125 Watts. This is because the power supply is able to provide 250 Watts if other internal devices (e.g. second hard drives, recordable CD drives, etc.) are connected. n general, the power supply nameplate rating of networking equipment was considerably - up to a factor of 2 - higher than the metered power, suggesting that all networking devices were overdesigned for power. This is similar to the difference in nameplate capacity and actual generation of an electric power plant, e.g. 300 Megawatts (MW), but it might operate the majority of the time at only 250 MW because not all turbines are in use. This is an important observation because several existing studies have relied on the nameplate power ratings of equipment instead of the actual use. Our research suggests that such methods could overestimate electricity use by a factor of two. There were four types of Cisco switches in use in the building wiring closets on campus with electricity consumption ranging fiom 39 to 370 Watts. Within wiring closets, there are switches that serve only the wired network and only the wireless network - as shown in Figure 1. For the switches associated with the wired network, we metered each switch and found average power consumption. nstead of metering each individual switch in each closet, a simplifying assumption was made. For the 128 closets, we estimated the energy consumption as if each type of switch were the only type of switch used, since it was impossible to visit all closets to identify the actual switches used. For wired network switches in closets this led to a 4-point estimate of wiring closet energy use, 32,000 to 305,000 kwh per year. While this is an order of magnitude difference, and merits fiuther analysis, compared to the main Cyert Hall usage of roughly 4 million kwh, this difference is small. Next, the impacts of ONE (e.g. +port hubs and 8-port switches) on campus were estimated. Again, some simplifying assumptions were made. nstead of inventorying the entire campus, we studied 2 specific departments, and did ONE inventories. We selected the Civil and Environmental Engineering (CEE) department and the English department. We felt that these 2 departments would well represent the high end and low end of such devices. The CEE department is a very computing and network-intensive department while the English department is much less intensive. After performing inventories and metering electricity use, results were extrapolated to the rest of campus in the form of low-end and high-end estimates. The CEE department is comprised of 19 faculty, 11 staff, and 60 graduate students. Only graduate students were counted because undergraduate students do not have campus office space. Based on a walkthrough of each room in the department noting the number and type of hubs and switches used, we estimated the electricity use from this equipment to be 1,800 kwh per year. This number was also disaggregated across faculty, staff, and graduate students to estimate per-capita energy use estimates for each category. These estimates were then combined with campus data on total faculty, staff, and graduate students to estimate total ONE use of 109,000 kwh. Using a similar method with the less-intensive English department yielded a campus wide estimate of 12,000 kwh per year. Again, while this is an order of magnitude difference, it is small compared to the other numbers. Finally, the electricity consumption of the wireless network on campus was estimated. This estimate comprises wireless switches in closets and WT. The wireless network on campus is made up of 355 Lucent Technologies wireless transceiver units that each consume only 8 Watts of power. Each transceiver can support approximately 25 users. The power consumed seems independent of the number of users. Scaling up by the number of transceivers and the BWC wireless network switches used to support them yielded an estimate of 30,000 kwh per year. t should be noted that the energy impacts of the wireless network are not only 30,000 kwh per year. The wireless network is only wireless at the point closest to the user. From the wireless transceiver, an Ethernet wire runs back to the BWC and into a Cisco switch and then eventually back to the CNS. Thus the total energy use of the wireless network is much greater than the 30,000 kwh above. A first-order approximation of the difference between the wired and wireless networks can be found by looking at only the end point electricity use. Assuming that only 46

4 Table 1: Estimates of Campus NC Electricity Use (all values annual, in millions of kwh) Scenario CNS ONE(high) BWC Wireless Total One Two Three Four CNS ONE (low) BWC Wireless Total Five Six Seven Eight the BWC switches, wireless transceivers, and ONE are relevant, then the wired network consumes between 200,000 and 500,000 kwh per year, and the wireless network only 30,000 kwh. This is a factor of 10 difference in energy use. But the wireless network right now cannot support the number of users as the wired network, even if accounting for network performance differences. The current wireless network is meant to complement the wired network. RESULTS AND CAVEATS As shown in Table 1, the total electricity used by NC on campus is estimated between 4.4 and 4.8 million kwh per year, or about 6% of campus electricity use. The scenarios represent a range of energy consumption of the BWC switches described above for both the high and low ONE cases. This number is consistent with the current state of knowledge about T electricity use (about 3% of all electricity use is devoted to T in the U.S.). This analysis shows network electricity use for a university campus that uses much more electricity for T than most universities and commercial buildings. These results therefore cannot be extrapolated to the U.S. as a whole. First, a college campus is not indicative of either a commercial, industrial, or residential building. Making comparisons to any of these categories would be inappropriate. Even amongst college campuses, this campus is much more computer-intensive than the average. t has been named the most wired campus by Yahoo! for several years. n addition, as with most college campuses, heating and ventilating is done with chilled water and steam as opposed to large scale air conditioning and heating systems. This represents a fairly large share of campus energy use which is not represented in the totals above. So while NC is estimated to be 6% of campus electricity use, it may be a much smaller percentage of total energy use when HVAC energy is included. n many commercial buildings, all cooling is done via electricity. Allocating the total NC electricity across the campus population of 10,000 people works out to be about 600 kwh per capita. f network electricity impacts per user were compared to other campuses the CMU campus could be a factor of 2 higher. Another important factor is to more broadly consider the impact of computing and networking. While this study only considered networking impacts, a study that also includes computer and network support personnel, use of office and cluster computers would show additional electricity use. Looking at computer electricity use would easily push this total to over 10%. On campus, about 35% of total energy cost comes from chilled water and steam. f the benchmark used for NC electricity use is against total energy cost, then at 5 cents per kilowatt-hour is $225,000 per year, or about 4% of total campus energy use. As noted above, the wireless network does not have the same bandwidth as the wired network. However both are shared resources - many users are using each one at the same time. Despite the large bandwidth difference, many more users are using wired networking and so its bandwidth is shared over many more users. n practice, users on the wireless network get only 1.4 the bandwidth of users on the wired network. Thus, to a first approximation, assuming functionally equivalent comparisons, we should consider a wireless network running faster that consumes 4 times as much energy as our model. This would consume 120,000 kwh, still significantly less than the existing wired network. But this is unrealistic. Faster la wireless transceivers will likely use only slightly more energy than the current models. Thus wireless networks will probably continue to dominate wired networks on energy consumption. SUMMARY The analysis has shown the contribution of network electricity use on a college campus to be about 6% of total electricity use. n several ways, this study is not intended to be representative of all commercial enterprises, or even 47

5 all campuses. f all networking and computer electricity use was considered, the total electricity from computing would be more than 10%. However, if total energy is the benchmark, NC electricity consumes 4% of total campus energy. From an end-use electricity perspective, there is nearly 10 times more energy used for wired networks compared to wireless, and this difference is unlikely to change significantly with emerging higher-bandwidth wireless standards. ACKNOWLEDGMENTS The authors thank the Camegie Mellon Computing Services department and staff (especially Pete Bronder), as well as support from the AT&T Foundation, AT&T Corporation, and the Green Design Consortium. Finally, to Jon Koomey, Michele Blazek, and Arpad Horvath for comments and modeling advice. REFERENCES [ 11 Salus, Peter H., Casting the Net: From ARPANET to NTERNET and Beyond, Addison-Wesley, [2] OHara, B. and Petrick, A., The EEE Handbook: A Designer s Companion, EEE Press, Kawamoto, Kaoru, Jonathan Koomey, Bruce Nordman, Richard E. Brown, Maryann Piette, Michael Ting, and Alan Meier Electricity Used by Office Equipment and Network Equipment in the U.S. Energy--The nternational Journal (also LBNL-45917). vol. 27, no. 3. March. pp Koomey, Jonathan G Rebuttal to Testimony on Kyoto and the nternet: The Energy mplications of the Digital Economy.Berkeley, CA: Lawrence Berkeley National Laboratory. LBNL August. < ects/nfotech.html>. Roth, Kurt Electricity Used by nformation Technology and Telecommunications Equipment in Commercial Buildings--Volume : Energy Consumption Baseline. Washington, DC: Prepared by Arthur D. Little for the U.S. Department of Energy. January. Chong, Hui Min, nventory of Electricity End-Use Categories on Campus, Senior Honors Thesis, Camegie Mellon,