Energy Consumption Of The Internet

COIN - ACOFT 2007 24 - 27 June 2007, Melbourme, Australia

Energy ns ne et e In"terne C,,onsumption 0ofthe .~~~~1

-

9

9

,

I/

._

1_

._

Jayant Baliga, Kerry Hinton and Rodney S. Tucker ARC Special Research Centre for Ultra-Broadband Information Networks (CUBIN) Department of Electrical and Electronic Engineering, University of Melbourne, Victoria 3010, Australia Abstract-As concerns about global energy consumption increase, the power consumption of the Internet is a matter of increasing importance. We present a network-based model that estimates Internet power consumption including the core, metro,

Core

Core

Core

Metro

Edge

Edge

Edge

and access networks.

I. INTRODUCTION

Continued expansion of the Internet based services has now become an integral contributor to sustained growth of the World economy. In addition to the expected bandwidth demand arising from new Internet there is also the growth due to new (high bandwidth) services. Current trends indicate the average global IP traffic annual growth rate is around 50% and possibly higher [1] As IP traffic increases, the amount of equipment required to route this traffic must grow at a corresponding rate. A consequence of this is a growth in power consumption of the equipment Energy consumption is becoming a key en vironmental, social and political issue. It is now accepted that electrical energy generation is a significant contributor to greenhouse gasses [2]. This raises the issue of whether the Internet may ultimately be constrained not by the speed of routers and affiliated electronics but rather by their power consumption [3] 17]. In this paper, we present a network based model of the power consumption of the Internet formulated with data from major equipment vendors. Using this model we estimate the power consumption of the Internet to be about U0 of the total electricity consumption in broadband enabled countries with average access rates on the order of 30 Mb/s. As the access rate rises, this percentage will also rise. The model focuses on power consumption in the telecommunications network, rather than the home network equipment and the PC of the end users. Our analysis does not include ATM, Frame Relay and similar legacy networks. In addition, it does not include power consumed by servers [8][9].

Network

Curb

Curb

Curb

Curb

Curb

Curb

Home

Home

HoHomeome

Home

Home

Access Network

Fig. 1. Schematic of network structure showing the core, metro and access networks.

the edge nodes, a passive splitter at the curb-side combines the signal from several homes into a single fibre which is directly connected to one of ten edge nodes. The edge routers together with the links that connect the edge routers to the central node comprise the metro network. The access network consists of the curb-side nodes, the network unit at each home that connects the home to the access network, the concentrators and the links that connect these three stages of the network. We now describe each of these three sections in more detail as well as the concept of oversubscription. A. Oversubscription The exact number of core and edge routers depends heavily on the oversubscription rate used by the ISP. If a given network can support N users simultaneously at a given peak access rate but MI > N users are connected, then the oversubscription rate is ( Ni- N)INV For example an oversubscription rate of 1 0 means that only half of the users can simultaneously access the Internet at furll speed, or all users at a half of the full speed In the past when the Interet was used mainly for website browsing, traffic demand. was very bursty. Therefore a user would load a webpage read it for a minute or more and then move on to a new webpage. In addition, the number II. NETWORK STRUCTURE of active users at any time was typically a small fraction of A standard Internet Service Provider's (ISP) network can be network subscribers. These scenarios allowed ISPs to heavily logically split into three main sections - the access network, oversubscribe the network with oversubscription rates of 24 the metro network and the network core. The overall structure not being unusual [10]. However, as the Internet is increasingly of the network is shown in Figure 1. The nodes at the top of used for multimedia applications such as streaming videos, the hierarchy represent core nodes. The core routers perform IPTV, video conferencing and other peer-to-peer applications, all the necessary routing and also serve as the gateway to the demand is mu.ch greater and, is also relatively more neighbouring core nodes. The next level of hierarchy repre- constant, thus restricting the level of oversu.bscription. sents the edge nodes which aggregate/concentrate the highly fluctuating traffic from the end user and serve as the interface B. Access Network between the access network and, core network. To reduce the To minimize both operational and installation costs, the number of fibres required to connect each home to one of future access network will have a Passive Optical Network

ISBN 978-0-9775657-3-3 Authorized licensed use limited to: Univ of Texas at Dallas. Downloaded on October 10, 2008 at 16:29 from IEEE Xplore. Restrictions apply.

(PON) [ 11] structure. Every curb-side node consists of one or more passive splitters that split a single fibre from an Optical Line Terminal (OLT), which resides in an edge node, into several fibres. Each of these fibres connects to an Optical Network Unit (ONU) attached, to each home. Each group of ONUs sharing the same connection to an OLT communicates with the OLT in a time multiplexed order with the OLT assigning time units to each ONU based on its relative demand. The oversubscription rate can be changed by changing the number of ONUs that share a connection to an OLT, which is currently limited by most manufacturers at 32 [12]. C. Metkro Network On the network side, the OLTs are typically connected by four Gigabit Ethernet lines to one of the edge routers which encapsulate the IP packets into a SONET/SDH format for transmission to the network core [13][14]. In our network model we us the Cisco 12816 edge rutur [13], this device is typical of routers available from other manufacturers. The edge router connects to the core node via fourteen 1O Gbls OC192c/STM-64c Packet-over-Sonet (POS) links, and to forty OLTs each via 4xl1 Gb/s links The oversubscription is negligibly small. The edge routers do not need to be interconnected, as they only perform forwarding. D. Core Network All the network routing is performed at the central node by several multishelf core routers [13][14]. An example of one is the Cisco CRS-1 Multishelf System [13] Though a single large router is preferable, there still exist several limitations that must be overcome before current routers can be scaled to very large capacities [15]. Each CRS-1 Multishelf system can scale up to 72 line card shelves and eight fabric card shelves to give a combined switching capacity of 92 Tbls fu'llduplex [13] The CRS 1 routers are interconnected through single port OC-768c/STM-256c POS line cards, which operate at 40 Gb/s, and connect to the edge routers through 4-port OC-1192c/STM-64 POS line cards, where each port operates at 10 Gb /s [13]. The core routers also have connections to neighbouring long distance locations and these connections are also through single port OC-768c/STM-256c POS line cards

[13].

11. POWER CONSUMPTION We now calculate the power consumption in each section of the network and analyse some of the factors that will affect the consequent total power consumption. As typical power consumption values are sometimes not readily available and the maximum power values stated in the data sheets are simply the rated power of the power supplies, we use the heat dissipation to estimate the power used by the various routers. The calculations do not include the extra power required to maintain redundant routers for reliability and they do not include the power consumption of data centers [8]. This paper therefore provides a lower bound on the power consumption of the Internet. To calculate cooling requirements, we assume

that for every watt of power consumed, another watt of power is required for cooling [16]. A. Metro and Access Network Power Consumption The ONU consumes approximately 10 W [12]. The curb side node does not consume any power as the access network is designed to be passive between the ONU and, the OLT. Each OLT consu.mes 100 W [12]. In our model the metro network consumes approximately 4 21 kW per Cisco 12816 router [1 3]. With these assumptions the power consumption of the metro and access network Pif a is given by:

N

F4.21 kW

Nx 10 W (1) 40 where p is the oversubscription rate and N is the number of homes. The parameter S is the number of ONUs that connect to one OLT when the oversubscription rate is 0. For example, if the peak access rate is 1 Gb/s then S is equal to 4 while if the peak access rate is 1L00 Mbls then S is equal to 40

S(p+ 1)

L

B. Core Network Power Consumption The network core consumes approximately 13 2 kW per line card shelf and 8.97 kW per fabric card shelf [13]. A complete Cisco CRS-1 Multi-shelf System with 72 line card,

shelves and 8 fabric card shelves therefore consumes 1020 k

As the interconnection structure of the core routers is highly dependent on the nature of the traffic, we approximate the total power consumed by the core routers by considering the total switching capacity required While this ignores the fact that if data needs to be sent from one core router to another then the switching required is effectively doubled, we assume that the number of occurrences of this is minimized if we ensure that each of the edge routers is connected to mu'ltiple core routers We also assume that 50% of the traffic into the router is internal, 25% is add/drop and finally 25% is bypass. This means that the total traffic being processed by the router is equal to 133% of the total metro traffic. Finally, core routers are built such that they are able to cope with future growth and so are usually built such that they can handle double the current peak demand [17] With these assumptions in place the power consumed by the core node P, is: 1020kW x N x 2R (2) (p + 1) x 92 Tbls where p is the oversubscription rate, N is the number of homes andd R is the peak access rate in each direction C. WDM Links The WDM terminal systems connecting the edge nodes to the core node consume 1.5 kW for every 64 wavelengths [18]. At most one multiple wavelength amplifier, which consumes 6 W per fiber [1 8], is required, between each pair of terminal systems provided the size of the city is less than 200 km in diameter. The WDM terminal systems connecting the core nodes consumes 811 W for every 176 channels, while each intermediate line amplifier (ILA) consumes 622 W for every

Authorized licensed use limited to: Univ of Texas at Dallas. Downloaded on October 10, 2008 at 16:29 from IEEE Xplore. Restrictions apply.

7OF-

5000 100 Mb/E Total

Powera,

50

350

> 40

5 300 Metro + Acce -

aL

0-

Total F

400

30

Mer,fAcs

250-

aL

° 2000150-

Core

20

100-

10

0&0

........ 20

DM

40

60

Average Acces Rate (Mb/s)

-Core

50-

Links

0'

100

80

400

WDM Links 600

Averane Acces Rate

(Mb/s)

800

Fig. 2. Power consumption of WDM links, core node, metro plus access network and total power consumption for the full range of average access rates when there are 2 million homes and the peak access rate is 100 Mb/s.

Fig. 3. Power consumption of WDM links, core node, metro plus access network and total power consumption for the full range of average access rates when there are 2 million homes and the peak access rate is 1 Gb/s.

176 channels [19]. For a distance of 1 500 km between core nodes, two terminal systems and 14 ILA systems are required. The power consumed by the WDM links P d, is therefore

IV. CONCLUSION We have presented a network based model of the power consumption of the Interet. We estimate the power consumption of the Internet to be 1% of electricity consumption in broadband enabled countries. This percentage could increase to over 4% as the access rate increases. The energy bottleneck in the Internet is the routers not the optical fibre links

Pwdm

16.44 W x N S(p + 1)

t

(p

12

A

GbPs

1)x4OGb/s

where p is the oversubscription rate, N is the number of homes R is the peak access rate in each direction and S is the number of ONUs that connect to one OLT when the oversubscription rate is 0.

D. Total Network Power Consutmption

Figures 2 and 3 show the total power consumption of the network, versus average access rate when there are 2 million homes and the peak access rate is 1L00 Mb/s and 1 Gbls per home, respectively. The average access rate is given by + 1) where R is the peak access rate and p is the oversubscription rate. Also included in the plots are the power core node and the metro plus consumption of the WDM access network.

R(p

links,

Let us first access rate of

consider the 100 Mb

Mbls Assuming 2 per

power

consumption for

s and an average access

persons per

home and using

capita electricity consumption

in OECD

rate

a

peak of

30

an average

countries of

kW [20], we conclude that the Internet consumes I% of total electricity consumption. In this scenario (30 Mb/s access rate) the majority

of

power

is

consumed by

the

metro and access

networks with the core node only consuming 2200 of the power. The WDM links consume only 2% of the power. Increasing the peak access rate to 1 Gbls with an average access rate of 300 Mbls, would increase the power consumption of the internet to 400 of total electricity consumption. The power consumption of a single core node serving 2 million users would be 78 MW. Supplying this power to a single building and dissipating the resultant heat will be a forrmidable engineering challenge.

REFERENCES [1] Telegeography Research & Primetrica Inc. 2005. [Online]. http :www.telegeography. com [2] "US 2002 Carbon Dioxide Emission from Energy Consumption". [Online]. http:Heed.llnl.gov/flow/carbon02.php [3] MA Gupta and S Singh, "Greening of the Internet" ACM SIGCOMM, Karlsruhe, Germany, Aug. 2003. [4] A. Vukovic, "Data Centers: Network Power Density Challenges," ASHREA Journanl vol. 47, p. 55, Apr. 2005 [5] K. Christensen, "The next frontier for communications networks: power management," Proc. SPIE - Performonce and Control of NextGeneration Communications Net'orks, vol. 5244, pp. 1-4, Sep. 2003. [6] P. McKenna, "The Waste at The Heart of the Web," Aew Scientist, vol.192, issue 2582, p.24, Dec. 2006. [7] D. T. Neilson, "Photonics for Switching and Routing," IEEE J Sel/ Topics Ouantum Electron., vol. 12, Jul./Aug. 2006 [8] J. G. Koomey, "Estimating Total Power Consumption by Servers in the U S and the World" Feb 2006. [Online]. http:Henterprise.amd.com/Downloads/svrpwrusecompletefinal .pdf [9] The Greening Grid. [Online]. http:Hwww.thegreengrid .org [10] P Ferguson and G Huston "Quality of Service on the Internet Fact Fiction, or Compromise9," INET'98, Geneva, Switzerland, 1998. [I 1] P. Chanclou, S. Gosselin, J. Palacios, V Alvarez, E. Zouganeli, "Overview of Optical Broadband Access Evolution," IEEE Comm. Maag p. 29, Aug. 2006. [12] NEC GE-PON Data Sheets. [Online]. http:Hwww.nec.cojp [13] Cisco Data Sheets. [Online]. http://www.cisco.com [14] Juniper Data Sheets. [Online]. http://wwwjuniper.net [15] H. J. Chao, "Next Generation Routers," Proc. IEEE, vol. 90, pp. 15181558, Sep. 2002. [16] K. Dunlap, "Cooling Audit for Identifying Potential Cooling Problems in Data Centers," 2006. [Online]. http://www.apcmedia.com/salestools/VAVR-5UGVCN R2 EN.pdf [17] E. Desurvire, "Capacity Demand and Technology Challenges for Lightwave Systems in the Next Two Decades," J Lightu ave Technol., vol. 24, p.4697, 2006. [18] Lucent Technologies Data Sheets. [Online]. http:Hwww.lucent.com/eon [19] Fujitsu Data Sheets. [Online]. http://www.fujitsu.com [20] NationMaster. [Online]. http:Hwww.nationmaster.com

Authorized licensed use limited to: Univ of Texas at Dallas. Downloaded on October 10, 2008 at 16:29 from IEEE Xplore. Restrictions apply.