Methodologies for Assessing the Use-Phase Power Consumption and

Dec 4, 2012 - Internet traffic has grown rapidly in recent years and is expected to continue to expand significantly over the next decade. Consequentl...
0 downloads 13 Views 2MB Size
Download PDF
Article pubs.acs.org/est

Methodologies for Assessing the Use-Phase Power Consumption and Greenhouse Gas Emissions of Telecommunications Network Services Chien A. Chan,*,† André F. Gygax,†,‡ Elaine Wong,† Christopher A. Leckie,†,§ Ampalavanapillai Nirmalathas,† and Daniel C. Kilper∥ †

Centre for Energy-Efficient Telecommunications, Department of Electrical and Electronic Engineering, The University of Melbourne, 3010, Victoria, Australia ‡ Department of Finance, The University of Melbourne, 3010, Victoria, Australia § Department of Computing and Information Systems, The University of Melbourne, 3010, Victoria, Australia ∥ Bell Laboratories, Alcatel-Lucent, Holmdel, New Jersey, 07733, United States S Supporting Information *

ABSTRACT: Internet traffic has grown rapidly in recent years and is expected to continue to expand significantly over the next decade. Consequently, the resulting greenhouse gas (GHG) emissions of telecommunications service-supporting infrastructures have become an important issue. In this study, we develop a set of models for assessing the use-phase power consumption and carbon dioxide emissions of telecom network services to help telecom providers gain a better understanding of the GHG emissions associated with the energy required for their networks and services. Due to the fact that measuring the power consumption and traffic in a telecom network is a challenging task, these models utilize different granularities of available network information. As the granularity of the network measurement information decreases, the corresponding models have the potential to produce larger estimation errors. Therefore, we examine the accuracy of these models under various network scenarios using two approaches: (i) a sensitivity analysis through simulations and (ii) a case study of a deployed network. Both approaches show that the accuracy of the models depends on the network size, the total amount of network service traffic (i.e., for the service under assessment), and the number of network nodes used to process the service. lifetimes,10−12 the development of assessment methods for use-phase GHG emissions represents an important step toward understanding the emissions from networks and services, thereby allowing telecom providers and users to achieve greater energy efficiency. Ideally, the CO2 emissions from a telecom network service should be accurately calculated using the bottom-up approach, which utilizes power consumption and service traffic measurements collected from each individual network equipment unit. However, telecom networks offer many different services that often use service-specific equipment and functionality, presenting the challenge of extracting the network GHG emissions associated with a specific service (a typical telecom network is shown in the Supporting Information (SI)). Furthermore, the use-phase energy consumption due to networks and services may not be easy to calculate due to limited network equipment monitoring

1. INTRODUCTION Currently, the information communications and technology (ICT) industry contributes approximately 2% of global CO2 emissions,1−6 which is equivalent to that contributed by the aviation industry.4,5 However, ICT emissions are expected to almost double by 2020,6 so without significant green investments in the telecom infrastructure that serves as the backbone of ICT services, the energy consumption of these services and the associated CO2 emissions are expected to grow significantly. For example, the network equipment in the United States is estimated to have used between 14 and 18 TWh in 2008, a figure that is expected to grow to 23 TWh in 2012 without improvements in energy efficiency.2,7 The total European telecom network is expected to increase from 14.3 TWh per annum in 2005 to 35.8 TWh per annum in 2020 without the adoption of green technologies.8,9 Furthermore, the “Smart 2020” report finds that telecom devices and infrastructure emissions are expected to grow from 300 Mt CO2 emissions in 2007 to 350 Mt CO2 emissions in 2020.6 Because the use-phase GHG emissions from telecom network equipment represent approximately 70−90% of the total lifetime GHG emissions due to long equipment © 2012 American Chemical Society

Received: Revised: Accepted: Published: 485

August 20, 2012 November 27, 2012 December 4, 2012 December 4, 2012 dx.doi.org/10.1021/es303384y | Environ. Sci. Technol. 2013, 47, 485−492

Environmental Science & Technology

Article

Figure 1. Summary of the power consumption and traffic measurement requirements for selecting appropriate models.

In general, the total amount of service traffic traversing through a telecom network varies over time. Therefore, the mean traffic (e.g., bits per second) is used to calculate a network service’s CO2 emissions. Averaging should be performed over a time period long enough (weekly or monthly) to include all short-term fluctuations but not longterm changes (e.g., due to capacity upgrades). However, the peak capacity should be used to determine the traffic fraction for cases in which network equipment is provisioned based on peak rates or provisioned bandwidth (e.g., access or metroaggregation networks). The energy consumption associated with a network service can be converted to CO2 emissions using regional- or countrybased carbon emission conversion factors. These conversion factors can be obtained from government agencies such as Defra in the U.K. and the International Energy Agency.26,27 The network measurement requirements for selecting an appropriate model based on the level of information granularity are summarized in Figure 1. 2.1. Coarse-Grained and Fine-Grained Models. The coarse-grained and fine-grained models are closely related to each other. The fine-grained model includes additional terms to account for the number of hops that network services take through each class of equipment. In contrast, the coarse-grained model does not require such information. An equipment class is a grouping of equipment performing similar operations that may be specific to a given network and must be defined in the calculation. A detailed example of this equipment is given by Phillips et al.28 The corresponding power consumption of the ith service (the service under assessment), PSi, can be defined as

capability, especially for legacy network equipment, and the complexity of how services are delivered within networks. Therefore, telecom providers often use the most basic topdown model, which takes into account power consumption and service traffic measurements at the network platform level, i.e., an entire network or network segment within a region or country. For example, AT&T reported that their network consumed 347 kWh to process one terabyte of data13 and Verizon estimated that their network produced 0.064 tons of CO2 to process one terabyte of data14 in 2011. However, in this study, our accuracy assessment results show that the top-down model can produce a relatively large estimation error when accounting for the energy efficiency of a given service, depending on the network size, especially when the network service is processed only by a certain number of network nodes (e.g., service-specific equipment). To combat this problem, we developed coarse-grained and f ine-grained models, which use slightly higher information granularity at the equipment class level compared to the top-down model but have better estimation accuracy. In addition, we expand the contribution by testing and comparing the accuracy of these models to the most basic top-down model available for various network scenarios using two approaches: (i) sensitivity analysis through simulation and (ii) a case study of a deployed network, the California Research and Education Network (CalREN).15 Using the results obtained from the two approaches, we present the network conditions under which these models are accurate and the circumstances under which the estimation error increases.

2. ASSESSMENT METHODS Assessing the use-phase GHG emissions requires information at the service level but not necessarily at the overall network level. For example, if we were to use telecom services such as teleconference services in lieu of automotive or airline travel,16−24 it would be useful to know the quantity of GHG emissions that are associated with that particular e-distance service to perform an accurate comparison of the two methods used to accomplish the same task. Furthermore, with the use of video and music streaming over the Internet in lieu of traditional shipping methods and printed newspaper delivery to customers,16,22,25 one may want to know the quantity of GHG emissions associated with the delivery of content to the endusers for an accurate comparison.

PSi = ηSi × TSi

(1)

where TSi denotes the mean network traffic of the ith service and ηSi is the efficiency of the network in delivering the ith service. ηSi is defined as the power required to process the total mean traffic of the ith service (W/bps) and can be expressed by ηSi =

PN × MSi TSi

(2)

where PN is the total network platform power consumption and MSi is the service power map, which determines the fraction of power for each class of equipment that is attributed to the 486

dx.doi.org/10.1021/es303384y | Environ. Sci. Technol. 2013, 47, 485−492

Environmental Science & Technology

Article

Table 1. Summary of Calculation Methods for Top-Down, Coarse-Grained, Fine-Grained, and Bottom-Up Models top-down

fine-grained

coarse-grained

service i CO2e, CSi

bottom-up

CSi = PSi × t × EF t = time/usage EF = emissions factor

service i power, PSi

PSi = ηSi × TSi TSi = mean service i traffic

service i efficiency, ηSi

ηSi =

PN TN

ηSi =

PN × MSi TN

PN = network platform power TN = mean network traffic l

MSi =

k=1

service power map, MSi

N/A

⎛

∑ ⎜fEk ⎝

×

TSi _Ek ⎞ ⎟ TEk ⎠

f Ek = fraction of equipment class power TEk = equipment class traffic TSi_Ek = service traffic in each equipment class

∑ ⎜fEk k=1

l

⎛

∑ ⎜fEk k=1

⎝

×

TSi _Ek ⎞ ⎟ TEk ⎠

l

⎛

∑ ⎜fEk k=1

⎝

× HEk _Si ×

(3)

TSi _Ek ⎞ ⎟ ThEk ⎠

(4)

where TSi_Ek is the mean traffic of the ith service traversing through the kth class of equipment and ThEk is the hit-weighted traffic, which can be determined by

d

v

ThEk =

∑ (HEk _Si × TSi _Ek) i=1

TSi _Ek ⎞ ⎟ ThEk ⎠

d

MSi =

⎛

∑ ⎜⎜fDj j=1

⎝

×

TSi _Dj ⎞ ⎟ TDj ⎟⎠

equipment when the customer access rate is low.29 As the access rates increase, the core network power consumption increases and ultimately surpasses the access power consumption.29 Therefore, each subnetwork will have a unique value of η in this model. Furthermore, these subnetworks can be service-dependent networks, such as mobile backhaul or content delivery networks. 2.3. Bottom-Up Model. In theory, the bottom-up model can be used as a reference model to compare the accuracy of the top-down, coarse-grained, and fine-grained models because it provides the most accurate estimations by utilizing power consumption and service traffic measurements collected from individual equipment units. Using the bottom-up model, eqs 1 and 2 are used to calculate the power of the ith service and a network’s efficiency in delivering the ith service. However, the service power map, MSi, is modified to determine the fraction of power for each individual network equipment unit attributed to the network service based on its traffic fraction. Therefore, eq 3 is rewritten as

where f Ek represents the fraction of power for the kth class of equipment, f Ek = PEk/PN, where PEk is the power consumption for the kth class of equipment, and l is the total number of equipment classes. When using the fine-grained model, the parameter HEk_Si (which denotes the mean number of times an individual equipment unit in a given equipment class is accessed within a network) is included in MSi. Therefore, eq 3 can be rewritten as MSi =

⎝

× HEk _Si ×

HEk_Si = number of hops in equipment class k f Dj = fraction of individual equipment unit power ThEk = total hit-weighted traffic for each TDj = individual equipment unit traffic equipment class TSi_Dj = service traffic in individual equipment unit

network service based on its traffic fraction. Using the coarsegrained model, MSi can be modeled as MSi =

⎛

l

MSi =

MSi = (5)

j=1

2.2. Top-Down Model. A top-down model can be used if power and traffic measurements cannot be collected at the equipment class level. This model requires only the overall network or platform power, PN, the total network traffic, TN, and the network traffic for the ith service, TSi. In contrast to eq 1, the efficiency calculation of the top-down model is not dependent on the service type. Therefore, the service power of the ith service can be determined using the expression PSi = η × TSi, where η = PN/TN. Consequently, the value of η is identical for every network service that traverses through the network. However, it should be noted that when using the top-down model to determine an end-to-end service, the overall network must be divided into multiple subnetworks (i.e., access, metro, and core) because the efficiency differences between these subnetworks can be substantial. For example, Internet power consumption is dominated by the access

⎛

∑ ⎜⎜fDj ⎝

×

TSi _Dj ⎞ ⎟ TDj ⎟⎠

(6)

where TSi_Dj is the mean traffic for the ith service that is traversing through the jth individual equipment unit, TDj is the mean traffic for the jth individual equipment unit, and d is the total number of individual equipment units in the network. f Dj represents the fraction of power consumed by the jth individual equipment unit, where f Dj = PDj/PN. It should be noted that equipment power consumption consists of marginal power and base power consumption. Therefore, PDj can be written as PDj(t) = (Pmax − Pbase) × (TDj(t)/Tmax) + Pbase, where Pbase is the power consumption when the network equipment is in idle mode and Pmax is the power consumption when the equipment is fully utilized. The term (Pmax − Pbase) denotes the marginal power consumption, and TDj(t) is the equipment utilization (bps) at time t. Furthermore, TSi_Dj in eq 6 has to be rewritten as TSi_Dj(t). The effect of marginal power consumption is significant for next 487

dx.doi.org/10.1021/es303384y | Environ. Sci. Technol. 2013, 47, 485−492

Environmental Science & Technology

Article

network. One such representation is the small-world network.33 Assuming that the small-world network parameter, p, is equal to 0.1,33 the simulation will generate a range of small-world networks with the number of nodes ranging from 5 to 50. In each simulation trial, a different small-world network topology will be generated. The reason for choosing a maximum network size of 50 nodes is that a large network should be divided into multiple subnetworks (i.e., access, metro-aggregation, and core) because the difference in the efficiency of these subnetworks can be substantial. 3.2. Equipment Class. Analyzing the global Multiprotocol Label Switching (MPLS) core network platform data of a large national operator,34 we find that network equipment units can be categorized into three main classes of network equipment: Class H, Class I, and Class L. We use the following power law parameters to determine the number of network units per equipment class in the simulation: 65.75% Class H equipment, 28.35% Class I equipment, and 5.90% Class L equipment. The associated amounts of energy consumed (CO2 emissions) for Classes H, I, and L are approximately 5961 MWh/year (3456 tons/year), 8406 MWh/year (4872 tons/year), and 4349 MWh/year (2521 tons/year), respectively. Class H represents small chassis routers and switches, which consume approximately 200−400 W.35 Class I represents medium chassis routers and switches, which consume approximately 1−3 kW.35 Class L represents large chassis routers and switches, which consume approximately 3−5 kW.35 3.3. Network Node Degree. The average node degree of a telecommunications network differs among topologies, ranging from 2.55 to 4.4.36 Therefore, it is reasonable to set the average node degree to 4 when generating small-world networks in the simulation. Furthermore, we find that the scenarios with average node degrees of 2, 6, and 8 have no significant difference in terms of model estimation error compared to the scenario with an average node degree of 4. 3.4. Network Service Traffic Weight Distribution. The network service under assessment may have a different weight in each of the network nodes due to traffic engineering within the network and the use of service-specific network equipment. For example, some nodes might process the network service heavily, while others might not be used to process that particular network service at all. Therefore, we model four cases for service traffic weighting distribution in the network using a log-normal distribution, as shown in Figure 2. The mean network traffic generated in the simulation is randomly and uniformly distributed.

generation network equipment because the Pbase is expected to be close to zero. However, the Pbase of current network equipment is approximately 80%−90% of Pmax.30,31 The power consumption versus traffic utilization characteristics of telecom network equipment are presented in the SI. For the coarsegrained model, PEk(t) = Σ PDk(t) and TEk(t) = Σ TDk(t), where Dk ∈ equipment class k. A summary of the calculation methods used for different models is shown in Table 1. 2.4. Model Estimation Error (MEE). In this study, the bottom-up model is used as the reference model. The estimation error of the model m (top-down, coarse-grained, or fine-grained) is determined using the formula MEEm (W) = PSi_m − PSi_BU, where PSi_BU is the service power consumption calculated from the bottom-up model. Hence, the MEE can be modified to a percentage value based on the determined mean service power using the equation MEEm (%) = MEEm (W)/ PSi_m × 100% to account for the fact that a network’s power consumption increases when the number of nodes increases by factoring out the total power. It is important to note that negative regions in the MEE indicate that the model is underestimating the actual service power consumption, while positive regions show that the model is overestimating this consumption. 2.5. Converting Service Energy Consumption to Service CO2 Emissions. Network service energy consumption can be converted to service CO2 emissions using the emissions factor provided by refs 26 and 27. It should be noted that the emissions factor is normally an average value of the various energy sources used to power a region’s network. For example, the state of California uses a mix of energy sources, i.e., natural gas (53.4%), nuclear (15.7%), large hydro (14.6%), renewable (14.6%), and coal (1.7%).32 If detailed emissions factor information for electricity generated from baseload and peak load power plants were available, telecom providers could calculate carbon allocation for telecom services based on the hours of the day or the days of the year to better analyze the impact of marginal energy consumption. Service CO2 emissions CSi are calculated by multiplying the power of the ith service PSi (W) by a specified period of time t (to convert to energy consumption) and the corresponding emissions factor EF (regionally based): CSi = PSi × t × EF.

3. SIMULATION CALIBRATION The accuracy of the top-down, coarse-grained, and fine-grained models is examined using two different approaches: (i) a sensitivity analysis of various network scenarios and (ii) a case study of a deployed network, the CalREN. The sensitivity analysis accounts for several important network criteria, such as network topology, network size, distribution of individual equipment units in each class of equipment, network traffic generation, and service traffic weighting in each node (the number of network nodes that are used to process the network service under assessment). Extensive simulations are conducted to perform the sensitivity analysis and each simulation trial generates a different network topology with a predefined number of nodes and average node degree. The simulation generates the mean traffic for individual equipment units using a uniform distribution, with the average utilization assumed to be 50%. Every scenario is then simulated over 100 000 trials to obtain accurate results. 3.1. Network Topology. Several studies indicate that the network topology of many telecommunications networks lies between that of a regular network and that of a random

Figure 2. Cumulative distribution function (CDF) of the service traffic weighting among nodes for a network size of 50 nodes. 488

dx.doi.org/10.1021/es303384y | Environ. Sci. Technol. 2013, 47, 485−492

Environmental Science & Technology

Article

Figure 3. Model estimation error (%) of the top-down (red bar), coarse-grained (blue bar), and fine-grained (green bar) models under various node service traffic weighting cases (refer to Figure 2).

should be noted that the width of the normal distribution curves changes under different network configurations, resulting in different MEEs. Therefore, the 95% confidence interval of the MEE is calculated to determine the ± error region in which 95% of the errors will fall. Figure 3 shows the MEEs of the top-down, coarse-grained, and fine-grained models for the four different node service traffic weighting distribution cases in the network (refer to Section 3.4). For an enlarged version of the figure that compares the coarse-grained and finegrained models, see the SI. It should be noted that the network size ranges from 5 nodes to 50 nodes, and the average node degree is 4. Other average node degree cases are not shown because we find that different average node degree cases have no significant impact on the MEE of the top-down model, mainly because the top-down model accounts only for the overall incoming traffic of the network TN, and the value of TN remains unchanged regardless of how well the network is connected. In general, the fine-grained model produces the lowest MEE among the three models. As the network size increases, the fine-grained model’s estimation error increases and saturates at a certain level (e.g., ± 6% and 5% beyond 25 nodes for Case 1 and Case 2, respectively) depending on the node service traffic weighting distribution. Therefore, the MEE difference between the coarse-grained and fine-grained models is large when the network size is small to medium. However, the difference decreases as the network size increases, mainly because the mean hit list (MHL; refer to Section 2.1) term in the finegrained model becomes less significant compared to the weight of the overall network traffic as the network size increases. It is also clear that as the skewness of the node service traffic weighting distribution decreases from Case 1 to Cases 2, 3, and 4, the MEEs of the two models decrease; hence, the difference between the two models also decreases. For the top-down model, the MEE decreases as the skewness of the node service traffic weighting distribution decreases from high to medium, low, and linear. The negative MEE of the topdown model is larger than the positive MEE in cases in which the network service traverses only through network nodes, which consume high power. The top-down model does not reflect this fact because it accounts only for the overall platform power and traffic. Furthermore, Figure 3 shows that as the

Case 1 represents a scenario in which the network service traverses only through a small number of nodes within the network and results in a highly skewed distribution. Cases 2 and 3 represent scenarios in which the network service traverses through a considerable number of nodes and results in medium and low skewness distributions, respectively. Case 4 represents a scenario in which every network node processes a similar amount of network service traffic, which produces a linear distribution among the network nodes. 3.5. Field Network DataCalREN. Collecting network data from the field to conduct a case study is a difficult task due to the lack of publicly available information on network power consumption and traffic tracing. The CalREN is one of the very few research and education networks that shares publicly available data on network topology, the type of network equipment used, and network traffic tracing. The CalREN is a combination of aggregation and core networks that consists of 16 sites and 57 nodes with an average node degree of 4.5. The number of network equipment units per equipment class can be modeled as 35% Class H equipment, 53% Class I equipment, and 12% Class L equipment. The associated energy consumption (CO2 emissions) for Classes H, I, and L are approximately 114 MWh/year (76 tons/year), 585 MWh/year (391 tons/year), and 288 MWh/year (192 tons/year), respectively. The mean network traffic and link utilization data for the month of February in 2012 are collected and employed in this case study.

4. RESULTS The traffic contribution of a network service varies in different networks. In the first part of the sensitivity analysis study, we assume that the network service of interest has a mean traffic of 10% of the overall network traffic. Figure 2 is used to determine the distribution of the network service traffic weight in each node, which corresponds to 10% of the overall network traffic. In the second part of the sensitivity analysis, we study the impact of varying the service traffic proportion from 10% to 50% of the overall network traffic. 4.1. Sensitivity Analysis Results. The MEE of each model behaves as a normal distribution with its tails toward both the positive and negative regions (the MEE probability distributions for different models are shown in the SI). However, it 489

dx.doi.org/10.1021/es303384y | Environ. Sci. Technol. 2013, 47, 485−492

Environmental Science & Technology

Article

4.3. Case Study of a Deployed NetworkCalREN. The entire CalREN network consumed an average of 169 kW and 118 MWh of power and energy, respectively, in the month of February 2012. In contrast, the network equipment used by the associated members of CalREN to access the CalREN consumes approximately 115 MWh per month (detailed calculations are presented in the SI). According to refs 26 and 27 the average electricity-to-CO2 emissions conversion factor in the state of California is 0.6682 tons/MWh; thus, the total CO2 emissions of the entire CalREN network in February 2012 are calculated to be approximately 79 tons. More details regarding the CalREN network topology, node power consumption (W), and average node utilization (Gbps) are provided in the SI. Since the parameters of the network size, network topology, average node degree, and network node utilization can be obtained, we set a range of mean service traffic from 10% to 50% of the overall network traffic (similar to Figure 4) to study the impacts of different service traffic proportions. Hence, Figure 2 (with the maximum number of nodes changed to 57) is used to choose the distribution of network service traffic weight in each node. Figure 5 shows the MEEs of the top-down, coarse-grained, and fine-grained models for different service traffic weighting cases and mean service traffic proportions. As in the results observed from Figure 4, the MEEs of the three models decrease as the skewness of the node service traffic weighting distribution decreases from high (Case 1) to medium (Case 2), low (Case 3), and linear (Case 4). Comparing Figure 4 and Figure 5 reveals that the deployed network case study produces a higher MEE than the simulation study. This difference is due mainly to the fact that the power consumption and utilization of individual network nodes within the same class of equipment are different in the CalREN network. In contrast to the CalREN network model, the simulation-based sensitivity analysis approach simply assumes that individual units of equipment within the same equipment class have similar power consumption and that the overall utilization of each node is 50%. The detailed MEEs of different models for a mean service traffic proportion of 10% are shown in Table 2. Nevertheless, both approaches show that the fine-grained model has significantly greater accuracy under conditions in which the network service traffic is concentrated in a small number of nodes (e.g., due to traffic engineering and the use of service-specific network equipment). The coarse-grained model, provided that the distribution of the service traffic weighting at each node is not highly skewed as in Case 1 (which means that a considerable number of network nodes are used to process the network service), can exhibit an accuracy similar to that of the fine-grained model and may be preferable because of its reduced complexity in obtaining network measurements. In contrast, the top-down model is accurate only under the network conditions in which the network service traffic weighting is similar among all the nodes in the network because it has a large MEE in other cases. The methodologies provided in this paper are ready to be used by environmental engineers and environmental scientists to assess use-phase emissions and to construct accurate life cycle assessment models for telecom services. Using the models presented in this paper, and subject to cost constraints, the CO2 emissions of telecom services can be minimized by optimizing three factors that have different sensitivities: power usage effectiveness (PUE), equipment energy efficiency improve-

network size increases, the MEE of the top-down model decreases, but the MEE of the top-down model remains large for a large network even under the network condition in which services are concentrated in a small number of nodes. 4.2. Impact of the Mean Service Traffic Proportion. Figure 4 shows the MEEs of the top-down, coarse-grained, and

Figure 4. Sensitivity analysis approach: MEE (%) of the (a) top-down, (b) coarse-grained, and (c) fine-grained models with different service traffic proportions and node service traffic weighting cases.

fine-grained models under various service traffic weighting cases and different mean service traffic proportions using a network size of 50 nodes and an average node degree of 4. As the mean service traffic proportion increases from 10% to 50% in the network, the corresponding service power consumption increases linearly. Furthermore, the number of nodes required to process the service traffic also increases, which directly decreases the skewness of the service traffic weighting distribution. Therefore, the MEE of the three models decreases as the mean service traffic proportion increases from 10% to 50%. However, it should be noted that despite the MEE percentage being low in high mean service traffic regions, the resulting change in Watts (unit for power) or tons (unit for CO2e) can be large. 490

dx.doi.org/10.1021/es303384y | Environ. Sci. Technol. 2013, 47, 485−492

Environmental Science & Technology

Article

backup, and uninterruptable power supplies.37 Reducing the PUE results in lower power consumption overheads, thus reducing CO2 emissions. Second, network equipment energy efficiency is increasing at a rate of approximately 10% to 20% per annum.38 Therefore, replacing legacy equipment with energy-efficient equipment could potentially increase the energy efficiency of telecom services, thus reducing overall network CO2 emissions. Third, telecom providers could reduce CO2 emissions by increasing the use of renewable energy sources to power equipment and associated facilities.

■

ASSOCIATED CONTENT

S Supporting Information *

Additional figures for the schematic of a typical network example, probability distribution of model estimation error, enlarged version of Figure 3 for coarse-grained and fine-grained models, CalREN network node power consumption and utilization, CalREN network topology, additional tables for common parameters used in the assessment models, telecom network equipment characteristics, model-estimation-error for different mean service traffic proportions (20%, 30%, 40%, and 50%), and detailed calculations of CalREN associates’ network equipment energy consumption. This information is available free of charge via the Internet at http://pubs.acs.org/.

■

AUTHOR INFORMATION

Corresponding Author

*E-mail: [email protected]. Phone: +61 3 8344 7682. Notes

The authors declare no competing financial interest.

■

ACKNOWLEDGMENTS This work contributes to the Greenhouse Gas (GHG) Protocol’s Product Standard for ICT Sector Guidance on Telecommunications Network Services Chapter. We thank Mark Shackleton (BT), Glyn Stacey (BT), and Louise Burness (BT) for sharing information on the BT Global MPLS network. We acknowledge The Corporation for Education Network Initiatives in California (CENIC) for sharing CalREN network information. We also thank Rodney S. Tucker (CEET), Kerry Hinton (CEET), Andie Stephens (Carbon Trust), Ewan Stephens (Carbon Trust), Tom Okrasinski (Alcatel-Lucent), Gabrielle Giner (BT), Graham Seabrook (BT), Hugh Barrass (Cisco), and Darrel Stickler (Cisco) for valuable discussions about the various network service CO2 emissions assessment models.

Figure 5. CalREN case study approach: MEE (%) of the (a) topdown, (b) coarse-grained, and (c) fine-grained models with different service traffic proportions and node service traffic weighting cases.

Table 2. Model-Estimation-Error of Top-Down, CoarseGrained, and Fine-Grained Models for (a) CalREN and (b) Hypothetical Network (57 Nodes Simulation Study), Assuming the Network Service of Interest Has a Mean Traffic of 10% of the Overall Network Traffica top-down MEE (%)

+

−

(a) CalREN Case 1 30 −53 Case 2 24 −35 Case 3 12 −14 Case 4 2 −2 (b) hypothetical network Case 1 17 −21 Case 2 14 −18 Case 3 7 −8 Case 4 2 −2

coarse-grained

fine-grained

+

−

+

−

18 14 8 1

−47 −23 −10 −1

13 10 6 1

−36 −18 −7 −1

7 6 3 1

−8 −7 −3 −1

6 5 2 1

−7 −5 −2 −1

■

REFERENCES

(1) Kilper, D. C.; Atkinson, G.; Korotky, S. K.; Goyal, S.; Vetter, P.; Suvakovic, D.; Blume, O. Power Trends in Communication Networks. IEEE J. Sel. Topics Quantum Electron. 2011, 17 (2), 275−284. (2) Baliga, J.; Ayre, R.; Hinton, K.; Sorin, W.; Tucker, R. S. Energy Consumption in Optical IP Networks. IEEE J. Lightwave Technol. 2009, 27 (13), 2391−2403. (3) Gunaratne, C.; Christensen, K.; Nordman, B. Managing energy consumption costs in desktop PCs and LAN switches with proxying, split TCP connections, and scaling of link speed. Int. J. Network Manage. 2005, 15 (5), 297−310. (4) T-systems International GmbH. Green ICT (White Paper); Frankfurt, Germany, 2010. (5) Gartner Inc. Gartner Estimates ICT Industry Accounts for 2% of Global CO2 Emissions. Gartner Press Release, 2007; http://www. gartner.com/it/page.jsp?id=503867.

a Service traffic proportions of 20%, 30%, 40%, and 50% are presented in the SI.

ment, and the use of renewable energy sources. PUE measures the power overheads of central offices due to cooling, battery 491

dx.doi.org/10.1021/es303384y | Environ. Sci. Technol. 2013, 47, 485−492

Environmental Science & Technology

Article

(6) Global e-Sustainability Initiative; Boston Consulting Group. Smart 2020: Enabling the low carbon economy in the information age, 2010; http://www.theclimategroup.org/publications/2008/6/19/ smart2020-enabling-the-low-carbon-economy-in-the-information-age/. (7) Lanzisera, S.; Nordman, B.; Brown, R. E. Data network equipment energy use and savings potential in buildings. Energy Effic. 2012, 5 (2), 149−162. (8) European Commission DG INFSO. Impacts of Information and Communication Technologies on Energy Efficiency, 2008; http://cordis. europa.eu/fp7/ict/sustainable-growth/studies_en.html. (9) Bolla, R.; Bruschi, R.; Davoli, F.; Cucchietti, F. Energy Efficiency in the Future Internet: A Survey of Existing Approaches and Trends in Energy-Aware Fixed Network Infrastructures. IEEE Commun. Surv. Tutorials. 2011, 13 (2), 223−244. (10) Scharnhorst, W. Life Cycle Assessment in the Telecommunication Industry: A Review. Int. J. LCA 2008, 13 (1), 75−86. (11) Mahadevan, P.; Shah, A.; Bash, C. Reducing lifecycle energy use of network switches. In Proceedings of IEEE International Symposium on Sustainable Systems and Technology, Palo Alto, 2010. (12) Raghavan, B.; Ma, J. The Energy and Emergy of the Internet. In Proceedings of the 10th ACM Workshop on Hot Topics in Networks, Cambridge, 2011. (13) AT&T Inc. AT&T Inc. 2011 Annual Report; http://www.att. com/Common/about_us/files/pdf/ar2011_annual_report.pdf. (14) Verizon. Verizon’s 2011 Annual Report; http://responsibility. verizon.com/assets/docs/verizon-annual-report-2011.pdf. (15) The Corporation for Education Network Initiatives in California (CENIC)’s California Research and Education Network Website; http://www.cenic.org/calren/. (16) Toffel, M. W.; Horvath, A. Environmental Implications of Wireless Technologies: News Delivery and Business Meetings. Environ. Sci. Technol. 2004, 38 (11), 2961−2970. (17) Kitou, E.; Horvath, A. Energy-Related Emissions from Telework. Environ. Sci. Technol. 2003, 37 (16), 3467−3475. (18) Kitou, E.; Horvath, A. Transportation Choices and Air Pollution Effects of Telework. J. Infrastructure Sys. 2006, 12 (2), 121−134. (19) Kitou, E.; Horvath, A. External Air Pollution Costs of Telework. Int. J. LCA 2008, 13 (2), 155−165. (20) Matsuno, Y.; Takashi, K. I.; Tsuda, M.; Nakamura, J.; Nishi, S. Eco-Efficiency for Information and Communications Technology (ICT): The State of Knowledge in Japan. In Proceedings of the 2007 IEEE International Symposium on Electronics & the Environment, Orlando, 2007. (21) Williams, E. D. Assessing the potential of telecommuting as an energy savings technology in Japan. In Proceedings of the 2003 IEEE International Symposium on Electronics & the Environment, Boston, 2003. (22) Takahashi, K. I.; Tatemichi, H.; Tanaka, T.; Nishi, S.; Kunioka, T. Environmental Impact of Information and Communication Technologies Including Rebound Effects. In Proceedings of the 2004 IEEE International Symposium on Electronics & the Environment, Phoenix, 2004. (23) Takahashi, K. I.; Tsuda, M.; Nakamura, J.; Matsuno, Y.; Adachi, T. Social Impact of ICT services − Different Impact in Different Countries. In Proceedings of the 2008 IEEE International Symposium on Electronics & the Environment, San Francisco, 2008. (24) Nakamura, J.; Tsuda, M.; Takahashi, K. I.; Nishi, S.; Takaoka, M. Analysis of the Potential Contribution of ICT Services to a Sustainable Society. In Proceedings of the 2006 IEEE International Symposium on Electronics & the Environment, San Francisco, 2006. (25) Seetharam, A.; Somasundaram, M.; Towsley, D.; Kurose, J.; Shenoy, P. Shipping to Streaming: Is this shift green?. In Proceedings of the 1st ACM SIGCOM workshop on Green Networking, New York, 2010. (26) Department for Environment Food and Rural Affairs. 2012 Greenhouse Gas Conversion Factors for Company Reporting, 2012; http://www.defra.gov.uk/publications/2012/05/30/pb13773-2012ghg-conversion/. (27) International Energy Agency. CO2 Emissions from Fuel Combustion 2011, 2011; http://www.iea.org/publications/ freepublications/publication/name,4010,en.html.

(28) Phillips, S.; Woodward, S. L.; Feuer, M. D.; Magill, P. D. A regression approach to infer electricity consumption of legacy telecom equipment. ACM SIGMETRICS Perform. Eval. Rev. 2010, 38 (3), 61− 65. (29) Hinton, K.; Baliga, J.; Feng, M.; Ayre, R.; Tucker, R. S. Power Consumption and Energy Efficiency in the Internet. IEEE Network 2011, 25 (2), 6−12. (30) Chabarek, J.; Sommers, J.; Barford, P.; Estan, C.; Tsiang, D.; Wright, S. Power Awareness in Network Design and Routing. In Proceedings of the 2008 IEEE Conference on Computer Communications, Phoenix, 2008. (31) Mahadevan, P.; Sharma, P.; Banerjee, S.; Ranganathan, P. A Power Benchmarking Framework for Network Devices. In Proceedings of the 2009 International IFIP-TC 6 Networking Conference, Berlin, Germany, 2009. (32) The California Energy Commission Website; http:// energyalmanac.ca.gov/overview/energy_sources.html. (33) Watts, D. J.; Strogatz, S. H. Collective dynamics of ‘small-world’ networks. Nature 1998, 393, 440−442. (34) British Telecom. BT Global Multi-protocol Label Switching Network equipment data given by British Telecom researchers. (35) The Greenhouse Gas Protocol Initiative. GHG Protocol Product Life Cycle Accounting and Reporting Standard ICT Sector Guidance, 2012; http://www.ghgprotocol.org/feature/ghg-protocol-product-lifecycle-accounting-and-reporting-standard-ict-sector-guidance. (36) Quoitin, B. Topology Generation Based on Network Design Heuristics. In Proceedings of the 1st International Conference on Emerging Networking Experiments and Technologies, Toulouse, 2005. (37) Belady, C.; Rawson, A.; Pfleuger, J.; Cader, T. Green Grid Data Center Power Efficiency Metrics: PUE and DCIE (White Paper); The Green Grid, Beaverton, OR, 2008. (38) Neilson, D. T. Photonics for Switching and Routing. IEEE J. Sel. Topics Quantum Electron. 2006, 12 (4), 669−678.

■

NOTE ADDED AFTER ASAP PUBLICATION There was a formula error in the third paragraph of section 2.3 of the version of this paper published December 12, 2012. The correct version published December 18, 2012.

492

dx.doi.org/10.1021/es303384y | Environ. Sci. Technol. 2013, 47, 485−492