Methodology To Model the Energy and Greenhouse Gas Emissions of

Nov 22, 2011 - Energy Star(15) publishes measured energy consumption data on specific enterprise servers based on different component configurations. ...
0 downloads 11 Views 3MB Size
ARTICLE pubs.acs.org/est

Methodology To Model the Energy and Greenhouse Gas Emissions of Electronic Software Distributions Daniel R. Williams* and Yinshan Tang Technologies for Sustainable Built Environments Centre, University of Reading, Reading RG6 6AF, United Kingdom

bS Supporting Information ABSTRACT: A new electronic software distribution (ESD) life cycle analysis (LCA) methodology and model structure were constructed to calculate energy consumption and greenhouse gas (GHG) emissions. In order to counteract the use of high level, top-down modeling efforts, and to increase result accuracy, a focus upon device details and data routes was taken. In order to compare ESD to a relevant physical distribution alternative, physical model boundaries and variables were described. The methodology was compiled from the analysis and operational data of a major online store which provides ESD and physical distribution options. The ESD method included the calculation of power consumption of data center server and networking devices. An in-depth method to calculate server efficiency and utilization was also included to account for virtualization and server efficiency features. Internet transfer power consumption was analyzed taking into account the number of data hops and networking devices used. The power consumed by online browsing and downloading was also factored into the model. The embedded CO2e of server and networking devices was proportioned to each ESD process. Three U.K.-based ESD scenarios were analyzed using the model which revealed potential CO2e savings of 83% when ESD was used over physical distribution. Results also highlighted the importance of server efficiency and utilization methods.

1. INTRODUCTION Electronic software distribution (ESD) has increased dramatically over the past 15 years,1 but little research into the energy consumption and greenhouse gas (GHG) emissions of this change has been published. For this research, we define ESD as the transmission of software to an end user device using the internet. Many online software retailers are offering ESD as an alternative to traditional physical distribution routes.1 Some may assume that ESD consumes less energy and thus emits less GHG compared to traditional distribution; however, little scientific assessment has been undertaken to ratify this. Those who claim ESD as green could be viewed as practicing what is known as “green washing”, claiming environmental credentials without substantiation. It is therefore essential to scientifically quantify the energy impact that the internet and its associated services have upon the environment in terms of ESD. Various life cycle analysis (LCA) studies26 indicated that e-commerce product distribution alternatives could reduce GHG emissions when compared to a physical commerce process. However, the studies also concluded that there was a large variance in model results making it difficult to measure the exact impact. Commercial organizations7,8 have developed simple LCA models demonstrating a ∼90% reduction in GHG emissions when using ESD instead of physical distribution. However, more attention and detail are needed to counteract the use of high level variables and assumptions. For example, some models2,7,8 which have similar aims to this study took a top-down approach and calculated energy consumption from data that had been inferred from r 2011 American Chemical Society

nonspecific devices with unspecified utilization levels. Additionally, individual hardware and data routes were not taken into account. Top-down approaches which decompose broad level data are commonly quicker to perform and easier to collect data for. However, they often have very high levels of uncertainty and cannot be applied to other situations. More importantly, the models do not analyze the efficiency of servers or network devices. These factors need to be addressed further. As with all LCA modeling, assumptions are made and uncertainties justified. Gard and Keoleain9 demonstrated how sensitive modeling parameters can be, and how they can wildly influence results. A common issue of setting different analysis boundaries26 makes comparisons difficult or irrelevant. A universal methodology and approach would go some way to alleviate this issue. The aim of this research is to develop a methodology to calculate the energy consumption and GHG emissions of ESD in terms of carbon dioxide equivalent (CO2e). The methodology acts to create a set of boundaries facilitating comparison between ESD and related alternative physical distribution methods. Previous modeling7,8 will be used as a foundation for this research. A detailed bottom-up methodology will be pursued which seeks to assess direct energy consumption from hardware used, thus reducing uncertainty. The primary deliverable will be a method Received: June 22, 2011 Accepted: November 22, 2011 Revised: November 21, 2011 Published: November 22, 2011 1087

dx.doi.org/10.1021/es202125j | Environ. Sci. Technol. 2012, 46, 1087–1095

Environmental Science & Technology that can be used across different ESD scenarios. The research will identify and highlight key variables and parameters to focus effort upon, minimizing variability and sensitivity issues. A model based upon the developed methodologies will be created and tested against three U.K. based distribution options offered by the Microsoft Store U.K.11 (MSS). The internationally recognized specification PAS 2050,10 used in the assessment of life cycle GHG emissions of goods and services, has no guidance or methodology for ESD. Therefore, this research aims to deliver a methodology that can be used in conjunction with PAS 2050, thus enabling valid comparisons. In order to create a relevant methodology, this research utilized operational and product data from the MSS online shop. Access to the MSS provided the opportunity to develop realistic methodology boundaries and provided data to model three online distribution scenarios. Some data collected from the MSS is highly confidential and cannot be published although model results are published. This scope for ESD analysis is defined in section 2. Each ESD methodology calculation and variable is explained in section 3. Section 4 describes the data used within the ESD scenario analysis. Section 5 explains the method used to model the physical distribution parts of the scenarios. Section 6 highlights the results of the scenario analysis, and section 7 provides an overall discussion. The Supporting Information document contains extra details, explanations, and an abbreviation list.

2. METHODOLOGY SCOPE The methodology is based upon three purchasing scenarios used by the MSS. The scenarios act as boundaries for the methodology and subsequent model. Established methodologies12 will be utilized for relevant physical distribution parts of the scenarios: scenario 1, a user purchasing and downloading a software package from an online store; scenario 2, a user purchasing from an online store but opting for a physical software package, sent by post; scenario 3, a user purchasing online, downloading the software but getting a physical backup DVD disk sent by post. In order to create an ESD methodology, this research performed a “cradle to grave” LCA13 of the MSS distribution scenarios. The functional unit was defined as the purchasing and distribution of one software program from an online store. This analysis allowed the formation of the following logical sections: data center, internet transfer, user device, and embedded carbon. In each section key variables of the distribution were then calculated. The MSS serves the European and worldwide market; however, this assessment focused upon services provided within the United Kingdom. The MSS has a wealth of resources and expertise which allow it to have an efficient distribution process and thus was viewed as a state of the art scenario. Figure S1 (please refer to Supporting Information) provides a basic overview of the three scenarios being analyzed. All scenarios begin with a user browsing and successfully purchasing software from the online store. The online store uses a website hosting server and servers that host customer relationship management (CRM) and finance systems. The user accesses the website via an internet enabled device, and after the successful purchase of software, the distribution method is selected. Downloading software (scenario 1) involves the use of download fulfillment servers that host and manage the transfer of software to a user device. Physical software distribution (scenario 2) involves the creation and subsequent transportation of packaging

ARTICLE

and a DVD disk to the user. Scenario 3 involves both physical and digital routes; however, the physical product contains less packaging. All processes conclude with an end of life analysis; this specifically includes servers (both browsing and fulfillment) and the packaging and media materials for physical products.

3. DEVELOPED ESD METHODOLOGY The ESD methodology consists of calculating the energy consumed at each stage of the distribution scenario. Process modeling was undertaken on each MSS scenario and is presented in Figure 1. In order to identify devices used and data routes, a step by step audit of the digital systems and processes used was undertaken. We defined a typical “download” and “browse” task from MSS user statistics, and all device touch points were noted. Four main sections of analysis were highlighted: the data center devices used, the transfer of data across the internet, the device operated by the user, and the embedded carbon of the data center devices. These sections form the basis of the presented methodology and each contains data variables and calculations related to the process mapping. Figure 1 highlights the data required and calculations undertaken with reference to the paper's sections. The variability in data accessibility in some sections resulted in more than one method being required to calculate energy consumption. The model used to calculate each of the MSS scenarios was created using Microsoft Excel. It comprised each of the variables and calculations laid out in Figure 1. 3.1. Data Center. The data center section quantified the amount of power consumed by data center devices that host the online store’s website and fulfills the software download, in regards to the analyzed scenario process. A basic data center process map (see Figure S2 in Supporting Information) highlights the devices that were considered in this section for the MSS. Any online process will have a unique data flow, and thus, a process audit must be undertaken. It is, however, common for an online store to have separate servers that directly interface with the internet to host the website and others that fulfill the software download. Additional auxiliary servers are used for CRM and transaction, order, and financial management. Also, storage servers devoted to fulfilling downloads are common. Additionally, each system may be integrated onto one high-end server or numerous volume servers. These additional networking and storage devices may be difficult to account for; this is resolved via a network or storage equipment to server ratio in section 3.1.1. We calculated a per device power consumption (kW) per process via eq 1 device kW process ¼ SEP  SPRU  PUE  AA  seconds to transmit data

ð1Þ

where SEP is the total server equipment power (section 3.1.1) used per server device, SPRU is server process utilization (section 3.1.2), PUE is power usage effectiveness (section 3.1.3), and AA is access attempts (section 3.1.4). To calculate the overall power consumption per process, all data center device results were added together. Each of the variables is explained in the following sections. 3.1.1. Server Equipment Power. In order to determine the power consumed by each server, hardware specifications of each model were required. The ideal method to measure power consumption is via a direct power monitoring tool. As experienced in previous studies,7,8,14 attempting to define or measure 1088

dx.doi.org/10.1021/es202125j |Environ. Sci. Technol. 2012, 46, 1087–1095

Environmental Science & Technology

ARTICLE

Figure 1. Process map of ESD indicating the data required and calculations used to calculate overall CO2e per ESD process. 1089

dx.doi.org/10.1021/es202125j |Environ. Sci. Technol. 2012, 46, 1087–1095

Environmental Science & Technology

ARTICLE

the exact server model being employed was difficult because of the equipment turnover, center size, confidentiality, and high security associated with modern data centers. For the MSS scenario only some server models were definable, and none were directly measured. For the majority of the remainder of servers only basic component information was available, and some servers were not definable. Therefore, the following section describes the methods used to determine power consumption for each case. Where the server model was unknown, an average server was defined using approximate information on particular server components. For the purpose of this research, we wanted to focus on certain components that most affected power consumption to reduce research time. Therefore, components which have a greater effect on power consumption were investigated. Energy Star15 publishes measured energy consumption data on specific enterprise servers based on different component configurations. Following Energy Star’s measurement method, it was found that the key components that affect energy consumption are the quantity and speed of processor cores, the capacity and quantity of on board hard disk drives, and system memory size. A statistical correlation test (see Table S1 in Supporting Information) on the Energy Star15 component data revealed that the amount of system memory and the number of hard drives have a greater impact on power consumption. From the statistical test, little variance within the processor or hard disk capacity data occurs because of industry standardization, and thus, it is important to understand these components. Servers are commonly grouped into three categories, volume, midrange, and high-end,14 based upon price and technology levels. Using these server categories and focusing upon the researched power affecting server components made it possible to select servers from the most popular manufacturers (IBM, HP, and Dell) and compile an average server model. Power consumption for servers that had either been definable or averaged was determined using the following method. Manufacturer server power consumption profiles were found to contain different reported power consumption types. Koomey14 describes four popular types that server manufacturers publish power consumption; server typical power (TP), server maximum measured electricity (MME), power supply maximum rated input power (MRIP), and power supply output power (OP). The most accurate being MME as it is directly related to the server’s individual hardware setup and was preferred in this research. Where the MME type was not available, methods for calculating the MME from TP, MRIP, or OP consumption types were derived from Koomey.14 The calculation array used to convert to the MME value is described within Table S2 of the Supporting Information. In most instances a server includes associated networking and storage devices which were accounted for via a network equipment ratio (NER) and storage equipment ratio (SER). Adding these to the MME resulted in overall SEP (eq 2). In some cases the ratios were not required as data on the storage sever was available and thus was calculated as a separate server device. SEP ¼ MME  ½ðNER þ SERÞ  1

ð2Þ 7

A NER and SER of 1.15 as calculated by WSP was determined to be a best fit value to the MSS scenario because of similarities in the data centers studied. WSP7 calculated the redundant equipment energy after accounting for server energy in a live data center environment. Further analysis of a range of studies determined that a range of values between 1.0517 and 1.4014 could be used as NER and SER depending upon scenario. The EPA17

suggests that if the data center or process being analyzed is data storage intensive, then a higher value needs to be applied to the SER. Similarly, if the process being analyzed is focused on quick and efficient networking speeds, then a higher value is placed onto the NER. Little bottom-up work in this area has been completed, and thus, the SER and NER values have less certainty. In the case where no server detail could be sourced, an average MME value based upon the top three shipped server models of 201016 was calculated at 0.49 kW for volume and 1.48 kW for midrange servers. This resulted in SEP values (kW) of 0.64 kW for volume and 1.93 kW for midrange servers (see Koomey14 for high-end analysis). Hardware manufacturer tools were used to estimate power consumption based upon the maximum hardware setup possible per server. 3.1.2. Server Process Utilization. The power consumption of each server can be affected by the level of power utilization, the degree of virtualization and the resource demand from the OS, and software used to support the process. The process being analyzed was focused upon to take account of nonscenario processes being run within each of the factors listed. Therefore, each of these factors was accounted for via eq 3. SPRU ¼ SPU  size of VS  ð% of VS by OS þ % of VS by softwareÞ

ð3Þ

Here, SPU is server power utilization (SPU) and VS refers to a virtual server. SPRU is calculated by multiplying SPU by the relative size of the VS session (% of SEP) running the process and by the addition of the OS size (% of VS session) and the supporting software size (% of VS session). In most situations the sum of the OS size and supporting software size came to 1; however, this was included to account for VSs that were running other software not specific to the task being analyzed. Auditing and modeling power utilization levels become difficult tasks due to variances in utilization and demand over time. This research defined two types of utilization: server utilization and server power utilization (SPU). A server at an idle power state may not be utilized by any meaningful process resulting in zero server utilization. SPU was defined as the entire power load of a server. Increasing server utilization will increase SPU.18 SPU can be obtained by measuring the overall power consumption of the server during a time period that represents average server use and dividing this by the MME value. Measurement was not possible; thus, average SPU values (suitable for MME values) of 48% for volume and midrange servers and 79% for high-end servers were derived.14 Koomey14 lists average utilization values according to the type of power rating method being used, see Table S2 of the Supporting Information. Features such as virtualization and increased power management have increased the efficiency of software;19 this research assumed a 20% increase in the SPU reported by Koomey.14 The average SPU value requires more indepth investigation and thus represents an area of high uncertainty. Virtualization is a technique which maximizes resource usage by creating multiple servers on one physical server. We therefore needed to identify the power consumption, as a % of SEP, of the VS running the analyzed process. Apportioning a specific VS size is unrealistic as virtualization can flexibly change resource requirements. VS size measurement is possible via monitoring over an average use time; this is rare data. VSs, however, are often setup with resource boundaries, for example not to use more than 50% of system memory or not to use more than 2 out of 24 processor cores. Therefore, to combat the lack of VS measurement 1090

dx.doi.org/10.1021/es202125j |Environ. Sci. Technol. 2012, 46, 1087–1095

Environmental Science & Technology

ARTICLE

data, the VS size was derived using resource boundary data. For example, one server may host five VSs, with four each using 12.5% and one using 50% of total system resources. Using boundary data is not robust as particular hardware components will consume more power than others resulting in a possible skew of the system resource level. Within each server or VS, an OS such as Windows Server 2008 R2 supports numerous processes that fulfill external requests. The OS and software used to run the analyzed process were viewed as a base load and thus accounted for. This research assumed, from monitoring similar systems, that the OS consumed approximately 20% of system power. This was increased to 25% for a VS to take into account virtualization management software. Attributing a size to the software used to support the process was completed by monitoring and measuring an OS’s processes over time. 3.1.3. PUE. PUE20 (eq 4) was included to account for any extra cooling and maintenance power loads used for the server. PUE (also known as data center infrastructure efficiency (DCiE)) is an industry standard measure which can be defined as PUE ¼

total facility power IT equipment power

ð4Þ

Energy Star21 collated 61 different PUE values that fell into the range 1.363.59, with a mean average of 1.92. Some larger data centers report low PUEs of 1.1622 which can be classed as state of the art. 3.1.4. Time Accessed and Access Attempts. To calculate the total power consumed by each server the total use time of each server and the amount of times each access was attempted were used. The engagement time(s) of a server is dependent upon the total size and content of data being accessed, across all data requests. For example, web pages are commonly small in size and do not require a constant connection, whereas video streaming may require a constant and more reliable stream. For this research, AA and three access activity times were used: browsing and downloading time, and the time that auxiliary servers were engaged. The process of browsing utilizes a web server engaged in fulfilling each page as it is requested. To calculate total browsing time, an average website visit and product purchase were assessed for pages visited and time spent on each. The size of each page, required for the internet transfer calculations, was also calculated using Microsoft Internet Explorer’s developer tools. The web server will not be engaged in data processing for the total browse time, because the user will be digesting each page’s information. For the MSS scenarios, an assumption was made that a web server was actively engaged for 50% of the total user browse time. Although this method asserts that the webserver would be idle for 50% of time, we recognize that others would be using the server, which the SPRU accounts for. The time that the download fulfillment server is engaged for was dependent upon the file size and network infrastructure (calculated in section 3.2). For auxiliary services such as customer management and finance, an average use time of 2 min was used. This time was determined via analysis of MSS11 servers. When data packets are sent over a network, errors and network issues are to be expected. In the case that data cannot be repaired on route, the data packet can be rerequested from the downloading machine. An average AA value of 1.2 and 1 for the web

facing and internal transaction servers, respectively, was derived from MSS11 statistics. 3.2. Internet Transfer. This research accounted for power consumed when data packets are transmitted over the internet from a data center to the end user. It is virtually impossible to model the actual path a packet of data takes across the internet because of the multitude of possible routes, network conditions, and specific network management. To estimate total power consumed an approximation of the number of hops that a packet of data takes was considered. At each hop the type and power consumption of each networking device was considered and assessed and a proportion of device power attributed to the transmitted data packet (device power per transmission (DPT)). Device transmission consumption (kW) (eq 5) was calculated using total transmission time, the devices' energy consumption, DPT and AA. Combined device transmission consumption was calculated from the sum of all devices used (or hops taken). device transmission consumption ¼ transmission time ðsÞ  device energy ðkWsÞ  DPT ð%Þ  AA

ð5Þ

Internet infrastructures vary by geographical location; this research analyzed a U.K. network infrastructure. A graphical illustration of the various steps that a packet of data takes across the internet to reach an end user is presented in Figure S3 of the Supporting Information. The following network data was sourced with assistance from a director of a U.K. broadband company.23 Not all steps and devices were accounted for because of the lack of standardization across networks; however, the basic layout of network connections presented here can be refined and detailed in future studies. 3.2.1. Data Packet Routes and Hops. When a connection between a server and end user device has been established, data center servers often use advanced routing and switch devices to direct data to the wider internet backbone. The internet backbone commonly consists of optical connections and switch devices and utilizes interconnections between national and international networks. Data is then routed via a national scale internet service provider (ISP) data center to a local scale “colocation” facility, such as a telephone exchange for DSL technology. The colocation facility transmits data via various methods such as copper based DSL (ADSL, VDSL, etc.) or fiberoptic connections (gigabit passive optical network (GPON)). Data is finally transmitted to the user device via a user network which can utilize a variety of devices. The average number of hops taken from a data center to a selected destination was estimated by using Microsoft Windows trace route software. The trace provides basic information on the different hops that data completes; Table S3 of the Supporting Information shows example destinations and the number of hops taken. A trace can also provide information on the IP address of the device used to switch the data packet, and thus, detailed investigations can be carried out using this method. 3.2.2. Device Power. The power consumption of each network device was researched and converted to kW/s using the data packet route information and estimated number of hops. Network size and confidentiality dictated that it was not possible to obtain actual hardware models and types. Consequently, average power consumptions (W) from popular manufacturers were researched for each transfer stage of the data packet routing (see Table S4 of the Supporting Information for results). 1091

dx.doi.org/10.1021/es202125j |Environ. Sci. Technol. 2012, 46, 1087–1095

Environmental Science & Technology

ARTICLE

3.2.3. Device Power per Transmission. The percentage of the device’s power that the data transmission consumed (DPT) was calculated via eq 6. This involved calculating each device's average data throughput and an average network data rate. DPT ð%Þ ¼

data rate ðMbit=sÞ loaded data throughput ðMbit=sÞ

ð6Þ

For each device, the average loaded data throughput (Mbit/s) was calculated by scaling its maximum data throughput by an average network loading value. Each network and thus each device will have network loading values according to location, bandwidth, and technology employed. For example high loading is encountered at peak usage times. However, specific data on the loading of a network was not available as a result of confidentiality considerations. Loading values for switch and router end user devices were set at 60% and 25%.18 Lower loadings were used for end user devices as local traffic is often less complex and dense. Data overheads24 dedicated to network operators were removed as this bandwidth would be unusable by data transmission. In order to determine a U.K. average network data rate (Mbit/s), the average U.K. broadband speed of 6.2 Mbit/s25 was used. This value varies by Internet connection type, distance, and service provider limitations. However, if, for example, an older dial-up modem was utilized then the data rate would be around 0.05 Mbit/s, while GPON connection rates would be higher.23 3.2.4. Transmission Time. The time taken to transmit data from a server to an end user device was calculated by dividing the total size of the data packet (Mbit) by the network data rate (Mbit/s). This calculation is shown in eq 7. transmission time ðsÞ ¼

file size ðMbitÞ data rate ðMbit=sÞ

ð7Þ

3.3. User Device. The power consumed by the device used to browse and download from the online store was calculated. This research did not single out a network component but instead incorporated the entire device including visual display unit. An array of devices can connect to the internet, and thus, power consumption of each device differs. Power consumption was identified in two activity states, browsing and downloading. In order to obtain activity state measurements accurately, the device’s power consumption (W) should be measured at the desired state over a representative time use. Accurate measurements were not available, and thus, the subsequent method was used. Power consumption was inferred by applying an activity state loading value (%) to the energy difference (W) between maximum and idle modes and adding this to the idle mode value. Using Energy Star26 data, the average power consumption of various device types in idle mode were calculated. Maximum power consumption was calculated by applying an idle to maximum ratio of 2.0 (see Table S5 of the Supporting Information). Activity state loading values for internet browsing and downloading data used in this research were ∼10% and ∼8%, respectively (for a Microsoft Windows 7 based PC). These values were obtained by averaging activity state CPU utilization from the device’s task monitor over a representative time period. CPU was used as a consequence of it commonly consuming proportionally more power per device than other components. It was recognized that these values would vary by different computer hardware and

software. The energy consumed by a visual display unit was also accounted for. The activity state power consumption (kW/s) was multiplied by the time (s) of the average browsing session and by the data transmission time (Section 3.2.4), respectively. This research attributed all idle power to the assessed activity. This therefore does not take into account that multiple activities or virtualization can be undertaken simultaneously on a device thus making hardware more efficient. 3.4. Embedded Carbon. This research included an apportionment of each server and network device’s embedded CO2e. Embedded CO2e can be defined as the total carbon consumed during the lifetime of a product. Including embedded carbon is not a permitted function of the PAS 2050 methodology. However, this research proposes that server and networking devices are an integrated and ultimately vital part of the electronic distribution process; without them there is no other way to distribute the electronic data. Embedded CO2e (kg) per device was calculated using eq 8. Each device was then added together to gain an overall embedded value. device embedded CO2 e per transmission ¼ device CO2 e   device use time   SPRU or PDT  ðNER þ SERÞ device lifetime

ð8Þ The device’s embedded CO2 e (kg) was multiplied by the previously calculated device utilization level (SPRU for servers and DPT network devices), and for the server by the NER and SER ratio to account for the extra equipment. This was then multiplied by the percent of the device’s lifetime that was being used in the analyzed data transmission. This required the already calculated device use time (s) being divided by the device's overall lifetime (s). Average device lifetimes (s) were sourced from hardware manufacturers. Although difficult to source, information on each server and networking device’s embedded CO2e (kg) was found from hardware manufacturers and commercial and academic LCA studies. Embedded CO2e values included transport, process energy, materials, and end of life where possible; see Table S6 of the Supporting Information for further detail. 3.5. Electricity Emission Factors. In order to relate each section’s calculated energy (kW) to CO2e (kg), a country specific CO2e electricity emission factors (including generation and distribution losses) based upon a unit (kWh) of consumed electricity were sourced. Emission factors vary by country due to the power generation plants used, losses, and fuel type, meaning that for each stage of the model the geographical location influenced the level of CO2e impact. This research used emission factors for the U.K. and Germany at 0.54 and 0.58 kg CO2e/kWh, respectively.27,28

4. ESD SCENARIO MODEL DATA Primary data presented within section 9 of the Supporting Information was used within the developed Excel model. Where primary data was unavailable, secondary data, set out in this methodology, was used. Some data was confidential and thus not available for publishing. 1092

dx.doi.org/10.1021/es202125j |Environ. Sci. Technol. 2012, 46, 1087–1095

Environmental Science & Technology

ARTICLE

Table 1. Scenario 1 (ESD Only) Results ESD

kg CO2e

% of total CO2e

process stage

technology group

a1, hosting servers

0.0944

33.63%

browsing

server

a2, fulfillment servers

0.1292

46.02%

download

server

b1, internet transfer

0.0118

4.21%

download

network

b2, internet transfer

0.0001

0.02%

browsing

network

c, user download

0.0261

9.31%

download

user device

d, user browsing

0.0141

5.01%

browsing

user device

transfer

network and server

e, embedded

0.0050

1.80%

total

0.2807

100%

5. PHYSICAL DISTRIBUTION Scenarios 2 and 3 involved a software package being sent via physical distribution. Consequently, a separate model was used to calculate GHG emissions. Calculating physical product carbon footprints is commonplace using the PAS 2050 method. Therefore, the PAS 2050 methodology10 was utilized which analyzes materials, distribution, and end of life. As scenario 2 involved the use of the online store to browse and order the software, the ESD methodology was added (browsing and hosting only). Scenario 3 involved all ESD scenario activities, and thus, scenario 1 was added to its physical model results. 5.1. Materials. Every material used within the packaging and software was identified to assess the impact of materials used in the physical distribution scenarios. Each material component was related to a material and processing emission factor to determine the CO2e impact (see Table S8 of the Supporting Information). Sourced material emission factors took into account the individual material’s lifecycle until it was ready to be processed into packaging or software. The processing emission factor took into account the energy consumed to transform the material into the final product. A specific emission factor for a DVD could not be sourced, and thus, an average was calculated2,7,8,29 at 0.231 kg CO2e per DVD (including DVD material and processing). Eight of the most popular MSS products were analyzed. By obtaining the product's confidential official component bill of material (BOM) and cross checking with physical copies, the packaging and shipping materials were componentized and accurately weighed. Shipping materials were included within the BOM and factored in a proportion for one software package within the whole shipper. 5.2. Distribution. The entire distribution chain was accounted which included the transportation of materials from creation to packaging manufacturing location and the final product to the end delivery point. Each stage required the following key data: relative origin to destination distance, mode of transport, fuel type, and product or material density and weight. Components sent via marine shipping required the average shipping container size. The final calculation for each stage in the chain was completed using a precise equation12 with U.K. specific fuel and transport emission factors.28 5.3. End of Life. The end of life (EOL) model section calculated emissions resulting from wastage and recycling. This section used a complex formula12 which required the weight and useful lifespan of each component from the package. The calculation took into account a credit for biogenic carbon and includes U.K. recycling, incineration, and landfill rates, and calculations. The transport of waste materials was also taken into account using section 5.2.

5.4. Online Browsing. The amount of energy consumed when a user browsed the online store was calculated using hosting servers (section 3.1), internet transfer (section 3.2), and user device browsing (section 3.3).

6. SCENARIO MODELING RESULTS Results for the three MSS distribution scenarios were calculated (see Tables S9 and S10 of the Supporting Information). ESD download (Table 1) emitted 83% and 55% less CO2e than physical distribution and physical and ESD distribution, respectively. Scenario 1, ESD download only, emitted 0.28 kg CO2e. The largest emission source was the downloading processes which accounted for 59.54% of total CO2e; browsing accounted for 38.66%. By technology grouping, server use accounted for 79.65%, user device use at 14.32%, and network use at 4.24%. Embedded CO2e emitted only 1.80% of total emissions. For scenario 2, physical distribution yielded 1.68 kg CO2e. Materials emitted 84.30% of the total CO2e, with PET plastic packaging accounting for 61% of the material total (see Figure S4 of the Supporting Information). Distribution and EOL emitted 7.89% and 1.31%, respectively, and browsing and hosting activities accounted for 6.50% of total CO2e emitted. For scenario 3, physical distribution of a smaller package along with ESD download emitted 0.63 kg CO2e. Packaging materials, distribution, and EOL accounted for 53.12%, 1.51%, and 0.87% of total CO2e. The ESD results did not change from scenario 1, and all ESD activities accounted for 44.51% of the total emissions. 7. DISCUSSION This research has presented a new methodology to calculate the energy consumption and GHG emissions (CO2e) of ESD, which takes into account devices used by the data center, internet transfer, and user, as well as the embedded CO2e of server and network devices. The methodology was created from a process mapping assessment of three distribution scenarios from the MSS online retailer. A physical software distribution scenario was also analyzed using the PAS 2050 methodology, listing key boundaries and variables. An Excel based model validated the methodology which analyzed each of the three MSS distribution scenarios. The modeling results indicated that, by utilizing ESD over physical distribution, GHG emissions can be reduced by up to 83%. MSS scenario modeling revealed that the most energy intensive section of the ESD scenario was the use of servers for fulfillment and hosting services. A focus was thus placed upon defining servers according to their type and utilization levels. Average server models were defined because gaining data on 1093

dx.doi.org/10.1021/es202125j |Environ. Sci. Technol. 2012, 46, 1087–1095

Environmental Science & Technology specific server hardware was difficult. The scenario analysis highlighted that less product packaging could significantly reduce CO2e by 76% (materials only). These results indicate an environmental benefit for both the use of ESD and reduced software packaging. Server utilization levels were accounted for, making this methodology unique and increasing its accuracy and transferability to real world usage compared to previous studies.7,8,14 The server utilization factors included were power supply efficiency, virtualization levels, OS and process size, and server use time. Servers and data centers that have high power consumption and utilization levels can now be assessed on a per process basis in preference to a per server approach or by PUE alone. This could be important when, for example, “cloud” and “streaming” services are analyzed. The power consumption of transmitting data packets across the internet was included. A focus was placed on the distance and number of hops that data was transmitted. Although device power consumption and loading levels were assumed for each hop, this section can provide a foundation for future studies. This research argued that because a server’s usage can be accounted for accurately, embedded CO2e could be included. Scenario 1 results revealed that the embedded CO2e accounted for only 1.80% of overall GHG emissions. This highlights that ESDs emissions come mainly from operational use. Reporting impact in terms of CO2e electricity factors may not fully indicate the environmental and indeed social and economic impacts of each scenario. For example, dematerialization of a physical supply chain may remove transportation vehicles and factories used to create materials. However, an increase in the manufacture of technology devices may occur. Economic impacts of ESD must also be assessed to understand the potential rebound effect. Potential economic (carbon or currency) efficiencies could allow more products to be sold, thus increasing overall energy use. However, in terms of distribution efficiency, software demand may not change significantly as software is commonly only purchased once. Software development was not included; nevertheless, it should be explored in future studies to reveal its environmental impact and to assess if distribution is significant in comparison. It is also recognized that CO2e electricity factors do not account for potential secondary impacts of power generation technologies. A goal of this research was to provide a foundation methodology upon which other ESD scenarios can be analyzed; this has been achieved. The research highlights devices that have been included via an analysis of the processes involved which are unique to the MSS. Although unique, alternative scenarios will include similar devices and data routes and thus the calculation methods described here are applicable to other ESD scenarios. Software varies in size and scenarios with large file sizes would potentially increase GHG emissions. This may be negated in the future because of an increasing focus by governments and network providers to deliver more efficient networks. Therefore, further scenario studies need to be carried out against this methodology to further develop and refine the applicability of the method. Uncertainties are present throughout the methodology. However, this research is intended as a foundation to be developed further. The highest power consumption of the ESD scenario arose from the use of servers. Therefore, both the use time of each server and the reliance upon the server CPU as a guide for overall power utilization are emphasized as having the largest levels of uncertainty. These areas would benefit from further

ARTICLE

attention. In particular, power utilization could be improved by either direct component utilization measurements or identification of the main components being utilized by an analyzed process.

’ ASSOCIATED CONTENT

bS

Supporting Information. Additional details, figures, and tables. This material is available free of charge via the Internet at http://pubs.acs.org.

’ AUTHOR INFORMATION Corresponding Author

*Phone: +44 118 909 5562; e-mail: [email protected].

’ ACKNOWLEDGMENT The lead author is supported by an EngD studentship provided by the U.K. Engineering and Physical Sciences Research Council (EPSRC) and Microsoft (U.K.) Ltd. Microsoft (U.K.) Ltd provided all data relating to the MSS scenarios. ’ REFERENCES (1) Konary, A. M. Worldwide Electronic Software Distribution 2010 2014 Forecast; Market Research for IDC;IDC: Framingham, MA, 2010. (2) Weber, C.; Koomey, J. G.; Matthews, S. The energy and climate change implications of different music delivery methods. J. Ind. Ecol. 2010, 14 (5), 754769. DOI 10.1111/j.1530-9290.2010.00269.x. (3) Seetharam, A.; Somasundaram, M.; Towsley, D.; Kurose, J.; Shenoy, P. In Shipping to Streaming: Is This Shift Green?; Proceedings of the First ACM SIGCOMM Workshop on Green Networking, New Delhi, India, Aug 30, 2010; ACM: New York, 2010. DOI 10.1145/ 1851290.1851304. ~ .; Johansson, M.; Finnveden, G.; Jonsson, A. Effects of (4) Moberg, A a Total Change from Paper Invoicing to Electronic Invoicing in Sweden. A Screening Life Cycle Assessment Focusing on Greenhouse Gas Emissions and Cumulative Energy Demand; Report from KTH Centre for Sustainable Communications; Stockholm, Sweden, 2007. (5) Toffel, M. W.; Horvath, A. Environmental implications of wireless technologies: News delivery and business meetings. Environ. Sci. Technol. 2004, 38 (11), 29612970. DOI 10.1021/es035035o. (6) Abukhader, S. M.; J€onson, G. The environmental implications of electronic commerce: A critical review and framework for future investigation. Manage. Environ. Qual.: Int. J. 2003, 14 (4), 460476. DOI 10.1108/147778303104886685. (7) Calculating Business Value and Environmental Benefit of Digital Software Distribution; Technical Report; WSP Environment and Energy: San Francisco, CA, 2007. (8) Demonstrating the Benefits of Electronic Software Distribution: A Study of Greenhouse Gas Emissions Reduction; Technical Report; WSP Environment and Energy: San Francisco, CA, 2009. (9) Gard, D. L.; Keoleian, G. A. Digital versus Print: Energy Performance in the Selection and Use of Scholarly Journals. J. Ind. Ecol. 2002, 6 (2), 115132. DOI 10.1162/108819802763471825. (10) PAS 2050:2008 Specification for the Assessment of the Life Cycle Greenhouse Gas Emissions of Goods and Services; British Standard Institution: London, 2008, ISBN 9780580509780. (11) Microsoft Store UK website: http://www.microsoftstore.co.uk. (12) Footprint Expert Software for LCA, version 2.2; Carbon Trust: London, 2010. (13) Finnveden, G.; Hauschild, M.; Ekvall, T.; Guinee, J.; Heijungs, R.; Hellweg, S.; Koehler, A.; Pennington, D.; Suh, S. Recent developments in life cycle assessment. J. Environ. Manage. 2009, 91 (1), 1–21. 1094

dx.doi.org/10.1021/es202125j |Environ. Sci. Technol. 2012, 46, 1087–1095

Environmental Science & Technology

ARTICLE

(14) Koomey, J. G. Worldwide electricity used in data centers. Environ. Res. Lett, 2008, 3 (3), 034008. DOI 10.1088/1748-9326/3/ 3/034008. (15) Energy Star: Enterprise Servers for Consumers; Energy Star: Product Dataset; Washington, DC, Mar 4 2011. (16) Quarterly Statistics: Servers, Worldwide, 4Q10 Update; Gartner Inc: Stamford, CT, 2011. (17) Report to Congress on Server and Data Center Energy Efficiency Public Law 109431; U.S. Environmental Protection Agency Energy Star Program: Washington, DC, 2007. (18) Ton, M.; Fortenbury, B. High Performance Buildings: Data Centers Server Power Supplies; Lawrence Berkeley National Laboratory: Berkeley, CA, 2005. (19) Hassell, J. Windows Server 2008: The Definitive Guide; O’Reilly Media: Sebastopol, CA, 2008. (20) The Green Grid Metrics: Data Center Infrastructure Efficiency (DCiE) Detailed Analysis; The Green Grid White Paper 14; The Green Grid: Beaverton, OR, 2008. (21) ENERGY STAR Performance Ratings Technical Methodology for Data Center; Energy Star Technical Report; Energy Star: Washington, DC, 2010. (22) Data Center Efficiency Measurements; Google Online Data Set 2011; Q1 2011. (23) Williams, T. L. Hyperoptic. Private communication, 2011. (24) Understanding SONET/ SDH; www.electrosofts.com/sonet/. (25) UK Fixed Broadband Speeds, November/December 2010; OFCOM Research Report; OFCOM: London, 2011. (26) Energy Star: Computers for Consumers; Energy Star Product Dataset; Washington, DC, 2011. (27) National Inventory Report for the German Greenhouse Gas Inventory 19902008; Federal Environment Agency (Umweltbundesamt): Dessau-Rosslau, Germany, 2010. (28) 2010 Guidelines to Defra/DECC’s GHG Conversion Factors for Company Reporting, version 1.2.1; DEFRA: London, 2010. (29) Bottrill, C.; Lye, G.; Boykoff, M.; Liverman, D. UK Music Industry Greenhouse Gas Emissions for 2007; Julie’s Bicycle: London, 2009.

1095

dx.doi.org/10.1021/es202125j |Environ. Sci. Technol. 2012, 46, 1087–1095