Is it really Greener in the Cloud?
An investigation of energy trends in Cloud Computing.
Chris Browne, Haley Jones, Paul Compston
Research School of Engineering, Australian National University
Contact: Chris.Browne@anu.edu.au
Abstract
Cloud computing has become a popular method of IT deployment in the last decade,
most visibly through the rise of online services and social media. Recent reports claim
that businesses can achieve significant reductions in greenhouse gas emissions by shifting
their processes to the cloud. These studies do not consider the change in user interaction
with data when it is in the cloud, such as through mobile devices. This paper investigates
this energy-saving claim by reviewing trends in and between four domains of this
problem and identifying their key drivers in a business setting. These domains considered
are on-site computing, cloud infrastructure, data transport, and device adoption. By
mapping a causal loop diagram to explain how the dominant trends are linked between
these domains, we conclude that overall energy consumption is likely increase when
moving computing processes to the cloud.
1. Introduction
Cloud computing is the latest paradigm in computing’ (Buyya, 2008), and is changing the
way that people and businesses interact. It has also been marketed as a ‘green’ computing
alternative for business, with claims that there are significant energy savings, and hence
cost savings, by moving traditional infrastructure to a cloud-based computing model
(CDP 2011, Google 2011).
This paper examines the energy-saving effect of cloud computing by examining trends in
energy consumption. In order to formulate our dynamical hypothesis, major trends have
been identified in the domains of on-site computing, cloud infrastructure, data transport
and device adoption. Previous studies examine and account for energy consumption in
these ‘problem’ domains (see Table 1), but do not consider a dynamic relationship
between them. We propose that an approach that considers feedback between the problem
domains may draw different conclusions from the previous work. A causal loop diagram
is used in this paper to propose these feedback mechanisms.
The utility of computing technology has improved significantly through the use of cloud
services, through new communication platforms such as those in social media, and
through new collaboration applications. However, the validity of energy-saving claims
needs to be discussed and debated. If cloud computing allows people to access more data
through a greater number of platforms, then per capita energy consumption is likely to
increase unless there are other mechanisms that reduce energy consumption. This
becomes critical if a perception that the cloud is ‘greener’ leads to a misinformed
adoption, where there is in fact higher overall energy demand.
In Section 2, we provide a brief background to the reports that prompted this paper. In
Section 3 we look at previous work and the dominant trends in each of the four domains.
In Section 4, a causal loop diagram has been mapped to examine the feedback structures
between the four domains. In Section 5 we look at ‘what if’ scenarios, and conclude our
study in Section 6 with our recommendations for a numerical model.
1. Cloud computing has many definitions, from full-blown application platforms to browsing the Internet. The
definition of cloud computing we have used in this investigation considers the cloud as interaction with data as a
unified resource that exists beyond the business’s immediate systems. Similar definitions call this the public cloud.
2
2. Problem Articulation
In 2011, the Carbon Disclosure Project (CDP) released a report finding that a typical,
large company transitioning their human resources system to the cloud could reduce their
carbon emissions significantly’ (CDP 2011).
This paper does not seek to question the specifics of the CDP paper and model. Our point
of concern is to question its simple cause-and-effect message: that cloud computing
reduces energy consumption. This message was picked up by many news sources and
technology blogs, and was used by cloud service provider Google to show that switching
to their cloud-based email service, Google Mail, was almost 80 times more efficient than
running in-house email (Google 2011).
By conflating the utility benefits of cloud computing with apparent energy saving
benefits, advocates of cloud computing may be misleading the public about the overall
environmental benefits of adopting cloud services. This could lead to a paradoxical
situation that, by moving business operations to the cloud to reduce energy consumption,
the overall consumption actually grows.
To investigate the claim further, we will look at a typical business setting transitioning
their services to the cloud. The intention is to match the CDP paper’s scenario’ by using
similar context. The trends of each of the problem domains will be examined in the next
section.
3. Problem Domains and Trends
The following domains were identified as areas to investigate the energy consumption of
cloud computing:
* On-site Computing: energy consumed by the user’s workstation and on-site servers
* Cloud Infrastructure: energy consumed by the user at the data centre*
+ Data Transport: energy consumed by the transport of the user’s data between the
site and the data centre
+ Device Adoption: energy consumed by new platforms available to access the data
2, The CDP report claimed that a typical Food and Beverage company could reduce CO2 emissions by 30,000 metric
tons over five years in the cloud, when compared with a business as usual scenario. In an economy-wide scenario, US
businesses with annual revenues over $1 billion could cut CO2 emissions by 85.7 million metric tonnes annually, when
compared with a business as usual scenario.
3, The key aspects of the CDP scenario are: a Food and Beverage firm of $10 billion annual revenue, with 60;000
employees across 30 countries.
4. Data centres have been chosen as the focus for this domain as they are the back-end of cloud services.
3
Recent research focuses on single domains of cloud energy consumption, rather than the
relationships and feedback between them. In this section we investigate the dominant
trends in each domain, and in Section 4 we propose how each of the domains listed above
interact through a causal loop diagram.
Table 1 maps how three recent reports consider energy consumption due to the cloud. A
column for this paper is included in this table as a comparison.
Table 1: Comparison of considerations in key recent studies into overall cloud energy consumption.
Considerations classed as (S)tatic, (D)ynamic or (N)ot Considered based on discussions in each paper:
Domains
On-site Workstations ,s) .
Computing _|Servers to zero No change on-site (x) PC or laptop js) Workstations i)
Cloud Energy per servers) Data centre efficiency »
: Google Mail si) |S ®)
Infrastructure |Data centre efficiency wp | °°8® WAH SCrVERSs)_| Server energy «sy Data centre demand (o,
Dat Size of fi F ; Ti rt effici
ines ise ree) No change in data) — |Range of scenarios, | -unSPOrt euteteney is)
Transport Energy/transaction (s) Change in volume jp)
Device Change in device
‘Adoption None (ny None (yy ‘None (x) adoption jp)
General model information
Cost of transitioning —_| Traditional email Relative energy change =
Report Focus |large business to cloud | compared to Google with different data pene between
service Mail service consumption patterns
Time period | Forecast 2011-2020 _| Point in time Point in time To be determined*
Scat8 Large business Small, medium and Single user/instance _| Single user in large
extrapolated to market _| large business business
Energy Cloud more efficient Cloud email 80 times | Varies based on scenario |To be determined
Conclusions than business as usual _| more efficient
Table 1 categorises the previous research into the domains considered in this paper. It
shows that the previous research has not examined all the domains that we consider to be
important when calculating the total energy consumption of the cloud. We have been
unable to find data that comprehensively illustrates the total energy consumed by the
cloud. With the data available to us, in this section we establish dominant trends for each
of the domains, and identify likely drivers within each domain. In Section 4 we look at
the relationships between these trends.
3.1. On-Site Computing
The first domain considered is on-site computing, which has been simplified in this
scenario as only the energy consumed by user workstations.
5. This is not determined at this stage of the study, but it is expected that data will be sufficient to examine trends and
propose a forecast for the period 2000-2025.
4
The energy consumed by computers within a business has been of concern for many
decades. This is demonstrated most visibly in the 1992 introduction of the Energy Star
program, which raises awareness of the energy efficiency of technology in a modern
office environment. However, Bray (2006) shows there is little technical compliance with
the program’, and user non-compliance with power management practices, such as
turning computers off in idle times, varies between organisations’. The effective
implementation of a power management strategy across a business could reduce energy
consumption considerably. The other driver we found for energy consumption is the
increase in computing utility, which will now be explained.
We have described the main driver of workstation energy consumption as computing
utility, which can be explained through two trends: improvements in computing
performance, and increasing workstation screen size. To demonstrate this trend, we have
compiled data on the maximum power draw’, screen size and release dates of Apple Inc.’s
series of all-in-one computers, which have been in production since 1984 (Apple 2012).
The reason for using an all-in-one machine is that it represents a consistent overall form
factor, rather than a highly variable energy consumption of a multitude of desktop and
monitor configurations.
6. The Energy Star programs largely require computers to meet Idle, Sleep and Standby settings. However, there is
low industry compliance, with only 21% of 141 surveyed desktops meeting all the requirements (EPA, 2006). An
Australian study showed that only 64% of a surveyed 22 computers advertised as meeting Energy Star requirements
actually met them (DEHWA 2009). In addition, the “In Use’ energy efficiency on Desktop computers relies on the
energy efficiency of the power pack, rather than a compliance to an amount of energy consumed.
7. Webber et al (2006) observed that 60% of workstations were left on overnight. The results, however, are quite
varied, with Bray (2006) observing that anywhere between 0%-91% of computers were turned off out of business
hours.
8. Koomey ef al (2009) notes that using nameplate power to draw conclusions on power consumption is erroneous,
with measured energy consumption generally significantly lower. This may be the case in these data in Figure 1;
however, the trend would still suggest that energy consumption is increasing over time.
5
Apple All-in-One Model Maximum Power Draw Over Time
sees 27-inch iMac
°
G3 All-in-One
300 Watts 2 e
°
‘ . :
200 Watts 2 20-inch LCD iMac, - A
Eg . .
e 3 . 1.
g . m5-inch LCD iMac
100 Watts . 2 +.
4 ae ™ Original iMac
Macintosh 512K .
0 Watts
1984 1991 1998 2005 2012
® Standard Macintosh Line © Top-End Machines
Figure 1: Apple all-in-one model maximum power draw over time (Compiled from Apple 2011). The trend
shows that maximum power draw is increasing over time.
The trend in power draw over time in Figure | is somewhat counter-intuitive, especially
as this product line has stated in its specifications that it meets Energy Star guidelines
since 2002. However, the Energy Star rating only considers idle and standby power draw,
which explains why there is no incentive to keep active power draw low.
We propose that this increased maximum draw is due to an increase in computing utility,
which we have identified as a continuous increase in processing performance (see
Moore’s Law, Moore 1965) and increasing screen size (Figure 2). The trend in increasing
energy consumption with larger screens is apparent, even though there was a significant
reported energy saving when monitors shifted from CRT’ to LCD" technology"'.
Apple All-in-One Model Maximum Power Draw v Diagonal Screen Size
30-Inch
8
e
@ 8 . 8
23-Inch = g sw 2
= 4 ee 2
2 § 3 =
6 e e@ a. = a
48-Inch aa = 28 .
g 2
Macintosh 512K I 3 z
‘oo s 5
BInch 2 g z
4 8
é
g
O4nch =
0 50 cr a i i
= CRT e LCD
Figure 2: Apple all-in-one model maximum power draw v diagonal screen size
(Compiled from Apple 2012). The trend shows all-in-one energy consumption increases with screen size.
9. cathode ray tube
10. liquid crystal display
11. Atypical LCD screen draws one-half to one-third of the power of an equivalent-sized CRT monitor (IEA 2009)
6
The combination of low adoption of power management strategies and the trends in
Figure | and Figure 2, leads us to the conclusion that as computers are replaced through
business cycles in the workplace, the overall energy consumption is likely to increase.
This is shown through an approximate four-fold increase in energy consumption between
the years 1984-2012 (see Figure 1), and has been included as a driving influence for this
reason.
Domain Drivers (On-site Computing)
In summary, the domain drivers for on-site computing energy consumption have been
identified as:
* pressure to improve energy efficiency leads to increased power management com-
pliance, which reduces on-site energy consumption
* pressure to improve workstation utility leads to investment in technology, which
increases energy consumed.
We also considered areas that may be of concern but for which available trend data was
not able to suitably address:
* user patterns on power management compliance, and whether this changes over
time
+ whether old workstations or servers are actually retired from the system, or new
applications are found for them
In the following section we investigate how the cloud infrastructure affects the overall
energy consumption.
3.2. Cloud Infrastructure
The cloud infrastructure in this study has been defined as the energy consumed by the
data centre, as data centres are the major back-end to cloud services. In Koomey’s (2008)
definition of energy consumption in data centres, the network that connects the data
centre to the user is not included’*. For this reason, we consider data transport as a
separate domain. However, Koomey’s (2008) definition does include the infrastructure in
the data centre, such as heating and cooling, and the internal network.
A breakdown of energy consumption trends for worldwide data centres shows that the
total energy required to power data centres more than doubled between 2000 and 2006
(EPA 2007). A more recent estimate suggests that this general trend has continued, though
12. We include data transport energy in Section 3.3.
it is thought that an increase in adoption of virtualisation — where multiple virtual
machines can run on a single physical server — and the recent economic downturn
stalled growth in the global installed base. These data are shown in Table 2.
Table 2: Energy consumption of data centres worldwide" (from data in Koomey 2008).
2010 Lower and Upper represent the best estimates to date (Koomey 2011).
Units 2000 2005 2010 2010
(lower estimate) | (upper estimate)
Total Energy BkWh" 70.8 152.4 203.5 271.8
Volume servers % 27.8% 33.1% 33.0% 30.0%
Mid-range servers % 9.5% 44% 2.5% 2.7%
High-end servers % 4.0% 2.8% 3.8% 48%
Storage % 4.0% 4.9% 9.4% 8.9%
Communications % 48% 4.8% 6.0% 5.7%
Infrastructure % 50.0% 50.0% 45.2% 47.9%
Calculated PUE ratio 2.0 2.0 1.82 1.91
These data suggest that there is a move away from mid-range servers, and an increase in
the need for storage capacity. The Power Usage Effectiveness (PUE) measures efficiency
through a ratio that compares the total energy, and the energy consumed by computing.
This metric is used to compare the efficiency of data centres.
Significant improvements can be made in the efficiency of data centres, with major cloud
providers reporting their PUEs below 1.5 (Katz 2009). However, data in Table 2 suggests
that there have only been modest improvements across the board on global data centre
PUE.
Domain Drivers (Cloud Infrastructure)
In summary, the domain drivers for cloud infrastructure energy consumption have been
identified as:
* an increased demand for cloud computing directly leads to increased off-site en-
ergy consumption
* increased demand also provides opportunity for data centres to improve their effi-
ciency through technologies that allow a data centre to achieve a lower PUE.
In the following section, we investigate the dominant trends in the energy required to
transport data between the user and the data centre.
13. PUE has been calculated from these data using the formula: PUE = Total Energy / Computing Energy. Columns
should total 100%, but may not due to rounding.
14, Billion kilowatt hours
3.3. Data Transport
The third domain we examined was data transport, which is a significant contributor to
off-site energy consumption. This was discounted in the Google (2011) analysis, which
states:
We would expect network energy to increase somewhat, as more
traffic must traverse the Internet in the cloud-based solution.
However, this effect is secondary to the large effect on server energy.
Baliga et al (2010) investigates the trends in data transport in an office setting. They
conclude that energy can be a large factor for energy consumption in cloud computing,
and describe scenarios where the cloud does and does not provide opportunities for
savings.
Using their example of the public cloud as a storage service, they show that transport
consumes the vast majority of energy, in a dynamic relationship with storage and the
servers, when almost any amount of data is transferred. This relationship is shown in
Figure 3a. Figure 3b compares the energy required access a 1.25MB file on a laptop hard
drive or through the cloud. Their analysis shows that the total energy consumed is less
through the cloud, unless the download rate exceeds approximately SMB/hour.
100 — 10°
x =
z =
5 8
i p°
5 Ft sep H00, |
i een en AP }
I : i
F Bl Puticecrage cone |
& E
§° 5 Prnate storage service |
is be \
3 |
2 al
10° 10" 10) 10°
Downloads per hour (1
Figure 3: a) Left: Relative percentages of total power consumption in the public cloud; b) Right: Total
power consumption per service per user from (Baliga et al 2010). In this example, the file downloaded is
1.25 MB. Transport makes up a significant percentage of total energy consumed when data is,
downloaded".
The analysis in Figure 3 (Baliga et al 2010) does not take into account the energy
required by the end user’s device, only the transport of data. Baliga et al (2010)
acknowledge that accessing data from the cloud with a laptop would consume the energy
15. Private cloud storage shown in Figure 3b consumes less energy in transport, as it passes through fewer switches,
routers and exchanges.
9
to power the server, transport the data, and run the device. This observations highlights
the justification behind the three domains discussed thus far. It does not include the
device adoption, which will be discussed in the following section.
A second major trend in data transport is that the volume of data transported is increasing
over time. Data averaged from a Nielson (2011) report in US smartphone data usage is
displayed in Figure 4. This displays a trend that consumption is increasing.
Average Quarterly Mobile Data Usage (MB)
400MB
325.25, Sg08
308
100MB
OMB
Qtr 12010 Qtr 2 2010 Qtr 3 2010 Qtr 4 2010 Qtr 12011
— Average Data Usage
Figure 4: Average Quarterly Mobile Data Usage compiled from data in Nielson (2011). Sources averaged
are Android OS, Apple iOS, Blackberry OS and Windows Mobile.
At the same time that data usage is increasing, the cost per megabyte of downloaded data
has decreased. In the same Nielson report, the cost per megabyte of data downloaded
dropped from US$0.14 at the beginning of 2010 to US$0.08 at the beginning of 2011.
This demonstrates that as the cost of data is decreasing, consumption is increasing.
Domain Drivers (Data Transport)
In summary, the domain drivers for data transport energy consumption have been
identified as:
+ as the volume of data increases, the transport energy increases
* transport accounts for a large percentage of relative power consumption for trans-
porting data to the cloud (Figure 3a)
+ there is a relationship between cost of data, and volume of data downloaded
As more devices become available to access the data, the data can be downloaded
multiple times. In the next section, we examine the effect of new devices on energy
consumption.
10
3.4. Device Adoption
This category has been discounted in both the CDP (2011) analysis, and the Google
(2011) response. A footnote from the Google analysis says:
There is no appreciable difference in client energy, since the user is
usually not changing the device they use to access email.
This statement appears somewhat at odds to their deployment strategy for accessing
Google Mail through mobile devices, which shows 16 methods available to access
Google Mail via a smartphone. This is included in Table 3 to demonstrate that there are
multiple methods on multiple devices to access the same data.
for Google Mail on various platforms shows that there are a number of ways to
ice on a mobile phone. Data from Google (2012).
Table 3: Supported phon
access their cloud email s
Access Choice) Android | Blackberry iPhone Nokia $60 | Windows Others
Web App Yes Yes Yes Yes Yes Yes
Native App Pre-installed Yes
Syne via IMAP Yes Yes Yes Yes Yes
Google Syne Yes Yes Yes
We propose that the availability of cloud-connected devices has a strong adoption effect.
Data that can be accessed through a number of devices encourages interaction with the
cloud-based services and, in turn, encourages the creation of more cloud-based
applications.
A user that accesses a cloud-based email service may simultaneously be accessing that
data on a number of devices, such as their desktop, laptop, tablet and phone. If this were
the case, data that travels between the cloud and the user could increase up to fourfold.
The availability of new devices, and the lowering of data costs, are significant reasons for
adoption of cloud-based services. For this reason, device adoption should be included in
the dynamic model.
Domain Drivers (Device Adoption)
The domain drivers for device adoption energy consumption have been identified as:
+ as the demand for cloud computing increases, an opportunity to improve data util-
ity opens up. As the device gap is filled, it reinforces the demand for cloud
computing
* as more devices are used, the energy required to power devices increases
In the next section, we combine these drivers from the four domains into a causal loop
diagram.
4. Model Mapping
In this section, the relationships between the four domains has been mapped into a model
boundary chart (Table 4) and causal loop diagram (Figure 5).
The model boundary chart in Table 4 states the boundaries that were considered when
constructing the proposed model. We have approached our study to look at the energy
scenario for a business that is not influenced by market and economy-wide fluctuations.
This is a deliberate, as it matches the situation described in CDP (2011), which could
also be described as a ‘business-as-usual’ scenario.
Table 4: Model boundary chart for energy consumption in cloud computing by domain
Endogenous Exogenous Excluded
On-site energy consumption Population Material consumption
Off-site energy consumption Competition Production energy
Data transport energy consumption | Market share Technology leaps
Data consumption Employment cycles Environmental constraints
Device energy consumption Green Accounting Company growth/decline
Investment in technology GDP Data access method variability
Improvement in technology capacity | Profit and loss
Business decision pressures Government Policies
Demand for services Energy Price Fluctuations
Demand on infrastructure
Cost of energy consumption
Work practices
Technology expectations
The definition of this model boundary helped in mapping the causal loop diagram, which
is our dynamical hypothesis. This is shown in Figure 4.
There are two aspects of the causal loop diagram that have not been explained in Section
3. The expected utility improvement drives provides opportunities to improve workstation
and data utility. This utility expectation is analogous to the expectations in Moore’s Law
(Moore 1965): that the performance of computing will continually improve.
The causal link between Off-site energy consumption and Energy efficiency programs
describes a link that does not directly exist. It may be the case that this occurs as an
increased financial cost, but we were unable to find any relationship between energy
consumption and cost of the cloud service. This may become more obvious if a price on
carbon is introduced, or if the direct accounting of energy was required to be reported.
For this reason, it is indicated with a delay.
12
& it ,
Device energy
Off-site energy * —— transport
Ener
: _ “cant enemy, se %
a a
+ , Volume of data
Energy efficiency transported to cloud
A) programs tes) i
Desired Efficoncy Efficioney Ara)
energy Pressure to improve
consumption Qn-site energy - Investment in data data centre efficiency Data Affordability E |
consumption centre ‘ecm a my of access
+
Workstation
utility Nd of =
Investment in ee en computing
~_—e LD available
to access data
& Rt «
Boosting
Util Expected utility * Prins
shortall ~~ __Improvement factor Data utility
"opportunities
Figure 5: Causal loop diagram ofdonate, as mapped from trend data in Section 3. Expected utility
improvement factor and the relationship between Off-site and On-site energy consumption is ssed
below.
Figure 5 maps our dynamical hypothesis in a causal loop diagram. We discuss the
relationships in the following section.
5. Model discussion
Our dynamical hypothesis leads us to observe that using the cloud appears to only add to
the total energy consumed in the system. In this section we look at the implications of the
causal loop diagram, and suggest possible ‘what if? scenarios. The relationships have
been mapped to reflect the behaviour described in the trends in Section 3.
The workplace efficiency loop (B1) reduces the on-site energy consumption indepen-
dently of the technology that the business has in use. The technology boosting loop (B2)
seeks to meet the desired level of utility, which is continually increasing as the expected
utility improvement factor pushes expectations higher. As the workstation utility
increases, the on-site energy consumption goes up.
The cloud efficiency loop in data centre efficiency (B3) reflects the attempts to improve
data centre PUE. Demand for cloud computing increases the pressure for cloud efficiency
as the demand grows.
13
The reinforcing behaviour mapped in the device adoption loop (R1) drives energy up
through an increased popularity of new devices. As the data affordability loop (R2)
drives the cost of access down, the energy required to power increasing data transport
drives the energy consumption up.
We expect that the energy consumption will increase significantly in the whole system in
this dynamical relationship. In our dynamical hypothesis, the on-site energy consumption
has no feedback from the other loops. There may be a relationship that reduces on-site
energy consumption through the removal of installed servers; however, we have not been
able to find sufficient evidence that this is significant.
We were also unable to find any evidence to suggest that an increase in off-site energy
consumption has any effect on the decisions made in on-site computing. This could be the
scenario if the use of the cloud lengthened the time that they replaced workstations, or if
they were required to account for off-site energy.
There are a number of ‘what if’ scenarios that could lead to a comparative reduction:
+ if the adoption of devices, which are generally of lower power consumption than
workstations, led to the discontinuance of the trend for desktops with large
screens.
+ if the processing power shifted to the cloud and the workplace policy on machine
replacement took place over a longer time
+ if the off-site energy footprint fed back to the decision-making process in house.
This could be through a requirement to publish or report the off-site energy
consumption.
When a numerical model is constructed, these ‘what if’ scenarios should be considered.
6. Conclusions
We investigated the trends around workplace computing and associated shift to cloud
computing. We found that previous studies did not consider all of the domains required to
properly account for the total energy consumption. Our dynamical hypothesis embodied
in our causal loop diagram indicates that there are reinforcing loops that encourage
increased data consumption, and an increase in the number of devices that can access the
data. Further work is required to understand this dynamical relationship, but there is a
likelihood that the total energy consumed by moving applications to the cloud actually
increases, in contrast to reports that suggest otherwise. Further work to investigate this
claim will be done through a numerical model, which is the next stage of this research.
14
7. Bibliography
Apple Inc., 2012. Tech Specs. Accessed 22 February 2012.
<http://support.apple.com/specs/#desktopcomputers>
Baliga, J., Ayre, R.W.A., Hinton, K., and Tucker, R.S., ‘Green Cloud Computing:
Balancing Energy in Processing, Storage, and Transport’, Proceedings of the IEEE, Vol.
99, No. 1, January 2011, pp149-167.
Bray, M., 2006, Review of Computer Energy Consumption and Potential Savings, White
Paper. Dragon Systems Software Limited.
Buyya, R., Yeo, C.S., and Venugopal, S., 2008, ‘Market-Orientated Cloud Computing:
Vision, Hype, and Reality for Delivering IT Services as Computing Utilities’, /0th IEEE
International Conference on High Performance Computing and Communications, 2008.
Carbon Disclosure Project Study 2011, Cloud Computing - The IT Solution for the 21st
Century, Verdantix, 2011.
DEHWA, 2009, ‘ENERGY STAR Computers in Australia 2009’, Department of
Environment, Heritage, Water and the Arts, Canberra.
EPA , 2007, Report to Congress on Server and Data Center Energy Efficiency, Public
Law 109-431. Prepared for the U.S. Environmental Protection Agency, ENERGY STAR
Program, by Lawrence Berkeley National Laboratory. LBNL-363E. August 2.
<http://www.energystar.gov/datacenters>
IEA, 2009, ‘Gadgets and Gigawatts - Policies for Energy Efficient Electronics’, Interna-
tional Energy Agency, Paris.
Katz, R.H., 2009, ‘Tech Titans Building Boom’, Spectrum, Vol. 26 Issue 2. pp40-54
Koomey, J.G., 2008. ‘Worldwide electricity used in data centers.’ Environmental
Research Letters. vol. 3, no. 034008. September 23.
<http://stacks.iop.org/1748-9326/3/034008>
Koomey, J.G., Berard, S., Sanchez, M., Wong, H., 2009, ‘Assessing Trends in the
Electrical Efficiency of Computation Over Time’, Final Report to Microsoft Corporation
and Intel Corporation.
Koomey, J., 2011, Growth in data center electricity use 2005 to 2010. Oakland, CA:
Analytics Press. July.
Google Inc., 2011, ‘Google’s Green Computing: Efficiency at Scale’. Accessed February
17 2012 <http://static.googleusercontent.com/external_content/untrusted_dlcp/
www.google.com/en/us/green/pdfs/google-green-computing.pdf>
Google Inc., 2012, ‘Gmail for Mobile’. Accessed 12 March 2012.
<http://www.google.com.au/mobile/mail/>
Moore, G.E., 1965, ‘Cramming More Components onto Integrated Circuits,’ Electronics,
vol. 38, no. 8, 1965, pp. 114-117.
15
Is it really Greener in the Cloud?
Chris Browne, Haley Jones and Paul Compston. Australian National University, Research School of Engineering. chris.browne@ anu.edu.au
Cloud computing has revolutionised modern communication.
We question claims that cloud computing saves energy in this changing landscape.
Problem Origin
Proponents of cloud computing argue that switching to Cloud-based services provides significant energy savings.
so Model derived net energy savings 2011-2020 Model derived net CO, savings 2011-2020 aie
$12,000
“At one location, we took
400 development servers
and consolidated them
into 8-10 physical servers
§ $8,000
: 6.7] 10.9] 17.0) 24.7 9 93.4 Bl 43.1 53.4 69.7 NM 74.6 I 85.7]
$6,000 to achieve significant
4.000 savings.
[ 2011 «2012 2013 2014 2015 2016 2017 2018 2019 2020
7 U
1 State Street,
ai
Madge Meyer
From Carbon Disclosure Project 2011 (p16).
Locally Hosted Email:
Inefficient use of servers
by individual businesses. *
age Cloud-Based Email:
Efficient use of servers by
collective use in the cloud. /
Million tons of CO2
Coc Soc Eel
From Google (p4)
From Carbon Disclosure Project 2011 (p17). From Carbon Disclosure Project 2011 (p16).
Problem Domains and Trends
We considered trends in a broad range of domains to identify system feedback
On-Site Computing Domain
Cloud Infrastructure Domain
Data Transport Domain
Device Adoption Domain
Energy consumed by workstations Energy consumed at data centres Energy consumed transporting data Energy consumed by devices
1
Apple All-in-One Model Maximum Power Draw Over Time 140 °°
ze 4 i U MB)
400 Watts 27-inch iMac € Historical energy use Future energy Historical x 90 Lpaceceeensneeece—aaanesenseseeenend Average Quarterly Mobile Data sage ( )
e — 120 use projections > “ 4 scenario § 80 ee 400MB
, £ so} 343.75
G3 All-in-O1 g 7 Current efficiency E a Transport 325.25
300 Watts i = @ 2 400 2 7 x “| trends scenario 70} Pi 4 308
° : oe “ 300MB 251.5
. . Reo woe ge | Seiezetoveston 605 Fa 4 pours
Ui 9 20-inch LCD iMac 3 a” ge : 50+ \ ;
g a? = 8 o . a 200MB
e s) o oe * 2 _~ = B 40 //
g - 15-inch LCD iMac 3 EE Best pra 5 forage
100 Watts a 2s MO, -p an-n os nannns cape a enna nnn snes ons fan nnn nae Orgs Se Sie ad 30} 100MB
a 8 an a ™ Original iMac 3 ~~" | scenario 20} s
Macintosh 512K < 2 .
oWetts 10 een “Qi 2010 Qr22010 au320I0 @r42010 arr 2011
1984 1991 1998 2005 2012 a — r
= Standard Macintosh Line © Top-End Machines 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 O52 10" 10° 10° 102 — Average Data Usage
Downloads per hour (1/hr)
Trends suggest that Data Transport is a large portion
of energy demand, ignored in the Google study
(graph from Baliga et al 2010)
Trends indicate that data consumption is
increasing on mobile devices. Applications are
also available on more platforms (Nielson 201 1)
wae ie ener
Energy Off-site energy ae wranspor . gy
- difference Ve consumption ~\
i LO ” N\ a
tar) aA Ae) = Volume of data
Energy efficiency -
Energy consumption for data centres is increasing
over time, though energy efficiency is improving
(graph from EPA 2007)
Trends show that computing energy consumption
is increasing over time (compiled from Apple 201 1)
Initial Model Mapping
Based on these trends, we mapped these relationships
in a Causal Loop Diagram
transported to cloud
Workplace programs Cloud
Desired Efficiency Efficiency hs
Model Boundary Chart energy Pressure to improve
consumption On-site energy - Investment in data data centre efficiency Data Affordability
Endogenous Exogenous Excluded consumption centre econo » Cost of access
+
On-site energy consumption Population Material consumption + \ =
Off-site energy consumption Competition Production energy
Data transport energy consumption Market share
Data consumption Employment cycles
Device energy consumption Green Accounting
Investment in technology GDP
Improvement in technology capacity Profit and loss
Business decision pressures Government policies
Demand for services Energy price fluctuations -
Demand on infrastructure
+
Technology leaps Workstation +
Environmental constraints utility ee as Demandfor +
Investment in cloud computing
Company growth/decline
(a2 technology / Devices available
x (Rid to access data
Data access method variability
Technology
Boosting
= Devi
Cost of energy consumption Utilit Expected utility + adage
Work practices Data utility
shortfall a factor
a re
+
Technology expectations opportunities
+
Our Causal Loop Diagram based on identified trends
Discussion and Development
We intend to move towards a Stock-and-Flow model to investigate the system behaviour
At this stage our research is showing that it is likely total energy consumption will increase
by switching to the cloud and a large portion of the energy consumption will be outsourced.
Likely What-if Scenarios...
Can external energy consumption be
disclosed to inform IT decision makers?
Would workstations be replaced less often
if processing shifted to the cloud?
What will happen if new devices displace
traditional methods of computing?
References
Apple Inc., 2012. Tech Specs. Accessed 22 February 2012. <http://support.apple.com/specs/#desktopcomputers>
Baliga, J., Ayre, R.W.A., Hinton, K., and Tucker, R.S., ‘Green Cloud Computing: Balancing Energy in Processing, Storage, and Transport’, Proceedings of the IEEE, Vol. 99, No. 1, January 2011, pp149-167.
Carbon Disclosure Project Study 2011, Cloud Computing - The IT Solution for the 21st Century, Verdantix, 2011.
EPA , 2007, Report to Congress on Server and Data Center Energy Efficiency, Public Law 109-431. Prepared for the U.S. Environmental Protection Agency, ENERGY STAR Program, by Lawrence Berkeley National Laboratory. LBNL-363E. August 2. <http://www.energystar.gov/datacenters>
Google Inc., 2011, ‘Google’s Green Computing: Efficiency at Scale’. Accessed February 17 2012 <http://static.googleusercontent.com/external_content/untrusted_dicp/ www.google.com/en/us/green/pdfs/google-green-computing.pdf>
Nielson Online, 2011. Average US Smartphone Data Usage Up 89% as Cost per MB Goes Down 46%. Accessed 28 February 2012. <http://blog.nielsen.com/nielsenwire/online_mobile/average-u-s-smartphone-data-usage-up-89-as-cost-per-mb-goes-down-46/>
