A CO2 Emissions Accounting Framework with Market-based Incentives

for Cloud Infrastructures

David Margery

, David Guyon

, Anne-C

ecile Orgerie

, Christine Morin

Gareth Francis

, Charaka Palansuriya

and Kostas Kavoussanakis

Inria, IRISA, Rennes, France

University of Rennes 1, IRISA, Rennes, France

CNRS, IRISA, Rennes, France

University of Edinburgh, EPCC, Edinburgh, U.K.

Keywords:

Cloud Computing, Energy Monitoring, Carbon Emissions, CO2 Accounting.

Abstract:

CO2 emissions related to Cloud computing reach nowadays worrying levels, without any reduction in sight.

Often, Cloud users, asking for virtual machines, are not aware of such emissions which concern the entire

Cloud infrastructures and are thus difﬁcult to split into the actual resources utilization, such as virtual ma-

chines. We propose a CO2 emissions accounting framework giving ﬂexibility to the Cloud providers, pre-

dictability to the users and allocating all the carbon costs to the users. This paper shows the architecture of our

accounting framework and ideas on how to practically implement it.

1 INTRODUCTION

Cloud computing’s wide adoption leads to a rising in-

crease of data center’s electricity consumption. This

major social issue will worsen with the explosion of

connected devices and Internet of Things (IoT), ask-

ing for always more computing and storage capacity

in the Cloud. In 2013, U.S. data centers consumed

an estimated 91 billion kWh of electricity; this con-

sumption is projected to increase to roughly 140 bil-

lion kilowatt-hours annually by 2020, the equivalent

annual output of 50 power plants, costing American

businesses $13 billion per year in electricity bills and

causing the emission of nearly 150 million metric tons

of carbon pollution annually (Natural Resources De-

fense Council, 2014).

This uncontrolled energy consumption of Cloud’s

data center causes increased greenhouse gas (GHG)

emissions. This important consequence is mainly de-

termined by the amount and sources of consumed en-

ergy (Bosse et al., 2016). Among GHG, carbon diox-

ide (CO

) is the major one in quantity produced by

human activities. Consequently, carbon taxes have

been proposed in order to reduce CO

emissions

and their negative effects on environment (Nordhaus,

2012). From an operational point of view, a carbon

tax requires a monitoring and accounting infrastruc-

ture in order to fairly distribute CO

costs among the

Cloud users. Even outside a carbon tax system, such

an infrastructure can provide useful information to

users about their real CO

emissions based on their

utilization of the Cloud system, and therefore, it can

raise their environmental awareness and incite them

to adopt more sustainable practices.

To build a carbon tax system, it is required to

precisely monitor the resource usage that can be at-

tributed to each user (computing, storage, communi-

cation), and to account for the resource cost induced

by the user’s utilization, like the data center air con-

ditioning cost for instance. While the live monitoring

issue has already been addressed in literature (Wajid

et al., 2015), the accounting issue has received little

attention.

The accounting problem consists in splitting the

indirect costs between the Cloud users (such as air

conditioning), and forecasting the direct costs for

each user. Indeed, Cloud computing is using a pay-

as-you-go model where users buy computing, stor-

age and network resources in the form of virtual ma-

chines (VM). Cloud providers exhibit prices per vir-

tual machine type, depending on the amount of vir-

tual resources included in the virtual machine. Such

Margery, D., Guyon, D., Orgerie, A-C., Morin, C., Francis, G., Palansuriya, C. and Kavoussanakis, K.

A CO2 Emissions Accounting Framework with Market-based Incentives for Cloud Infrastructures.

DOI: 10.5220/0006356502990304

In Proceedings of the 6th International Conference on Smart Cities and Green ICT Systems (SMARTGREENS 2017), pages 299-304

ISBN: 978-989-758-241-7

299

a model involves an a priori cost which is known by

the user upon purchase as opposed to an a posteriori

cost based on a precise monitoring of the resources

really used and thus, provided to the user at the end

of its Cloud resources utilization. Such an account-

ing model has to be ﬂexible enough for the Cloud

providers to be attractive, and it should provide to the

users a predictable cost. From an external third-party

organization, the carbon tax accounting system needs

to be certiﬁed: for a given period of time, all the car-

bon emissions of the data center must be equal to the

overall carbon emissions charged to the users.

In this paper, we propose a CO

emissions ac-

counting framework giving ﬂexibility to the Cloud

providers, predictability to the users and allocating all

the carbon costs to the users. We provide the architec-

ture of our accounting framework and ideas on how to

practically implement it. We argue that instead of try-

ing to keep the difference between predicted and real

emissions as low as possible at any time, an ef-

fective framework could consider this difference as a

ﬂexible capital to support an economical approach for

users’ energy-awareness.

The paper is organized as follows. Section 2 pro-

vides motivational examples and the context of this

work. The related work is presented in Section 3. Our

proposed architecture is described in Section 4. Sec-

tion 5 discusses the advantages and drawbacks of our

approach and provides ideas for implementing it in

real Cloud infrastructures. Section 6 concludes this

work.

2 CONTEXT AND MOTIVATION

2.1 From the Cloud User Point of View

Cloud users are renting virtual machines (VM) on

a pay-as-you-go basis. The energy consumption of

their virtual machine depends on the resource utiliza-

tion (CPU, memory, disk, network) and on the infras-

tructure power management (cooling cost, resource

allocation management, etc.) (Kurpicz et al., 2016).

The CO

emissions depends on the energy consump-

tion and on the electricity mix (Wajid et al., 2015).

One could compute these costs at each time and di-

vide them proportionally to the number of resources

booked by each VM. However, it would mean that

identical VMs running the same computation could

have really different costs. Indeed, if a VM is alone on

the infrastructure at a given time, then it would sup-

port the entire infrastructure cost, while this same VM

during a busy period would account for a much lower

cost. Such an accounting model, with high variations

over the time, would provide great unpredictability to

users, and is thus not desirable.

On the contrary, we argue for a shift of the pre-

diction responsibility from the user to the provider.

The Cloud provider gives a CO

cost for a VM upon

its purchase. This cost depends on the VM size and

can vary over time, but it cannot change for VMs al-

ready paid. So, the provider has to carefully monitor

resource consumption, infrastructure costs and elec-

tricity mix to entirely attribute the CO

emissions to

its users for a given period of time (a month for in-

stance).

2.2 From the Cloud Provider Point of

View

The Cloud provider is responsible for assigning all the

costs to the users over a long period (a month for

instance). A third-party, like a governmental organi-

zation could be in charge of certifying the summary

of provider’s CO

accounts. The accounting model

described above does not aim at being as accurate as

possible. Indeed, the provider could compute the dif-

ference between invoiced and real costs at the end of

each VM booking and it could directly pass on this

difference to the next client. However, such a system

does not give any ﬂexibility to the provider. Instead,

we argue for a ﬂexible model, ﬁxed by the provider

itself, and following market opportunities.

The provider is then responsible for dealing with

the difference between invoiced and real costs. It

ﬁxes its own CO

cost model for its VMs, and it can

choose not to make direct adjustments, but to use this

difference as a capital to invest. For instance, this

capital can be reinvested to reduce the cost of VMs

when the electricity mix is better or when cooling

costs are lower (at night for instance with free cool-

ing). This capital with its associated cheap offers

would constitute a market-based incentive to increase

users’ energy-awareness. This accounting framework

also favors energy-aware behavior from the Cloud

providers as they need to invoice all the CO

costs

to the users. So, in order to be competitive, they need

to have CO

costs as low as possible. It creates then

a strong incentive to switch off unused servers or per-

form over-commitment on servers hosting VMs with

low workload.

3 RELATED WORK

The carbon emissions of users’ virtualized resources

mainly depend on their power consumption. The

power consumption attributed to a user is not equal

SMARTGREENS 2017 - 6th International Conference on Smart Cities and Green ICT Systems

300

to the total server power consumption when the user

is not using all the server’s resources. A ﬁne-grained

monitoring of the power consumed by each VM on

a server is necessary in order to be able to estimate

their carbon emissions. Several VM power models

have been proposed in literature with different imple-

mentations. They are usually based on counters (hard-

ware or software) in order to monitor the resource

usage. Their accuracy thus depends which resources

are selected, how they are monitored and which for-

mulas are used to estimate the VM power consump-

tion from the monitoring data, such as linear regres-

sion (Kim et al., 2011) (Wu et al., 2016), polynomial

regression (Xiao et al., 2013), machine learning (Yang

et al., 2014) or tree regression based approach (Gu

et al., 2015). In these studies, estimation errors typi-

cally ﬂuctuate from 2 to 5%.

Research studies start to include ecological-

related factors in their optimization algorithms.

Bosse et al introduces GHG emissions into the system

availability and cost optimization problem of fault-

tolerant IT services (Bosse et al., 2016). Experiments

show that, for a slightly increased cost, a signiﬁcant

reduction in GHG emissions is possible. Similarly,

Khosravi et al propose a VM placement algorithm

taking into account Cloud sites’ PUE and their carbon

footprint (Khosravi et al., 2013). While maintaining

the same level of QoS, their solution manages to re-

duce the power consumption and CO

emissions.

Workload predictions as well as green energy

availability predictions bring an important contribu-

tion in reducing CO

emissions as it offers the ability

to adapt system conﬁgurations in order to face future

trends. Cloud resources usage can be predicted for

a given Cloud by using Extreme Learning Machine

algorithm on VM usage traces and user behavior (Is-

maeel and Miri, 2016). Sharma et al present a pre-

diction model for green energy availability (Sharma

et al., 2010). The model is able to predict next day en-

ergy harvesting based on weather forecasts. They im-

proved accuracy by 27% with machine learning tech-

niques (Sharma et al., 2011).

The existing studies show the ability to moni-

tor VMs power consumption, the inclusion of GHG

emissions factor in algorithms and also the possibil-

ity to predict availability of green energy as well as

Cloud users workload. However, to the best of our

knowledge, no work is handling the difference be-

tween predictions and calculated estimations of ef-

fective CO

emissions. Enabling quotes would allow

Cloud providers to bill the CO

emissions of a Cloud

VM to the ﬁnal user.

4 PROPOSED ARCHITECTURE

Figure 1 presents the high level architecture for en-

abling a provider to attribute CO

emissions to end-

users. This system allows users to access informa-

tion about resource usage (past and present), CO

emissions (estimated and attributed) for the VMs they

run, and to quotes for CO

emissions that will be

attributed to their future usage. Moreover, external

services named Third Party Cloud Brokers can select

platforms emitting the smallest amount of carbon be-

tween several Cloud providers to execute an applica-

tion.

Figure 1: High level architecture of a CO

emissions ac-

counting framework.

In order to accurately predict the carbon emis-

sions, the system needs to compute current and histor-

ical data. The current data is retrieved by communi-

cating with the Infrastructure Monitoring Service and

Third Party Information Services components. The

historical data comes from the Database and Third

Party Information Services. These components are

presented in details below.

4.1 Infrastructure Monitoring Service

The Infrastructure Monitoring Service accesses the

underlying hardware and software infrastructure

which belongs to the Cloud provider. Thus, it knows

the hardware data of each server: its energy efﬁciency,

its availability and down times (planned and unex-

pected). As for the software infrastructure, it collects

the amount of resources used by each VM. It is used

to compute a cost per resource type as presented to

the user.

A CO2 Emissions Accounting Framework with Market-based Incentives for Cloud Infrastructures

301

4.2 Third Party Information Services

To increase the accuracy of the carbon emissions pre-

diction, the model gathers information from external

services. They can be of different types and sources.

The electricity power grid service provides informa-

tion about the current Carbon Intensity Factor (CIF).

The CIF represents the amount of carbon emitted for

a given quantity of electricity provided by the power

grid. The value of CIF varies over time depending on

the electricity’s origin (renewable or not). Other in-

formation sources can be used depending on the elec-

tricity used. For instance, weather forecast services

can help predicting the production of solar panels.

4.3 Application Programming Interface

The Application Programming Interface (API) gives

clients access to resource usage metrics for their VMs.

It also provides a quote system that informs the clients

about the CO

emissions that will be attributed to

their VMs. The quote system publishes the period for

which the quote is valid, binding the Cloud provider

to attribute CO

emissions according to that quote.

This allows clients, who are in control of resource us-

age (i.e. they can deploy additional VMs), to predict

emissions that will be attributed to their usage,

enabling them to make provisioning decisions based

only on parameters under their control.

4.4 Database and Data Mining

Information

The data from the Infrastructure Monitoring Service

and Third Party Information Services is gathered and

stored in the Database. The former stores static in-

formation such as the energy efﬁciency of servers as

well as non-static information such as planned down

time of servers. The Database also saves periodically

information about the current CIF value and the other

third party-related information.

The Data Mining Information provides the data

needed by CO

Emission Predictor. It can give the

variation trend of the CIF associated to a speciﬁc time

period for instance.

4.5 CO

Emissions Predictor

In order to offer quotes, the provider must be able to

forecast resource usage and energy provisioning. The

forecast and quote calculations are based on a mix

of current data and historical data patterns retrieved

by mining data stored in the Database. For example,

records such as the past variation trend of the CIF for

a speciﬁc time period can have a signiﬁcant impact

on the accuracy of the prediction. Prediction accu-

racy can also be increased by including factors that

have an effect on CO

emissions. Such factors can

be seasonal changes or the prediction of green energy

availability (wind energy, solar energy).

Internal information is as important as external

information. The Infrastructure Monitoring Service

component allows to ﬁgure out which servers are

likely to be available to satisfy a request and their

footprint. Some servers may be less energy ef-

ﬁcient and therefore have a greater carbon footprint.

The past carbon emission attribution is gathered from

the Database records. Users can ask for a collection

of their historical carbon emissions for a speciﬁc time

period. After data is retrieved from the Database, the

collection is created and sent to the user through the

API.

The prediction algorithm is left to Cloud

providers’ discretion as it is part of the market-based

incentive. They can indeed choose to underestimate

or to overestimate their CO

emissions predictions at

some point in order to attract clients. The difference

between the billed cost (depending on the predictions

and policy of the Cloud provider at a given time) and

the real cost is managed by the CO

Capital.

4.6 CO

Capital

At the infrastructure level, the provider can mea-

sure the total power consumption, and using external

sources, the total CO

emissions attributable to its in-

frastructure. Comparing this latter value with the sum

of CO

emissions attributed to users provides a way

to measure the difference induced by the Cloud pol-

icy. These differences are accumulated over time in

the CO

Capital. In order to keep this capital under

an acceptable threshold, it can be taken into account

by the prediction algorithm, or by a business policy

(not described in the architecture) sitting between the

API and the Predictor. Please note that this capital can

be negative when estimation is too high and thus users

are attributed more CO

emissions than attributable to

the infrastructure.

5 DISCUSSION

In this section, we discuss the pitfalls to avoid for im-

plementing the proposed accounting framework.

SMARTGREENS 2017 - 6th International Conference on Smart Cities and Green ICT Systems

302

5.1 Architecture Difﬁculties

The architecture presented in this paper relies on the

capacity to get real-time information about the energy

sources and their CO

intensity at the level of each

Cloud infrastructure. While it is quite easy to get

that information country-wide or to rely on contrac-

tual promises made by the energy provider, it might

be difﬁcult in practice to get that information in the

general case. Indeed, some green energy accounted

for in the country-wide energy mix will probably al-

ready have been sold as such.

A second issue with this architecture is that it does

not include a mechanism to protect an infrastructure

provider from users massively changing their behav-

ior to adapt their workload to the current quote for

emissions. This in turn would impact the ac-

curacy of the workload prediction component of the

Emission Predictor, always increasing the error

made during prediction. Some form of user behavior

modeling might be necessary to be able to keep pre-

dictions accurate, or some business model component

to link quotes to clients using Service Level Agree-

ments and base quotes not on periods but on expected

resources usage.

5.2 Certiﬁcation

The presented architecture exhibits good properties

only if the CO

Capital stays very low. This is what

ensures that globally all CO

emissions estimated for

the infrastructure are passed on to users.

On the one hand, it seems relatively easy to accu-

rately measure power consumption at the infrastruc-

ture level, using the same perimeter as in PUE calcu-

lations, and to certify that value. In the same vein,

a certiﬁcation authority could independently monitor

power sources, and certify the CO

emissions that

need to be attributed to users.

On the other hand, getting a certiﬁable view of all

emissions attributed to users is more complex.

Some form of publicly auditable record of CO

emis-

sions attributed to every client must be made avail-

able, raising conﬁdentiality issues. Some form of

block-chain usage might help here (Swan, 2015).

The difference between the CO

emissions to at-

tribute and those attributed to clients builds up over

time in the CO

Capital. Further work is needed to

understand the properties required for the CO

capi-

tal so as to limit side effects in the way CO

emissions

are attributed.

At a global scale, our initial thoughts are that the

Capital must stay within a few percent of total

emissions over a year. Maybe a higher threshold

should be required at the scale of a week, a month,

etc. Another threshold could be set per client, so that

eco-aware clients do not beneﬁt from the presence of

other users that don’t care about CO

emissions. This

would avoid shifting all CO

emissions to clients that

do not report their carbon footprint, so as to offer un-

realistic reports to eco-aware clients.

5.3 Quotes and Business Logic

We have said very little until now about the contents

of quotes, other than the fact that they give users the

amount of attributed CO

per unit of usage of re-

sources over a period of time in the future. We an-

ticipate that providers might want to offer different

quotes for different periods, different quotes for dif-

ferent VM sizes, or for different hardware zones or

regions. The provider could even attempt to sell at a

higher price usage of the part of energy he gets from

renewable sources, in an attempt to partition its user

base between clients for which environmental impact

is important and others. As long as it can stay cer-

tiﬁed while doing this, it is possible, thus enabling

a dynamic and competitive eco-system of eco-aware

providers.

5.4 VM CO

Emissions

As seen in the related work, the literature has focused

on modeling the power consumption of a single VM.

These models are seldom able to take into account

the infrastructure costs of the VM, for example, the

amount of unusable memory on the VM’s host be-

cause of the effective size of the VM.

With the presented architecture, it only impor-

tant to have an accurate enough VM power model

so that the clients have little opportunity to change

their behavior to beat the system to be attributed less

emissions than the system would do. It is impor-

tant to note that there is no value for real CO

emis-

sions. The difference does not lie between an ob-

jective value measured after the fact and the value

attributed to a VM according to the quote given to

the client by the provider. It lies between the value

attributed to a VM and calculated using the power

model at the VM level and the one using the power

model at the infrastructure level. As the complete

model takes into account the infrastructure contribu-

tion to power consumption, and that the quote sys-

tem does not, there are optimization opportunities for

clients.

Because the optimization opportunities come

from infrastructure costs more than from inaccu-

rate power modeling, the focus of an infrastructure

A CO2 Emissions Accounting Framework with Market-based Incentives for Cloud Infrastructures

303

provider should be to make its infrastructure energy

proportional, rather than to provide accurate VM

power modeling.

5.5 Passing the Cost Up to the End-user

We have discussed here how CO

emissions are at-

tributed to VMs by the infrastructure provider. Be-

cause the VM user has the ability to predict CO

emis-

sions as she knows in advance how they will be com-

puted from resource usage counters, she has the abil-

ity to apply the same techniques to pass the costs up

to the different users of her VMs. This can be applied

recursively up to the end user, who is then empowered

with information about the CO

emissions attributed

to her usage of computing resources.

6 CONCLUSION

We present in this paper an architecture that allows

users of a Cloud infrastructure to have predictable

emissions attributed to their usage while taking

into account the difference between predictions and

estimations of effective CO

emissions. If this differ-

ence is kept under a pre-deﬁned threshold, it opens the

way to an eco-system where infrastructure providers

can be certiﬁed as providing reasonable CO

emis-

sions certiﬁcates to users while at the same time giv-

ing predictability to users. This creates a fair playing

ﬁeld where infrastructure providers compete to attract

eco-aware users in a way such that the complete in-

frastructure costs are taken into account. This should

increase the adoption of green technologies in all as-

pects of datacenter provisioning and therefore, con-

tribute to limiting the impact of IT on GHG.

ACKNOWLEDGEMENTS

This work has been supported by the ECO2Clouds

project (http://eco2clouds.eu/) and was partially

funded by the European Commission under grant

agreement number 318048.

REFERENCES

Bosse, S., Jamous, N., Kramer, F., and Turowski, K.

(2016). Introducing Greenhouse Emissions in Cost

Optimization of Fault-Tolerant Data Center Design. In

IEEE Conference on Business Informatics (CBI), vol-

ume 01, pages 163–172.

Gu, C., Shi, P., Shi, S., Huang, H., and Jia, X. (2015). A

Tree Regression-Based Approach for VM Power Me-

tering. IEEE Access, 3:610–621.

Ismaeel, S. and Miri, A. (2016). Multivariate Time Series

ELM for Cloud Data Centre Workload Prediction. In

International Conference on Human-Computer Inter-

action (HCI), pages 565–576.

Khosravi, A., Garg, S. K., and Buyya, R. (2013). Energy

and Carbon-Efﬁcient Placement of Virtual Machines

in Distributed Cloud DataCenters. In International

Conference on Parallel Processing (EuroPar).

Kim, N., Cho, J., and Seo, E. (2011). Energy-Based

Accounting and Scheduling of Virtual Machines in

a Cloud System. In IEEE/ACM International Con-

ference on Green Computing and Communications

(GreenCom), pages 176–181.

Kurpicz, M., Orgerie, A.-C., and Sobe, A. (2016). How

much does a VM cost? Energy-proportional Account-

ing in VM-based Environments. In Euromicro In-

ternational Conference on Parallel, Distributed, and

Network-Based Processing (PDP), pages 651–658.

Natural Resources Defense Council (2014). Data Cen-

ter Efﬁciency Assessment. NRDC Issue Paper,

https://www.nrdc.org/sites/default/ﬁles/data-center-

efﬁciency-assessment-IP.pdf.

Nordhaus, W. D. (2012). Carbon Taxes to Move Toward

Fiscal Sustainability, pages 208–214. Columbia Uni-

versity Press.

Sharma, N., Gummeson, J., Irwin, D., and Shenoy, P.

(2010). Cloudy Computing: Leveraging Weather

Forecasts in Energy Harvesting Sensor Systems. In

IEEE Conference on Sensor, Mesh and Ad Hoc Com-

munications and Networks (SECON), pages 1–9.

Sharma, N., Sharma, P., Irwin, D., and Shenoy, P. (2011).

Predicting Solar Generation from Weather Forecasts

Using Machine Learning. In IEEE International Con-

ference on Smart Grid Communications (SmartGrid-

Comm), pages 528–533.

Swan, M. (2015). Blockchain: Blueprint for a new econ-

omy. ” O’Reilly Media, Inc.”.

Wajid, U., Cappiello, C., Plebani, P., Pernici, B., Mehand-

jiev, N., Vitali, M., Gienger, M., Kavoussanakis,

K., Margery, D., Garc

ıa-P

erez, D., and Sampaio, P.

(2015). On Achieving Energy Efﬁciency and Reduc-

ing CO2 Footprint in Cloud Computing. IEEE Trans-

actions on Cloud Computing, PP(99):14.

Wu, W., Lin, W., and Peng, Z. (2016). An Intelligent

Power Consumption Model for Virtual Machines Un-

der CPU-intensive Workload in Cloud Environment.

Soft Computing, pages 1–10.

Xiao, P., Hu, Z., Liu, D., Yan, G., and Qu, X. (2013). Virtual

Machine Power Measuring Technique with Bounded

Error in Cloud Environments. Journal of Network and

Computer Applications, 36(2):818–828.

Yang, H., Zhao, Q., Luan, Z., and Qian, D. (2014). iMe-

ter: An integrated VM power model based on perfor-

mance proﬁling. Future Generation Computer Sys-

tems, 36:267 – 286.

SMARTGREENS 2017 - 6th International Conference on Smart Cities and Green ICT Systems

304