From Reducing Energy Consumption to Reducing CO

Emissions: The ECO

Clouds Approach

Usman Wajid

and Pierluigi Plebani

University of Manchester, Manchester, U.K.

Politecnico di Milano, Milano, Italy

usman.wajid@manchester.ac.uk

{pierluigi.plebani, barbara.pernici}@polimi.it

Abstract. While the topic of energy efficiency (in cloud and large scale data-

centers) has been attracting significant interest from academia and industry ow-

ing to mainly economic and somewhat environmental implications, reducing

emissions is often side-lined as a consequent benefit. This paper presents a

more direct approach adopted in the ECO

Clouds project for reducing CO

emissions in cloud and scale datacenters. The ECO

Clouds approach relies on

(a) definition of key metrics to enable quantification of CO

footprint at appli-

cation and infrastructure level, (b) CO

aware deployment strategies, and (c)

adaptive management of workloads to constantly minimize the CO

footprint of

applications as well as underlying resources/infrastructure.

1 Introduction

With rapid proliferation of large scale datacenters and cloud facilities, tackling the

issue of energy consumption has emerged as one of the critical research challenges to

be dealt with. In response, a plethora of approaches have proposed ways to achieve

energy efficiency and harvest associated benefits that mostly translate into economic

advantages.

Besides energy efficiency, a related stream of research focuses on evaluating and

reducing the environmental impact of cloud computing by means of minimizing CO

emissions from cloud and datacenter facilities. The issue of CO

emissions resulting

from vast energy consumption is a growing concern that is generating economic,

social and political pressure. This pressure from different sectors of society is likely to

result in regulations as well as influence consumer selection criteria for outsourcing.

Existing work in this area mostly ranks CO

reduction or environmental impact as a

consequent benefit of achieving energy efficiency. However, CO

derived from the

different energy sources is often not addressed directly, and relatively little attention

is diverted towards effective utilization and optimization of available energy sources.

Hence there are opportunities for addressing environmental implications by adopting

a more direct approach for reducing CO

emissions.

In this background, the European Commission funded ECO

Clouds (Experimental

Awareness of CO

in Federated Cloud Sourcing) project aims to addressing the envi-

ronmental impact of cloud computing by not only developing methods for quantifica-

tion of energy consumption and CO

emissions at different levels of cloud computing

Pernici B. and Plebani P.

From Reducing Energy Consumption to Reducing CO2 Emissions - The ECO2Clouds Approach.

DOI: 10.5220/0006183200610070

In European Project Space on Information and Communication Systems (EPS Barcelona 2014), pages 61-70

ISBN: 978-989-758-034-5

(e.g. testbed, host, virtual machine and application level) but also allowing for proper

consideration of gathered data at application deployment and execution lifecycle.

ECO

Clouds develops and advocates a reinforcement learning model that relies on

monitoring of ecological parameters in cloud computing and using this information to

improve the way applications are designed and how infrastructure responds to chang-

es occurring at physical hosts, running applications, down to energy mix consumed.

The ECO

Clouds solution is particularly designed for federated clouds that (a) of-

fer opportunities of utilizing different cloud resources, and (b) enable minimizing CO

footprint of applications by considering energy mix of different testbeds in the appli-

cations deployment decision making model. In this respect, the availability of differ-

ent and geographically separated infrastructure or testbeds allow distribution of appli-

cations on eco-friendly resources, where permitted by testbed and availability of suit-

able resources required by applications.

The ECO

Clouds is described in Section 2. Section 3 describes how the quantifi-

cation of environmental impact of cloud computing is carried out in ECO

Clouds.

Section 4 presents the deployment and adaptation techniques devised in ECO

Clouds

in order to minimize the environmental impact of cloud computing. Section 5 presents

discussion of some preliminary results of ECO

Clouds solution. Section 6 concludes

paper with directions for future work.

2 An Overview of ECO

Clouds Approach

ECO

Clouds adopts an iterative development approach involving three key phases

namely Measure, Create and Test, shown in

Fig. 1

Fig. 1.

Three phases of ECO

Clouds approach.

The Measure phase focuses on quantification of energy consumption and envi-

ronmental impact of cloud computing, the Create phase develops techniques and

software artefacts to help realize awareness of the environmental impact of cloud

computing by making improvements in the reduction of energy consumption and CO

emissions, and the Test phase tests the outcome of previous two phases on an existing

FIRE (www.ict-fire.eu) facility known as BonFIRE (www.bonfire-project.eu).

The three phased approach addresses important research questions such as quanti-

fication of environmental impact of cloud computing, enacting deployment and

runtime adaptation actions that can decrease the energy consumption and CO

foot-

Measure

Create

Te st

EPS Barcelona 2014 2014 - European Project Space on Information and Communication Systems

print of cloud computing and considering environmental implications in the design

and subsequent execution of cloud applications.

Fig. 2. Implementation of ECO

Clouds approach.

Fig. 2 provides an overview of implementation and execution aspects of

ECO

Clouds approach where the workflow is triggered by an application deployment

request. A deployment request typically entails the availability of certain number of

virtual machines with certain characteristics (e.g. CPU and Memory requirements).

The deployment of virtual machines is determined by the ECO

Clouds Scheduler, a

key decision making component in the system. The subsequent step involves the col-

lection of eco-metrics by Scheduler from different levels of the cloud infrastructure.

Eco-metrics is a term used in ECO

Clouds to refer to parameters that provide infor-

mation concerning energy consumption and CO

footprint at different levels of cloud

infrastructure. The information provided by eco-metrics is used in the eco-aware

scheduling of applications on the federated cloud resources, as shown in step 3. Once

deployed the applications are constantly monitored along with the way the underlying

infrastructure behaves e.g. how spikes in resource utilization, deployment of new

applications and termination of already running applications provide opportunities for

further improvements in energy consumption and CO

footprint of available re-

sources. The monitoring information is taken into account while deciding and trigger-

ing certain adaptation actions that can reduce the energy consumption and CO

foot-

print of running applications as well as the underlying infrastructure. After the execu-

tion of applications, the detailed monitoring information is compiled in an eco-report

for user.

The report informs the user (e.g. application designer and developer) about total

energy consumption and CO

footprint of their applications. Based on the reinforce-

ment learning model of ECO

Clouds the user is encouraged to use the information

from eco-report to improve the design and execution aspects of their applications in a

From Reducing Energy Consumption to Reducing CO2 Emissions - The ECO2Clouds Approach

bid to make them more energy efficient and eco-friendly. Furthermore, it is envi-

sioned that the awareness about energy consumption and CO

footprint can enable the

application designers and developers to better tune their applications e.g. by more

effectively utilizing the available resources or by choosing adequate resources for the

deployment and execution of application.

3 Quantification of Environmental Impact

As the final goal of the ECO

Clouds project is to reduce the environmental impact of

applications running on a federated cloud infrastructure, it is crucial to properly quan-

tify and monitor such an impact.

The energy consumed by an application is usually taken as a primary metric: the

more the application consumes, the more impact the application has on the environ-

ment 2. Although calculating the energy consumption is important, this captures only

a fraction of the real impact. Indeed, it is also important to evaluate which are the

sources that provide such energy: Nuclear plants? Coal plants? Renewable sources?

As a consequence, CO

emissions are usually taken as a good metric for evaluating

the real impact, so that, an application that consumes an amount of energy coming

from a renewable source is preferable to an application that consumes the same

amount of energy coming from a coal plant.

It is worth noting that, although considering the CO

emissions during the execu-

tion is reasonable, the computation of CO

emission could be tricky. For instance,

renewable sources are correctly considered as zero-emission sources. This is true if

we consider the emission during the energy provisioning. On the contrary, if we con-

sider the complete life cycle of renewable sources as hydro-power plants or wind

farms, the CO

emission cannot be considered zero as some energy coming from CO

emitting sources is required to build, and it will be required to dismiss, the plants.

This issue becomes more important when considering nuclear plants. Indeed, CO

emissions of this kind of plants are almost zero, by contrast the energy required to

build for the decommission of a nuclear plant is incredibly high 3. As at this stage it is

very difficult to really compute this holistic CO

emission, the ECO

Clouds approach

quantifies only the CO

emitted when producing the required energy. Anyway, infor-

mation about the energy sources are considered in order to increase the awareness of

the user about the kinds of energy sources involved when the application required are

used.

On this basis, the ECO

Clouds project proposes a layered approach to monitor the

environmental impact of the applications, the virtual machines, and the sites as shown

in Fig.

3. For each of the layers, a set of metrics to be monitored has been identified 1.

Each of these sets of metrics captures not only aspects about the environmental im-

pact, but also the performances of the system. This allows our solution to mediate

between improvements with respect the environmental sustainability and the perfor-

mances.

More specifically, starting from the bottom, the infrastructure layer includes met-

rics that measure the behavior of the single physical machines and the entire site.

Especially for the energy metrics, this layer is fundamental, as it is the place in which

the power can be physically measured using PDU (Power Distribution Units). Along

EPS Barcelona 2014 2014 - European Project Space on Information and Communication Systems

Fig. 3. Layered architecture in ECO

Clouds.

with the classical eco-metrics as PUE (Power Usage Effectiveness), the environmen-

tal impact is taken into account by measuring the GEC (Green Energy Efficiency) 4:

i.e., the percentage of the power coming from renewable source. Yet, availability of

the resources (CPUs, memory, and storage) is also measured to monitor the possibil-

ity to deploy application on a site.

Moving to the next layer, the VM layer focuses on the metrics able to evaluate the

status of the VMs running on the physical host. At this layer, most of the metrics are

derived from the metrics measured at the infrastructure layer. Focusing on the envi-

ronmental impact, here the energy consumed to run a VM is computed by considering

the amount of resources reserved and really used by the VM in order to properly sub-

divide the energy consumed by the physical host among the VMs running on it.

At the top-most layer, there are the metrics that evaluate the applications running

on the VMs. Assuming that an application can be distributed among several VMs and

these VMs can live on different sites, the metrics included at this layer, compute the

environmental impact as well as the performances starting from the metrics at the

lower levels. For instance, the energy consumption of the application is obtained

considering the execution time and the energy consumed by all the VMs required by

the application. At this level, it is worth noting that metrics inspired by the usual met-

rics included in the infrastructure layer are also included. For instance, the A-PUE

(Application PUE) has been inspired by the classical PUE. As the latter is defined as

the quantity of power used by the site divided by the power really used by the IT

devices, the A-PUE is defined as the quantity of power used by all the VMs involved

in the application divided by the power really used for running the application. This

parameter has been considered as, usually, when running an application on a VM

several other processes are running as well. Some of them are required (i.e., operating

system processes), some other starts automatically when the VM boots but they are

not really needed (i.e., a mysql deamon or an http server even if the application does

not need them). Goal of this metric is to make the user aware of this fraction of power

Application layer

VM layer

Infrastructure layer

Site

Host

CPU utilization

Availability

Energy

Consumption

IOPS/

EnergyConsumed

Site utilization

Availability

Storage

utilization

GEC

Site infrastructure

Efﬁciency

PUE CUE

Site saturation

CPU usage

I/O Usage

Storage usage

Memory Usage

Energy

consumption

IOPS/

EnergyConsumed

Task execution

time

Application

execution time

Response time

Throughput

Energy

consumption

A-PUE

AeP

A-GEC

Energy

consumption

Energy

Consumption

Energy

consumption

From Reducing Energy Consumption to Reducing CO2 Emissions - The ECO2Clouds Approach

that is not consumed specifically for the application. More details on the set of metrics

and the facilities involved to monitor them can be found in 1.

4 ECO-Aware Application Deployment and Adaptation

At primitive level the ECO

Clouds solution is driven by the Scheduler, a central enti-

ty or service responsible for determining eco-friendly deployment of application and

later runtime adaptation of deployment configurations based on the availability of

eco-metrics.

4.1 ECO-Aware Deployment

The ECO

Clouds Scheduler implements different application deployment policies and

allows system administrator to switch between the policies using a simple REST

interfaces. The deployment policies that can be characterized as eco-aware heuristics

perform decision making at two different level i.e. at testbed and physical host level.

Keeping in view the federated cloud infrastructure, once a deployment request is

received, the first Scheduling step is to select a suitable testbed. This step is realized

by the switching the Scheduler in one of the following modes:

Individual Deployment Mode performs selection of a testbed for each individual

VM in the deployment request. In this respect, the VMs in a single deployment re-

quest (representing a distributed application) may be deployed on different testbeds,

for example after the allocation of a single VM the suitability of a testbed may

change, thus the next VM in the same deployment request may be allocated to another

testbed that fits the suitability criteria.

Bulk Deployment Mode performs selection of testbed for all VMs in a particular

deployment request. In this mode, all VMs (belonging to an application) will be de-

ployed on a single suitable testbed.

In the above two modes, the suitability of a testbed is determined by a combination

of multi-criteria optimization and load balancing functions. The multi-criteria func-

tion takes into account a number of eco-metrics at testbed level with particular em-

phasis on energy consumption, CO

footprint (determined from energy mix consumed

by the geographically distributed testbeds), PUE and GEC. Once a testbed is selected,

the following deployment policies can be used to determine physical host level de-

ployment configuration of new VMs/applications.

Max-Utilization or Task Consolidation tries to maximize the utilization of individ-

ual physical hosts by deploying VMs on minimum number of hosts (e.g. hosts with

most CPU utilization and highest energy consumption) while keeping as many hosts

idle as possible. The idle hosts can be switched off or switched to hibernate mode

provided adequate support by the underlying cloud infrastructure.

Min-Utilization or Task Dispersal, which tries to minimize the utilization of indi-

vidual hosts by deploying VMs on least used hosts (e.g. hosts with most free-CPU

EPS Barcelona 2014 2014 - European Project Space on Information and Communication Systems

and lowest energy consumption) in order to balance the workload across available

hosts.

The combination of deployment modes and policies help in determining an eco-

ware deployment configuration for cloud applications. The experimental evaluation of

the above deployment modes and policies is yet to determine their effectiveness over

each other, however the consideration of environmental impact (by means of consid-

ering energy consumption as well as CO

emission derived from energy sources) is a

step forward in realizing ecologically aware cloud computing.

4.2 Eco-Aware Adaptation

In ECO

Clouds adaptation is referred to a process of change that makes applications

and underlying cloud infrastructure more energy efficient and environmentally friend-

ly. In ECO

Clouds adaptation occurs at two different levels:

 Infrastructural level adaptation: the ECO

Clouds Scheduler drives the adap-

tation by properly managing the VMs.

 Application level adaptation: the Application Controller, an application de-

pendent module, drives the adaptation by distributing the workload among

the VMs or by requesting the allocation and the release of VMs.

4.2.1 Adaptation at Infrastructure Level

In ECO

Clouds the deployment decision making entity i.e. Scheduler only works at

the VM level and has no control over what go on inside a VM (or an application). In

this respect, the adaptation driven by Scheduler tries to achieve the minimum energy

consumption and CO

footprint at the infrastructure level without interfering directly

with the running applications. The overall objectives of adaptation at infrastructure

level are:

 Efficient utilization of available resources in order to keep the energy con-

sumption and CO

footprint of cloud infrastructure to a minimal level

 Fulfilment of requests from running applications

The above objectives are achieved by enacting the following actions at infrastruc-

tural level:

 Supporting applications in starting, freezing and restarting VMs based on

applications’ internal adaptation criteria

 On request (from running applications) allocation and termination of VMs

 Improve the detail of observation of one metric (e.g. change granularity of

metric by varying the frequency of sampling)

 During the runtime of applications, informing users about the potential of

moving certain VMs on more energy efficient and/or CO

friendly resources

such as testbeds and physical hosts.

From Reducing Energy Consumption to Reducing CO2 Emissions - The ECO2Clouds Approach

4.2.2 Adaptation at Application Level

Adaptation at infrastructure level is based on information regarding the status of the

physical hosts and virtual machines. Another source of useful information that needs

to be considered is at application level. For example, at infrastructure level, the

Scheduler notices that a VM has a low CPU load and it might need to ask, the infra-

structure layer, to move the VM to another physical host so the one in which the VM

is currently running can be switched off. Although this choice sounds reasonable, the

Scheduler cannot be aware of the fact that the status of low CPU load is only tempo-

rary and in no more than 10 minutes the CPU load will drastically increase.

To realize application level adaptation ECO

Clouds prescribes that cloud applica-

tions implement an internal management module or Application Controller (AC). AC

is envisioned as a complementary component responsible for providing a set of facili-

ties for accessing the eco-metrics related to the application and enabling the adapta-

tion at application level. The application designers can implement the AC based on

the nature of their applications and best possible adaptation scenarios.

The role of the Application Controller is to read the application level metrics and

to find the best configuration in order to optimize the energy efficiency, reduce CO

footprint and avoid wasting of resources. This can be done by properly distribute the

workload among the several VMs reserved for an application, changing the state of

running VMs based on application specific conditions, or to ask to change the number

of VMs in case they are either too much or not enough. In this way, the application

controller tries to avoid as much as possible, situations in which energy is wasted and

resources are underused.

5 Discussion

The ECO

Clouds solution is implemented as an extension of the BonFIRE platform.

This BonFIRE platform provides a federated cloud infrastructure for deploying and

running VMs of different sizes (e.g., according to the usual nomenclature: small,

medium and large). The platform also provides a monitoring system based on Zabbix

to evaluate the metrics at infrastructure and VM level. In ECO

Clouds, this set of

metrics has been extended to cover the requirements introduced in Section 3.

When running an application, ECO

Clouds platform is able to monitor and collect

the values of all the metrics presented before. Some of the values can be viewed in

Fig.

4. Here two application level metrics, i.e., Application Throughput and the Appli-

cation Energy are compared to identify possible energy wasting. Indeed, the user is in

charge of improving the application and can figure out when there is a discrepancy

between the trend of the energy consumed and the effective work done. In this case,

the evaluation can be only qualitative as these two metrics are not directly compara-

ble.

In some other cases, the metrics can be compared and some adaptation decisions

can be taken. For instance, Fig.

5 shows the trend of the power consumed by the two

VMs used by an application. In this case, we have again some wastage as the

http://www.zabbix.com/

EPS Barcelona 2014 2014 - European Project Space on Information and Communication Systems

Fig. 4. Throughput against energy chart (best viewed in color).

Fig. 5. VMs and Application power chart (best viewed in color).

power of the application is constantly zero (as shown in the legend). As evident in the

figure, during in the reported period the VMs are running but no useful work for the

application is done since the application parts deployed in the VM are not running. In

such situations, when the VMs are running uselessly, the Application Controller can

take an adaptation decision e.g. by requesting the Scheduler to change the state of the

VMs or switch them off in order to conserve energy and make resources free for other

applications. In both cases, the adaptation actions will directly contribute towards

reducing the energy consumption and subsequent CO

footprint of applications.

6 Conclusion and Future Work

The ECO

Clouds project adapts a more direct approach towards tackling environmen-

tal impact of cloud computing by considering not only energy consumption but also

From Reducing Energy Consumption to Reducing CO2 Emissions - The ECO2Clouds Approach

footprint in its cloud sourcing model. The emphasis on CO

footprint and envi-

ronmental impact (primarily derived from energy sources) makes ECO

Clouds ap-

proach stand out from many other efforts that only focus on energy efficiency and

single site cloud infrastructures.

Current ECO

Clouds is going through final stages of the development and imple-

mentation phase. The initial testing of some components has been already initiated.

The next step will be to deploy system level tests, which will prove the effectiveness

of the overall solution and its ability to achieve the ambitious objectives of reducing

the energy consumption and CO

footprint of cloud computing, as set out in the pro-

ject plan.

For further details and updates concerning ECO

Clouds, the readers are referred to

project website: www. eco2clouds.eu

Acknowledgements

This work has been partially supported by the ECO

Clouds project

(http://eco2clouds.eu) and has been partly funded by the European Commission's IST

activity of the 7th Framework Program under contract number ICT- 318048. This

work expresses the opinions of the authors and not necessarily those of the European

Commission. The European Commission is not liable for any use that may be made of

the information contained in this work. The authors thank all the participants in the

project for common work and discussions.

References

1 C. Cappiello, S. Datre, M.G. Fugini, P. Melia, B. Pernici, P. Plebani, M. Gienger, A.

Tenschert: Monitoring and Assessing Energy Consumption and CO2 Emissions in Cloud-

Based Systems. SMC 2013: 127-132

2 Greenpeace: How Clean Is Your Cloud? 17th April 2012. Available on line:

http://www.greenpeace.org/international/en/publications/Campaign-reports/Climate-

Reports/How-Clean-is-Your-Cloud/

3 M. Lenzen, Life cycle energy and greenhouse gas emissions of nuclear energy: A review.

Energy Conversion and Management 49: 2178–2199, 2008

4 The Green Grid Consortium: Harmonizing global metrics to measure data centre energy

efficiency. Oct 2012. Available on line: http://iet.jrc.ec.europa.eu/energyefficiency/sites/

energyefficiency/files/files/documents/ICT_CoC/harmonizing_global_metrics_for_data_

center_energy_efficien cy_2012-10-02.pdf

EPS Barcelona 2014 2014 - European Project Space on Information and Communication Systems