Towards Energy-aware IaaS-PaaS Co-design

Alexandra Carpen-Amarie

, Djawida Dib

, Anne-C

ecile Orgerie

and Guillaume Pierre

INRIA - IRISA, Rennes, France

CNRS - IRISA, Rennes, France

University of Rennes 1 - IRISA, Rennes, France

Keywords:

Cloud Computing, Energy-awareness, IaaS, PaaS.

Abstract:

The wide adoption of the cloud computing paradigm plays a crucial role in the ever-increasing demand for

energy-efﬁcient data centers. Driven by this requirement, cloud providers resort to a variety of techniques

to improve energy usage at each level of the cloud computing stack. However, prior studies mostly consider

resource-level energy optimizations in IaaS clouds, overlooking the workload-related information locked at

higher levels, such as PaaS clouds. In this position paper, we argue that cross-layer cooperation in clouds is

a key to achieving an optimized resource management, both performance and energy-wise. To this end, we

claim there is a need for a cooperation API between IaaS and PaaS clouds, enabling each layer to share speciﬁc

information and to trigger correlated decisions. We identify the drawbacks raised by such co-design objectives

and discuss opportunities for energy usage optimizations. Moreover, we outline the design of a set of extension

modules for Libcloud to serve as building blocks for cross-layer information sharing and cooperation.

1 INTRODUCTION

Energy efﬁciency is a major concern in large-scale

cloud infrastructures. In 2010, data centers world-

wide reportedly consumed 203.4 to 271.8 TWh,

which accounted for 1.1% to 1.5% of the global elec-

tricity use. This number is expected to grow even fur-

ther (Koomey, 2011). To help make the planet greener

while reducing their energy-related costs, cloud op-

erators use a variety of techniques such as workload

consolidation and dynamic virtual machine (VM) re-

sizing. These strategies are however fundamentally

limited by the lack of knowledge by the data center

infrastructure about the applications that generate the

workload. If the data center infrastructure understood

better the full software stack it supports, it would be

able to use more aggressive consolidation policies and

derive further energy savings.

Platform-as-a-Service (PaaS) environments de-

liver users complete runtime environments for build-

ing, deploying and hosting software applications,

while dispensing them from the complexity of man-

aging the underlying software and hardware re-

quired for their applications. PaaS systems rely on

computational resources obtained on-demand from

Infrastructure-as-a-Service (IaaS) clouds, by acting as

clients that can create/remove, suspend/resume, shut-

down/restart, or migrate VMs.

This position paper claims that signiﬁcant energy

gains could be obtained by creating a cooperation API

between the IaaS layer (in charge of handling ele-

mentary computing resources) and the PaaS layer (in

charge of hosting applications). We discuss two com-

plementary approaches for establishing such cooper-

ation:

Cross-layer Information Sharing. Each of the

two layers contains information which is potentially

relevant for the other in order to reduce the energy

footprint of a given application. For example, an

energy-aware IaaS may have an estimate of the en-

ergy footprint of individual virtual machine types,

which may be useful for the PaaS to preferentially se-

lect energy-efﬁcient VM instance types. Conversely,

a PaaS layer may build short-term trafﬁc predictions

which constitute useful hints for IaaS-level consolida-

tion algorithms.

Cross-layer Coordination. Both layers should

coordinate their reconﬁguration actions to reach a

state where they help each other in achieving their

goals rather than potentially take mutually detrimen-

tal decisions. For example, if the IaaS layer decides to

migrate a speciﬁc VM, then the PaaS could in many

cases temporarily reduce the load of the concerned

203

Carpen-Amarie A., Dib D., Orgerie A. and Pierre G..

Towards Energy-aware IaaS-PaaS Co-design.

DOI: 10.5220/0004961402030208

In Proceedings of the 3rd International Conference on Smart Grids and Green IT Systems (SMARTGREENS-2014), pages 203-208

ISBN: 978-989-758-025-3

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

VM to facilitate its migration. Conversely, if the PaaS

layer gives early warnings to the IaaS about future

requests for creating/destroying resources, the IaaS

layer can prepare in advance for these changes.

This paper proposes a research agenda towards the

co-design of IaaS and PaaS layers. Section 2 ﬁrst

discusses the state of the art and motivation for this

work. Section 3 details the issues created by the lack

of IaaS-PaaS co-design. Then, Section 4 presents

our early ideas about opportunities for improvement,

while Section 5 outlines the design of interfaces to en-

able cross-layer cooperation in clouds. Finally, Sec-

tion 6 concludes this work.

2 STATE OF THE ART

Many recent research works have been proposed to

take into account the consumed energy in data cen-

ters. These works can be classiﬁed into two cate-

gories: (1) energy-aware systems and (2) energy ef-

ﬁcient systems. The ﬁrst category aims at inform-

ing cloud providers and users about the consumed

energy for running their applications. For instance

in (Singh et al., 2013), the authors propose a cloud-

based customer-centric architecture that allows con-

sumers to own and control access to their energy us-

age data and have it analyzed using algorithms of their

choice. The second category aims at optimizing the

consumed energy for running applications. Different

techniques are used to achieve energy efﬁciency. In

(Phan et al., 2012), the authors propose to increase

the use of renewable energy by dynamically moving

services across data centers. The works in (Jaianti-

lal et al., 2010; Takouna et al., 2011) propose to ad-

just CPU voltage and frequency according to the load

for saving energy. In (Gadaﬁ et al., 2010), the au-

thors dynamically resize the active server set accord-

ing to varying workload conditions. Workload con-

solidation on a limited number of servers is addressed

in (Duy et al., 2010; Meisner et al., 2009), in order to

allow idle servers to be switched off and save energy.

Other works have focused on analyzing the impact

of various optimization techniques on energy con-

sumption. For instance, in (Orgerie et al., 2010), the

authors investigate the tradeoffs and limitations of ex-

isting energy models for large-scale systems.

3 ISSUES

3.1 Energy-awareness in IaaS and PaaS

PaaS frameworks provide application developers with

runtime environments where applications can be eas-

ily deployed and managed in the cloud (Pierre and

Stratan, 2012; Dib et al., 2013). They do not provide

cloud resources directly, but rather make use of an

underlying IaaS layer in charge of resource manage-

ment. Consequently, PaaS does not have access to in-

formation about the underlying resources, while IaaS

does not have access to information about the run-

ning applications. However, both IaaS and PaaS sys-

tems target multi-objective optimizations including

primarily cost, performance and energy consumption.

These parameters are intrinsically related, thus requir-

ing complex trade-offs to be made between them.

Traditionally, energy-awareness and energy-

efﬁciency have been addressed at the IaaS level with

techniques presented in Section 2. To our knowledge,

no research work has been targeting energy-efﬁciency

or energy-awareness at the PaaS level. However,

studies on real cloud infrastructures show that servers

are considerably underutilized (Zhang et al., 2011),

thus wasting large amounts of energy (Acc, 2010).

This information about resource utilization is not

available at the PaaS level, and nor is the energy

consumption of resources. So, the PaaS layer has

no means to provide energy-related metrics to users

(e.g. energy-efﬁciency of their VMs, energy budget

of their applications), and thus no way to increase

their energy-awareness.

Conversely, the IaaS layer has no information

about the applications and the users. As previously

discussed, such knowledge would be useful to pre-

dict future resource utilization, thus leading to better

energy-related decisions at the IaaS level.

3.2 The Case for IaaS-PaaS Co-design

We believe that the separation of the cloud stack in

two distinct IaaS and PaaS layers, while having great

advantages for portability and separation of concerns,

is detrimental in terms of energy awareness. If each

layer is allowed to take energy-related decisions in-

dependently, these uncoordinated actions can lead to

signiﬁcant resource waste and performance degrada-

tion, possibly negating the beneﬁts of energy aware-

ness altogether.

For instance, the IaaS layer can decide to migrate

a virtual machine (VM) in order to perform a bet-

ter consolidation for energy-efﬁciency purposes. Yet,

this same VM may end a few seconds later because it

SMARTGREENS2014-3rdInternationalConferenceonSmartGridsandGreenITSystems

204

gets released by the PaaS layer. The decision to shut-

down this VM may have been taken several minutes

in advance by the PaaS layer. If this information is

not communicated to the IaaS layer, we take the risk

that IaaS will invest previous resources (for example

by migrating the VM) without seeing any beneﬁt from

this action (because the VM gets shut down just after).

Conversely, the PaaS layer may help the IaaS layer

in performing its VM management actions. For ex-

ample, it is often easy at the PaaS level to temporarily

redirect one VM’s workload to another (by redeﬁning

load balancing parameters for example). Ofﬂoading a

VM for just a few tens of seconds may greatly facili-

tate IaaS-level management tasks such as VM migra-

tion.

In order to avoid counterproductive independent

optimizations, we believe that the IaaS and the PaaS

should share their energy-related information and co-

ordinate their reconﬁguration actions. This coordina-

tion aims at allowing system-level optimizations and

trade-offs.

4 OPPORTUNITIES

In order to design efﬁcient cloud frameworks, both

performance and energy-wise, it is essential to under-

stand how the cloud stack layers can interact and co-

operate. We argue that there is a need to extend the

traditional boundaries of each class of cloud comput-

ing frameworks to ensure an effective management of

resources, while delivering seamless application per-

formance and SLA compliance to higher-level users.

In this section we identify a set of design opportuni-

ties for both PaaS and IaaS providers to help address

these goals. The research directions we investigate

are twofold: ﬁrst, we focus on the bidirectional in-

formation sharing opportunities between these cloud

layers and the coordinated decisions they can trigger.

Second, we discuss security concerns that arise from

such interactions.

4.1 Workload-aware IaaS Frameworks

The emergence of a wide array of open-source IaaS

solutions led to a gradual adoption of open standards

and interfaces. This was a key incentive for enter-

prises to deploy private clouds in their data centers,

allowing them to take advantage of the ﬂexibility

and performance of virtualized infrastructures, with-

out the vendor lock-in shortcomings of proprietary so-

lutions.

Most open IaaS cloud frameworks deliver simi-

lar services, typically focusing on resource manage-

ment, VM lifecycle and on providing framework us-

age statistics. However, little has been done to achieve

energy-efﬁcient resource usage in such infrastructure

clouds, despite the fact that lowering energy con-

sumption is becoming a growing concern for data

center management. This context provides a valu-

able opportunity for IaaS cloud providers to consider

techniques that take into account the requirements of

higher-level services, such as PaaS offerings, which

traditionally build on infrastructure clouds. To this

end, we believe that PaaS clouds can deliver a set of

insights to the infrastructure level to guide its resource

allocation and energy optimization decisions.

4.1.1 Passing Workload Information

VM allocation strategies typically rely on information

concerning physical machine capabilities and their

usage across the data center. Nevertheless, several

factors may impact the performance of VMs, with

critical side effects on application performance, user

costs and power consumption. For instance, intensive

network trafﬁc between a few VMs may lower the

available bandwidth across an entire cluster. On the

other hand, PaaS virtual resource management ser-

vices use the IaaS APIs to specify the number and

type of VMs required upon application deployment.

No information about the application class or its ex-

ecution parameters is forwarded to the underlying in-

frastructure.

In this context, the answer to achieving better re-

source management at the infrastructure level is to en-

able IaaS schedulers to analyze and exploit a wider

spectrum of parameters associated with the applica-

tion behavior. Thus, as such workload properties are

generally opaque to the underlying VM manager, a

possible approach is to enable PaaS services to expose

them to the infrastructure layer, as detailed below.

Resource Usage Patterns. The resource needs of

VMs have a potentially heavy impact on the VM mi-

gration duration. As an example, it was shown that

memory-bound VMs lead to very inefﬁcient migra-

tion times (Liu et al., 2011). IaaS systems may avoid

such operations by ﬂagging affected VMs at deploy-

ment time. A more ﬁne-grained approach may enable

the PaaS services to trigger such ﬂags dynamically, as

the application enters a memory-intensive stage, and

allow migration outside these intervals.

Execution Time. Cloud services can accommo-

date a large set of application types. For instance,

web servers represent long-running jobs with speciﬁc

access patterns. Oppositely, a short-lived VM, such

as a MapReduce job, may raise speciﬁc scheduling

constraints. Thus, as many MapReduce jobs run for

short periods of time, migrating the VMs during ap-

TowardsEnergy-awareIaaS-PaaSCo-design

205

plication execution may degrade their performance to

unacceptable levels (e.g., migrating the VM may take

longer than its actual runtime).

Elasticity needs. Unlike the infrastructure man-

ager, the PaaS services may be able to estimate the

application requirements in terms of workload peaks.

As elasticity is one of the most appealing cloud

features, anticipating application needs and passing

them on to the IaaS layer may result in more timely

workload adaptation mechanisms, which cannot be

achieved by the PaaS services alone.

4.1.2 Using Workload Information to Reduce

Energy Consumption

Such workload knowledge may allow IaaS providers

not only to optimize resource allocation decisions,

but also to improve the overall energy-consumption

levels. This goal is typically achieved by adapting

the number of used physical nodes to the workload

and shutting down underutilized nodes (Orgerie et al.,

2010). Taking into account additional PaaS-level as-

pects may lead to signiﬁcant and accurate energy-

saving decisions, as exempliﬁed below.

Resource Scheduling. Resource allocation in

IaaS clouds is generally based on the VM type and

straightforward scheduling policies. Beyond the VM

type in terms of required CPUs and memory, an es-

sential parameter is represented by the association of

VMs with a speciﬁc virtual cluster, and thus, a sin-

gle application. By taking into account this type of

hint made available by PaaS-level services, an IaaS

provider may change the allocation process of similar

VMs. For example, VMs belonging to the same job

may be deployed on the same physical machines to

beneﬁt from data locality and avoid unnecessary and

performance-degrading network trafﬁc.

Node shutdown Policies. A widely studied mech-

anism to minimize energy consumption is shutting

down unused nodes. However, resource managers

usually lack any knowledge related to the type of

workload or about the real-time evolution of resource

requirements. Such information provided by higher-

level services such as PaaS providers may allow for

more aggressive shutdown policies, without incurring

performance degradations. When the infrastructure

manager is aware of the short-term evolution of its

workload, it can adopt less conservative policies to

reduce the powered-on resources and the total energy

usage.

Management of Workload Peaks. Cloud ser-

vices have developed on the premise that they can

provide on-demand resource adaptation to workload

bursts, relieving the client from the burden of over-

provisioning resources to cope with unexpected work-

load ﬂuctuations. Without any means to foresee the

resource requirements of its scheduled jobs, an IaaS

cloud is constrained to maintain a pool of unused

running physical machines to accommodate possible

workload peaks. To minimize the impact of such an

approach on energy consumption, PaaS services may

anticipate workload trends and, furthermore, forward

them to the infrastructure level. Thus, IaaS schedulers

can redirect small bursts to public clouds when the

cost of switching nodes on is more important than the

cost of obtaining remote VMs, both performance and

energy-wise. When facing longer-duration workload

peaks, the IaaS cloud can automatically adapt the pool

of available physical machines, by enabling nodes in

advance.

4.2 Optimizing PaaS Resource

Management

Whereas advice from application-aware PaaS ser-

vices may improve the efﬁciency of resource man-

agement decisions, PaaS frameworks can also beneﬁt

from infrastructure-level insights.

4.2.1 Exposing Low-level Knowledge to PaaS

Clouds

Infrastructure Details. PaaS management services

interact with the underlying infrastructure through the

interfaces exposed by the IaaS provider, which typi-

cally hide low-level details from the user. As this in-

formation is provided in private clouds for administra-

tive purposes, it can be also made available for higher-

level entities, such as PaaS frameworks running on the

infrastructure. For instance, knowledge about cluster

utilization and node properties may allow the PaaS

services to select the most suitable VM types to ex-

ecute an application, enabling the use of large VMs

when the infrastructure comprises high-performance

physical machines.

Energy Awareness. The ﬁrst step towards ener-

gy-efﬁcient services at the platform level is user

awareness, in particular if lowering the energy con-

sumption is associated with monetary incentives. The

user may be presented with several execution plans

and the corresponding energy consumption estima-

tions for the same application. He can thus choose

the best execution schedule in terms of energy, regard-

less of a performance loss or a longer execution time.

Nevertheless, the PaaS layer has no means to accu-

rately estimate the energy consumption of a given ap-

plication. It is at the IaaS level where such predictions

can be made, based on the load of the infrastructure

and the application requirements.

SMARTGREENS2014-3rdInternationalConferenceonSmartGridsandGreenITSystems

206

4.2.2 IaaS-PaaS Coordinated Decisions

Since PaaS services do not have full control over the

management of their virtual resources, the key to an

optimized resource usage and energy consumption

lies in cross-layer coordination. Despite the fact that

all VM allocation and placement decisions are taken

at the infrastructure level, such a collaboration can

lead to both improved scheduling mechanisms and a

reduced incidence of conﬂicting actions.

Adaptive Resource Management. Equipped

with detailed information about the user job, the IaaS

layer can yield multiple scheduling options to handle

it. According to user set thresholds, such as an ex-

ecution deadline or a given energy budget, the IaaS

resource manager may be extended to delay jobs ac-

cording to the availability of energy-efﬁcient or re-

newable energy sources (e.g., wind energy). PaaS

may facilitate certain VM management actions as

well. To reduce the time spent during VM migration,

the PaaS can be instructed to temporarily decrease the

load of the VM, which may entail performance and

energy gains.

Preventing Conﬂicting Optimizations. Cloud

cross-layer communication contributes to another un-

expected, yet signiﬁcant optimization angle for VM

management, that is to prevent conﬂicting optimiza-

tions. As an example, the PaaS layer may detect an

idle VM and redirect a slice of the job workload to

it. Simultaneously, the infrastructure manager can de-

cide to migrate the same idle VM, resulting in both an

inefﬁcient migration and a reduced performance for

the PaaS application. To prevent such conﬂicting or

counter-productive optimizations, it is essential to de-

ﬁne a standardized interface and a set of parameters to

be passed between the two layers, which can ensure

portability and usability across cloud providers.

4.3 Privacy and Security Issues

As detailed before, cross-layer coordination can help

increasing both energy awareness and energy efﬁ-

ciency of the entire cloud system. However, this coor-

dination needs to be clearly deﬁned as it can raise ma-

jor concerns about the security and privacy of cloud

providers and users. Indeed, it has been shown that

having access to the energy consumption of a cloud

server can allow people to guess with high probabil-

ity what type of application (among various possible

ones) is running in its virtual machines (Hlavacs et al.,

2011). So, passing the information about energy con-

sumption of servers is not possible for privacy’s sake.

It is therefore necessary to consider a coordination

of these platforms that guarantees the privacy of the

users and of the applications, while providing useful

information to both layers in order to save energy. To

do so, it is required to identify sensitive information

that cannot be communicated between the two lay-

ers: energy consumption of physical servers, location

of virtual machines (allowing users to target speciﬁc

physical machines in order to be co-located with spe-

ciﬁc virtual machines). Moreover, the IaaS provider

and the PaaS provider can be different, and they may

not want to share sensitive information about the uti-

lization of their platform.

5 PROPOSAL

To facilitate the interaction between cloud layers

while preserving the separation and the interoperabil-

ity across the cloud stack, we argue there is a need for

an abstraction layer proposing coordination APIs. For

instance, let us take the example of the Apache Lib-

cloud (Libcloud, 2014) library. Libcloud consists of a

set of Python libraries providing a uniﬁed abstraction

across IaaS cloud APIs such as Amazon EC2, Open-

Stack or OpenNebula. It hides the differences be-

tween the various cloud providers by offering generic

APIs for compute, storage, authentication, load bal-

ancers and DNS services. We believe such interfacing

libraries can be the key of cross-layer interoperabil-

ity in the cloud stack. The development of client li-

braries that expose APIs to promote collaborative de-

cisions between IaaS clouds and higher-level frame-

works may come as an incentive for cloud providers

to exploit the potential beneﬁts of inter-layer cooper-

ation.

We aim at deﬁning a set of APIs to complement

existing Libcloud features with new mechanisms for

designing IaaS-PaaS cooperation strategies. Our goal

is twofold. First, we plan to enable IaaS clouds to ex-

pose extended infrastructure information to upper lay-

ers or applications. In this context, two sets of APIs

are required to enable PaaS providers to optimize re-

source management, as detailed in Section 4.2. On the

one hand, the IaaS provider can supply data about the

infrastructure status, including cluster utilization and

node properties, which can be accessed on-demand

through the Libcloud API. On the other hand, the

Libcloud suite may include a set of APIs dedicated

to exposing infrastructure energy consumption at the

higher levels, as the ﬁrst step towards energy-aware

PaaS and IaaS applications.

Our second goal is to add abstract triggers at

the Libcloud level to enable PaaS managers to push

workload-related information to the IaaS clouds. As

the applications executed on VMs rented from IaaS

TowardsEnergy-awareIaaS-PaaSCo-design

207

clouds are opaque to the infrastructure level, such

APIs are essential to make application information

available for the IaaS resource allocation and manage-

ment mechanisms. Consequently, such APIs should

allow PaaS managers to ask the IaaS provider to ﬂag

the activity of speciﬁc VMs or groups of VMs as

CPU-, IO- or memory-intensive, to predict workload

peaks or to provide hints regarding the execution time

of groups of VMs that belong to the same job.

Such an API-driven approach to cross-layer co-

operation achieves several beneﬁts. First, each cloud

provider may choose to implement only the required

plugins according to the amount of information it

intends to share. While private clouds may imple-

ment the whole set of energy-related and coopera-

tion APIs, public providers may decide to share only

non-sensitive data. Furthermore, an abstraction layer

such as Libcloud fosters interoperability and porta-

bility between clouds, such that PaaS providers can

make data available and collect infrastructure infor-

mation according to predeﬁned formats, while each

IaaS cloud can provide its own interface implementa-

tion according to its internal structure and APIs.

6 CONCLUSION

The growing demand for “green” and energy-efﬁcient

data centers is a key driver in the design of cloud

computing frameworks, which target not only perfor-

mance, but also cost efﬁciency. In particular, design-

ing virtualized environments for energy efﬁciency has

led to a wide spectrum of optimization techniques,

most of which are intended for IaaS cloud offerings.

Whereas higher-level PaaS services hold key infor-

mation related to workload properties and execution,

they do not have direct control over the resource-

management mechanisms that could trigger energy

gains. In this work, we argue that there is a need

for coordinated actions between IaaS and PaaS cloud

layers. We believe that the conventional boundaries

of the cloud stack layers can be extended to promote

energy-awareness at higher levels, while enhancing

low-level resource managers with detailed workload

proﬁles to achieve ﬁne-grained energy saving mecha-

nisms.

We investigated a set of opportunities for both

IaaS and PaaS providers to address these objectives

and highlighted the tradeoffs of cross-layer informa-

tion passing. As future work, our goal is to design the

proposed API suite by extending Libcloud, in order to

enable cross-layer cooperation between open-source

IaaS and PaaS cloud implementations and to develop

prototypes of cooperating cloud frameworks.

REFERENCES

Acc (2010). Cloud Computing and Sustainability: The En-

vironmental Beneﬁts of Moving to the Cloud. Techni-

cal report, Accenture, Microsoft and WSP Report.

Dib, D., Parlavantzas, N., et al. (2013). Meryn: open, SLA-

driven, cloud bursting PaaS. In Proc. of ORMaCloud

’13, pages 1–8, New York, NY, USA. ACM.

Duy, T., Sato, Y., et al. (2010). Performance evaluation of

a Green scheduling algorithm for energy savings in

cloud computing. In 2010 IEEE Int. Sym. on Par. Dist.

Proc., Workshops (IPDPSW), pages 1–8.

Gadaﬁ, A., Tchana, A., et al. (2010). Autonomic energy

management in clusters. In Proc. of the 1st ACM

Workshop on Green Comp., pages 16–21, NY, USA.

Hlavacs, H., Treutner, T., et al. (2011). Energy consumption

side-channel attack at virtual machines in a cloud. In

IEEE Int. Conf. on Cloud and Green Comp., pages

605–612.

Jaiantilal, A., Jiang, Y., et al. (2010). Modeling cpu energy

consumption for energy efﬁcient scheduling. In Proc.

of the 1st Workshop on Green Computing, GCM ’10,

pages 10–15, New York, NY, USA. ACM.

Koomey, J. (2011). Growth in data center electricity use

2005 to 2010. Oakland, CA: Analytics Press, 1.

Libcloud (2014). Apache Libcloud.

http://libcloud.apache.org/.

Liu, H., Xu, C.-Z., et al. (2011). Performance and energy

modeling for live migration of virtual machines. In

Proc. of the 20th ACM Int. Sym. on High Performance

Dist. Comp. (HPDC ’11), pages 171–182, USA.

Meisner, D., Gold, B. T., and Wenisch, T. F. (2009). Power-

nap: eliminating server idle power. SIGPLAN Notices,

44(3):205–216.

Orgerie, A.-C., Lefevre, L., and Gelas, J.-P. (2010). Demys-

tifying energy consumption in grids and clouds. In

Proc. of the Int. Green Comp. Conf., pages 335–342.

Phan, D. H., Suzuki, J., et al. (2012). Evolutionary multi-

objective optimization for green clouds. In Proc. of

the 14th ACM Int. Conf. on Genetic and Evolutionary

Computation Conf. Companion, GECCO Companion

’12, pages 19–26, New York, NY, USA.

Pierre, G. and Stratan, C. (2012). ConPaaS: a platform

for hosting elastic cloud applications. IEEE Internet

Computing, 16(5):88–92.

Singh, R. P., Keshav, S., et al. (2013). A cloud-based

consumer-centric architecture for energy data analyt-

ics. In Proc. of the 4th ACM Int. Conf. on Future en-

ergy systems, e-Energy ’13, pages 63–74, USA.

Takouna, I., Dawoud, W., et al. (2011). Dynamic conﬁg-

uration of virtual machine for power-proportional re-

source provisioning. In Proc. of the ACM Int. Work-

shop on Green Comp. Middleware, pages 1–6, USA.

Zhang, Q., Hellerstein, J., et al. (2011). Characterizing

Task Usage Shapes in Google’s Compute Clusters. In

Int. Workshop on Large Scale Distributed Systems and

Middleware.

SMARTGREENS2014-3rdInternationalConferenceonSmartGridsandGreenITSystems

208