Cost-efﬁcient Datacentre Consolidation

for Cloud Federations

Gabor Kecskemeti

, Andras Markus

and Attila Kertesz

Department of Computer Science, Liverpool John Moores University, U.K.

Software Engineering Department, University of Szeged, 6720 Szeged, Dugonics ter 13, Hungary

Keywords:

Cloud Computing, Datacentre Consolidation, Simulation.

Abstract:

Cloud Computing has become mature enough to enable the virtualized management of multiple datacentres.

Datacentre consolidation is an important method for the efﬁcient operation of such distributed infrastructures.

Several approaches have been developed to improve the efﬁciency e.g. in terms of power consumption, but

only a few attention has been turned to combining pricing methods with consolidation techniques. In this

paper we discuss how we introduced cost models to the DISSECT-CF simulator to foster the development

of cost efﬁcient datacentre consolidation solutions. We also exemplify the usage of this extended simulator

by performing cost-aware datacentre consolidation. We apply real world traces to simulate cloud load, and

propose 7 strategies to address the problem.

1 INTRODUCTION

Cloud computing enabled the virtualized manage-

ment and sharing of software and hardware solutions,

including computing and storage resources and ap-

plication runtimes. The elasticity of Infrastructure as

a Service (IaaS) clouds allows commercial providers

to better exploit their datacentres as well as increase

their incomes. Datacentre consolidation is a techni-

que that helps achieving these goals. Related works

in datacentre load balancing and consolidation have

already shown that multi-objective proposals can hin-

der performance and increase the problem complex-

ity, therefore innovative solutions are needed to deal

with multiple and complex aims.

In this paper, we investigate novel, cost-aware Vir-

tual Machine (VM) consolidation methods for cloud

datacentres using the DISSECT-CF simulator (Kec-

skemeti, 2015), which is a generic tool for investi-

gating infrastructure clouds. We discuss how to in-

troduce cost models to the DISSECT-CF simulator

to investigate cost efﬁcient datacentre consolidation

techniques. We also propose 7 different strategies to

perform VM consolidation, and present how to use

the extended simulator by performing cost-aware da-

tacentre consolidation by evaluating these algorithms.

We apply real world traces to simulate cloud load du-

ring the experiments.

The structure of the paper is the following. First,

in Section 2, we discuss state of the art approa-

ches in the ﬁeld. In Section 3, we introduce the

cost models applied in the DISSECT-CF simulator.

Section 5 discusses our experiments and measure-

ment results achieved with the extended simulator. Fi-

nally, Section 6 concludes our work.

2 RELATED WORK

Ahmad et al. presented a survey on datacenter con-

solidation solutions in (Ahmad et al., 2015). They

argued that virtual machine (VM) migration and dy-

namic voltage frequency scaling (DVFS) methods

are generally used to achieve server consolidation,

which help to achieve resource management goals

like load balancing and power management, though it

also affects application performance. They concluded

the survey that the unpredictable nature of worklo-

ads and the inability to accurately predict application

demands call for dynamic, lightweight and adaptive

VM migration designs to improve application perfor-

mance. Filho et al. (Filho et al., 2018) published anot-

her survey in this ﬁeld. They state that VM placement

and migration are the major challenging issues in ma-

nagement of virtualized datacenters, and many pro-

posals apply different approaches ranging from linear

programming, to genetic algorithms. They also sho-

wed that multi-objective proposals can reduce perfor-

Kecskemeti, G., Markus, A. and Kertesz, A.

Cost-efﬁcient Datacentre Consolidation for Cloud Federations.

DOI: 10.5220/0006775302130220

In Proceedings of the 8th International Conference on Cloud Computing and Services Science (CLOSER 2018), pages 213-220

ISBN: 978-989-758-295-0

213

Figure 1: The architecture of the DISSECT-CF simulator .

mance and increase the problem complexity, therefore

innovative solutions are needed to deal with multiple

and complex aims. Our proposed simulation environ-

ment aims at providing a way for investigating certain

policies to achieve these goals.

Concerning simulations, the CloudSim toolkit has

been widely used to propose and evaluate certain heu-

ristics for datacenter consolidation, such as in (Abdul-

lah et al., 2017) and (Kertesz et al., 2016). Though

these solutions provide load balance improvements,

they do not take into account and do not apply provi-

der pricing.

3 OUR PROPOSED COST MODEL

FOR CLOUD DATACENTRE

MANAGEMENT

DISSECT-CF is a compact open source (DISSECT-

CF, 2017) simulator focusing on the internals of IaaS

systems. Figure 1 presents its architecture including

our extensions (denoted with grey colour). There are

six subsystems (encircled with dashed lines) imple-

mented, each responsible for a particular functiona-

lity: (i) event system – the primary time reference;

(ii) uniﬁed resource sharing – models low-level re-

source bottlenecks; (iii) energy modelling – for the

analysis of energy-usage patterns of resources (e.g.,

NICs, CPUs) or their aggregations; (iv) infrastructure

simulation – for physical/virtual machines, sensors

and networking; (v) cost modelling – for managing

IoT and cloud provider pricing schemes, and (vi) in-

frastructure management – provides a cloud like API,

cloud level scheduling, and IoT system monitoring

and management.

In a recent work (Markus et al., 2017), we intro-

duced the following new components to model IoT

Cloud systems: Sensor, IoT Metering and IoT Con-

troller. Sensors are essential parts of IoT systems,

and usually they are passive entities (actuators could

change their surrounding environment though). Their

performance is limited by their network gateway’s

(i.e., the device which polls for the measurements and

sends them away) connectivity and maximum update

frequency. Our network gateway model builds on

DISSECT-CF’s already existing Network Node mo-

del, which allows changes in connection quality as

well. In our model, the Sensor component is used

to deﬁne the sensor type, properties and connections

to a cloud system. IoT Metering is used to deﬁne and

characterize messages coming from sensors, and the

IoT Controller is used for sensor creation and mana-

gement.

To incorporate cost management, we enabled de-

ﬁning and applying provider pricing schemes both for

IoT and cloud part of the simulated environments.

These schemes are managed by the IoT and Cloud

Pricing components of the Cost modeling subsystem

of DISSECT-CF, as shown in Figure 1.

3.1 Conﬁgurable Cost Models based on

Real Provider Schemes

In order to enable realistic datacentre consolidation

simulations, we considered four of the most popu-

lar, commercial cloud providers, namely: Amazon,

MS Azure, IBM Bluemix and Oracle. Most providers

have a simple pricing method for VM management

(beside thaditional virtual machines, some provide

containers, compute services or application instances

for similar purposes). The pricing scheme of these

providers can be found on their websites. We conside-

red the Azure’s application service (Azure, 2017), the

Bluemix’s runtime pricing sheet under the Runtimes

section (IBM, 2017), the Amazon EC2 On-Demand

prices (Amazon, 2017), and the Oracle’s compute ser-

vice (Oracle, 2017) together with the Metered Servi-

ces pricing calculator (Oracle Calculator, 2017). The

cloud-related cost is based on either instance prices

(Azure and Oracle), hourly prices (Amazon) or the

mix of the two (Bluemix) provider uses both type of

price calculating. For example, Oracle charges depen-

ding on the daily uptime of our application as well as

the number of CPU cores used by our VMs.

Figure 2 shows the XML structure and the cost va-

lues for the applied categories we designed to be used

in the simulator. This conﬁguration ﬁle contains some

providers (for example the amazon element starting in

the second line), and the deﬁned values are based on

the gathered information from the providers’ public

websites discussed before. We speciﬁed 3 different

sizes for applicable VMs (named small, medium and

large).

CLOSER 2018 - 8th International Conference on Cloud Computing and Services Science

214

<instance-price>18.15</instance-price>

<hour-per-price>0.094</hour-per-price>

</medium>

</amazon>

<instance-price>139</instance-price>

<hour-per-price>0</hour-per-price>

</medium>

<large>

<instance-price>268</instance-price>

<hour-per-price>0</hour-per-price>

</large>

</oracle>

<large>

<instance-price>0</instance-price>

<hour-per-price>0.296</hour-per-price>

</large>

</bluemix>

</cloudproviders>

Figure 2: Cost model of Cloud providers.

This XML ﬁle has to contain at least that size ca-

tegory to be used for the experiments. As we can see

from the fourth line to the seventh line, a category de-

ﬁnes a virtual machine with the given ram and cpuco-

res attributes, and we state the virtual machine prices

with the instance-price and hour-per-price attributes.

If we select the amazon provider with small category,

then in the scenarios a virtual machine will have 1

CPU core and 2 GB of RAM, and the usage of this

virtual machine will cost 0.296 Euro per hour.

4 CONSOLIDATOR

ALGORITHMS

Data-centre consolidation techniques are heavily used

in commercial clouds. Consolidation is built on the

migration capabilities of virtual machines, where vir-

tualised workload is moved around in the data-centre

according to the cloud operator’s goals. In the past ye-

ars, there were several approaches proposed for con-

solidating the virtualised workloads of clouds. Most

of them were evaluated with simulations. When

analysing cost models, the effects of consolidation

could not be avoided. Although, the foundations for

these consolidator algorithms were laid down in our

DISSECT-CF simulator from the beginning (Kecske-

meti, 2015). Even with the addition of more precise

live-migration modelling (Maio et al., 2016), the con-

solidation algorithms were not present in the simula-

tor.

There are two distinct approaches possible to im-

plement a consolidation algorithm in DISSECT-CF:

(i) create an alternative physical machine (PM) con-

troller which utilizes consolidator related techniques

as well or (ii) create an independent consolidator

which builds on top of the other infrastructure ma-

nagement components of the simulator. While both

approaches could apply the same policies and enact

the same goals of a cloud provider, they should be im-

plemented differently. In the ﬁrst case, the PM con-

troller should extend its possible actions from swit-

ching on/off PMs to migrating VMs as well. In the

second case, the consolidator is dedicated to only de-

cide on migration related actions. This is beneﬁcial

as the consolidator algorithm could collaborate with

multiple PM controller strategies without the need for

a complete rewrite of the consolidation approach. As

this second approach is more generic, thus we present

it in this paper in more detail. Note, the source of the

presented approach is publicly available in the source

repository of DISSECT-CF (DISSECT-CF, 2017).

Figure 3 shows how the extension was imple-

mented. The main addition of the simulator is the

Consolidator class, which is to be extended by any

new consolidation policies in the future. This abstract

class handles the basic connection of the future con-

solidators to the IaaSService by monitoring the VM

related activities on the cloud. It is also responsible

for managing the frequency with which the consoli-

dation policy is run (to be implemented by third par-

ties in the doConsolidation() function). In general,

it ensures that the custom consolidator policy is only

invoked if there are any VMs in the cloud at any par-

ticular time. To do so, the consolidator monitors the

PMs and observes how they are managed by PM con-

trollers and utilised by the VM schedulers.

The simulator also offers a consolidation policy

called SimpleConsolidator. This policy packs the

VMs to the smallest amount of PMs as follows.

1. Creates an ordered list (P := {p

, p

, ...p

}) of

the PMs (e.g., p

) currently running in the IaaS

(where the number of running PMs in the IaaS is

n). This list has the least used PMs in the front and

the heaviest used ones at the tail: u(p

) ≤ u(p

) ≤

... ≤ u(p

). Where we denote the utilisation of a

Cost-efﬁcient Datacentre Consolidation for Cloud Federations

215

IaaSService Consolidator

SimpleConsolidator

0..1toConsolidate

PhysicalMachine

VirtualMachine

managed

hosted

observed

#doConsolidate()

Abstract, expects

implementors to

provide consolidation

policy and goals

Figure 3: Consolidation related extension of DISSECT-CF.

PM with the function : u : P → R. Note: the utili-

sation is determined solely on the resource alloca-

tions for the VMs hosted on each PM and it is not

dependent on the instantaneous resource usage of

any of the VMs in the cloud.

2. Picks the least used not yet evaluated PM (p

). If

there are no more PMs to evaluate, we terminate

the algorithm.

3. Picks a VM (v

) hosted by p

. Where v

∈ h(p

)

and the function h : P → 2

deﬁnes the set of VMs

which are hosted by a particular PM. This set is a

subset of all VMs (V ) in the IaaS service.

4. Picks the heaviest used (but not completely utili-

sed) and not yet tested PM (p

). Where we have

the following limits for k: i + 1 ≤ k ≤ n.

5. Checks if the new PM has enough free resour-

ces to host the VM: r

,t) ≥ r(v

,t), where

: P × R → R

and r : V × R → R

. The r

function tells the amount of free resources avai-

lable at the speciﬁed host at the speciﬁed time

instance t. Also, the r function tells the amount

of resources needed by the virtual machine at the

speciﬁed time instance. The resource set is mo-

delled by a triplet of real numbers: (i) number of

CPU cores, (ii) per core processing power and ﬁ-

nally (iii) memory.

• If the check was successful, then the VM is re-

quested to be migrated from the host p

to p

Then continue on with a new VM pick.

• If the check fails, we repeat with all untested

PMs. If no more PMs are around to test, we

pick another VM from the list of h(p

). If there

are no more VMs to pick, we return to step 2.

Thus we can summarize the algorithm as packing the

VMs to the heaviest loaded PMs with a ﬁrst ﬁt appro-

ach. This approach is efﬁcient with the PM controller

called SchedulingDependentMachines which swit-

ches off all unused machines once they become freed

up (in this case once all their VMs migrate away).

5 EVALUATION

During our implementation and evaluation, where ap-

plicable, we used publicly available information to

populate our experiments. In the next subsection we

introduce the applied workloads, then discuss the pro-

posed algorithms and scenarios, and the achieved re-

sults.

5.1 Workloads

Though virtual machine management log-based tra-

ces would be the best candidates for analysing cloud

characteristics, traces collected from other large-scale

infrastructures like grids could also be appropriate.

Generally two main sources are used for this purpose:

the Grid Workloads Archive (GWA (GWA, 2017))

and the Parallel Workloads Archive (PWA, 2017). For

this study we used traces downloadable from GWA

(namely: AuverGrid, DAS2, Grid5000, LCG, Nor-

duGrid and SharcNet).

We used the JobDispatchingDemo from the

DISSECT-CF examples project

, to transform the

jobs listed in the trace to VM requests and VM activi-

ties. This dispatcher asks the simulator to ﬁre an event

every time when the loaded trace prescribes. Also, the

dispatcher maintains a list of VMs available to serve

job related activities (e.g., input & output data trans-

fers, cpu and memory resource use). Initially the VM

list is empty. Thus the job arrival event is handled

with two approaches: (i) if there is no unused VM in

the VM list that has sufﬁcient resources for the pres-

cribed job, then the dispatcher creates a VM accor-

ding to the resource requirements of the job; alterna-

tively, (ii) if there is an unused VM with sufﬁcient

resources for the job, then the job is just assigned to

the VM. In the ﬁrst approach, the job’s execution is

delayed until its corresponding VM is spawned. In

both cases, when the job ﬁnishes, it marks the VM

as unused. This step allows other jobs to reuse VMs

pooled in the VM list. Finally, the VMs are not kept

for indeﬁnite periods of time, instead they are kept

in accordance with the billing period applied by the

cloud provider. This ensures, that the VMs are held

for as long as we paid for them but not any longer. So

if there is no suitable job coming for a VM within its

billing period, then the VM is terminated and it is also

removed from the VM list.

5.2 Scenarios

In the following we list the pricing strategies availa-

ble at the moment. They are applicable alone or in

https://github.com/kecskemeti/dissect-cf-examples

CLOSER 2018 - 8th International Conference on Cloud Computing and Services Science

216

combination as required.

S1 - Fixed Pricing. It uses a constant price for every

VM request. This pricing strategy does not consi-

der any factors in its price:

f ix

= m

, (1)

where M

f ix

is the price (i.e money) returned, and

is the constant base price which is conﬁgurable

for every simulation.

S2 - Resource Constraints Aware Pricing. It im-

plements a linear relationship between the price

of a VM and the amount of resources the VM

needs. The higher the resource needs are, the

more the user should pay.

rcaw

cores

, r

mem

, r

proc

) =

cpu

∗ r

proc

∗ r

mem

MAX

cpu

∗ r

MAX

proc

∗ r

MAX

mem

(2)

where the triple < r

cores

, r

mem

, r

proc

> represents

the resources requested by the customer for its

VM. The triple < r

MAX

cpu

∗ r

MAX

proc

∗ r

MAX

mem

> repre-

sents the properties of the largest resource amount

any PM has in the cloud provider. Note that all

the resource values are represented as the provi-

der sees them ﬁt, for the purpose of the paper we

assumed they are all positive real numbers (e.g.,

cores

∈ R

). Thus, this pricing model, charges

if the user requests the largest still serviceable

resource set.

S3 - Quantized Pricing. It applies a pricing strategy

similar to M

rcaw

. But instead of scaling the price

by a continuous function, we apply a transforma-

tion which transforms (T : R

→ R

) the original

request from the user to some preset values. When

deﬁning a quantized pricing, one must deﬁne this

transformation only, then we can apply the M

rcaw

model to ﬁnd out the actual price.

quant

cores

, r

mem

, r

proc

) =

rcaw

(T(r

cores

, r

mem

, r

proc

))

(3)

This is the technique that is used by most of cloud

providers nowadays. In those cases, the providers

are often restricting the amount of resources one

can request as well. An example transformation

function could be:











if r

mem

<= 2 ∧ r

cores

<= 1

mem

= 2, r

cores

= 1, r

proc

= 1

if 2 < r

mem

<= 8 ∧ 1 < r

cores

<= 2

mem

= 8, r

cores

= 2, r

proc

= 1

otherwise

mem

= 32, r

cores

= 8, r

proc

= 1

(4)

The simulator implements a pricing model which

can be conﬁgured to load a particular transforma-

tion function for a particular cloud provider. The

limits for the transformation functions are stored

in an XML ﬁle representing certain commercial

provider cost models. Later in the measurements

we apply the cost model presented in Section 3.

S4 - PM Utilization Aware Pricing. This strategy

also offers a linear pricing approach. In contrast

to the resource constraints aware pricing model,

this time, we adjust the price based on the number

of PMs in use at cloud provider:

utaw

= m

|P|

, (5)

where the P

is the set of PMs that host any VMs:

= {∀p

∈ P : u(p

) 6= 0}. Thus the more ex-

ploited the cloud provider is, the more the user

should pay.

S5 - Load Dependent Pricing. This works similarly

to the PM utilization aware pricing. At the cost

of additional monitoring requirements, it imple-

ments the same policy with a more ﬁne grained

utilization calculation:

= m

∑

∀p∈P

R(p)

∑

∀p∈P

MAX

(p)

, (6)

where R(p) represents the average amount of re-

sources utilised in the last hour from particular

physical machine, while R

MAX

(p) deﬁnes the to-

tal amount of resources the PM could offer in

the same hour. Thus, this pricing model consi-

ders how well the VMs actually use the resources

and if the VMs are not highly used (even though

they are hosted at the cloud at the moment), then

the prices will be lowered (this will attract further

users and enable the provider to use under provisi-

oning for those VMs that are just paid for but not

used at the moment).

S6 - Reliability Aware Pricing. It alters the price

based on the ratio of successfully and unsuccess-

fully hosted VMs at the cloud. A VM is classiﬁed

unsuccessfully hosted if it is terminated because

Cost-efﬁcient Datacentre Consolidation for Cloud Federations

217

of a physical machine failure, and not because of

a user’s request.

rel

= m

|V |

, (7)

where V

is the set of VMs which failed due to a

hardware issue at the provider side.

S7 - Proﬁt Margin Focused Pricing. It tries to price

resources so the proﬁt margin index (i) of the

cloud provider stays in a predeﬁned range (i.e.

min

< i < I

max

margin

) = m

(8)

margin

) =

margin

x−1

) ·











0.9, if i(t

) < I

min

1.1, if i(t

) > I

max

1, otherwise

(9)

where the function i(t) determines the current (or

at a given time, represented by t) proﬁt margin

index of a provider. This technique tries to adopt

the prices to make sure the provider is proﬁtable

even in competitive environments.

5.3 Results

As mentioned before, we investigated how policies

considering pricing information can affect consolida-

tion processes. We used 6 different trace ﬁles from

real world distributed systems to simulate load on

the cloud datacentres we aim to consolidate. We

also designed 7 different strategies to perform cost-

aware consolidation. In overall, the consolidation al-

gorithms succeeded to balance the load over the sy-

stem, and in most cases energy and money can be sa-

ved by applying them. In the following we highlight

the most interesting results.

We have performed numerous experiments by

executing the above listed strategies for all previously

mentioned trace ﬁles. Concerning experiments run

on the Grid5000 load, Figure 4 and Figure 5 depict

the tradeoff of energy gains and runtime expansions

for the given strategies (”S2 + S6” means we applied

both strategies, ”S5 + selling” means we applied the

S5 strategy and sold the shut down PMs to gain mo-

ney). By migrating certain VMs to other physical ma-

chines to balance the load, we managed to reduce the

power consumption, however the migration processes

took some time which appears in the overall runtime.

From the results we can see that the S6 strategy is the

Figure 4: Energy consumption of experiments with the

Grid5000 trace.

Figure 5: Runtime of experiments with the Grid5000 trace.

Figure 6: Energy consumption of experiments with

Grid5000 for different cloud provider pricing with the S3

strategy.

most efﬁcient for reducing power consumption, and

still it is the fastest solution.

Figure 6 and Figure 7 depicts the results of our S3

strategy that enables to load and apply different pro-

vider pricing schemes. From these results we can see

that the highest energy gains could be achieved with

the Amazon pricing scheme for this load condition,

while the worst result came from applying the Oracle

pricing.

We also experienced that the load types represen-

ted by the traces highly affect the results. Figure

8 presents measurements performed under different

CLOSER 2018 - 8th International Conference on Cloud Computing and Services Science

218

Figure 7: Runtime of experiments with Grid5000 for diffe-

rent cloud provider pricing with the S3 strategy.

Figure 8: Load and energy balance for different load condi-

tions with the S2 + S6 strategy.

load conditions with the combined S2 + S6 strategy.

The depicted balance represents the possible gains of

using consolidation in terms of cost (i.e. money) and

energy.

6 CONCLUSIONS

In this paper we addressed the problem of datacen-

tre consolidation. Though several approaches have

been developed to improve the utilization efﬁciency

of datacentres, only a few attention has been tur-

ned to combining pricing methods with consolidation

techniques.

We presented an extension of the DISSECT-CF

simulator to foster the development of cost efﬁcient

datacentre consolidation solutions. We also showed

how to apply real world traces to simulate cloud load,

and proposed 7 different cost-based strategies to ad-

dress the problem. Our results have approved that

cost-aware datacentre consolidation is a valid appro-

ach and can result in signiﬁcant cost and energy gains.

ACKNOWLEDGEMENTS

The research leading to these results was supported by

the UNKP-17-4 New National Excellence Program

of the Ministry of Human Capacities of Hungary,

and by the Hungarian Government and the European

Regional Development Fund under the grant num-

ber GINOP-2.3.2-15-2016-00037 (”Internet of Living

Things”).

REFERENCES

Amazon pricing website. Online:https://aws. amazon.com/

ec2/pricing/on-demand/.

DISSECT-CF website. Online: https://github.com/

kecskemeti/dissect-cf. Accessed at January, 2017.

Grid Workloads Archive website. Online: http://gwa.ewi.

tudelft.nl/. Accessed at September, 2017.

IBM Bluemix pricing sheet. Online: https://www.ibm.com/

cloud-computing/bluemix. Accessed at January,

2017.

MS Azure price calculator. Online: https://azure.

microsoft.com/ en-gb/pricing/calculator/. Accessed at

January, 2017.

Oracle pricing website. Online: https://cloud.oracle.com/

en US/ opc/ compute/compute/pricing. Accessed at

January, 2017.

Orcale Metered Services pricing calculator. Online:

https://shop.oracle.com/cloudstore/index.html?

product=compute. Accessed at January, 2017.

Parallel Workloads Archive website. Online: http://

www.cs.huji.ac.il/labs/ parallel/workload/. Accessed

at September, 2017.

Abdullah, M., Lu, K., Wieder, P., and Yahyapour, R. (2017).

A heuristic-based approach for dynamic vms conso-

lidation in cloud data centers. Arabian Journal for

Science and Engineering, 42(8):3535–3549.

Ahmad, R. W., Gani, A., Hamid, S. H. A., Shiraz, M., Yous-

afzai, A., and Xia, F. (2015). A survey on virtual ma-

chine migration and server consolidation frameworks

for cloud data centers. Journal of Network and Com-

puter Applications, 52(Supplement C):11 – 25.

Filho, M. C. S., Monteiro, C. C., Inacio, P. R., and Freire,

M. M. (2018). Approaches for optimizing virtual ma-

chine placement and migration in cloud environments:

A survey. Journal of Parallel and Distributed Compu-

ting, 111(Supplement C):222 – 250.

Kecskemeti, G. (2015). DISSECT-CF: a simulator to foster

energy-aware scheduling in infrastructure clouds. Si-

mulation Modelling Practice and Theory, 58P2:188–

218.

Kertesz, A., Dombi, J. D., and Benyi, A. (2016). A pliant-

based virtual machine scheduling solution to improve

the energy efﬁciency of iaas clouds. Journal of Grid

Computing, 14(1):41–53.

Cost-efﬁcient Datacentre Consolidation for Cloud Federations

219

Maio, V. D., Kecskemeti, G., and Prodan, R. (2016). An

improved model for live migration in data centre si-

mulators. In 2016 IEEE/ACM 9th International Con-

ference on Utility and Cloud Computing (UCC), pages

108–117.

Markus, A., Kertesz, A., and Kecskemeti, G. (2017). Cost-

aware iot extension of dissect-cf. Future Internet,

9(3).

CLOSER 2018 - 8th International Conference on Cloud Computing and Services Science

220