Future Generation Computer Systems ( ) –

Contents lists available at ScienceDirect

Future Generation Computer Systems

journal homepage: www.elsevier.com/locate/fgcs

A coral-reefs and Game Theory-based approach for optimizing elastic

cloud resource allocation
Massimo Ficco a,∗ , Christian Esposito b , Francesco Palmieri b , Aniello Castiglione b
a
Department of Industrial and Information Engineering, Second University of Naples, Via Roma 29, I-81031 Aversa (CE), Italy
b
Department of Computer Science, University of Salerno, Via Giovanni Paolo II, 132 I-84084 Fisciano (SA), Italy

highlights
• Bio-inspired coral-reefs optimization paradigm to model cloud elasticity.
• Game theory-based approach to identify the best cloud resource reallocation schema Fuzzy linguistic SLA formalization.

article info abstract

Article history: Elasticity is a key feature in cloud computing, which distinguishes this paradigm from other ones, such as
Received 15 January 2016 cluster and grid computing. On the other hand, dynamic resource reallocation is one of the most important
Received in revised form and complex issues in cloud scenarios, which can be expressed as a multi-objective optimization problem
16 May 2016
with the opposing objectives of maximizing demand satisfaction and minimizing costs and resource
Accepted 22 May 2016
Available online xxxx
consumptions. In this paper, we propose a meta-heuristic approach for cloud resource allocation based on
the bio-inspired coral-reefs optimization paradigm to model cloud elasticity in a cloud-data center, and
Keywords:
on the classic Game Theory to optimize the resource reallocation schema with respect to cloud provider’s
Cloud computing optimization objectives, as well as customer requirements, expressed through Service Level Agreements
Elasticity formalized by using a fuzzy linguistic method.
Live-migration © 2016 Elsevier B.V. All rights reserved.
Coral-reefs optimization
Game Theory
Fuzzy linguistic SLA

1. Introduction
In recent years, cloud computing has attracted attention from
industry, government and academic worlds. An increasing amount
of applications make extensive usage of cloud resources, accord-
ing to an on-demand, self-service, and pay-by-use business model,
with the progressive adoption of cloud-based services by many
sectors of modern society. One of the main reasons of this suc-
cess is the possibility of acquiring virtual resources in a dynamic
and elastic way. In particular, the emerging virtualization tech-
nologies allow multiple virtual machines (VMs) to run concur-
rently on a single physical host, called host machine (HM). Each
VM, in turn, hosts its operating system, middleware and applica-
tions, by using a partition of the underlying hardware resources
∗ are used in the implementation of cloud elasticity solutions: repli-
Corresponding author.
E-mail addresses: massimo.ficco@unina2.it (M. Ficco),
christian.esposito@dia.unisa.it (C. Esposito), fpalmieri@unisa.it (F. Palmieri),
castiglione@ieee.org (A. Castiglione).
http://dx.doi.org/10.1016/j.future.2016.05.025
0167-739X/© 2016 Elsevier B.V. All rights reserved.
capacity (CPU power, memory, store capability and network band-
width) [1]. Cloud elasticity is the key feature for implementing
server consolidation strategies. It allows for on-demand migration
and dynamic reallocation of VMs [2]. NIST defines elasticity as the
ability for customers to quickly purchase on-demand and auto-
matically release as many resources as needed, ideally giving to
the user the feeling that the cloud resource capabilities are unlim-
ited [3]. Specific elasticity control mechanisms are implemented to
decide when and how to scale-up or scale-down virtual resources,
in accordance with the provider’s own optimization objectives, as
well as with user-defined settings and requirements formalized
within the context of a specific Service Level Agreement (SLA) be-
tween the cloud provider and the customer. In order to do this,
resource usage information, such as CPU load, free memory and
network traffic volumes, for all the available HMs, has to be contin-
uously collected and analyzed on-line. Three different techniques
cation, migration and resizing [4], resulting in different strategies
and approaches for handling resource allocation within a cloud in-
frastructure.
2 M. Ficco et al. / Future Generation Computer Systems (
The dynamic resource allocation problem describes the deci-
sion of how many HMs are required overall to cope with the cur-
rent demand, and how VMs are (re-)allocated to each HM in the
individual time intervals. The optimization of resource allocation
consists in minimizing the number of HMs necessary to host all the
VMs associated to users’ demands. This is a typical multi-objective
optimization task, where the most important goals are the mini-
mization of the needed hardware resources and the satisfaction of
SLAs contracted with the customers. This problem, widely simpli-
fied, is closely related to a commonly known NP-hard combinato-
rial optimization problem, the ‘‘bin packing’’ problem [5], where
the items to be packed can be viewed as the VMs and the bins as
the HM nodes with their multi-dimensional capacity represented
by the available computing power, free memory, storage capabil-
ity and network bandwidth. Consequently, our elastic cloud re-
source allocation problem, that is inherently more complex, will be
NP-hard, and thus, a reasonable heuristic solution, achieving near-
optimal results is strongly desirable.
In such a complex multi-objective scenario, traditional opti-
mization methods are not able to efficiently provide good results,
due to their inherent difficulties in exploring huge solution spaces.
Hence, novel approaches based on bio-inspired meta-heuristic
schemes, seem to be the most promising options to achieve a bet-
ter trade-off between the complexity of the search process and the
optimization of the solutions found. Such schemes rely on models
and strategies borrowed from the observation of behaviors, and
evolution mechanisms available in nature, where only the fittest
individuals are able to survive in organizations characterized by
a high degree of competition. In this direction, to model the elas-
tic resource reallocation dynamics of a typical cloud environment,
we adopted a Coral-Reefs Optimization (CRO) approach, which ar-
tificially simulates the reefs evolution processes, including corals’
reproduction and competition for the space in the reefs [6], which,
by analogy, model the continuous demands for resource re-sizing,
migration and replication, characterizing the operational life-cycle
of VMs hosted within a cloud data center. This approach presents
extremely interesting and promising features in achieving conver-
gence towards global optima.
On the other hand, cloud customers can be seen as non-
cooperative entities, whose fundamental interest is not the op-
timization of the overall system performance, but rather the
maximization of their individual benefit. Therefore, a natural hint
for carefully driving the VM allocation process comes from classic
game theory. Accordingly, we used a game theory-based approach
to identify the best VM reallocation schema with respect to the crit-
ical requirements specified by the cloud customers in their SLAs,
formalized by using fuzzy linguistic label sets, and typically char-
acterized by different and even conflicting performance objectives
and optimization criteria. This approach is able to drive the com-
plex interactions between the customers and the cloud provider
towards a social optimum (i.e., an equilibrium point between their
own demands and goals), by optimizing a global objective function
that takes into account the individual interests of all the involved
entities.
The rest of the paper is organized as follows. In Section 2,
the strategies and methods currently adopted to implement cloud
elasticity capabilities are presented. Section 3 defines the problem
of interest. The proposed approach for elastic resource reallocation
is presented in Section 4. The related experimental performance
evaluation results are shown in Section 5. Finally, Section 6
presents some concluding considerations and remarks.
2. Background and related work
Elasticity represents a dynamic property of the cloud paradigm,
which allows the system to scale on-demand within an operational
) –
system context [7]. It enables applications to evolve, by follow-
ing the users’ demand, without needing traditional ‘‘fork-lift’’ up-
grades, and hence, introduces the degree of adaptiveness that is
fundamental for modern big data-centric environments.
Generally, Infrastructure as a Service (IaaS) architectures pro-
vide an elasticity controller, which is responsible for monitoring
information like CPU load, memory and network traffic, and for
making decisions on whether or not the virtual resources must
be scaled or migrated, in accordance with user-defined rules and
settings [8]. From a user perspective, elasticity allows perfect
matching between instantaneous customer’s needs and resources
available to him, by avoiding the unnecessary reservation of re-
sources that are not needed for most of time, with obvious im-
pacts on the overall system scalability and performance. From the
provider perspective, elasticity ensures better use of computing
resources, by providing economies of scale, and allowing a much
larger number of users to be served simultaneously [4]. More-
over, it can be used to increase the local resources capacity, by si-
multaneously reducing operational costs and energy consumption
[9–11].
Several strategies have been employed in order to support
the implementation of elasticity capabilities, including resizing,
replication, and migration:
• Resizing (vertical scaling): according to such strategy, CPU,
memory and storage resources can be re-sized on a running
virtual instance, that can be a virtual machine or a container.
Some of the most significant implementation of resizing
mechanisms are PRESS [12], ElasticVM [13], and Kingfisher [10].
• Replication (horizontal scaling): consists of adding/removing
instances from users’ virtual environments. It is currently
the most widely used technique for providing elasticity in
cloud environments [14–16]. Several cloud platforms, such as
Amazon and AzureWatch, offer auto-scaling and load-balancing
capabilities in order to split the load between the various
instances. Specific thresholds can be set by customers to add
or remove instances depending upon the actual usage.
• Live migration: consists in transferring the entire VM runtime
status (CPU and memory pages) from the actual host machine
to a new destination host, without preempting its execution
[17,18]. It is performed in two steps, in which at first, a pre-
copying of the VM status on the target host is performed (while
the VM is still running); then, the execution of the source
machine is stopped, a re-copy of the modified pages on the
destination host is performed, and finally, the migrated VM is
resumed.
A cloud provider’s resource management facility must
orchestrate the VM migration in order to simultaneously
minimize resource usage and maximize SLA adherence [19].
Specifically, it must provide:
– Cold-spots migration support: that is related to over-provi-
sioned resources with low utilization (server sprawl). In
order to reduce severe financial losses for cloud providers,
live migration can be used to consolidate VMs (optimize
VMs placement) with the aim of minimizing the number of
powered-on hosts and saving energy [20]. When the amount
of resource usage of a host drops below a given threshold,
VMs are migrated to another host providing enough available
resources capacity (the freed-up HMs are switched-off).
– Load balancing migration support: VM migration can be used
to dynamically move VMs among the available hosts with the
aim of rebalancing the utilization of resources over them. A
specific decision-making process is enabled when the load
imbalance within the cloud (among heavily loaded hosts and
lightly loaded ones) exceeds a given threshold [21,22]. Its
objective could achieve equal residual resource capabilities
across all the available HMs in order to optimize local
resource allocations in presence of increasing demands.
 M. Ficco et al. / Future Generation Computer Systems (
– Hot-spots mitigation support: if the HM does not own
enough additional locally available resources to meet the SLA
requirements of one of its VMs, the involved VM is migrated
to another host whenever possible (hot-spot). Specifically,
application-level monitoring of VMs must be adopted to
detect when the expected performance of the VMs drops
below a given threshold. In this case, in order to mitigate the
hot-spots, additional resources are either added locally, or if
the total resource usage of the host running the VM exceeds
a predefined threshold, a decision-making process is enabled
to migrate the VM to another HM [23,24].
From a cloud provider’s perspective, the migration process con-
sists in alleviating hot-spots and improving load-balancing in
order to meet SLAs of customers, whereas reducing cold-spots is
necessary to reduce power consumption (to minimize the number
of physical machines necessary to host all the VMs). The migra-
tion process addresses each of the previous objectives, but some
constraints should be considered: the overhead of the migration
process, the impact on applications during migration, the degree
of improvement in the specific resource utilization performance
goals [19,25,26].
3. Definition of the resource allocation optimization problem
In order to describe and evaluate the proposed approach,
we defined a simplified model for resource management in the
cloud computing environment. We consider a scenario, in which
a cloud data-center is composed by several computing nodes,
characterized by a set of parameters, such as CPU, memory,
and storage capacity. Specifically, we denote H = (h1 , . . . , hn )
as the set of HMs in the data-center, and respectively, C =
(c1 , . . . , cn ) their available (free) resource capacity, where cj is
described in terms of CPU, RAM, network bandwidth, and storage,
with cj = (κc ,j , λc ,j , µc ,j , νc ,j ) respectively. Moreover, we denote
V = (v1 , . . . , vm ) as the set of VMs to be assigned to HM and
R = (r1 , . . . , rm ) the set of required virtual resources, where
ri = (κr ,i , λr ,i , µr ,i , νr ,i ) represents the amount of resources for
running the VM i (in terms of CPU, RAM, bandwidth, and storage).
At each step of the allocation process, a set of virtual machines
W could be involved in the allocation process. Such VMs can
be: (i) new VMs to be allocated in the cloud; (ii) already hosted
VMs that require additional resources, and (iii) already hosted
VMs that must be reallocated either to minimize the resource
waste, or because of a hardware or software failure. Each involved
VM requires an amount of resources for a time duration d. We
assume κr ,i , λr ,i , µr ,i , νr ,i and dr ,i are random variables, each one
characterized by a specific probability distribution function pdf (·).
Specifically, the optimal allocation of the virtual resources required
by each VM is a problem that can be addressed by using stochastic
approaches. To address the heterogeneity of applications deployed
in the cloud, several approaches (and cloud providers) consider
application tasks as classified in different categories [27]. For each
type of task, they evaluate the probability distribution function
from the empirical data related to the usage of resources ri and the
task duration di .
Regarding live-migration, the optimization problem for cloud
resource reallocation must address three important questions:
1. When to migrate a VM?
2. Which VMs must be migrated?
3. Which are the set of destination HMs for migration, replication,
and placement of new VMs?
1. Migration of a VM in a data-center can be triggered:
• periodically: for example, data centers may be heavily used
in the morning, whereas they may be underloaded during the
night;
) – 3
• due to the overloaded condition of a HM (hot-spot);
• due to the low utilization of a HM (cold-spot);
• due to the imbalance in resource utilization levels of
different HMs, leaded from VMs that change their resource
requirements dynamically;
• due to the addition/removal of VMs and HMs, which can
affect the availability of the resources and may require a
change in the overall VM placement/scheduling plan.
2. Which VM must be migrated?: Selecting one or more VMs for
migration is a crucial decision. The migration process makes
the involved VM unavailable for a certain amount of time, as
well as consumes resources like network and CPU on source and
destination HMs, thus, the performance of co-resident VMs are
temporarily affected due to increased resource requirements
during migration. Therefore, the aim of VM selection is to
minimize the migration effort. However, different approaches
can be adopted to select the candidate VM for migration.
• The VM whose resource requirements cannot be locally
fulfilled is selected for migration.
• It may not always be efficient to select the overloaded VM.
In particular, if such VM uses a large amount of memory, the
time required for migration will be high. Therefore, the VM
with smaller memory could be selected for migration, so that
the memory freed can then be allocated to the other hosted
VMs.
• The VM whose SLA requirements are less critical is selected.
3. Where to locate the VM?: During migration, the destination
HM should have enough resources available, so that, it can
support the incoming migrating VM. However, considering the
availability of resources at the destination could be not enough.
For example, it should be considered how the performance of
VMs that are already hosted on the destination HM will be
affected by the migration, as well as it is necessary trying to
achieve some consolidation by co-allocating VMs that have high
memory sharing potential.
In order to solve the described elastic resource reallocation
problem by using Game Theory, we simulate the cloud elasticity
in a cloud-data center through a Coral-Reefs bio-inspired meta-
heuristic scheme (CRO) [6]. The CRO scheme represents the reefs
structure by using a grid, where each element of the grid (x, y)
can host a single or multiple corals (representing components of
the Game Theory-based solution). Therefore, in order to re-map
our resource allocation problem within such a scheme, we assume
that each component hx,y of the grid represents a HM of the set
H. Thus, we can model the cloud resource allocation problem, by
using a hypercube A = {ax,y,z }, in which, for each HM hx,y , the
vector Ax,y,i = (ax,y,1 , . . . , ax,y,m ) represents the mapping of the
VMs to hx,y , where ax,y,i is equal to 1, if the VM vi ∈ V has been
assigned to HM hx,y , otherwise it is equal to 0. The same holds for
the capacity set C represented by using a cubic matrix C, in which
each component cx,y is a capacity vector of the HM hx,y equal to
[κc ,x,y , λc ,x,y , µc ,x,y , νc ,x,y ]. The elements of the matrix C have the
same size, if the HMs belong to a homogeneous cluster, otherwise,
they have different dimensions (if the HMs are heterogeneous).
Briefly, the resource allocation problem exhibits three objective
functions:
1. resource minimization: minimize the total amount of data-
center resources used to execute the VMs;
2. resource balancing: allocate VMs on top of each host, so that, the
usage rate of host’s resources is kept around a certain threshold;
3. SLA optimization: the allocation process has to be orchestrated,
so that, the VMs associated to more critical SLA requirements
are privileged.
4 M. Ficco et al. / Future Generation Computer Systems (
According to the above described schema, the allocation
problem can be formulated in the following way. The minimum
number k of physical hosts needed to satisfy the VMs resource
requirements must be determined:
n 
m
 and Terminology [34].
k= ax,y,i , for each kind of resources (κ, λ, µ, ν).
x,y=1 i=1
The resource capacity cx,y on each involved HM hx,y must satisfy
all the VMs assigned to it, so that, the sum of resource demands
of all the VMs allocated on each HM hx,y should not exceed cx,y .
Moreover, a small amount ϵ of free resources should be guaranteed
in each machine, in order to offer a certain degree of vertical
scaling to applications without requiring any migration. Thus, for
each host hx,y ∈ HM, the physical resource constraints must be
expressed as:
m
 far from being concluded, and the need of a standardized structure,
ax,y,i ∗ ri < (cx,y − ϵ), for 1 < x, y < n.
i =1
Finally, each VM vi is associated with a weight ωi , which
represents the value of the criticality associated to the SLA
requirements to be met. The allocation process has to optimize the
VMs that have the highest weight ωi .
4. The meta-heuristic approach
This section, first, introduces a fuzzy linguistic model for
formalizing a generic SLA [28], established between the cloud
customer and the cloud provider (Section 4.1). Then, the CRO
algorithm, based on the coral-reefs behavior, is adopted to simulate
the continuous requests of new cloud resources needed to trigger
the resize, replication and migration processes characterizing
cloud data center activities (Section 4.2). Last, in Section 4.3, we
describe how to use a game theory-based approach to identify
the best VM allocation scheme with respect to the requirements
formalized in the previous section.
4.1. SLA formalization by using linguistic label sets
When a customer gains access to cloud-based services accord-
ing to a pay-by-use model, starting from its specific requirements,
it fills out a SLA with the service provider, whose satisfaction
is continuously monitored by the SLA management system [29].
Such a set of requirements formally documents the service(s)
needed, performance expectations, responsibilities and limits be-
tween cloud service providers and their customers. At the moment,
there is no standard that univocally determines the structure, for-
mat, syntax and semantic of the SLA requirements; neither the list
of important criteria within a SLA has been established.
Currently, the only available standard supporting both a for-
mal representation of SLAs and a protocol for their automation
is WS-Agreement (WSAG) [30], which was born in the context
of GRID computing and has been widely adopted, in the con-
text of many Cloud-oriented FP7 projects (e.g., SPECS, Contrail,
mOSAIC, Optimis, Paasage), to represent SLAs in the Cloud en-
vironment. The interest in WSAG has led in the last years to
the introduction of several extensions to the devised SLA life-
cycle and to the development of techniques for its automation
[31,32]. Nevertheless, WSAG only prescribes the main elements
of the agreement and leaves many details to custom implementa-
tions. Therefore, disputes often arise on the structure of the agree-
ments, and on how the provided quality of service is monitored
against the agreed SLA requirements. The European Commission
has issued a set of guidelines for SLAs in cloud computing [33],
) –
developed by the Cloud Select Industry Group (CSIG), as a strate-
gic way to help EU businesses in using the Cloud. Another well
known initiative in the area of Cloud SLA standardization is being
carried out by ISO/IEC JTC 1/SC38 on 19 086 Information Technol-
ogy (Cloud Computing) Service Level Agreement (SLA) Framework
(1)
Following these initiatives, several projects have been recently
funded by the EU to further investigate on the formalization
and management of SLAs. In the last years, some efforts have
been devoted both from the academic and industrial organizations
towards the introduction and formalization of security-related
parameters in the SLAs [35]. Among these, it is worth mentioning
the SPECS project [36], which focuses on Security SLAs (i.e., SLAs
containing security guarantees) and aims at building a framework
for the development of secure cloud applications, according to
the SLA life-cycle (negotiation, enforcement and monitoring).
Despite these and other similar initiatives, the current debate is
(2) terminology and syntax for SLAs in clouds is still an open issue.
Since, there is no common agreement on which properties
to be considered in a SLA, we have decided to keep such
properties generic, without discriminating if they are related to
performance, dependability or security. Therefore, we assume a
generic formulation of SLA contracting as a set of criteria, each
representing one of the possible quality properties that a customer
is interested to monitor and obtain, such as a performance grade,
a reliability degree or security level. Correspondingly, we have a
vector of n available QoS properties for a SLA binding between a jth
customer, namely Cj , and a provider, for a given task, namely τCj :
QoS (τCj ) = {q1 , q2 , . . . , qn }. We assume that each customer, before
submitting a task, has provided to the cloud platform an agreed
SLA binding in the form of a set of values, under which, the quality
properties should not fall during the task execution. Specifically, if
we indicate with fi the monitored value assumed by the ith quality
property, at run time, we have that fi ≥ qi .
The generality of our SLA formalization does not involve the
number and type of the quality properties, but also the values
that they can assume. Specifically, there are some properties that
naturally may have numeric expressions, such as the performance
grade indicated in terms of the latency in completing some
particular actions, or the reliability degree (expressed in terms
of the number of nines of reliability) of the cloud platform
in completing the tasks. Such a quantitative expression of the
quality properties is naturally obtained by the monitoring activities
operated by the SLA management system, but it is difficult to be
handled by human operators, especially if they are not experts in
these contexts. Not all the operators may be aware of the meaning
of nines for reliability, or are able to express an acceptable value of
nines for their applications. Considering that SLAs are specified by
humans, it is necessary to have a more human-friendly qualitative
approach to the expression of these properties. For this reason, in
this work, we have adopted linguistic labels, such as adjectives
like LOW or HIGH, to express the value of a particular quality
properties required by the cloud customer. In addition, we use the
fuzzy linguistic modeling theory, where a fuzzy set is associated to
a given linguistic label [37], taken from a properly-defined set of
terms, such as the following one of five terms:
S = s0 : N (NONE ), s1 : L(LOW ), s2 : M (MEDIUM ),
s3 : H (HIGH ), s4 : P (PERFECT ). (3)
In any linguistic approach to information modeling, in addition
to the linguistic descriptors of the adopted terms, we need to define
the associated semantics, which is typically done by means of fuzzy
sets defined in the [0, 1] interval, which are usually described by
a membership function µÃ with a triangular, trapezoidal or other
shapes. If we consider a numeric value for a given quality property,
 M. Ficco et al. / Future Generation Computer Systems (
Fig. 1. The set of five terms in Eq. (3) and its semantics as fuzzy sets.
then the function µÃ specifies the membership in the vaguely
defined set A of such a number, with an output within the interval
[0, 1]. An example of a balanced linguistic label set of five terms,
as in Eq. (3), i.e., with terms that are uniformly and symmetrically
distributed in the [0, 1] interval, and the underlying fuzzy sets are
depicted in Fig. 1.
Typically, the tuning of the parameters characterizing the
membership functions for the linguistic terms within a fuzzy
variable can be made by experts and must be tailored to the specific
characteristics of the measure it represents. Since, we assume a
balanced linguistic term sets, we adopt a partition of the interval
[0, 1] in even parts among the fuzzy sets of the adopted linguistic
terms.
Since, we have formulated the set of requirements in a SLA
as fuzzy linguistic labels, we have the problem of comparing
crisp numeric values returned by the SLA management system
with these requirements. Such a problem can be resolved by a
proper process of converting a fuzzy linguistic label into a crisp
number, called defuzzification [38], namely φ(l), where l is a fuzzy
linguistic label in the label set Λ. Defuzzification of a linguistic
label is performed by using the centroid method, by considering
the membership function of the linguistic term as the input to
the operation. Such a method determines the geometric center
of the membership function over the x-axis. Typically, a required
property can assume a single linguistic label, or even a set of
them, each with a given membership degree. In the second case,
defuzzification is obtained by computing the weighted median
point among the centroids obtained for each linguistic label, where
the weights are the membership degrees for each label.
Both the two crisp numbers must be within the [0, 1] interval,
if the values of fi are not within such an interval, they can be
normalized as follows. Let us assume that, the values admissible for
(i) (i)
a given monitored property fi lie between Blower and Bupper , we can
convert it into an attribute with values between 0 and 1, according
to the following transformation:
 (i)
fi − Blower
for a positive property

 B(i) − B(i)


upper lower
f¯i =
 B(upper

i)
− fi

 (i) (i)
for a negative property
Bupper − Blower
where, a positive property means that high values for such a
characteristic indicates a better services (e.g., higher number
of nines indicates a higher reliability degree), while a negative
property has an opposite interpretation (e.g., a higher values in
milliseconds as completion time indicates worse performances).
After, we have brought all the involved measures, i.e., fi and qi
for all i = 0, . . . , n, in a form that is comparable, we are able to
compute a satisfaction degree, which is a measure of the satisfac-
tion of the SLA properties by the current task settings in the cloud
platform. If such an index is negative, then the quality of the pro-
vided cloud service does not respond to the customer requirements
) – 5
and some changes are needed. Specifically, such an index can be
formulated as follows:
n

σ (τCj ) = φ(wi ) · (φ(qi ) − f¯i ) (5)
i=0
where, ωi is a weight given to the ith property. Such a weight is a
measure of the importance of the satisfaction of a given propriety
for the consumer, and it is formulated in terms of a fuzzy linguistic
label provided by the consumer.
4.2. Simulating cloud elasticity by Coral Reefs-based ecosystem
approach
Coral-Reefs Optimization (CRO) is a meta-heuristic approach
classified as a bio-inspired algorithm, similarly to ant colony
optimization [39] and artificial bee colony (ABC) [40] schemes. It
artificially simulates the behavior of a coral reefs ecosystem to
tackle optimization problems. The algorithm mimics the process of
the sexual and asexual coral reproduction, as well as the process of
coral reefs formation, where a fight for space occurs. The scope is to
find a place in the reefs where the coral can be settled, considering
that the space in the reefs is a limited resource compared with the
high reproduction rate of the corals. Therefore, corals must fight
to obtain a place in the reefs. In particular, in each step of the CRO
algorithm, a set of coral larvae is generated, and each larva must
fight to obtain a place in the reefs. The result of this fight is that
some corals die, because they cannot defend the place where they
are located, as well as not all larvae find any place in the reefs where
they can be settled [41]. It depends on how strong the larva is, i.e.,
how good the solution to the optimization problem is.
Specifically, CRO algorithm simulates the life of the corals in
a reef Λ modeled as a grid of NxM elements, during different
generations. We assume that each grid element (x, y) of Λ, is
able to allocate a micro-colony of corals Φ(x,y) , representing a
solution to the optimization problem, which is encoded as a string
of numbers in a given alphabet Γ . According to the considered
scenario, the reefs represent the cloud infrastructure, where each
(4)
grid element (x, y) is a HM, and the micro-colony Φ(x,y) represents
the set of hosted VMs A = {ax,y,i }.
The initialization of the CRO algorithm consists in randomly
assigning some elements in the reefs to be occupied by corals and
some other elements to be empty, where new corals can freely
settle and grow in the future. The rate between free/occupied
elements determines the initial population density ρ0 (with 0 <
ρ0 < 1). By analogy, during the initialization process, only some
HMs host one or more VMs, whereas the other are free.
The CRO is based on the fact that the reefs will progress, as long
as stronger corals survive, whereas less healthy corals perish. This
means that, the cloud resource utilization changes over time. The
VMs are replicated and migrated, as well as new VMs are added,
and others are removed (turned-off).
6 M. Ficco et al. / Future Generation Computer Systems (
After the initialization process, the reefs formation in terms
of the corals reproduction is artificially simulated by the CRO
algorithm until an objective criteria is achieved. Several operators
are used to imitate corals reproduction, which model the following
five different reproduction procedures and phenomena:
1. Brooding (internal sexual reproduction): Selected couples will
form a coral larva by sexual crossover. It consists in the
generation of new larvae by copying and mutation of existing
corals. This reproduction procedure generates a coral larva,
which is a copy of its parent and it is randomly mutated. The
produced larva is then released to be located in the reefs.
2. Budding (asexual reproduction): The budding model consists of
the formation of a coral larva by means of a random mutation of
the brooding-reproductive coral (self-fertilization considering
hermaphrodite corals). It consists in sorting all the corals
according to their health. From this sorted list, a fraction of the
best corals are duplicated and tried to be located in the reefs.
3. Broadcast Spawning (external sexual reproduction): New larvae
could come from outside the reefs, and try to settle in the reefs.
4. Catastrophic events in the reefs: at the end of each reproduction
step p, a small number of corals in the reefs can be depredated
by polyps. The depredation operator is applied with a very small
probability Pp at each step p.
5. Natural death: Corals may die during the reefs formation phase,
thus liberating space in the reefs for next coral generation.
Note that, the corals that have been reproduced by using a method
in the current generation cannot be selected to reproduce by using
other methods.
After the sexual and asexual reproduction, the larvae produced
have to be placed to grow in the reefs. Therefore, in the considered
cloud context, the larvae represent either a request of additional
cloud resources from a hosted VM, or a replication request of VM
instance, or a request of resources from a new VM. Therefore,
the previous phenomena simulate, respectively, the following
behaviors:
1. brooding: the co-residency of a VM i with other tenants can
require additional resources, activating a decision process for
choosing whether re-size VM i or migrate a hosted VM from the
host machine. In the last case, it is necessary to decide which
VM migrates, and on which HM it has to be deployed;
2. budding: reducing or increasing the load can trigger an auto-
scaling process, which can involve either the turn-off of a
VM instance, or the activation of a replicated VM instance,
respectively; in the second case, it is only necessary to decide
on which HM it has to be deployed;
3. broadcast: the new larvae represent requests to host new VMs,
which have to be allocated on the more appropriate HMs;
4. catastrophic events: a software or hardware failure is generated
on a HM, i.e., one or all the VMs running on the host machine
fail; in this case, all the involved VMs must be replicated on
other hosts; eventually the failed HM is removed from the
cloud;
5. natural death: a VM has completed its work or is turned off.
Therefore, the reefs are defined by VMs, where the set of VMs
deployed on the cloud HMs represent a possible solution I for
the modeled problem. Each coral i is labeled with an associated
health evaluation function that measures the fitness of such coral
within each micro-colony f (Φ(x,y)i ) : I ⇒ R, which represents
the problem’s objective function. Such a function is equal to the
satisfaction index σ (τCj ), that we have defined in Eq. (5). It is
used for comparing corals during the selection or reproduction
processes. This aspect means that each VM requires a certain
amount of cloud resources (e.g., CPU, memory and storage) to
satisfy the SLA contracted with the cloud customer. f (Φ(x,y)i )
) –
measures if the VM i can be hosted by a physical machine that
satisfies such requirements. The described optimization problem
has been faded by a Game Theory approach described in the next
section.
According to the coral model, each larva located either in a
free space, or in an occupied one, has to fight against the existing
corals, and only the strongest individuals will survive. A VM can be
migrated from its hosting HM, either to acquire more resources, or
to make available resources to other VMs which have more critical
SLA requirements, as well as to minimize the number of active
HMs. Moreover, in the considered cloud context, if the cloud is
public, we can assume that the resources are unlimited and all the
requests can be always satisfied (larvae cannot be deprecated). If
the objective is the optimization of the resource utilization in a
private cloud, we can assume that the resources are limited and
some requests could not be successfully allocated into a place
(whose capacity eventually grows), during the current allocation
step (larvae are deprecated). The deprecated larvae will be queued
for the next allocation step. Finally, the allocation process must
minimize the number of migrations per step.
4.3. Distributed strategic resource allocation with Game Theory
The abstraction provided by the CRO algorithm perfectly
disciplines and models the competing behavior of the VMs within
a cloud infrastructure, to obtain the resources needed to satisfy
both the consumers’ SLA requirements and the cloud provider’s
benefit. Therefore, according to the problem formulated in the
previous section, the issue to be treated is determining: (1) which
VM has to be migrated from a certain HM, and towards which HM
such VM should be migrated or replicated, and (2) where allocate
new VMs. These objectives have been performed with the intent
of optimizing the satisfaction degree for all the consumers and
minimizing the number of active HMs.
As a consequence, our game-theoretic scheme is applied in
the two mentioned situations. In the following, we resolve the
migration case, since both the replication and allocation of new
VMs are a naive relaxation of the first one (the problem is to
determine only the destination of the VM and not which VM
must be replicated or created, while in the migration case both
elements must be determined). Such a problem has been resolved
in a strategic manner by means of the Game Theory [42], which
has been already identified as a really promising framework for
managing resource reallocation within clouds [43]. The overall
goal of the approach, called in the literature as the social welfare,
is to maximize the satisfaction of customers’ SLA requirements
by the capabilities provided of the selected HMs, and minimizing
the number of globally involved HMs in the cloud infrastructure.
Specifically, we assume that the set of VMs in a HM are the players
of a one-shot game, which has a set of strategies among which
they can make an action in the game. Such strategies for the
ith player are binary
n values being elements of the set indicated
with S pi = {s∅ , j=0 si }, where s∅ indicates the strategy of not
migrating, while si is the strategy of migrating towards the ith HM
of the cloud that hosts n usable HMs. When a player selects a given
strategy, namely s̄ within its strategy set, it has to pay a certain cost,
namely Φ p (s̄). Dually, the player receives the payoff φ p (s̄), which
is the gain achievable by a player when following a given strategy,
considering the costs implied by the chosen strategy. In our case,
the game is defined non-cooperative since players are selfish, i.e.,
there is no direct communication among the players, and each one
only cares to minimize its own cost, or dually to maximize their
payoff, without considering the state of the other players (with the
eventuality of damaging them, even if it is not intentional). In our
game, we have to define a way so that, the strategy that maximizes
the payoff of each player is the one achieving the optimal social
 M. Ficco et al. / Future Generation Computer Systems (
welfare of our optimization problem. Specifically, the payoff of a
given strategy can be formalized as follows:
n
 is high and the weight of the VM is high, or a value ϵ when the
φ p (s̄) = s∅ · σ (s∅ ) + s∅ sj · [σ (sj ) − δ+HM − Φ p (sj )],
j =0
where all these contribution to the payoff function have values
within the interval [0, 1]. σ (·) indicates the satisfaction degree
expressed by Eq. (5), considering the strategy in input, i.e.,
migration or no migration, δ+HM is the cost to be paid if a new
HM is activated as the result of the migration, and Φ p (si ) expresses
the cost for the migration from the current HM to the target one.
The cost value, i.e., δ+HM , has to be properly selected in order
to discourage the activation of a new HM and can be modeled
as a constant, or can be assumed as a function of the actual
number of active HMs, in order to make it higher as the number
of HMs increases. This is needed to keep the number of active HMs
minimal. In addition, the cost Φ p (si ) is composed of three different
contributions to be properly considered:
• The VM that has to migrate experiences a temporary unavail-
ability, and this causes an inefficiency and a delay in the comple-
tion of the allocated task. Such an effect is modeled by a proper
cost function, which depends on the VM to be migrated and the
criticality of the task running on the VM. In fact, the bigger is
the VM to be migrated, the higher is the inefficiency to be expe-
rienced. Analogously, more critical is the task allocated to the
VM and more serious would result the inconvenience caused
by the migration. Such a cost can be formulated by assuming a
constant value for the inefficiency, weighted by the sum of the
importance and the weight of the VM to be migrated:
Φ (sj )contr1 = (ω1 + ω2 ) · cost migr ,
p
where ω1 and ω2 are real numbers in the [0, 1] interval given
by the defuzzification of a fuzzy linguistic label respectively
indicating the importance and the weight of the player. The
first one is inherited by the player from the task allocated to its
VM, while the second one is assigned at the allocation time of
the VM and is determined according to threshold-based rules:
if the request of resource for a VM is within a given interval
i
[rdown , rup
i
], then the VM has assigned li to ω2 .
• There is a cost to pay in order to physically move the VM, and
allocate it to the destined HM (e.g., in terms of power energy
and network bandwidth). Such a cost is a proper function of the
weight of the VM to be migrated:
Φ p (sj )contr2 = ω2 · cost move .
• When a VM is migrated to a new HM, it causes an inefficiency
to the other VMs already hosted by the destination HM. Such
a situation is modeled by a proper cost function, depending on
the number and type of the hosted VMs. Moreover, each HM
is characterized by a capacity C , which has to be split among
the hosted VMs. We can have a situation where the residual
capacity is not sufficient for the migrated VM and/or the
migration may cause a considerable reduction in the capacity
available for an eventual resizing of one of the hosted VM. Such
a case can be considered in computing the cost of migration,
by having a function that augments its value as the residual
capacity decreases (due to the migration):
Φ p (sj )contr3 = (ω2 − κ) · cost alloc + κ · ϵ(κ),
where κ indicates the residual capacity of the jth HM at the time
of decision, and cost alloc is a constant. Such a formulation implies
the highest value, when the weight of the migrating VM is high
and the residual capacity is low. On the contrary, the cost is
negative, i.e., we have an incentive, when the weight of the VM
) – 7
is low and the residual capacity is high (this is done to promote
the selection of a scarcely used active HM). In the other two
cases, the cost assumes a null value, when the residual capacity
(6)
residual capacity is low and the weight of the VM is low. ϵ can
be properly selected in order to discourage a migration towards
a HM with a limited residual capacity, despite the weight of
the migrating VM is low; for this reason ϵ is a function of the
κ parameter.
Definitely, the cost for the migration is formulated as follows:
Φ (sj ) = (ω1 + ω2 ) · cost migr + ω2 · cost move
p
+ (ω2 − κ) · cost alloc + κ · ϵ(κ). (10)
The scope of the game is achieved by identifying a point of
equilibrium among players. It is named Nash equilibrium in the
case of non-cooperative games, based on the concept of domain
strategy, i.e., given a certain strategy s̄ ∈ S, it is not profitable
for a player to select a different strategy than the one in the
current profile, assumed dominant, since a different strategy will
not change or even augment the needed costs, so a player has no
incentive to change strategy:
∃s ∈ S : ∀p ∈ P , ∀x ∈ Y , φ̄ p (sp , s−c ) ≥ φ̄ p (x, s−p )
(11)
→s is a Nash equilibrium.
Specifically, if we consider that the first decision to be taken is to
migrate or not, and then, to identify the most suitable HM towards
which the deployment must take place, we have that no migration
is profitable if the following condition holds:
σ (s∅ ) ≥ σ (s∗ ) − δ+HM − Φ p (s∗ ), (12)
(7) where s∗ is the strategy of the migration towards the most
convenient HM, which is the one that resolve such an optimization
problem:
s∗ = max si · [σ (si ) − δ+HM − Φ p (si )]. (13)
i=0,...,n
Eq. (12) can be further improved in order to find a condition for the
migration or not:
σ (s∗ ) − σ (s∅ ) ≥ δ+HM + Φ p (s∗ ), (14)
which means that is convenient to migrate, only if, the improve-
ment obtained in terms of improved satisfaction of the customer’s
SLA requirements is higher than the costs to be paid for this op-
eration. Moreover, such a solution, minimizes the number of HMs
jointly activated, thanks to the cost of the additional HM inserted
(8)
in Eq. (13).
5. Experimental results
According to the proposed Coral Reefs-based ecosystem ap-
proach, each grid element (x, y) of the reef (modeled in our context
as a vector A = {ax,y,i }) represents a HM hx,y . We assume that, the
space available (resource capacity cx,y ) within each element hx,y is
different, i.e., the HMs are heterogeneous. Moreover, the HM hx,y
can host a set of VM instances (micro-colony), such that Eq. (2) is
verified.
Thus, at each step of the simulation process, the CRO-algorithm
computes which VM instances require more resources (brooding),
activating a decision process for choosing whether to re-size the
(9)
VM (if there are still resources available on the HM), or migrating
one of the instances hosted on the involved HM. We assume an
increase or a reduction of resources, which is no greater than
30% compared to the resources already used by the involved
VM. In order to make the simulated scenario more realistic, we
assume that the increase of resources may also be negative. In
8 M. Ficco et al. / Future Generation Computer Systems ( ) –

Fig. 2. The behavior of the decision process.

Fig. 3. Number of migrations and replications requests to be satisfy, during each Fig. 4. Time need to reallocate the allocation requests computed at each step pi .
simulation step pi .

this case, the decision process implies a resize (reduction) of the

resources used by the involved VM. Moreover, assuming that
some kind of auto-scaling policy has been contracted in the SLA,
the algorithm computes which instances are subjected to a load
increase or reduction (budding), which triggers either the turn-off
of a replicated VM instance (scaling-down), or the activation of a
new VM instance (scaling-up). Finally, the algorithm decides which
VM instance has to be removed (for natural death or catastrophic
events), and how many added (broadcast). Then, by using Game
Theory, we identify which VMs of the involved HMs must be
migrated, and what are the destination HMs on which he migrated
or replicated instances have to be deployed, as well as what are the
destination HMs on which the new VMs can be deployed. During
each simulation step, the amount of free resources available on Fig. 5. Number of active HMs and VMs instances hosted in the cloud ecosystem.
each HM is updated according to the decision process.
In order to make the experimental results more clear, we Finally, Fig. 5 represents graphically how the number of the
assume that, at each step, only a new VM is added to the cloud, involved physical hosts (i.e., the number of active HMs that host
and no VMs are removed. Therefore, the number NM /R of VMs to at last a VM) and the number of VM instances hosted in the cloud
be reallocated is the sum of migration and replication requests ecosystem (such as VM (pi ) = VM (pi−1 ) + R + 1) increase during
NM + NR + 1. Moreover, at step 0, we assume that five VMs are the simulation process.
already hosted in the cloud, and they are deployed on some HMs,
as well as the allocation of each new VM i on the host hx,y requires
6. Conclusions
an amount of resources ri included in the range:

(cm ∗ 0.1) < ri < (cm ∗ 0.3),
where cm represents the average resource capacity ci,j of the HMs
available in the cloud ecosystem.
The results reported in Fig. 2 show the behavior of the decision
process. In particular, for each simulation step, the number of
migrations (M) and replications (R) are reported, together with the
number of the involved instances (VMs) and the physical host on
which they are hosted (HMs).
Fig. 3 shows, at each step pi , the total number of migrated
and replicated requests (NM /R ), generated by the Coral-Reef bio-
inspired algorithm, which must be satisfied.
In Fig. 4 is shown, for each simulation step pi , the simulation
time necessary for computing the allocation schema (i.e., to
reallocate in the cloud the VMs allocation requests computed by
the Coral-Reef algorithm).
(15) In this work, a new bio-inspired and game theoretic-based
meta-heuristic scheme for managing elastic resources reallocation
in cloud infrastructures has been presented.
Specifically, the proposed solution leverages an evolutionary
algorithm based on the observation of the reefs structure and coral
reproduction, showing a very interesting behavior to simulate the
continuous requests of new cloud resources, which trigger the
resize, replication and migration processes characterizing cloud
data-center activities. It also exploits the competition dynamics
among the strategic users (with their SLAs) and service providers
(with their revenue maximization objectives), to converge towards
socially optimal equilibrium solutions that are able to satisfy
the apparently contradicting interests of the involved parties.
In particular, a game theory-based approach is used to identify
the best VM reallocation schema with respect to the conflicting
performance objectives of the customers and the cloud provider,
 M. Ficco et al. / Future Generation Computer Systems (
by optimizing a global objective function that takes into account
the individual interests of all the involved entities.
The CRO-based model strongly characterizes the proposed
solution by simulating interesting elasticity properties of the cloud
paradigm. Due to its evolutionary nature, in which the collective
emergent behavior of the ‘‘reefs’’ exhibits a great deal of global
intelligence, capable of complex tasks, the resulting schema can
provide a good scalability in presence of a large number of virtual
machines and host-related resources available into the cloud.
We analyzed the performance of the proposed approach by
implementing a simple simulation-based proof of concept scenario
that considered both providers and users together with their
available resources, capacities and SLAs. The experimental results
validated our hypothesis that a combined bio-inspired and game
theoretic-based meta-heuristic approach, does not only achieve a
satisfactory solution in terms of adaptiveness and elasticity, but it
also can lead to significant performance improvements in terms
of convergence time, when the problem scales towards very large
clouds with plenty of machines and VMs to be reallocated.
References
[1] Weiwei Lin, Baoyun Peng, Chen Liang, Bo Liu, Novel resource allocation model
and algorithms for cloud computing, in: Proc. of the 4th Int. Conf. on Emerging
Intelligent Data and Web Technologies, 2013, pp. 77–82.
[2] M. Ficco, C. Esposito, Henry Chang, Kim-Kwang Raymond Choo, Live migration
in emerging cloud paradigms, IEEE Cloud Comput. (2016) 12–19.
[3] L. Badger, R. Patt-Corner, J. Voas, Draft cloud computing synopsis and
recommendations recommendations of the national institute of standards and
technology, NIST Spec. Publ., vol. 146, Available at:
http://csrc.nist.gov/publications/drafts/800-146/Draft-NIST-SP800-146.pdf.
[4] Gu. Galante, L.Carlos E. de Bona, A survey on cloud computing elasticity, in:
Proc. of the IEEE/ACM 5th Int. Conf. on Utility and Cloud Computing, 2012,
pp. 263–270.
[5] M. Garey, D. Johnson, Computers and Intractability: A Guide to the Theory of
NP-completeness, W.H. Freeman and Co, 1979.
[6] S. Salcedo-Sanz, A. Pastor-Sanchez, D. Gallo Marazuela, A. Portilla-Figueras,
A novel coral reefs optimization algorithm for multi-objective problems,
in: Proc. of the Intelligent Data Engineering and Automated Learning, in: LNCS,
vol. 8206, Springer Berlin Heidelberg, 2013, pp. 326–333.
[7] D. Agrawal, A. El Abbadi, S. Das, A.J. Elmore, Database scalability, elasticity, and
autonomy in the cloud, in: Proc. of the 16th Intl. Conf. on Database Systems for
Advanced Applications, Springer-Verlag, 2011, pp. 2–15.
[8] L.M. Vaquero, L. Rodero-Merino, R. Buyya, Dynamically scaling applications in
the cloud, SIGCOMM Comput. Commun. Rev. 41 (2011) 45–52.
[9] R.N. Calheiros, C. Vecchiola, D. Karunamoorthy, R. Buyya, The aneka platform
and qos-driven resource provisioning for elastic applications on hybrid clouds,
Future Gener. Comput. Syst. 28 (6) (2011) 861–870.
[10] Z. Shen, S. Subbiah, X. Gu, J. Wilkes, Cloudscale: elastic resource scaling
for multi-tenant cloud systems, in: Proc. of the 2nd Symposium on Cloud
Computing, (SOCC), ACM, 2011, pp. 1–14.
[11] S. Ricciardi, F. Palmieri, J. Torres-Vials, B. Di Martino, G. Santos-Boada, J.
Sol-Pareta, Green data center infrastructures in the cloud computing era,
in: Handbook of Green Information and Communication Systems, 2013,
pp. 267–293.
[12] Z. Gong, X. Gu, J. Wilkes, Press: Predictive elastic resource scaling for cloud
systems, in: Proc. of the 6th Int. Conf. on Network and Service Management,
(CNSM), IEEE, 2010, pp. 9–16.
[13] W. Dawoud, I. Takouna, C. Meinel, Elastic vm for cloud resources provisioning
optimization, in: Advances in Computing and Communications, Vol. 190,
Springer Berlin, Heidelberg, 2011, pp. 431–445.
[14] J.O. Fito, I.G. Presa, J.G. Fernandez, Sla-driven elastic cloud hosting provider,
in: Proc. of the 18th Euromicro Conf. on Parallel, Distributed and Network-
Based Processing, (PDP), IEEE, 2010, pp. 111–118.
[15] N. Roy, A. Dubey, A. Gokhale, Efficient autoscaling in the cloud using predictive
models for workload forecasting, in: Proc. of the 4th Int. Conf. on Cloud
Computing, (CLOUD), IEEE, 2011, pp. 500–507.
[16] N. Vasic, D. Novakovic, S. Miucin, D. Kostic, R. Bianchini, Dejavu: accelerating
resource allocation in virtualized environments, in: Proc. of the 17th Intl. Conf.
on Architectural Support for Programming Languages and Operating Systems,
(ASPLOS), ACM, 2012, pp. 423–436.
) – 9
[17] T. Knauth, C. Fetzer, Scaling non-elastic applications using virtual machines,
in: Proc. of the 4th Int. Conf. on Cloud Computing, CLOUD, 2011, pp. 468–475.
[18] K. Lazri, S. Laniepce, J. Ben-Othman, When dynamic VM migration falls under
the control of vm users, in: Proc. of the IEEE 5th Int. Conf. on Cloud Computing
Technology and Science, CloudCom, 2013, pp. 395–402.
[19] M. Mishra, A. Das, P. Kulkarni, A. Sahoo, Dynamic resource management using
virtual machine migrations, IEEE Commun. Mag. 50 (9) (2012) 34–40.
[20] S. Ricciardi, D. Careglio, G. Santos-Boada, J. Sol-Pareta, U. Fiore, F. Palmieri, Sav-
ing energy in data center infrastructures, in: Proceedings—1st International
Conference on Data Compression, Communication, and Processing, CCP 2011,
2011, pp. 265–270.
[21] E. Arzuaga, D.R. Kaeli, Quantifying load imbalance on virtualized enter-
prise servers, in: Proc. of the 1st Int. Conf. on Performance Engineering,
(WOSP/SIPEW), ACM, 2010, pp. 235–242.
[22] A. Gulati, A. Holler, M. Ji, G. Shanmuganathan, C. Waldspurger, X. Zhu,
Vmware distributed resource management: Design, implementation, and
lessons learned, 2012.
[23] X. Zhang, Z.-Y. Shae, S. Zheng, H. Jamjoom, Virtual machine migration in
an over-committed cloud, in: Proc. of the IEEE Network Operations and
Management Symposium, NOMS, 2012, pp. 196–203.
[24] N.M. Calcavecchia, O. Biran, E. Hadad, Y. Moatti, VM placement strategies for
cloud scenarios, in: Proc. of the IEEE 5th Int. Conf. on Cloud Computing, CLOUD,
2012, pp. 852–859.
[25] Andrei Sfrent, Florin Pop, Asymptotic scheduling for many task computing in
big data platforms, Inform. Sci. 319 (20) (2015) 71–91.
[26] Mihaela-Andreea Vasile, Florin Pop, Radu-Ioan Tutueanu, Valentin Cristea,
Joanna Kolodziej, Resource-aware hybrid scheduling algorithm in heteroge-
neous distributed computing, Future Gener. Comput. Syst. 51 (2015) 61–71.
[27] M. Ficco, B. Di Martino, R. Pietrantuono, S. Russo, Optimized task allocation
on private cloud for hybrid simulation of large-scale critical systems,
Future Gener. Comput. Syst. (2016) http://dx.doi.org/10.1016/j.future.2016.
01.022. available online at: http://www.sciencedirect.com/science/article/pii/
S0167739X16300061.
[28] C. Esposito, A. Castiglione, F. Palmieri, M. Ficco, Trust management for
distributed heterogeneous systems by using linguistic term sets and
hierarchies, aggregation operators and mechanism design, Future Gener.
Comput. Syst. (2015) available online at:
http://www.sciencedirect.com/science/article/pii/S0167739X15003878.
[29] V. Casola, A. Mazzeo, N. Mazzocca, M. Rak, A SLA evaluation methodology in
service oriented architectures, Adv. Inf. Secur. 23 (2006) 119–130.
[30] A. Andrieux, K. Czajkowski, A. Dan, K. Keahey, H. Ludwig, T. Nakata, J. Pruyne,
J. Rofrano, S. Tuecke, M. Xu, Web services agreement specification (WS-
Agreement), in: Global Grid Forum, The Global Grid Forum (GGF), 2004.
[31] R. Kubert, G. Katsaros, T. Wang, A RESTful implementation of the WS-
agreement specification, in: Proceedings of the Second International Work-
shop on RESTful Design, Ser. WS-REST’11, ACM, New York, NY, USA, 2011,
pp. 67–72.
[32] A. De Benedictis, M. Rak, M. Turtur, U. Villano, REST-based SLA management for
cloud applications, in: Proceedings of REST-based SLA Management for Cloud
Applications, WETICE 2015, Larnaka, Cyprus, pp. 93–98.
[33] European Commission. Cloud Service Level Agreement Standardisation
Guidelines, 2014, available online at
https://ec.europa.eu/digital-agenda/en/news/cloud-service-level-agreement-
standardisation-guidelines.
[34] International Organization for Standardization (ISO). ISO/IEC DIS 19086-1,
June, 2015, available online at http://www.iso.org/iso/home/store/catalogue_
tc/catalogue_detail.htm?csnumber=67545.
[35] V. Casola, A. De Benedictis, M. Rak, On the adoption of security SLAs in the
cloud, in: Accountability and Security in the Cloud, in: LNCS, vol. 8937, 2014,
pp. 45–62.
[36] M. Ficco, M. Rak, SLA-oriented security provisioning for cloud computing,
in: Cloud Computing and Services Science, in: LNCS, vol. 367, Springer-Verlag,
2013, pp. 230–244.
[37] D. Bacciu, M.G. Buscemi, L. Mkrtchyan, Adaptive fuzzy-valued service
selection, in: Proceedings of the 2010 ACM Symposium on Applied Computing,
SAC, March 2010, pp. 2467–2471.
[38] T.J. Ross, Fuzzy Logic with Engineering Applications, third ed., Wiley, 2010.
[39] M. Dorigo, V. Maziezzo, A. Colorni, The ant system: optimization by a colony
of cooperating ants, IEEE Trans. Syst. Man Cybern. 26 (1) (1996) 29–41.
[40] D. Karaboga, B. Basturk, On the performance of the artificial bee colony (ABC)
algorithm, Appl. Soft Comput. 8 (2008) 687–697.
[41] S. Salcedo-Sanz, J.E. Sanchez-Garcia, J.A. Portilla-Figueras, S. Jimenez-
Fernandez, A.M. Ahmadzadeh, A coral-reef optimization algorithm for the
optimal service distribution problem in mobile radio access networks, Trans.
Emerg. Telecommun. Technol. 25 (11) (2013) 1057–1069.
10 M. Ficco et al. / Future Generation Computer Systems (
[42] M.J. Osborne, A. Rubinstein, A Course in Game Theory, The MIT Press, 1994.
[43] F. Palmieri, L. Buonanno, S. Venticinque, R. Aversa, B. Di Martino, A distributed
scheduling framework based on selfish autonomous agents for federated cloud
environments, Future Gener. Comput. Syst. 29 (6) (2014) 1461–1472.
Massimo Ficco received the degree in computer engi-
neering from the University of Naples Federico II (IT) in
2000 and the Ph.D. degree in information engineering
from the University of Parthenope in 2010. He is an assis-
tant professor in the Department of ‘‘Ingegneria Industri-
ale e dell’Informazione’’ at the Second University of Naples
(SUN). From 2000 to 2010, he was a senior researcher at
the Italian University Consortium for Computer Science
(CINI). His current research interests include software en-
gineering architecture, cloud computing, security aspects
of critical infrastructure, and mobile computing.
Christian Esposito received the graduate-degree in com-
puter engineering in 2006, and the Ph.D. degree in 2009
from the University of Naples Federico II. Currently, he
is a postdoc researcher in the Institute of High Perfor-
mance Computing and Networking (ICAR) at the National
Research Council (CNR). His main interests include mobile
computing, benchmarking, aspects of publish/subscribe
services, and reliability strategies for data dissemination
in large-scale critical systems.
) –
Francesco Palmieri received the M.S. degree and the
Ph.D. degree in computer science from the University
of Salerno. He was an Assistant Professor at the Second
University of Napoli. Currently he is an Associate Professor
at the University of Salerno. His research interests
include advanced networking protocols and architectures
and network security. He has been the director of the
Networking Division of the Federico II University of Napoli
and contributed to the development of the Internet in Italy
as a senior member of the Technical-Scientific Advisory
Committee and of the CSIRT of the Italian NREN GARR.
He serves as the editor-in-chief of an international journal and participates to the
editorial board of other ones.
Aniello Castiglione received a Ph.D. in Computer Science
from the University of Salerno, Italy. Actually he is an
adjunct professor at the University of Salerno (Italy)
and at the University of Naples ‘‘Federico II’’ (Italy). He
received the Italian National Academic Qualification (ASN)
as Associate Professor in Computer Science. He serves
as a reviewer for several international journals and is
Managing Editor of two international journals. He also
acts as a reviewer in around 50 international journals. He
served as Program Chair and TPC member in around 90
international conferences. He acted as a Guest Editor in
several journals and serves as Editor in several editorial boards of international
journals. He published more than 100 papers in international journals and
conferences. One of his papers has been selected as ‘‘Featured Article’’ in the
IEEE Cybersecurity initiative. He has been involved in forensic investigations,
collaborating with several Law Enforcement Agencies as a consultant. He is a
member of several associations, including ACM, IEEE. His current research interests
include Information Forensics, Digital Forensics, Security and Privacy on Cloud,
Communication Networks and Applied Cryptography.