Cluster Computing

https://doi.org/10.1007/s10586-020-03053-x (0123456789().,-volV)(0123456789().,-volV)

Efficient dynamic resource allocation method for cloud computing

environment
Ali Belgacem1 • Kadda Beghdad-Bey1 • Hassina Nacer2 • Sofiane Bouznad1

Received: 4 June 2019 / Revised: 29 November 2019 / Accepted: 20 January 2020

 Springer Science+Business Media, LLC, part of Springer Nature 2020

Abstract
The dynamic resource allocation is a good feature of the cloud computing environment. However, it faces serious problems
in terms of service quality, fault tolerance, and energy consumption. It was necessary, then, to find an effective method that
can effectively address these important issues and increase cloud performance. This paper presents a dynamic resource
allocation model that can meet customer demand for resources with improved and faster responsiveness. It also proposes a
multi-objective search algorithm called Spacing Multi-Objective Antlion algorithm (S-MOAL) to minimize both the
makespan and the cost of using virtual machines. In addition, its impact on fault tolerance and energy consumption was
studied. The simulation revealed that our method performed better than the PBACO, DCLCA, DSOS and MOGA algo-
rithms, especially in terms of makespan.

Keywords Cloud computing  Dynamic resource allocation  Makespan  Cost  Task scheduling

1 Introduction management requires a good strategy to deal with the

problems resulting from their dynamic allocation and
Recent advances in information technologies and release [6, 7].
telecommunications have a major impact on lifestyles. It The virtual machine (VM) is a fundamental component
affected many areas of life, including employment, edu- of the cloud environment. It is a type of software that can
cation, health, and so on [1, 2]. This prompted researchers run an operating system and applications on the operating
to focus more on the cloud paradigm, as it is an important system of the physical machine [3]. This technology makes
part of modern technologies [3]. Cloud is the collection of it possible to create virtual networks, servers, devices, etc;
heterogeneous resources known for their on-demand self- creating a pool of rentable, replaceable, and reassignable
service, scalability, resource pooling, broad network resources. This renders the provision of resources dynamic,
access, rapid, elasticity, and higher availability [4, 5]. fair, profitable, and most importantly, the virtual machines
These features allow flexible and dynamic sharing resour- become available to customers, at any time. This model of
ces among a large number of customers. However, resource cloud computing is called Infrastructure as a Service (IaaS)
[5]. Accordingly, the state of the system needs to be
& Ali Belgacem dynamically adjusted according to resource requirements
al.BLgacem@gmail.com and availability for a better performance in the cloud
Kadda Beghdad-Bey computing environment. Like so, task scheduling is the
k.beghdadbey@gmail.com most critical aspect within this issue. Yet, it is a difficult
Hassina Nacer problem to solve (NP-hard) [6].
sino_nacer@yahoo.fr Dynamism in the cloud lies in the ability to automati-
Sofiane Bouznad cally adjust resources according to the needs of the user.
bouznad.sofiane@gmail.com This management is subject to some constraints of cloud
1
providers and cloud users, which can pose problems for the
École Militaire Polytechnique (EMP), Bordj El Bahri, delivery of cloud services. Moreover, cloud dynamism,
Algeria
2
requiring continuous monitoring of requests and resources,
MOVEP Laboratory (USTHB), Algiers, Algeria

123
 Cluster Computing

handling of ever-changing requirements, schedules and allocation model. Section 4 presents the S-MOAL algo-
prices, selecting appropriate services and plans to meet rithm to schedule tasks over virtual machines. Section 5
overall objectives of the cloud [8, 9]. Dynamic resource shows the simulation results of the proposed method, and
allocation is really apparent in the case of the public cloud Sect. 4 concludes the paper.
due to the large number of users. This problem becomes
more acute in the case of infrastructure as a service.
Since computing resources are entities in which tasks 2 Related works
are executed, the diversity of their capabilities causes a
disparity in the execution time of tasks. In other words, The cloud computing paradigm provides easy and cheap
virtual machines are configured with various processing access to measurable and billable computing resources. As
speeds and memory, which indicates that a task executed such, customers can get different services on-demand,
on different VMs may result in varying execution costs. anytime and anywhere. These characteristics confer a
This results in a conflict of objectives because the virtual dynamic resource allocation capacity in this environment
machine that runs the task in a short time becomes more and also raise the question of the effectiveness of task
expensive, which raises the question of service quality to scheduling methods used for this purpose. Indeed, various
customers. Besides, fault tolerance is one of the major studies were held about the issue, yet, each addressed one
challenges to ensure reliability, robustness, and availability or two problems without generating an optimal solution
of services, as well as the applications’ execution in the [16]. Furthermore, the dynamic aspect of the cloud envi-
cloud computing system, especially because service ronment has been approached from different angles.
downtime is a major problem facing cloud providers in the A number of researchers highlighted dynamicity
last years [10, 11]. Typically, task failure may be due to through proposed scheduling approaches that employ the
VM malfunctioning or power failure. Likewise, regarding information on resources available and properties of
the efficiency of data centers, studies have concluded that scheduled tasks to overcome the diverse issues of resource
in average about 55% of the energy consumed in a data allocation. For example, the study [17] used the Cloud
center, is used up by the computing system, while the rest Information Service (CIS) to identify the services needed
used by supporting systems such as cooling, uninterrupted to perform the tasks received from the user. It presented the
power supply, etc. [12–15]. Discrete Symbiotic Organism Search (DSOS) algorithm to
At present, current works deal with these issues sepa- improve the makespan. In the work [18], authors proposed
rately, which leads to the improvement of each aspect a task scheduling algorithm to minimize both makespan
independently without achieving an optimal solution. and energy consumption. However, they did not consider
Furthermore, the dynamic aspect of the cloud has not been cost and fault tolerance issues. In addition, the Pareto-
explicitly addressed, which raises doubt on the effective- based multi-objective scheduling algorithm introduced in
ness of existing methods in real-world applications. For [19] targeted to allocate different on-demand virtual
this reason, this research focuses on dynamic virtual machines with optimal makespan and various prices by
machine scheduling to achieve efficient resource alloca- neglecting the energy consumption of the system. Whereas
tion. At the same time, we must maintain service quality to in [20] authors optimized a task scheduling solution by
customers by minimizing the makespan and cost of using proposing an improved genetic algorithm which allows
virtual machines; ensuring the permanence and durability finding an optimal assignment of subtasks on the resources
of the service over time by fault prevention and fault tol- with a minimum of makespan. However, the proposed
erance. We also offer the Cloud Resource Provider (CRP) method did not compare with existing scheduling methods.
the possibility to consume energy consciously. More pre- Also, the virtualization characteristic of cloud computing
cisely, the contributions of this paper are summarized as environment was not considered.
follows: Instead, in [4], researchers developed a dynamic
scheduling algorithm that balances the workload among all
– Construct a dynamic resource allocation model to
the virtual machines by neglecting resources usage cost. In
manage the energy consumption of unused VMs with
[10], they proposed a dynamic clustering algorithm
fault tolerance.
(DCLCA) to update and reflect on the status of cloud
– Propose a new approach named Spacy Multi-Objectives
resources. The given solution produces a group of inde-
Antlion algorithm S-MOAL to minimize makespan and
pendent tasks as a set of suitable computational size or task
cost.
clusters, using task characteristics focusing on priority or a
In light of the above, the paper discusses the related work partition. This work did not also consider resource usage
in resource allocation. Afterwards, Sect. 3 defines the cost constraint. In paper [21], the authors used a method to
problem and describes the proposed dynamic resource analyze the flow of information about resources to classify

123
Cluster Computing

the attributes of computing resources. This method speeds time. Other researchers proposed an online resource allo-
up the allocation of 14.65% and 7.43%, respectively, cation algorithm using VMs consolidation to achieve
compared to the traditional HEFT algorithm. Other work energy efficiency and reduce service-level agreement
based on the ant colony algorithm called PBACO took into (SLA) violations of cloud. For this end, authors considered
account the makespan, cost, deadline violation rate, and the power usage of servers, the number of migrations, and
resource utilization as constraints for multi-objective the path length of migrations in data centers [35]. Yet, the
optimization. However, this solution did not deal with the problem of fault tolerance was not addressed.
dynamic change of the cloud environment [22]. There are Instead, the authors of [36] combined placement con-
other studies, which allow enhancing the performance of trollers and periodic reassignments of virtual machines to
cloud, but need to be more adaptive with the dynamic achieve the highest energy efficiency subject to predefined
resource allocation, as with the works [23–26]. service levels. Additionally, to optimize the deployment of
On the other hand, some studies used predicting meth- virtual machine image replicas in cloud storage clusters, a
ods to monitor the alterations of resources consumed and new mechanism is proposed in [37]. The latter helps
required, which change dynamically. As in the case of improve the overall throughput of a storage cluster and
work [27], the authors used a Markov decision process to speeds up image installation operations. Alternatively,
estimate the execution time of user requests based on the another work aimed to achieve energy reduction for cloud
state of the environment. In addition, the system uses a providers and payment saving for cloud users at the same
genetic algorithm to develop a task scheduling method. time [15]. Nevertheless, this work required more evalua-
Furthermore, in [28], they proposed a new scheduling tions to prove its effectiveness in a cloud environment.
approach named PreAntPolicy. It consists of a prediction In addition to the above, there is a research category that
model based on fractal mathematics and a scheduler based addresses the dynamic resource allocation in the cloud, but
on an improved ant colony algorithm. Also, in the article in an inefficient way. In this regard, some researchers
[29], the authors presented a new adaptive scheduling attempted to mix the three trends mentioned above to adapt
algorithm based on the auction mechanism taking into the dynamic changes in the cloud environment. For
account the network bandwidth and the auction deadline. example, in the Ref. [38], the authors used a solution to
The proposed algorithm has good results in terms of cloud correctly map virtual machines to a group of tasks in a data
service providers profits and the resource utilization with- center. To our knowledge, this work did not optimize
out dealing with execution time constraints and power energy consumption efficiency. Furthermore, in [13],
consumption of the VMs. Furthermore, in [30], researchers researchers sought to find a good trade-off of energy con-
combined the Taguchi method with a differential evolution sumption and the makespan without considering monetary
algorithm (DEA) to minimize total cost and makespan. costs associated with the use of VMs. The same problem is
However, this solution still need to be improved to adapt to discussed in the work [39], the authors proposed a PD-LBP
the allocation of resources. The proposed algorithm in the method for the allocation of multi-agent tasks in dynamic
study [31] uses estimated execution time and various grid and cloud environments. This solution aims to accel-
optimization techniques to determine an almost optimal erate online response for improving the performances of
resource allocation plan. However, it did not take into task allocation. Evaluation results show that PD-LBP has a
account operating costs and energy consumption factors. good ability to assign large-scale tasks in a dynamic
Another class of researchers using the properties of environment, although this is not always feasible.
virtualization technology, such as migrating and consoli- In the same vein, a hybrid algorithm presented in [40]
dating virtual machines, as well as modifying the state of for load balancing, and maximizing the throughput of the
virtual machines (awake, sleep, listening, etc.) to cope with cloud provider network. Furthermore, In [41] authors used
dynamic changes in the cloud environment. For instance, a new GA-based approach (MOGA) to improve the accu-
researchers at [32] exploited the virtualization technique to racy of resource state prediction in the data center, thereby
solve the dynamic resource allocation problem. They pro- reducing energy consumption. Yet they have neglected
posed a Dynamic Energy Saving Resource Allocation QoS constraints. Another work presented a dynamic
(DPRA) mechanism to place the virtual machines properly resource management system’s design for cloud services.
on the physical machines to minimize VM migrations. In Their system seeks to adapt to the changing demands of
the work [33], they focused on dynamic virtual machine users in order to minimize energy consumption. However,
scheduling to achieve energy efficiency and satisfy dead- the results were poor in terms of optimizing resource
line constraints with heterogeneous physical machines. allocation [42]. In the paper [43], the authors suggested a
Similarly, in [34] presented a task-scheduling model for new approach for the dynamic autonomous management of
cloud computing to minimize energy consumption of data resources, without including the VMs state criteria. The
center servers, without paying attention to task execution proposed method gave acceptable results in terms of

123
 Cluster Computing

Table 1 The overview of the dynamic resource allocation aspect in the cloud
References Issue addressed Dynamic aspect

[38] Minimizing the power consumption Grouping tasks into defined categories (clusters) dynamically to assign them to an
appropriate group of VMs
[12] Decrease the power consumption Proposed a scheme capable of scaling up and down the amount of virtualized resources
[13] Minimize the energy consumption Real-time dynamic task scheduling
[39] Improving the performances of task Giving an indicator and a binary variable at each system state, while the dynamism and
allocation openness of the environment are simulated by modifying the values of the indicators
[4] Quality of service (QoS) The cloud users requirements vary dynamically
[40] Improving the performances of Grouping tasks into defined categories (clusters) dynamically to assign them to an
dynamic resource allocation appropriate VMs
[41] Decreases the energy consumption The resource utilization is dynamically predicted accurately, no matter, the characteristics
of consumers request
[42] Minimizing skewness Predicting the future resource needs of VMs
[11] Fault tolerance Real-time resource allocation taking into account the volatility of system performance
[43] Acheave a less SLA violations (QoS) Dynamically placing the VMs over hosts using a migration mechanism
[44] Minimizing the energy cost Controlling the system state with the goal of dynamic adjusting the capacity of VM
services

extensibility, feasibility, and flexibility. The researchers in limited amount of resources. The Cloud Resources Provi-
[44] adopted an MPC technique to design the dynamic der (CRP) also needs to conserve energy consumption and
algorithm and adjust the supply and price according to provide fault tolerance for the cloud system to ensure
customer demands. Nevertheless, dynamic scheduling in service continuity at any time.
this work is not intended to achieve the goal of fault
tolerance. 3.2 Formulation of problem
Table 1 below summarizes existing mechanisms for
dealing with dynamic resource allocation and the key In each allocation, the monitor detects the number of
issues in this area. resources available to specify the number of tasks that can
participate in the scheduling operation (Fig. 1). To tackle
this problem, the following formulations are used:
3 Representation of the studied problem Input
– Set of tasks T ¼ fT1 ; T2 ; . . .; Tn g represented by an
identifier id, length ln, file size file and arrival time.
Tasks arrive in the queud continuously, also considered
3.1 Description of the problem
independent, where:
This study focuses on the dynamic allocation of resources – Ti 6¼ ; i ¼ 1; 2; . . .; n,
in infrastructure as a service. In IaaS, the computing – Ti \ Tj ¼ ; i; j ¼ 1; 2; . . .; n and i 6¼ j
resources are dynamically delivered in a virtualized envi-
– Set of VMs ¼ fVM1 ; VM2 ; . . .; VMm g characterized by
ronment via the Internet (cloud public). So, when cus-
an identifier id, millions of instructions per second mips
tomers send requests to determine their resource
parameter, number of CPUs (k) and RAM size.
requirements via customers interface (Fig. 1), the cloud
resources provider makes a set of virtual machines avail-
Output Achieving a better mapping of Ti to VMj ðTi ; VMj Þ,
able to them. Each virtual machine is configured with its
in manner to reduce the makespan and the cost.
own processors and memory resources, which gives an
Makespan Equal to the latest finish time among all the
overview of a resource pool. These resources are dynami-
execution tasks set, which is equivalent to the maximum
cally reserved and released.
finish time. The makespan f(t) is obtained using Eq. 1.
In addition, the customer pays the virtual machine used  
according to use time. However, the available resource lni
f ðtÞ ¼ max ð1Þ
capacity should always meet the demands of customers, mipsj
which prompts the administrators to customize properly the

123
Cluster Computing

Fig. 1 Dynamic resource allocation architecture model

X
Cost The cost of the VM utilization is denoted by g and is Tcpu  kj for each VMj ð3cÞ
defined as the sum prices of running all the tasks on the
VMs, the cost is expressed as follows:
x0 ð3dÞ
X
N
gðtÞ ¼ gi ðtÞ ð2Þ Constraints 3a and b mean that the CPU and memory
i¼0 configuration of the selected virtual machine must be
gi ðtÞ ¼ pcj  tj þ Cij ð2aÞ greater than or equal to user requirements. While 3c means
that the number of tasks loaded by the virtual machine is
filei less than or equal to its number of processors. Moreover, to
Cij ¼  e; e ¼ 0:023 ð2bÞ
mipsj  32 determine the feasible solutions, the variable xij is con-
where gi is the cost associated with the execution of task i sidered equal to 1 if the task Ti is assigned to VMj , other-
on the resource j; pcj is the price of using the resource j, tj wise xij take 0, as illustrated in Eq. 4.

is the time of running the task i on resource j; Cij is the 1 if Ti allocated to the VMj
transfer rate cost for a bus of 32 bits wide with a clock xij ¼ ð4Þ
0 Otherwise
speed of mipsj ; e is the transfer price, N is the number of all
tasks and filei is the file size occupied by the task i.
Function objective The aim is to minimize the value of 3.3 The proposed dynamic resource allocation
the objective function F by optimizing the makespan and model
the cost, as shown in Eq. 3.
The proposed dynamic resource allocation (DRA) model
F ¼ Min ½f ðxÞ; gðxÞ ð3Þ
permits to deal with the main interesting issues facing both
Subject to : customers and CRP, which are: minimizing both the
makespan and cost, reducing the energy consumption and
VMram  Tfile ð3aÞ improving fault tolerance. Moreover, making the cloud
VMcpu  Tcpu ð3bÞ system elastic, gives the possibility of expanding or
shrinking their infrastructure by acquiring/releasing

123
 Cluster Computing

resources [19]. That is why, the proposed method takes into available resources. This aims to select the same types of
account the adjustment of VMs aware task scheduling. tasks to increase the availability of resources because of the
Henceforth, we performed the DRA model under the fol- convergence of the execution time of the tasks (resources
lowing assumptions: released in the convergence time). This constraint is sig-
nificant in the case of large arrival. While the constraint 6b
– The scheduling follows the change of the system
permits to ensure the possibility of scheduling the tasks on
according to the arrival rate of the tasks and the
the current ready VMs. Furthermore, the ready VMs are
availability of the resources (the set of tasks or the set
those check this condition:
of VM is not fixed).
– Each task is scheduled once, but if it fails, it can be – 8t 2 ½0; 1½9t ¼ ti : ðVMcpu ðti Þ  1 ^ VMram ðti Þ 20;
rescheduled. 100%Þ, it is a VM that has an availability in CPU and
– A task is scheduled to a VM as soon as it arrives with RAM at time t.
respecting the deadline. Since the resources are provided according to an agreement
To meet the above constraints, we designed our model between a CRP and its customers, the deadline (D) must be
architecture to automatically meet customers’ on-demand respected. As such, the sorting of pool tasks is based on
needs. As a result, the resource becomes accessible directly their arrival time, with tasks with tight deadlines being
and easily once released. More precisely, to make our processed earlier than others to avoid SLA1 violations.
system ready to respond to different inside and outside Therefore, the Realtime Task (RTi ) that belongs to VMj is
events of the cloud changing environment, four parameters given as Eq. 7. The task with minfRTi g is selected firstly.
are defined (a; b; l; c), which provide information about the
RTi ¼ Di  ðfij þ tc Þ; ð7Þ
number of selected tasks, availability of RAM resources,
arrival rate and the number of failed tasks respectively where tc represents the current time and fij is the execution
(Fig. 1). Besides, the monitor analyzes the status of system
changes, such as currently used virtual machines, actual time of task Ti on VMj .
task processing time, and task patterns, etc. Accordingly,
we use two mechanisms (Fig. 2), the first is mainly con-
cerned with service quality, while the second offers the 3.3.2 Dynamic VMs adjustment mechanism
possibility to increase and reduce resources’ quantity;
minimizing the energy consumption of unused virtual In this study, we deal with two types of VMs that can be
machines, such as explained in next: encountered through the dynamic resource allocation in a
cloud environment, as is done in the following paragraphs:
3.3.1 Dynamic task selection mechanism Unused VMs (UVMs): Adjusting the situation of unused
virtual machines is considered an important step in the
The primary purpose of this mechanism is to use resource allocation of resources to enable their efficient use.
availability information to dynamically select the appro- Therefore, decrease the energy consumption (EC) tax for
priate tasks from the arrival tasks and the failed tasks (pool the CRP, knowing that the EC of the running VM (used
of tasks). The creation of the selected task set is as follows: VM) is payable by the customer. Accordingly, a standard
Markov decision process (MDP) mechanism is adopted to
– Calculate the selected tasks set size (a): overcome this issue. It is formalized as a tuple M = (S, A,
X
a¼ VMcpuj D) where:
ð5Þ
– S is the set of all possible VM states.
a represents the number of released CPUs – A is the set of all possible VM actions.
– Measure the common average RAM between the – D is a transition function D : S  A7!DðSÞ.
available resources (b): This mechanism is designed to reduce the energy con-
1 X sumption of the UVMs where:
b¼  VMramj ð6Þ
a • S ¼ fVa ; Vs; Vt ; Vl g : Va is the awake state, Vs is the
Trami  b ð6aÞ sleep state, Vt is the switching from the sleep state to
  the awake state and Vl is the listening period (Fig. 3).
Trami  min VMramj ð6bÞ
From the perspective of the whole system, the energy
conservation level of the system is achieved according to
The constraint 6a means that resource requirements of the
1
selected tasks must be less than the average of RAM Service level agreement.

123
Cluster Computing

Fig. 2 Dynamic task selection

and VMs adjustment
mechanisms

minimizing the EC of UVMs, which is calculated accord- if l [ 1 DðVMs Þ ¼ Min pð

s j s00 Þ
ing to Eq. 8.
X X X X if l [ 1 DðVMs Þ ¼ Max pð
s j s00 Þ
EC ¼ ECa þ ECs þ ECt þ ECl ð8Þ if l ¼ 1 keep the same state ð10Þ
number of arrival T þ c
l¼
number of runing T
– Where ECa is the EC of an awakening VMs, ECs is the
EC of a sleeping VMs, ECt is the EC of a switching Knowing that, c is the number of failed tasks.
VMs and ECl is the EC of a listening VMs. Failed VMs (FVMs): Actually, resource allocation in a
cloud environment suffers from service downtime. It is a
The transition function (D) as defined in the MDP is major problem facing cloud providers, which leads to a low
Markovian, that is the probability of reaching the next state level of service quality provided to customers. Because
depends only on the current state and action and not on the once task failure occurs, it impacts the execution of the
history of earlier states [45]. pð s j s00 ; ECs00 Þ represents the tasks scheduled to the failed components [10]. In this work,
probability that the VM will change to state s00 , with energy the executed task failed due to a failure of the virtual
consumption ECs00 . We impose classically that machine, while the virtual machine is considered failed if:
P
8s0 ; ECs00 pðs j s00 ; ECs00 Þ ¼ 1 and pð s j s00 Þ [ 1. In addi-
tion, a matrix representation of these transition probabili- – The execution time of one of its loaded tasks exceeds
ties is conventionally used noting Pv and its dimensions are the expected execution time on this virtual machine.
j4j  j4j (matrix 9). This requires rescheduling these tasks and replacing the
0 1 failed virtual machine. However, we do not know the
Va Vs Vt Vl
B C actual task execution time before scheduling, but its
Va B 1 0 pða j tÞ  pða j tÞ C lower and upper bounds can be obtained (Eq. 11) [15].
B C
Pv ¼ V s B B 0 2 1 0 C
C It is determined as follows:
B C    
Vt @ pðt j sÞ pðt j sÞ 1 0 A fi 62 ½min VMmipsj ; max VMmipsj  ð11Þ
Vl 1 0 0 2
ð9Þ We measure the disparity in the failed task scheduling
performance (DPSF) by dividing the time spent in the case
Hence, the transition of an unutilized VMs status is of no failed tasks by the time delay in the case of failed
accorded with the variation of arrival rate (l) as follows: tasks, as shown in the Eq. 12.
makespanj ðc ¼ 0Þ
DPSF ¼  100 ð12Þ
makespani ðc [ 0Þ

123
 Cluster Computing

elasticity and scalability of the system over time. More-

over, the value of l and c is related to the (task-VM)
mapping achieved through the scheduling approach. It can
also affect the other parameters (b; a); this is what we seek
to address in the next section.

4 Proposed scheduling approach based

on ALO

Ant Lion Optimizer (ALO) is a new heuristic approach

proposed by Mirjalili and Seyedali to solve optimization
problems. As its name implies, the ALO algorithm mimics
Fig. 3 VMs status change the intelligent behaviour of antlions in hunting ants in
nature [46]. In such a way, when ants move over a search
space, antlions hunt them using traps. In this study, ALO
The DPSF indicates the system’s ability to provide fault
operators (Fig. 4) are used as follows :
tolerance. The pseudo-code of dynamic resource allocation
is presented in Algorithm 1.
4.1 Coding
Algorithm1: Dynamic resource allocation algorithm
In order to solve the current problem, the ant is considered
Input: Alc ; readyV M List; T List;
a task seeking to find a suitable ready virtual machine and
Output: An optimized dynamic scheduling
the antlion as a targeted ready virtual machine. Each
combination (task-VM) refers a solution of S-MOAL
1: While (t → ∞) algorithm. While, the solutions are presented in a vector, as
2: Alc ← f alse shown in Fig. 5. For example, task three is placed in the
3: t0 ← t virtual machine of index eight and the size of the vector
4: calculate α using the equation 5 used is equal to the number of tasks.
5: calculate μ using the equation 10
4.2 Initialization
6: If (α = 0)
7: readyV M List ← V M s First, an initial ant population is randomly created and is
8: calculate β using the equation 6 the main search agent in S-MOAL approach. After that, the
9: Alc ← true fitness value of each ant is evaluated by the steps described
10: end If below, using the objective function to find the best ants and
//selected Tasks consider them an elite (optimal ants), while the interaction
between ants and antlions in the trap generates a xi solu-
11: While (Alc=true)
tion. The initial set of ants used during the optimization is
12: t ← ++
preserved in the following matrix:
13: For Each (Ti ) 0 1
a1;1 a1;2 :::: a1;m
14: If (Trami ≤ β and T List.size() < α) B C
T List ← Ti
B a2;1 a2;2 :::: a2;m C
15: B C
16: end If
an ¼ BB : : :::: : C
C ð13Þ
B C
17: end For @ : : :::: : A
18: end While an;1 an;2 :::: an;m
19: Call Algorithm2 Where: ai;j 2 ½0; 1
20: t + +; It is further assumed that the antlions are hiding some-
21: end While where in the search space where their positions are saved in
the matrix 14.
It is worth noting that the previous parameters contribute
to forming the ready VMs set to execute the selected
tasks. In addition, a good adjustment of these ensures the

123
Cluster Computing

Fig. 4 Flow chart of proposed dynamic resource allocation method

0 1
A1;1 A1;2 :::: A1;n WðkÞ ¼ ½0; cumð2nðk1 Þ  1Þ; . . .; cumð2nðkh Þ  1Þ;
B C
B A2;1 A2;2 :::: A2;n C ð15Þ
B C
AL ¼ B
B : : :::: : C C ð14Þ
B C where cum calculates the cumulative sum, h is the maxi-
@ : : :::: : A mum number of iteration, nðkÞ is a stochastic function, k
Am;1 Am;2 :::: Am;n shows the step of random walk (iteration in this study) and
rand is a random number generated with uniform distri-
Where: Ai;j 2 ½0; 1
bution in the interval of [0, 1].

4.3 Random walks of ants 1 if rand [ 0:5
nðkÞ ¼
0 if rand  0:5
The ants move over the search space using random walks
around the antlions. To mimic such behavior, the initial In order to keep the random walk in the boundaries of the
population provided by the previous step (matrix 13) is search space and prevent the ants from overshooting, the
manipulated, as follows: random walks should be normalized using the following
equation:
rk1  rk2
wðkÞ ¼ rk2 þ WðkÞ ; rk1 [ rk2 ð16Þ
a
if wðkÞ [ a wðkÞ ¼ j1  aj ð16aÞ
if wðkÞ\0 wðkÞ ¼ 0 ð16bÞ

The borders rk1 and rk2 of W(k) updating are defined as

Fig. 5 Coding scheme proportional to the current number of iterations using the

123
 Cluster Computing

roulette wheel method, and a provides the number of CPUs 4.6 Elitism
available in the cloud.
It should be mentioned that the population of antlions is At this step, the produced solutions are gathered to form an
never evaluated, whereas the locations of ants are assumed elitist set.
in the first iteration, then they are relocated to new posi-
tions of antlions in the remaining of iterations until the ants 4.7 Extract the non-dominant solutions
become better. Therefore, S-MOAL simulates the entrap-
ment of ants in antlions pits by changing the random walks The purpose of this step is to examine the solutions
around antlions. This way, the likelihood of choosing the acquired from the previous step (elitist set) that exists in the
random walk is proportional to its fitness value, which archive (Fig. 4). In doing so, the Pareto method is used,
corresponds to the act of ant trapping in antlion pits. which is an effective mechanism to help find a compromise
between two objectives. In this search, the Pareto is defined
4.4 Building a trap for sliding ants as follow:
towards the antlion
– Definition The vector FðxÞ _ is said to dominate another
vector Fð€ xÞ denoted Fi ðxÞ
_ Fi ð€xÞ if only if fi ðxÞ\g
_ i ð€xÞ
In nature, bigger antlions construct bigger pits to increase
for all i 2 f1; 2; . . .ng and fi ðxÞ\g
_ j ð€
x Þ for some j 2
their chance of survival. Then, they shoot sand outwards
the center of the pit once they realize that an ant is in the f1; 2; . . .ng [48].
trap. For that reason, the roulette wheel operator is This way, non-dominant solutions are sorted and evaluated
employed to find q1 and q2 from the matrix 14 and according to their suitability.
therefore choosing a fitted antlion that is able to catch the
ant. In this regard, S-MOAL simulates this antlion attitude 4.8 Treat the non-dominant solutions
by formulating the following equations:
To determine the relationship between two non-dominant
qk1
qk ¼ ; ð17Þ solutions, a technique based on Spacing Distance is
L
applied. This latter aims to remove the worst solutions
qk2 from the archive and replace them with the best to
pk ¼ ; ð18Þ
L accommodate new solutions. Equations from 20
where L is a ratio, qk and pk are the best locations for the through 24 are used to determine which solutions from
k non-dominants are worth preserving in the archive and
antlions at k-the iteration. In Eqs. 17 and 18, L ¼ 10Gs_
which should be discarded.
where s_ is the stability factor and G is defined based on the
i i
current iteration (G = 2 when k [ 0.1s,_ G = 3 when k [ fmax  fmin
pif ¼ ; ð20Þ
_ G = 4 when k [ 0.75s,_ G= 5 when k [ 0.9s_ and G= 6
0.5s, m1
when k [ 0.95s). _ These parameters allow for forming the gimax  gimin
solution xki (Fig. 6). pig ¼ ; ð21Þ
m1

4.5 Catching prey and re-building the pit where pif and pig are the average distance for the non-
dominant set in terms of makespan and cost respectively, m
In the final part of hunting, the ant is attracted to the bottom is the number of non-dominant solutions. The following
of the pit and caught in the antlion’s jaw. After that, the equations show how to calculate the distance per individual
antlion pulls the ant inside the sand and consumes its body from its neighbors (hif and hig ):
[47]. So that, S-MOAL simulates this process by using  i   
f  f i1  þ f iþ1  f i 
Eq. 19. i
hf ¼ ð22Þ
k k
2
Fðxki Þ ¼ ef ðxi Þ þ egðxi Þ ð19Þ  i   
g  gi1  þ giþ1  gi 
i
hg ¼ ð23Þ
Where 2
– F k is the fitness of executing a task i on a VM j. Thus, the new fitness of each individual is:
– The solution xki is an outcome of the preceding steps ½hif  pif  þ ½hig  pig 
   ¼ ð24Þ
xki 2 qk ; pk ; wðkÞ , that gives a good result according FðiÞ
  2
to Eq. 19 (xki is the best of qk ; pk ; wðkÞ ) (Fig. 6).

123
Cluster Computing

iterations). The pseudo-code of S-MOAL is presented as

Algorithm 2.

5 Experimental setup and results

This section presents the experiments carried out to eval-

uate the proposed solution. First, we describe the simula-
tion setup considered in the experiments. Then, three
experiments are adopted during the evaluation, empirical
experiment, statistical experiment and the DRA model
experiment. Second, we give the results of the proposed
Fig. 6 The production of a solution x dynamic resource allocation method and discuss its effect
on the different optimized metrics.

Finally, the non-dominant solutions that have the best FðiÞ
are retained. Thence, the algorithm continues to loop until
reaching the stop criterion (maximum number of

Algorithm 2: S-MOAL
Input: T1 , T2 , ..., Tm , V M1 , V M2 , ..., V Mn , k, h, Arch lis
Output: Return the best found solutions

1: For Each(Ti )
2: V Mj ← rand
3: xi ← (Ti , V Mj ) //correspond to ant agent
4: end For
5: While(k ≤ h)
6: For Each(ant)
7: Create a random walk and normalize it using the equations (15) and (16)
8: Update the the ant’s position using the equations (17) and (18)
9: end For
10: For Each(ant)
11: Evaluate fitness using the equation (19)
12: end For
13: Produce the solutions xi (elitism set)
14: For Each(xi )
15: Evaluate fitness using the equation (19) (ẍ represents the solutions of the archive)
16: If (F (xi ) ≺ F (ẍ))
17: Arch list ← xi
18: Else Evaluate fitness using the equations from 20 through 24
19: If (F̄ (xi ) > F̄ (ẍ))
20: Arch list ← xi
21: end If
22: end If
23: end For
24: k++
25: end While

123
 Cluster Computing

5.1 Simulation setup Table 4 Energy consumption values

Energy consumption Values (mW)
5.1.1 Tools and platform
EC of Vs 3
To check the performance of our method, we used a EC of Vt 2.3
Cloudsim 3.0.3 toolkit. It is a simulation tool to mimic the EC of Va 20
cloud computing scenarios. The experiments are done on EC of Vl 1.1
Intel (R) core (TM) i5 3320M Processor 2.6 GHz equipped
with 4GB RAM, Windows 7 platform, using the Eclipse
provides better quality fault tolerance scheduling. Instead,
IDE Luna release 4.4.0.
the MOGA algorithm is mainly used to minimize energy
consumption. Nevertheless, these algorithms are designed
5.1.2 Cloud infrastructure
as mentioned in the works [17, 10, 41, 22], but in a con-
structive manner to solve a multi-objective problem taking
The experiments are conducted for one data center using
into consideration the parameters and formulations of the
the parameter setup of VMs as shown in Table 2, where the
issues treated in this paper.
configuration of the three types of virtual machines is
In addition, to examine whether the optimization metrics
inspired by the configuration of Google instances [49].
obtained by S-MOAL are significantly more efficient than
While the parameter setup of tasks is shown in Table 3.
those of the other methods, a one sided t-test is used. It is a
Besides, energy consumption values (Table 4) of different
statistical test to find an observed or extreme result. This
virtual machine states have been chosen according to the
test conforms to two hypotheses:
assumption (ECa [ ECs [ ECt [ ECl ). Moreover, the
_ of S-MOAL is set at 4.
stability factor (s) – The null hypothesis (H0 ) is defined as a no difference
between the optimization metrics for the proposed
method and the other methods.
5.1.3 Evaluation methods – The alternative hypothesis (H1 ) is the opposite of the
H0 and assumes that some difference exists between
In order to evaluate the effectiveness of the proposed optimization metrics.
method, it has been compared with other existing methods On the other hand, to evaluate the impact of the parameters
in the literature, including heuristic algorithms DSOS, (a; b; l) on the behavior of the proposed DRA model,
PBACO, DCLCA, and MOGA. The DSOS algorithm has S-MOAL is compared with the DSOS, PBACO and
the advantage of converging faster as the search increases, MOGA algorithms. Furthermore, this comparison aims to
making it suitable for large-scale scheduling problems. The appreciate the scalability and elasticity of the proposed
PBACO method provides better results with regard to the model. In other words, this experiment is used to preview
makespan, cost, and deadline violation. The comparison how the DRA model responds to dynamic system change
with DCLCA makes it possible to evaluate S-MOAL in and customer requirements.
terms of fault tolerance, especially because DCLCA
5.2 Discussion of the performance optimization
Table 2 The types of existing VMs in data center results
VMs MIPS CPU RAM GO Price $/h
In this experiment, the number and type of used VMs are
Standard 1000 1 3.75 0.0475 mentioned in Table 5. Each test is performed 10 times,
High-mem 2000 8 52 0.4736 then the average value is taken. It should also be noted that
High-cpu 2500 16 14.40 0.5672 S-MOAL refers, in this part, to the implementation of
Algorithm 1. The obtained results are detailed below.

Table 3 The parameters setup of tasks in data center 5.2.1 Empirical experiment
Parameter Value

The number of tasks [100, 1000] Figure 7 shows the makespan variation for different
Length [200, 1000] tasks. This figure shows that the makespan increases as the
File size [400, 1000] MB number of tasks increases. It is clear that S-MOAL per-
Output size [20, 40] MB forms better than the other algorithms in terms of

123
Cluster Computing

Table 5 The VMs and arrival

Experiment Number of VMs l
rate used in experiments
Standard High-ram High-cpu

Empirical and statistical experiments 10 5 22 0.3

DRA experiment 114 500 64 0.2
0.5
1

Fig. 7 Makespan results for different number of tasks

Fig. 9 The violation rate of deadline for number of tasks = 600

Fig. 8 VMs usage costs for different number of tasks

Fig. 10 The energy consumption of the UVMs

makespan. The examination of the proposed approach in

effectiveness of the mechanisms of task selection and vir-
terms of cost is presented in Fig. 8. It is apparent from the
tual machine adjustment; chiefly because the main purpose
comparison of the three scheduling algorithms that the
of both mechanisms is to form adequate sets of ready
proposed approach decreases the value of the cost, due to
virtual machines and selected tasks in each allocation.
the completion of each task in an appropriate virtual
Furthermore, the third test checks the S-MOAL deadline
machine that ensures a minimum price. Therefore, S-
violation percentage for 600 tasks at different deadlines.
MOAL allows a better match between the tasks and the
The deadline violation rate (Dviolation ) is calculated by the
virtual machines. These results also reflected the
formula 25.

123
 Cluster Computing

compared methods in terms of makespan as well as for

MOGA in terms of energy consumption, we confine the
comparison of S-MOAL with the these two latter. The
choice of significance level at which we reject H0 is 1%
with critical values obtained from statistical Tables 6 and 7
as 2.738 and as 3.052 respectively.
Table 6 shows the calculated values of the average
makespan metric obtained statistically at ten times. It
demonstrates that for samples 300 to 1000, the calculated t-
value is greater than 2.738 (critical value), which indicates
that there is a significant difference between the S-MOAL
makespan and that of the algorithms. Henceforth, H1 is
accepted for samples 300 through 1000. However, for the
rest of the samples, the calculated value t is less than the
critical value, which means that the difference between
performances is not significant. This leads to the accep-
Fig. 11 Fault tolerance tance of H0 for samples 100 and 200. In addition, it can be
deduced that S-MOAL outperforms DSOS when the search
V space is larger because, for samples 300 and above, the
Dviolation ¼  100%; ð25Þ makespan is significantly lower than that of DSOS.
a
On the other hand, Table 7 represents the comparative
whereas V is the number of executing tasks that violated results of energy consumption for the proposed method and
the deadline. The real execution time of tasks is calculated the MOGA algorithm. The same interpretation says about
using the Eq. 7. The S-MOAL algorithm gives a reducing the energy consumption where the calculated t-value is
range which is from 8 up to 23.5% compared to the other greater than the critical value 3.052 for all the samples,
algorithms (Fig. 9). This is due to the ability to select which indicates that there is a significant difference
parallel tasks with convergent deadlines. Besides, our between energy consumption performance of S -MOAL
method dynamically starts and stops virtual machines, as and MOGA. This is due to the efficiency of virtual machine
needed, which dynamically increases and decreases the tuning compared to the MOGA algorithm.
number of resources according to the number of tasks. As a
result, the execution of the tasks becomes faster without 5.2.3 The dynamic model experiment
violating the deadline.
Figure 10 shows the total energy consumption for
unused VMs, measured as shown in Eq. 8. We can see that Since our model is dynamic, sets of allocations are
energy consumption of all algorithms is proportional to the obtained in each test. The purpose of this experiment is to
number of tasks, but the slope of the S-MOAL graph is the assess the response of the proposed DRA model with
least. That is due to a good adjustment of the unused VMs changes in the rate of requests’ arrival
state, which helps to save energy. In this work, fault tol- (l ¼ 1; l ¼ 0:5; l ¼ 0:2) and customer needs. Figure 12
erance ratio is measured as shown in Eq. 12. From Fig. 11, shows the number of allocations for different arrival rates.
the S-MOAL shows an improvement in dealing with the Visually, the S-MOAL curve located above the DSOS and
failed tasks. In comparison with the DSOS and DCLCA MOGA curves in the three graphs, indicates that S-MOAL
algorithms, it presents a lower DPSF, which permits to is better than the other algorithms.
deduce that S-MOAL provides better fault tolerance with Figure 13 presents the number of CPUs that are made
the increased number of tasks. available to customers in a different l. S-MOAL released a
large number of CPUs compared to DSOS and MOGA
5.2.2 The statistical experiment algorithms. For l ¼ 1 the CPUs resources (a) are normally
released as soon as the running task is completed. Instead,
In order to ensure the optimization performance of the in a case, l ¼ 0:5 the a is minimized, but approximately
proposed method, the statistical analysis in the Tables and stable. Whereas a great amount of resources is released
helps prove the effectiveness of the results obtained from when the arrival rate of tasks achieves l ¼ 0:2. In terms of
the empirical experience in terms of makespan and energy memory, few amounts are released in case of l = 1 and l =
consumption. However, since the empirical experiment 0.5. However, when l achieves 0.2, the released RAM
shows that DSOS has better results than the other increases (Fig. 14). Furthermore, the results in Fig. 15

123
Cluster Computing

Table 6 Comparison of makespan obtained by DSOS and S-MOAL for statistical experiment
Tasks size DSOS S-MOAL Improvement t-test
a
Best Worst Avg Best Worst Avg Best (%) Worst (%) Avg (%) Calculate t-value

100 58.52 86.3 68.87 78.06 261.38 165.38 - 33.39 - 66.99 - 58.36 - 1.91
200 126.64 151.52 139.41 51.89 180.07 114.38 53.62 - 15.85 17.95 1.19
300 218.1 250.17 233.86 42.83 149.54 96.89 74.95 46.14 58.57 7.04
400 313.28 331.9 320.65 46.16 362.08 221.7 83.31 - 8.33 30.86 3.37
500 391.55 434.03 416.92 52.73 183.27 117.20 81.27 57.78 71.89 13.01
600 468.22 537.55 510.02 39.96 139.74 87.33 83.97 74 82.88 21.43
700 576.83 651.63 612.3 38.71 135.40 86.71 87.88 79.22 85.84 34.29
800 654.91 763.11 708.04 38.23 146.87 91.78 87.09 80.75 87.04 29.43
900 751.48 868.11 809.73 42.58 148.56 96.34 87.55 82.89 88.10 32.34
1000 863.72 1015.56 934.05 43.26 151.31 96.64 87.84 85.1 89.65 36.33
a
Average

Table 7 Comparison of energy consumption obtained by MOGA and S-MOAL for statistical experiment
Tasks size MOGA S-MOAL Improvement t-test
Best Worst Avg Best Worst Avg Best (%) Worst (%) Avg (%) Calculate t-value

100 0.483 1.32 0.9 0.33 1.09 0.16 45.45 21.11 4.62 3.89
200 0.52 1.23 0.88 0.5 1.1 0.104 4 11.82 7.46 9.24
300 0.467 1.27 0.84 0.43 1.12 0.07 8.6 13.39 11 15.89
400 0.462 1.29 0.88 0.42 1.14 0.06 10 13.16 13.67 20.92
500 0.46 1.26 0.87 0.4 1.3 0.05 15 - 3.08 16.4 26.06
600 0.48 1.28 0.87 0.411 1.261 0.03 16.79 1.51 28 47.89
700 0.48 1.28 0.88 0.41 1.02 0.032 17.07 25.49 26.5 45.07
800 0.48 1.3 0.89 0.419 1.025 0.029 14.56 26.83 29.69 51.08
900 0.47 1.29 0.88 0.417 1.029 0.026 12.71 25.36 32.85 57.02
1000 0.48 1.3 0.89 0.415 1.034 0.023 15.66 25.73 37.69 66.15

Fig. 12 The number of allocations

123
 Cluster Computing

Fig. 13 Released CPUs (a)

Fig. 14 Released RAM ðbÞ

show a few amounts of unused VMs from the S-MOAL of resources used is adjusted according to the arrival rate
algorithm compared to the other methods, which reflect the (number of customers). This reflects the scalability and
efficient use of resources. flexibility of the system. This experiment used a large
These results are convincing because for l = 1 the number of tasks (between 10,000 and 50,000 tasks), which
number of allocation is regular due to the overhead of the means that S-MOAL is able to serve on a large scale.
system and with almost no released resources. Further-
more, for l ¼ 0.5, the number of allocations has become
important because there is a balance between current and 6 Summary and conclusion
incoming tasks. Like so, the number of released resources
increases. For l ¼ 0.2, the number of allocations is small, In this paper, we proposed an efficient dynamic resource
reflecting the high availability of resources. allocation method for cloud computing environments. To
Additionally, the results show that the system can con- simulate the interaction between CRP and customers, we
trol the change of resources according to the variation of scheduled a set of independent tasks on a set of available
the arrival tasks rate, using the parameters (a; b; l; c). virtual machines consisting of RAM and CPU resources.
Although the number of tasks changes, the system main- The proposed DRA model highlights the key issues faced
tains better responsiveness and behavior, while the number by resource providers in the cloud, including fault tolerance

123
Cluster Computing

Fig. 15 The average of UVMs

and energy consumption. Instead, the Spacing Multi-Ob- effective mechanisms for task selection and virtual
jective Antlion algorithm (S-MOAL) aimes at improving machine tuning.
service quality that is offered to customers by solving the
conflict between makespan and cost. This method combi-
nes the dynamic change of the cloud system state and the
task scheduling to solve various issues, which makes it References
effective.
1. Gupta, B., Agrawal, D.P., Yamaguchi, S.: Handbook of Research
To evaluate the performance of the proposed method, on Modern Cryptographic Solutions for Computer and Cyber
three experiments were used. The first shows good Security. IGI Global, Hershey (2016)
S-MOAL results in terms of makespan, cost, deadline, fault 2. Gupta, B.B.: Computer and Cyber Security: Principles, Algo-
tolerance, and energy consumption compared to the DSOS, rithm, Applications, and Perspectives. CRC Press, Boca Raton
(2018)
PBACO, DCLCA and MOGA algorithms. Likewise, the 3. Hamdaqa, M., Tahvildari, L.: Cloud computing uncovered: a
statistical experiment confirms the significant improvement research landscape. In: Advances in Computers, vol. 86,
of the S-MOAL for makespan and energy consumption. pp. 41–85. Elsevier (2012)
While the last experiment proved that our dynamic model 4. Kumar, M., Sharma, S.C.: Deadline constrained based dynamic
load balancing algorithm with elasticity in cloud environment.
resulted in an efficient responsiveness compared to the Comput. Electr. Eng. 69, 395–411 (2018)
DSOS and MOGA algorithms, regarding the number of 5. Mell, P., Grance, T., et al.: The Nist Definition of Cloud Com-
allocations, released resources and adjustment of unused puting. NIST, Gaithersburg (2011)
virtual machines. 6. Madni, S.H.H., Latiff, M.S.A., Coulibaly, Y., et al.: Resource
scheduling for infrastructure as a service (iaas) in cloud com-
We believe that the improved performance achieved by puting: challenges and opportunities. J. Netw. Comput. Appl. 68,
the S-MOAL algorithm is due to its attractive manner of 173–200 (2016)
hunting. Using the Pareto method led to improve the 7. Zhan, Z.-H., Liu, X.-F., Gong, Y.-J., Zhang, J., Chung, Henry
archive and thus, obtaining a near-optimal solution. The Shu-Hung, Li, Yun: Cloud computing resource scheduling and a
survey of its evolutionary approaches. ACM Comput. Surv.
proposed method is also valuable for the dynamic man- (CSUR) 47(4), 63 (2015)
agement of resources through adjusting the unused VMs, 8. Chowhan, S.S., Shirwaikar, S., Kumar, A.: Predictive modeling
which economizes the energy. This study is significant of service level agreement parameters for cloud services. Int.
since it takes into account fault tolerance performance that J. Next-Gener. Comput. 7(2), 115–129 (2016)
9. Sadashiv, N., Dilip Kumar, S.M.: Broker-based resource man-
makes the execution of tasks flawless on cloud computing agement in dynamic multi-cloud environment. Int. J. High Per-
systems. form. Comput. Netw. 12(1), 94–109 (2018)
In the future, we would like to apply this method in a 10. Latiff, M.S.A., Madni, S.H.H., Abdullahi, M., et al.: Fault tol-
real cloud computing system. In addition, we plan to erance aware scheduling technique for cloud computing envi-
ronment using dynamic clustering algorithm. Neural Comput.
implement it for workflow. We will also try to use other Appl. 29(1), 279–293 (2018)

123
 Cluster Computing

11. Yan, H., Zhu, X., Chen, H., Guo, H., Zhou, W., Bao, W.: Deft: 29. Kong, W., Lei, Y., Ma, J.: Virtual machine resource scheduling
dynamic fault-tolerant elastic scheduling for tasks with uncertain algorithm for cloud computing based on auction mechanism.
runtime in cloud. Inf. Sci. 477, 30–46 (2019) Optik Int. J. Light Electron. Opt. 127(12), 5099–5104 (2016)
12. Chou, L.-D., Chen, H.-F., Tseng, F.-H., Chao, H.-C., Chang, 30. Tsai, J.-T., Fang, J.-C., Chou, J.-H.: Optimized task scheduling
Yao-Jen: Dpra: dynamic power-saving resource allocation for and resource allocation on cloud computing environment using
cloud data center using particle swarm optimization. IEEE Syst. improved differential evolution algorithm. Comput. Oper. Res.
J. 12(2), 1554–1565 (2018) 40(12), 3045–3055 (2013)
13. Juarez, F., Ejarque, J., Badia, R.M.: Dynamic energy-aware 31. Wang, W.-J., Chang, Y.-S., Lo, W.-T., Lee, Y.-K.: Adaptive
scheduling for parallel task-based application in cloud comput- scheduling for parallel tasks with qos satisfaction for hybrid cloud
ing. Future Gener. Comput. Syst. 78, 257–271 (2018) environments. J. Supercomput. 66(2), 783–811 (2013)
14. Yong, L., Sun, N.: An effective task scheduling algorithm based 32. Chou, L.-D., Chen, H.-F., Tseng, F.-H., Chao, H.-C., Chang, Y.-
on dynamic energy management and efficient resource utilization J.: Dpra: dynamic power-saving resource allocation for cloud
in green cloud computing environment. Clust. Comput. 22(1), data center using particle swarm optimization. IEEE Syst. J.
513–520 (2019) 12(2), 1554–1565 (2016)
15. Zhang, Y., Cheng, X., Chen, L., Shen, H.: Energy-efficient tasks 33. Ding, Y., Qin, X., Liu, L., Wang, T.: Energy efficient scheduling
scheduling heuristics with multi-constraints in virtualized clouds. of virtual machines in cloud with deadline constraint. Future
J. Grid Comput. 16, 459–475 (2018) Gener. Comput. Syst. 50, 62–74 (2015)
16. Belgacem, A., Beghdad-Bey, K., Nacer, H.: Task scheduling in 34. Dong, Z., Liu, N., Rojas-Cessa, R.: Greedy scheduling of tasks
cloud computing environment: a comprehensive analysis. In: with time constraints for energy-efficient cloud-computing data
International Conference on Computer Science and its Applica- centers. J. Cloud Comput. 4(1), 5 (2015)
tions, pp. 14–26, 24–25 April, Springe in Algiers, Algeria (2018) 35. Jiang, H.-P., Chen, W.-M.: Self-adaptive resource allocation for
17. Abdullahi, M., Ngadi, M.A., et al.: Symbiotic organism search energy-aware virtual machine placement in dynamic computing
optimization based task scheduling in cloud computing environ- cloud. J. Netw. Comput. Appl. 120, 119–129 (2018)
ment. Future Gener. Comput. Syst. 56, 640–650 (2016) 36. Wolke, A., Tsend-Ayush, B., Pfeiffer, C., Bichler, M.: More than
18. Azad, P., Navimipour, N.J.: An energy-aware task scheduling in bin packing: dynamic resource allocation strategies in cloud data
the cloud computing using a hybrid cultural and ant colony centers. Inf. Syst. 52, 83–95 (2015)
optimization algorithm. Int. J. Cloud Appl. Comput. (IJCAC) 37. Cong, X., Yang, J., Weng, J., Wang, Y., Hui, Yu.: Optimising the
7(4), 20–40 (2017) deployment of virtual machine image replicas in cloud storage
19. Durillo, J.J., Prodan, R.: Multi-objective workflow scheduling in clusters. Int. J. High Perform. Comput. Netw. 10(4–5), 423–435
amazon ec2. Clust. comput. 17(2), 169–189 (2014) (2017)
20. Keshanchi, B., Souri, A., Navimipour, N.J.: An improved genetic 38. Alsadie, D., Tari, Z., Alzahrani, E.J., Zomaya, A.Y.: Dynamic
algorithm for task scheduling in the cloud environments using the resource allocation for an energy efficient vm architecture for
priority queues: formal verification, simulation, and statistical cloud computing. In: Proceedings of the Australasian Computer
testing. J. Syst. Softw. 124, 1–21 (2017) Science Week Multiconference, p. 16, January 29–February 02,
21. Wei, J., Zeng, X.: Optimal computing resource allocation algo- 2018, ACM in Brisband, Queensland, Australia (2018)
rithm in cloud computing based on hybrid differential parallel 39. Kong, Y., Zhang, M., Ye, D.: A belief propagation-based method
scheduling. Clust. Comput. 22, 7577–7583 (2018) for task allocation in open and dynamic cloud environments.
22. Zuo, L., Shu, L., Dong, S., Zhu, C., Hara, Takahiro: A multi- Knowl. Based Syst. 115, 123–132 (2017)
objective optimization scheduling method based on the ant col- 40. Mousavi, S., Mosavi, A., Varkonyi-Koczy, A.R., Fazekas, G.:
ony algorithm in cloud computing. IEEE Access. 3, 2687–2699 Dynamic resource allocation in cloud computing. Acta Polytech.
(2015) Hung. 14(4), 83–104 (2017)
23. Belgacem, A., Beghdad-Bey, K., Nacer, H.: Enhancing cost 41. Tseng, F.-H., Wang, X., Chou, L.-D., Chao, H.-C., Leung, Victor
performance using symbiotic organism search based algorithm in C.M.: Dynamic resource prediction and allocation for cloud data
cloud. In: 2018 International Conference on Smart Communica- center using the multiobjective genetic algorithm. IEEE Syst. J.
tions in Network Technologies (SaCoNeT), pp.s 306–311, 27–31 12(2), 1688–1699 (2018)
Oct 2018, IEEE in El Oued, Algeria (2018) 42. Xiao, Z., Song, W., Chen, Q.: Dynamic resource allocation using
24. Belgacem, A., Beghdad-Bey, K., Nacer, H.: A new task virtual machines for cloud computing environment. IEEE Trans.
scheduling approach based on spacing multi-objective genetic Parallel Distrib. Syst. 24(6), 1107–1117 (2013)
algorithm in cloud. In: International Conference on Computer 43. Onat Yazir, Y., Matthews, C., Farahbod, R., Neville, S., Guitouni,
Science and Information Systems, pp. 189–195, 9–12 September, A., Ganti, S., Coady, Y.: Dynamic resource allocation in com-
in Pozna, Poland (2018) puting clouds using distributed multiple criteria decision analysis.
25. Belgacem, A., Beghdad-Bey, K., Nacer, H.: Task scheduling In: 2010 IEEE 3rd International Conference on Cloud Comput-
optimization in cloud based on electromagnetism metaheuristic ing, pp. 91–98, 5–10 July 2010, in Miami, FL, USA (2010)
algorithm. In: 2018 3rd International Conference on Pattern 44. Zhang, Q., Zhu, Q., Boutaba, R.: Dynamic resource allocation for
Analysis and Intelligent Systems (PAIS), pp. 1–7, 24–25 October spot markets in cloud computing environments. In: 2011 Fourth
2018, IEEE in Tebessa, Algeria (2018) IEEE International Conference on Utility and Cloud Computing,
26. Zhang, F., Cao, J., Li, K., Khan, S.U., Hwang, Kai: Multi-ob- pp. 178–185, 5–8 December 2011, in Victoria, NSW, Australia
jective scheduling of many tasks in cloud platforms. Future (2011)
Gener. Comput. Syst. 37, 309–320 (2014) 45. Doshi, P., Goodwin, R., Akkiraju, R., Verma, K.: Dynamic
27. Barrett, E., Howley, E., Duggan, J.: A learning architecture for workflow composition: using markov decision processes. Int.
scheduling workflow applications in the cloud. In: Web Services J. Web Serv. Res. (IJWSR) 2(1), 1–17 (2005)
(ECOWS), 2011 Ninth IEEE European Conference on, pages 46. Mirjalili, S.: The ant lion optimizer. Adv. Eng. Softw. 83, 80–98
83–90, 14–16 Sept 2011, IEEE in Lugano, Switzerland (2011) (2015)
28. Duan, H., Chen, C., Min, G., Yu, W.: Energy-aware scheduling of 47. Mirjalili, S., Jangir, P., Saremi, S.: Multi-objective ant lion
virtual machines in heterogeneous cloud computing systems. optimizer: a multi-objective optimization algorithm for solving
Future Gener. Comput. Syst. 74, 142–150 (2017) engineering problems. Appl. Intell. 46(1), 79–95 (2017)

123
Cluster Computing

48. Das, I., Dennis, J.E.: Normal-boundary intersection: a new Hassina Nacer currently serves
method for generating the pareto surface in nonlinear multicri- as an Associate Professor in
teria optimization problems. SIAM J. Optim. 8(3), 631–657 Department of Computer Sci-
(1998) ence, University of Science and
49. On line: Tarifs de google compute engine. https://cloud.google. Technology Houari Boumedi-
com/compute/pricing. Accessed 17 Jan 2019 ene (USTHB). She received her
PhD degree from Bejaia
Publisher’s Note Springer Nature remains neutral with regard to University in 2010 and her
jurisdictional claims in published maps and institutional affiliations. Habilitation to direct Research
(HDR) in 2015 from USTHB.
From 1995 to 2001, she worked
Ali Belgacem is a third year in the Algerian customs. From
Ph.D. student in the computer 2001 to 2014, she was a teacher
science department at ’Ecole researcher in Bejaia University.
millitaire polytechniqu’. He Furthermore, she has taught in
holds a master’s degree in several universities (U.Bejaia, USTHB, EMP, U.Algiers 1). Her
computer science from the research interests include Web services Technology, Web semantic,
University of 8 May 1958 Knowledge base systems, Distributed and Cooperative computing,
Guelma, Algeria. His doctoral Web services security, and Cloud Computing. She refereed/ published
research investigates to deter- various papers for/in reputed international journals and conferences
mine an optimal solution for
dynamic resource allocation in Sofiane Bouznad is a teacher
the cloud computing. researcher in DCS Laboratory at
Military Polytechnical School,
Algeria. He is also a last year
Ph.D. student in LISSI Labora-
tory at Paris Est University in
France. He received B.E. degree
Kadda Beghdad-Bey received
in computer science from Mili-
his M.S. degree in Industrial
tary Polytechnical School,
Computer Science from Mili-
Algeria, in 2002, and M.Sc.
tary Polytechnic School in 2003.
degree in complex systems,
He received his Ph.D. in Com-
communication and control
puter Science from USTHB
technologies, from Paris Est
University, Algiers in 2010.
University, France, in 2013. His
Current, He is the Professor and
research interests include intel-
Head of the Laboratory of Dis-
ligent systems, activity recognition, wireless sensor network, Internet
tributed and Complex Systems
of things, distributed and Cloud computing.
at M. P. School. His areas of
research includes biometrics
identification, parallel and dis-
tributed computing, Meta-
heuristics techniques and task
scheduling, and resources allocation in cloud commuting.

123