Neural Computing and Applications (2020) 32:14817–14838

https://doi.org/10.1007/s00521-020-04834-6(0123456789().,-volV)(0123456789().
,- volV)

ORIGINAL ARTICLE

Cloud customers service selection scheme based on improved

conventional cat swarm optimization
Danlami Gabi1,2 • Abdul Samad Ismail3 • Anazida Zainal3 • Zalmiyah Zakaria3 • Ajith Abraham4 •

Nasiru Muhammed Dankolo3

Received: 15 February 2019 / Accepted: 5 March 2020 / Published online: 17 March 2020
 The Author(s) 2020

Abstract
With growing demand on resources situated at the cloud datacenters, the need for customers’ resource selection techniques
becomes paramount in dealing with the concerns of resource inefficiency. Techniques such as metaheuristics are promising
than the heuristics, most especially when handling large scheduling request. However, addressing certain limitations
attributed to the metaheuristic such as slow convergence speed and imbalance between its local and global search could
enable it become even more promising for customers service selection. In this work, we propose a cloud customers service
selection scheme called Dynamic Multi-Objective Orthogonal Taguchi-Cat (DMOOTC). In the proposed scheme,
avoidance of local entrapment is achieved by not only increasing its convergence speed, but balancing between its local
and global search through the incorporation of Taguchi orthogonal approach. To enable the scheme to meet customers’
expectations, Pareto dominant strategy is incorporated providing better options for customers in selecting their service
preferences. The implementation of our proposed scheme with that of the benchmarked schemes is carried out on
CloudSim simulator tool. With two scheduling scenarios under consideration, simulation results show for the first scenario,
our proposed DMOOTC scheme provides better service choices with minimum total execution time and cost (with up to
42.87%, 35.47%, 25.49% and 38.62%, 35.32%, 25.56% reduction) and achieves 21.64%, 18.97% and 13.19% improve-
ment for the second scenario in terms of execution time compared to that of the benchmarked schemes. Similarly,
statistical results based on 95% confidence interval for the whole scheduling scheme also show that our proposed
scheme can be much more reliable than the benchmarked scheme. This is an indication that the proposed DMOOTC can
meet customers’ expectations while providing guaranteed performance of the whole cloud computing environment.

Keywords Cloud computing  Scheduling  Cat swarm optimization  Pareto dominance

& Danlami Gabi 1

Department of Computer Science, Faculty of Science, Kebbi
[email protected] State University of Science and Technology, Aliero,
Kebbi State, Nigeria
Abdul Samad Ismail
2
[email protected] Department of Computing Science, Umea University, Umeå,
Sweden
Anazida Zainal
3
[email protected] Department of Computer Science, School of Computing,
Universiti Teknologi Malaysia, Johor, Malaysia
Zalmiyah Zakaria
4
[email protected] Machine Intelligence Research Labs, Scientific Network for
Innovation and Research Excellence, Auburn, WA 98071,
Ajith Abraham
USA
[email protected]
Nasiru Muhammed Dankolo
[email protected]

123
14818 Neural Computing and Applications (2020) 32:14817–14838

1 Introduction development, high-level research conducted in cloud

computing task and resource scheduling focuses on time
Cloud computing is a platform of choice providing dis- and cost models. These models capture customers’ QoS
tributed resources (e.g., virtual machine, storage and experience, where techniques of metaheuristic are exploi-
bandwidth) to meet increasing demand of customers. The ted in their evaluations. Although the metaheuristic tech-
environment provides cost-effective solution for running niques are attributed with certain limitations (e.g., local and
business applications through virtualized technologies global imbalance, local optimality) [14, 23], they have
[1–4]. Services that are made available by cloud environ- been proven to be more efficient in reducing the com-
ment are affordable using the concept of pay-per-use (PPU) plexity of a task scheduling problem when finding an
pricing models. Three types of service model are associated optimum or near-optimal solution [14, 24, 25]. Several
with the cloud computing environment: Software as a researchers [26–31] also focus on investigating relationship
Service (SaaS), Platform as a Service (PaaS) and Infras- between local optimality and slow convergence and prof-
tructure as a Service (IaaS) [5]. The cloud customers usu- fered solutions to addressing these concerns.
ally interact remotely with the SaaS layer to run their The conventional cat swarm optimization (CSO) is a
applications (e.g., datacenter) [5]. The SaaS layer thus metaheuristic optimization technique put forward by Chu
functions with the support of PaaS layer. The PaaS layer and Tsai in 2007. This technique mimics the behavior of
provides interactive mechanisms between cloud customers natural cat and has relatively proven better in terms of
and service providers. The PaaS thus allows cost-efficient convergence speed than the particle swarm optimization
development and deployment of applications [7, 8]. With (PSO) [32]. It has both the global and local search (also
the provision of numerous advantages (e.g., deployment of known as the seeking and tracing mode) and a control
PPU pricing model, maintaining application having the variable called the mixed ratio (MR) that determine whe-
same integrated environment) by the PaaS layer, the cloud ther the current position at which the cat is standing is in
customers can now have the means of using same inte- either seeking or tracing mode. As attributed to most
grated software application development environment metaheuristic techniques, the CSO suffers slow conver-
[9, 10]. On the other hand, the IaaS layer is responsible for gence speed that can lead to its entrapment at the local
providing services to the cloud customers in terms of the optima, and global and local imbalance causing instability
infrastructures. The cloud customers have no control over of the solution. Orthogonal Taguchi approach is a greedy-
managing the infrastructures; rather, they only utilize the based technique that when applied can increase the con-
available resources situated at the IaaS layer through ser- vergence speed of the CSO to avoid being trapped at the
vice request [6]. In another development, the IaaS provides local optima [13, 20]. In this paper, we proposed Dynamic
pool of resources of varied types which are leased Multi-Objective Orthogonal Taguchi-Cat (DMOOTC)
according to customers’ request. This paper provides cus- scheme. In the proposed scheme, we exploit advantages of
tomers with techniques for optimal mapping of their choice the orthogonal Taguchi approach to avoid the DMOOTC
of resources at the IaaS layer to meet their computation from being entrapped at the local optima. To provide
need [7, 11, 12]. customers with choices of service preferences, we incor-
At the IaaS layer, virtual machines are heterogeneous. porated the Pareto dominance strategy in the proposed
Some requests are said to have a very high demand for DMOOTC scheme. The simulation results show that our
virtual machines, while others require more storage proposed DMOOTC scheme can provide customers with
[13, 14]. In most instances, the customers demand resour- better service choices compared to the benchmarked
ces that have minimum costs of execution. Virtual schemes.
machines with high processing speed usually incur high The contribution of this paper is as follows:
processing cost [15–18]. In Amazon EC2 for instance, a • Development of a multi-objective task scheduling
cloud customer has the privilege of accessing and con- model for cost and computation time.
trolling set of virtual machines that run inside the data- • An improvement is proposed to the CSO algorithm
center of the service provider, while being charged for a using the orthogonal Taguchi approach and Pareto
specific time the virtual machine has been allocated. At this dominance.
instance, the customer satisfaction is measured based on • Development of a DMOOTC scheme to solve the multi-
the Quality of Service (QoS) (e.g., execution time, cost and objective task scheduling model.
storage) he/or she experiences [19, 20]. To provide cus-
tomers of cloud computing resources with better options of The rest of this article is organized as follows: Review
selecting their resource preferences, there is need for an on related work is discussed in Sect. 2. Section 3 discusses
optimal resource scheme [15, 16, 21, 22]. In recent the CSO technique. Orthogonal Taguchi approach is dis-
cussed in Sect. 4. Section 5 provides discussion on Pareto

123
Neural Computing and Applications (2020) 32:14817–14838 14819

dominance strategy. The proposed models are discussed in can generate better VM-provisioning plan for customers to
Sect. 6. Section 7 discusses the proposed DMOOTC achieve minimum task execution time and monetary cost
scheme. Discussion on simulation results is provided in compared to the benchmarked techniques.
Sect. 8. Section 9 discusses the performance metrics used In Zuo et al. [14], a Multi-Objective Ant Colony Opti-
in the evaluation of the schemes. Discussion on the simu- mization (MOACO) algorithm is proposed. Their objective
lation results is provided in Sect. 10. Section 11 provides is to minimize the makespan time and budgetary cost. A
discussion on the statistical results, and finally, Sect. 12 cost model that reflects the relationship between cus-
concludes the paper. tomers’ resource and budgetary cost is introduced in
evaluating the efficiency of their proposed algorithms.
According to the researchers, simulation results show their
2 Related works proposed algorithm has achieved minimum execution cost
compared to that of the benchmarked algorithms. However,
Task scheduling is one of the most research areas in cloud the updating process of pheromone exhibited by the ants
computing. Researchers have shown that scheduling along the path can lead to local trapping. Duan et al. [38] in
strategies such as heuristic and metaheuristic can be their part proposed a communication and storage-aware
promising when applied to deal with scheduling problems multi-objective task scheduling algorithm, which is based
in cloud datacenters. Metaheuristics such as particle swarm on sequential cooperative game. Their goal is to optimize
optimization (PSO), genetic algorithm (GA), ant colony the execution time and economic cost. In their proposed
optimization (ACO) and cat swarm optimization (CSO) are approach, the individual players are considered behaving
few examples of metaheuristic techniques that can handle selfishly. The global knowledge of all players is computed
large scheduling problems. Their improvement with tra- by ordering each customer’s task in decreasing order. The
jectory-based (e.g., simulated annealing (SA)) and greedy- simulation results show their proposed algorithm can
based techniques such as the orthogonal Taguchi approa- achieve better solution in terms of makespan, cost, system-
ches can further improve their performances toward pro- level efficiency and fairness in less execution time com-
viding more efficient solutions [33–35]. Few among pared to Grid-Min–Min, Grid-Max–Min and Grid-
researchers that exploit these advantages are discussed in Suffrage.
the following: In Verma and Kaushal [39], a Hybrid Particle Swarm
In Wei et al. [36], a Compounded Local Mobile Cloud Optimization (HPSO) algorithm is proposed. Their objec-
Architecture (LMCpri) with dynamic priority queue is tive is to provide resources that can guarantee customers
proposed to solve a multi-objective scheduling problem. In with minimum execution time and processing cost under
their proposed approach, a priority-based positioning deadline and budget constraint. In their studies, trade-off
technique based on auction processing is incorporated to values are adopted to provide customers with the choices of
store jobs upon arrival from the cloud customers. Then, a selecting their service preferences. According to the
Non-Static Genetic Algorithm II (NSGA-II) is introduced researchers, simulation results shows their proposed HPSO
for scheduling tasks on resources to achieve minimum scheduling approach can reduce execution time and exe-
processing time and decrease the request cost. According to cution cost as compared to the Non-Static Genetic Algo-
the researchers, simulation results show their proposed rithm II (NSGA-II), Multi-Objective Particle Swarm
algorithm can provide better performance than PSO and Optimization (MOPSO) and e-Fuzzy Dominance sort-
sequential scheduling algorithms in terms of minimum based Discrete Particle Swarm Optimization (e-FDPSO)
total execution time and cost. However, improvement in scheduling algorithm. However, the global optimization
their proposed method is still possible since mutation process exhibited by the PSO may not always guarantee the
process exhibited by the GA can lead to local trapping due required optimum solution. Panda and Jana [34] in their
to slow convergence speed. In another development, Liu part proposed a Multi-Objective Task Scheduling (MOTS)
et al. [37] proposed a Single Site Virtual Machine Provi- optimization algorithm for heterogeneous multi-cloud
sioning (SSVP) approach and ActGreedy to minimize task environment. Their goal is to minimize both makespan
execution time and monetary cost. In their proposed time and total execution cost. In their method, two phases
approach, a single-site initialization module is used to of task scheduling processes were adopted. The first phase
ensure virtual machine provisioning and multisite data goes through normalization process to scale values
transfer. At the instance of task execution, a virtual between 0 and 1. In the second phase, the normalization
machine is not allowed to restart for the execution of any process is performed by dividing the Expected Time to
continuous activities on the site due to the fact that they are Compute (ETC) matrix and cost matrix element into their
grouped and scheduled as a fragment. According to the corresponding maximum values. As put forward by the
researchers, simulation results show their developed SSVP researchers, simulation results show their proposed method

123
14820 Neural Computing and Applications (2020) 32:14817–14838

has outperformed the two benchmarked algorithms in Cuckoo Optimization Algorithm (MOSCOA). In their
achieving minimum execution time, minimum total cost of proposed approach, each cuckoo is used as a scheduling
execution and improves the average cloud utilization. solution. As tasks are placed in order of their arrivals, the
Ramezani et al. [15] introduced a Multi-Objective Particle cuckoo technique in turn does the mapping to the most
Swarm Optimization (MOPSO) algorithm to provide cus- appropriate virtual machines as it enables the movement of
tomers with better service choices to deal with the chal- tasks toward the global optima region using a target
lenges of high computation time and cloud performance. immigration operator. The technique was later evaluated
As stated by the researchers, their simulation results show using large number of random graphs and real-world
they can achieve an optimal solution in a reasonable application. The simulation results show their proposed
amount of time. However, incorporating a novel technique MOSCOA algorithm is much more superior in terms of
can further improve their solution in terms of generating performance compared to their previously proposed task
trade-offs for consumers’ service preferences during task scheduling algorithm. Voicu et al. [43] introduced Multi-
scheduling. Objective and Multi-Constrained (MOMC) task scheduling
Gabi et al. [33] put forward an Orthogonal Taguchi- algorithm for scheduling tasks in Hadoop system. Their
Based Cat Swarm Optimization (OTB-CSO) algorithm to objective is to minimize deadline and budget. The simu-
improve the performance of cloud environment. The lation results according to the researchers show their pro-
researchers exploited the advantages of Taguchi method to posed MOMC method can provide better performance in
achieve task mapping on appropriate virtual machines. The Hadoop system.
simulation results according to the researchers show their In Bilgaiyan et al. [44], a Multi-Objective Cat Swarm
proposed algorithm can reduce makespan compared to Optimization (MOCSO) algorithm is proposed. Their goal
Min–Max, Hybrid Particle Swarm Optimization with is to improve the performance of cloud environment in
Simulated Annealing (HPSO-SA) and Particle Swarm terms of minimum execution cost, makespan and CPU idle
Optimization with Linear Descending Inertia Weight time. In their task scheduling process using the proposed
(PSO-LDIW). However, the improvement in their OTB- MOCSO, a control variable known as the mixed ratio is
CSO to handle a multi-objective optimization problem is used to decide the best virtual machines to assign tasks.
required to meet customers’ expectation. Liu et al. [40] The experimental results according to the researchers show
dwelt on Improved Min–Min algorithm for cloud com- their proposed MOCSO algorithm can achieve minimum
puting environment. Their objective is to achieve QoS, execution cost, makespan time and CPU idle time than
dynamic priority model and minimum cost of service Multi-Objective Particle Swarm Optimization. On the other
delivery to customers. In their scheduling process, static hand, Xu et al. [45] in their part proposed a Multi-Objec-
priority rule and dynamic changing factors are used for tive Genetic Optimization Algorithm (MOGA). Their goal
providing the scheduling of higher priority tasks first. The is to minimize the average completion time, total com-
results of their simulation as indicated by the researchers pletion time and ensure load balancing on virtual machines.
show their proposed algorithm can increase resource uti- In their scheduling process, large tasks are divided into
lization, ensure longer task is executed at reasonable time multiple sub-tasks using the chromosomes encoding. Each
and meet customers’ QoS requirement compared to the chromosome length signifies length of the sub-tasks with
benchmarked algorithm. In another development, Beegom smaller tasks mapped on virtual machines. To determine
and Rajasree [41] in their part put forward a new variant of the performance of their proposed MOGA, three different
continuous Particle Swarm Optimization (PSO) algorithm fitness functions models were designed to evaluate the
called the Integer-PSO. The Integer-PSO adopts Pareto fitness of each chromosome according to their objectives.
optimality using a weighted sum approach. Their goal is to The simulation results presented by the researchers show
minimize the makespan and execution cost. In their their proposed MOGA can achieve the minimized average
scheduling process, a model as a constraint biobjective completion time, total completion time and ensure load
optimization for the makespan and cost is developed. The balancing on virtual machines with faster convergence than
efficiency of their proposed algorithm is tested using the the benchmarked single-objective genetic algorithms. In
developed model. Simulation results as shown by the their part, Milani and Navin [46] proposed a multi-objec-
researchers indicate that their proposed Integer-PSO has tive scheduling algorithm based on PSO technique. Their
outperformed the Smallest Position Value (SPV) rule- objective is to minimize the total execution time, average
based PSO technique in terms of achieving minimum waiting time and number of missed tasks. The researchers
makespan and execution cost. exploit the PSO technique to propose a scheduling
To ensure effective scheduling on heterogeneous virtual approach that can allocate tasks on the best virtual
machines and reduce task execution time, Akbari and machines. To investigate how efficient is their solution, a
Rashidi [42] proposed a Multi-Objective Scheduling fitness function model is developed. The experimental

123
Neural Computing and Applications (2020) 32:14817–14838 14821

results as put forward by the researchers show their pro- techniques. Although the metaheuristics are promising than
posed algorithm can achieve minimum execution time, the heuristic techniques, their improvements using a tra-
waiting time and missed tasks compared to First Come jectory-based technique like the simulated annealing as
First Served (FCFS), Shortest Process Next (SPN) and well as the greedy-based techniques like the orthogonal
Highest Response Ratio Next (HRRN). Taguchi approach can enable the metaheuristics to become
Jena [47] proposed a multi-objective Two-State PSO a potential solution in solving a multi-objective task
(TSPSO) algorithm. Their aim is to reduce the energy scheduling problem in cloud computing environment.
consumption and makespan at the cloud datacenters. In Therefore, this paper addresses the concern of customers
their scheduling method, selection of best virtual machine service selection strategy using an improved metaheuristic
is introduced using the nondominance strategies. As the algorithm to meet customers’ QoS expectations with a
tasks are scheduled across virtual machines, the two-stage focus on multi-objective task scheduling.
PSO generates two fitness functions and are compared to
determine the solution that is nondominant. The nondom-
inant solution is then chosen to represent the optimum 3 Cat swarm optimization
solution of the task scheduling process. According to the
researchers, the simulation results show their proposed Chu and Tsai [32] proposed the CSO technique. The
TSPSO algorithm can minimize the energy and makespan technique mimics the common behavior of natural cat. The
compared to Best Resource Selection (BRS) and Random CSO technique has two modes of operation: resting
Scheduling Algorithm (RSA). In Khajehvand et al. [48], (seeking) and chasing (tracing) mode. The two modes are
the researchers introduced a hybrid First-Fit Cost-Time also referred to as the global and local search. A control
Trade-Off (FCTT) and Workflow Planning Cost-Based factor within CSO known as the mixed ratio (MR) is used
(WPC) model to minimize the runtime and execution cost to determine the current position of the cat. The cat posi-
of scheduled tasks on virtual machines. In their scheduling tion also signifies solution (fitness) set. The velocity of the
process, large task is divided into sub-tasks which are cat is associated with a dimension and a fitness value. As
sorted in nonincreasing manner. In their proposed algo- the cat progresses closer to the solution (fitness), it updates
rithms, a bottom-up traversal technique is later incorpo- itself each time with better results at the memory until all
rated to assign each sub-task a rank. The child sub-tasks are the cats achieve the best solution (fitness) [20, 32, 49]. The
first allocated virtual machines for their execution. The following sections explain the seeking and tracing modes
parent tasks are later executed only when the child tasks [20, 49, 50].
completed their execution on the virtual machines.
According to the researchers, their simulation results show 3.1 Seeking mode
the proposed FCTT can reduce task runtime and execution
cost compared to MOGA and Best Effort (BE) algorithms. The seeking mode is known as the global search process of
However, task updates method exhibited by their proposed the CSO technique [50]. Algorithm 1 shows the pseu-
WPC technique can lead to longer execution time since the docode for the seeking mode [49, 50].

performance of the algorithm depends on its update 3.2 Tracing mode

process.
Form the literature reviewed so far, high complexity, The tracing mode corresponds to the local search process
slow convergence and imbalance between global and local of the CSO. The pseudocode for the tracing mode is shown
search are some of the drawbacks of the metaheuristic in Algorithm 2 [20, 49, 50].

123
14822 Neural Computing and Applications (2020) 32:14817–14838

 
3.3 The need for cat swarm optimization Ln 2n1 ; ð3Þ
improvement
where n  1 represents the number of columns in two-level
To provide efficient scheduling in cloud datacenters with orthogonal array; n = 2k is the number of experiments
the goal of meeting customers’ expectations, the CSO then corresponding to the n rows and columns; 2 represents the
needs to be improved. However, global search of the CSO number of levels required for each factor Z; and k is a
does not always assure superior solutions when the search positive integer ðk [ 1Þ: The matrix in Table 1 shows the
space increases. The number of cats that always move into values in the column are mutually orthogonal. According to
the global search mode of the CSO always exceeds that of Taguchi, for any pair of columns, the combinations of all
the local search. Thus, its convergence toward a stable so- factors at each level occur at an equal number of times. As
lution becomes difficult, leading to its entrapment at the described in [45], to allocate six factors each with two
 
local optima [32]. On the other hand, for each iteration, the levels ‘‘L8 26 ,’’ only six columns are needed for the run of
global and local search modes exhibited by the CSO are the experiment. Hence, L8 ð27 Þ orthogonal seems sufficient,
independently carried out. These also cause its velocity and since there are seven columns. The L8 is an indication that
position update to perform similar process. This can lead to eight experimental runs will be conducted by studying
high computation time during task scheduling on cloud seven variables at two levels. The value ‘‘7’’ represents the
computing environment. In another concern, the CSO can dimension of the problem. The main objective of adopting
only handle a single-objective optimization problem. the Taguchi approach is to find an optimal solution in a
Imbalance between global and local search of the CSO reasonable amount of time [51, 55]. Detail about the
becomes another challenge. Hence, there is a need for Taguchi method can be found in [33].
improving CSO to make it efficient for service provision- Here, the value 1 at each column represents the first set
ing in cloud computing [33]. of factors to be considered for the experiment, while the
value 2 represents the second set of experiments.

4 Taguchi orthogonal array Table 1 L8 ð27 Þ OA [27]

Experiment number Factors
Taguchi method is a greedy approach put forward by
Genichi Taguchi [51]. The Taguchi method uses an A B C D E F G
Column numbers
orthogonal array (OA) matrix representation for its exper-
iment. It involves the study of a large number of design 1 1 1 1 1 1 1 1
variables to achieve efficient results using few numbers of 2 1 1 1 2 2 2 2
simulation runs [52]. According to Taguchi, for any two- 3 1 2 2 1 1 2 2
level orthogonal array (2OA) with Z factors where Z rep- 4 1 2 2 2 2 1 1
resents the number of designed factors, each factor will be 5 2 1 2 1 2 1 2
based on two levels. Taguchi formulated a general symbol 6 2 1 2 2 1 2 1
shown in Eq. (3) for the establishment of an OA with n 7 2 2 1 1 2 2 1
levels of Z factors [53, 54]: 8 2 2 1 2 1 1 2

123
Neural Computing and Applications (2020) 32:14817–14838 14823

4.1 The Taguchi optimization algorithm the cloud customers’ attention can be drawn to the trade-off
points P* known as the Pareto front, where customers are
The Taguchi method can serve as a better approach in allowed to select their service preferences in terms of vir-
reducing the execution time of a task when used for solving tual machines that can provide them with the minimum
task scheduling problem in cloud computing. In cloud task execution time and cost [48, 58]. At this instance, the
scheduling, the total cost of execution is mostly influenced customers are left to optimize their service preferences by
by amount of time tasks are executed on a virtual machine. selecting the best trade-offs [57, 58]. The main goal of this
Hence, incorporation of Taguchi optimization algorithm research is to make sure that the Pareto optimal set is
into a conventional CSO can be a potential solution to discovered in a minimum amount of time for all the tasks
achieve the desired results. The pseudocode for the Tagu- scheduled on virtual machines using our proposed tech-
chi optimization algorithm is shown in Algorithm 3 nique. This study is based on the following definitions in
[54, 55]. solving the multi-objective task scheduling problem.

5 Pareto dominance strategy Definition 1.1 Multi-objective Optimization problem.

A typical multi-objective optimization problem can be
For any task scheduling, the chances of locating an optimal expressed as a minimization of a K components of a vector
solution that can meet customers’ expectations in terms of function fi in the form [15, 49]:
minimum execution time and cost are becoming harder in a
!   ! ! !  
large search space like the cloud computing environment Min f ! x ¼ f1 x ; f2 x ; f3 x ; . . .; fk !
x ; ð4Þ
[56]. Due to multi-criteria requirements associated with
where ! x 8!
x ¼ fx1 ; . . .;g represents the vector of decision
customers’ request, the concept of optimality needs to be
variable such that fi : Rn ! R; i ¼ f1; 2; . . .; kg are the
achieved. Multi-objective optimization approach is a ! 
potential solution to solving this kind of problem. It is objective functions in a universe U. f ! x is the multi-
characterized with trade-off factors, where each trade-off objective function.
solution corresponds to a specific order of importance of Definition 1.2 Pareto Dominance
the objectives [18]. Currently, the Pareto optimization
strategy is the most widely adopted strategy for solving In Pareto dominance, two different participants can be
several multi-objective problems. Individual objectives can judged on how good they are in terms of their perfor-
be combined using Pareto dominance strategy to achieve mances. Given two candidate solutions !
x1 and !x2 from U;
their Pareto front [50]. The Pareto optimal solution that vector x1 is said to dominate vector !
! x2 (denoted as
 ! !
represented best possible among these objectives without x1 Þ  ðx2 Þ if and only if:
worsening another objective is chosen as the best candidate
!  
solution [24, 39, 53]. fi ð x1 Þ  fi !
x2 ; 8i 2 f1; . . .; kg ð5Þ
Task scheduling to meet the expectation of customers
! !
involves dealing with multi-objective problems. The Pareto fi ð x1 Þ \fi x2 ; 9i 2 f1; . . .; kg ð6Þ
optimization approach can be adopted to provide customers
as many nondominant solutions as possible by allowing a Equation (5) shows that !x1 dominates !x2 in all objec-
!
tives, while Eq. (6) shows x1 dominates for at least one of
set of trade-offs in terms of execution time and execution
cost [37, 57]. Although study shows that the actual cus- the objectives [7].
tomers’ service preferences are quite difficult to predict, Definition 1.3 Pareto Optimal

123
14824 Neural Computing and Applications (2020) 32:14817–14838

If for all means ! x1 is not dominated by any other forwarding to the scheduler. The Pareto generator then
!
solution, x1 is then considered to be the nondominant generates a set of trade-offs according to the customer’s
(Pareto optimal) solution. requirement which are presented to the customer to select
his/or her choice. The customer opts for his/or her choice of
Definition 1.4 Pareto Optimal set service preference (i.e., the best virtual machine), and the
Set of all solutions !
x1 2 U correspond to the Pareto process continues with the scheduler. The DMOOTC task
optimal. scheduling algorithm then judges whether available
resource has met requirement of the task in terms of time and
Definition 1.5 Pareto Front (P*) cost. Then, the proposed DMOOTC algorithm dispatches
The fitness value of the solution is called trade-off or P* the customer’s task on the chosen virtual resources. The
if and only if Eq. (7) holds. This is the optimal solution of a main contribution of this work lies in that the proposed
multi-objective optimization problem that is comprised of a DMOOTC task scheduling algorithm can address uncer-
set of solutions: tainties by allowing customers to realize better performance
!   to cost and time ratio in cloud computing environment.
f ð x1 Þ  f ! !2U
x2 ; :9x 1 ð7Þ
6.1 Problem description
where : is true if !
x1 2 U; 9 means there exist !
x1 2 U; 2
represents an element of U; and U is the universal set.
The problem is first represented by using a set of inde-
The framework that describes the scheduling process
pendent tasks waiting to be scheduled on sequence of
using the developed scheme is discussed in the following.
heterogeneous virtual machines. V ¼ fvk m  k  1g is a set
of virtual machines, where m is the number of virtual
machines. T ¼ fti n  i  1g represents the tasks’ groups,
6 The system scheme
and n is the overall number of tasks [37]. Our goal is to
dynamically assign each task ti 8i ¼ f1; 2; . . .; ng as cus-
Cloud computing consists of several datacenters that are
tomers’ requests on appropriate virtual machines vk 8k ¼
usually managed by the cloud service providers. For any
f1; 2; . . .; mg in order to determine the timing and execu-
cloud, virtual machines are dynamically created and
tion cost of the tasks. We assume the following in our
deployed in datacenters based on task availabilities. These
scheduling problem: (1) Two datacenters are used for the
virtual machines are heterogeneous in nature, having dif-
tasks scheduling; (2) the datacenters are said to belong to
ferent characteristics in terms of memory and sizes. Our
the same service provider; (3) tasks are assigned to virtual
assumption is that one datacenter is not sufficiently enough
machines dynamically where the total number of all pos-
to handle our task scheduling problem. Therefore, two dat-
sible schedules is ðn!Þm for the problem with n number of
acenters each with 20 virtual machines are sufficiently
tasks and m number of virtual machines; (4) preemptive
enough for our task scheduling problem. Our proposed
scheduling allocation is not allowed; and (5) the cost of
system scheme is illustrated in Fig. 1. The scheme integrates
using a virtual machine for a time quantum varies from one
Pareto optimization strategy for generating a set of trade-
to the other. By adopting the Expected Time to Compute
offs in finding the best schedule that can minimize the
(ETCÞ matrix as shown in Eq. (8), our goal is to dynami-
execution time and cost. Three modules (customers, Pareto
cally assign each virtual machine vk with the right com-
generator and scheduler) are defined in the proposed
puting capacity to appropriate customers’ request in order
scheme, where each of the modules consists of sub-modules
to find the optimum value of the total execution time and
that carry out the scheduling process. The proposed
the total execution cost [34, 53]:
scheduling scheme adopts the DMOOTC algorithm to make 2 3
the scheduling decisions. The global and local resource t 1 v1 t 1 v2 : : : t 1 vk
managers within the scheme work together with the sched- 6 t 2 v1 : : : : : 7
6 7
uler to achieve near-optimal solution. In the scheduling ETC ¼ 6 : 6 : : : : : 7 ð8Þ
7
process, customers submit their request based on certain 6 : : : : : : 7
6 7
6 t n v1 : : : : t n vk 7
resource requirements. The task manager is responsible for
estimating the amount of resources to execute the cus-
tomer’s requests. On the other hand, the local resource
6.2 The proposed multi-objective task
manager is responsible for monitoring and managing local
scheduling model
virtual nodes and to obtain information about processing
elements and memory information and bandwidth which are
Our proposed multi-objective time and cost model is for-
later submitted to the global manager for subsequent
mulated from the problem description. The model reflects

123
Neural Computing and Applications (2020) 32:14817–14838 14825

Fig. 1 System model

the relationship between time and cost [14]. A combined 6.2.1 Execution time model
method put forward in [14, 15, 25] was used in the for-
mulation of the multi-objective time–cost model. In our Let T ¼ fti n  i  1g denote the set of tasks and V ¼
assumption for the formulation of the time and cost model, fvk m  k  1g the set of heterogeneous virtual machines.
all virtual machines are said to belong to the same service Assuming ti 8i ¼ f1; 2; . . .; ng is to be scheduled on
provider, ignoring the cost of data transfer [39]. vk 8k ¼ f1; 2; . . .; mg, the execution time execk of all tasks

123
14826 Neural Computing and Applications (2020) 32:14817–14838

processed on a virtual machine is computed using Eq. (9) The multi-objective task scheduling mathematical
[15, 25]. The total execution time Texek of all tasks ti processed model can be expressed as follows:
on all virtual machines vk is computed using Eq. 10 [25]: Min F ð X Þ ¼ ½Texeck ð X Þ; TTexecos tk ð X Þ ð14Þ
X n
ti vk
execk ¼ xik  ; ð9Þ subject to
i¼1
npe k  vkmips X
xik ¼ 1; 8i ¼ 1; 2; . . .; n; xik 2 f0; 1g; 8i; k
m X
X n
ti vk k¼1
Texeck ¼ xik  ð10Þ
npek  vkmips
k¼1 i¼1 Equation (14) is the proposed multi-objective opti-
where execk is the execution time of running tasks on one mization time–cost model that captured customers’ QoS
virtual machine; xik is equal to 1 if task ti is assigned to a requirement.
virtual machine, otherwise xik ¼ 0; ti is the task whose
length is given in million instructions (MIs); vkmips is the
virtual machine speed whose unit is given in million 7 The multi-objective scheduling method
instructions per second (MIPS); and npek is the number of based on the proposed DMOOTC scheme
processing elements of a virtual machine.
The proposed DMOOTC scheme consists of two phases
6.2.2 Execution cost model (global and local search) that are combined to solve the
multi-objective task scheduling optimization problem. The
The proposed cost model is a multi-objective task following attributes were considered to arrive at an optimal
scheduling model that captures the customers’ QoS solution: the tasks number, the number of virtual machines
requirement. The model permits charging a customer based and other relevant parameters such as count dimension to
on the amount of time the virtual machine spent executing change (CDC). Each cat symbolizes the choice of virtual
his/or her request [35]. The time quantum [37] of a virtual machine used for the task schedule. This is encoded in
machine is the smallest discrete unit use by the service [1 9 n] vector, with n belonging to a number of tasks. We
providers to define the cost of a virtual machine in either per also assume that each virtual machine in a datacenter has
second or on hourly basis. In this study, we assume the cost different cost per time quantum.
of memory and central processing unit (CPU) are all Based on expected time to compute (ETC), when tasks
included in the monetary cost of a virtual machine [25]. For are scheduled on a virtual machine by our proposed
instance, assume for every one-minute N of using a virtual DMOOTC algorithm, it uses two-level orthogonal array
machine, the price specified by the service provider is given Ln ð2n1 Þ8n  N þ 1; where N represents task number.
as 0.5 dollars per hour. For time quantum in minutes of Each task is assigned to a cat (also known as the virtual
using a virtual machine, the execution cost can be computed machine). Each of the cats has a dimension D, and the models
as N0:5 associated with each cats are based on two objective func-
60 dollars [37]. Assuming the cost vkcost of executing
tions: the total execution time ðTexeck ð X ÞÞ and total execution
tasks on a virtual machine per hour (/h) is known, Eq. (11)
cost ðTTexecostk ðXÞÞ. When a cat traverses all tasks, it formed
holds for the execution cost execostk of tasks ti on a virtual
a feasible solution. Each cat has both position and the velocity
machine per time quantum in second [25, 37]:
vector. The position of the cat symbolizes the solution attained
1 Xn
ti vk by the cat. A mixed ratio (MR) is a control factor that is used to
execostk ¼  vkcost  xik  ð11Þ
3600 i¼1
npek  Vmipsk specify two groups of cats. The cats are moved into either
seeking or tracing mode at random using the value of the MR.
where vkcost is the monetary cost of a virtual machine per When the cat reaches its desired targets, its fitness value is
time quantum in US dollar per hour: computed based on the defined objective function (Texeck and
1; if ti is assign on a virtual machine: TTexecostk Þ. This process of assigning tasks to virtual machine
xik ¼ ð12Þ mimics the process of the orthogonal approach. As the
0; otherwise
velocity of cat points the cat toward achieving near-optimal
When more than one vk 8k ¼ f1; 2; . . .; m } is used by a ~set1 ðtÞ and
solution, two sets of candidate velocity vectors V
service provider to execute many tasks, the total tasks k;d

~set2 ðtÞ are generated as follows:

V
execution cost TTexecostk by all virtual machine in a data- k;d

center in a datacenter can be computed using Eq. (13): (  

~set1 ðtÞ ¼ w1  V
V ~k;d ðt  1Þ þ ðc1 r1  X~gbsetd ðt  1ÞX
~k;d ðt  1Þ
k;d
 
1 Xm X n
ti vk ~set2 ðtÞ ¼ w2  V
V ~k;d ðt  1Þ þ ðc1 r1  X~ibsetd ðt  1ÞX
~k;d ðt  1Þ
TTexecostk ¼  vkcost  xik  ð13Þ k;d

3600 k¼1 i¼1

npek  Vmipsk ð15Þ

123
Neural Computing and Applications (2020) 32:14817–14838 14827

such that: ~k;d represents the position of the cat; and t; is the number
X
( of iterations. To update the velocity, the velocity among
! ~set1 ðt  1Þ; if orthogonal array element is ‘‘1’’
V
Vo k;d ðtÞ ¼ k;d
the two velocity sets with the best optimum solution is
~set2 ðt  1Þ; otherwise
V k;d
selected using the condition in Eq. (17):
ð16Þ ( h i
max v; if ~k;d ðt  1Þ þ !
V Vo k;d ðt  1 Þ [ maximum velocity;
! ~k;d ðtÞ
V
where Vo k;d ðtÞ represents two candidate velocity sets; d is ~k;d ðt  1Þ þ !
V Vo k;d ðt  1Þ otherwise
dimension of the solution space; X ~gbsetd represents the ð17Þ
global best position attained by the cat; X ~lbsetd represents
where maxv is the maximum velocity attained by the cat;
the local best position of the cat; w1 ; w2 are the controlled
~k;d represents the velocity attained by the cat; and
V
factors; r1 represents uniform random number in the range
!
of [0, 1]; c1 represents a constant value of the acceleration; Vo k;d ðtÞ represents the two candidate solutions. A domi-
nant strategy is used to compare the optimum solution and

Fig. 2 Flowchart of the DMOOTC scheduling algorithm

123
14828 Neural Computing and Applications (2020) 32:14817–14838

is stored at the archive where the final velocity that should   Xm ! n ! o
formulate the latest velocity is selected. This velocity ~ ¼
QoS X Wj fj Xi ; 8Xi  Archieve ð19Þ
j¼1
returns optimal solution which is used to compute the new
position of the cat as indicated in Eq. (18): where m is the number of objective functions and Wj is the
!
~k;d ðtÞ ¼ X
X ~k;d ðt  1Þ þ V
~k;d ðt  1Þ: ð18Þ preference weight for every objective function ðfj Xi Þ.

The quality of solution is evaluated using a fitness Algorithm 4 provides the pseudocode for the developed
function. Every cat is assessed based on the value of the DMOOTC task scheduling algorithm, while Fig. 2 illus-
  trates the flowchart of the scheduling algorithm.
~ in Eq. (19):
fitness function QoS X

123
Neural Computing and Applications (2020) 32:14817–14838 14829

Table 2 Computer and utility

Windows specifications Software
software specifications
Intel CoreTM i5-5200U CPU @ 3.60 GHz 3.60 GHz Eclipse-java-luna-SR2-win32-x86-64
Window 10 (64-bit), x64-based processor CloudSim 3.0.3 [41, 54, 55]
4 GB DDR3L RAM
1000 GB(1 TB) SATA-3G HDD
Intel CoreTM i5-5200U CPU @ 3.60 GHz 3.60 GHz

Table 3 Configured properties

Parameter Values
for the cloud datacenter
Datacenter No. of datacenters 2
No. of hosts in a datacenter 1
Host RAM 2 GB
Storage ITB
Bandwidths 10 GB/s
Accumulated host processing power 1,000,000 MIPS
Tasks No. of tasks [20–100]
Lengths [100, 1000] Mis
File size [200, 400] MB
Output size [300]
Number of CPUs [1–4]
Virtual machine Virtual machine ID [1–20]
Virtual machines monitor Xen
Accumulated RAM 0.5 GB
Accumulated storage 10 GB
Bandwidth 1 GB/s
Virtual machine processing power 1000–10000 MIPS
Number of processing elements 1 to 4
Virtual machine policy Time-shared
Cost per unit of virtual machine 0.17$ - 1.25$/hour
Cost of using memory 0.05$/hour

8 Simulation environment the datacenters, tasks and virtual machines, while Table 4
indicates the parameter settings for the scheduling
Table 2 shows specifications of the computer system and algorithms.
utility software that we used in the simulation of our pro-
posed DMOOTC scheme. The proposed DMOOTC
scheme is benchmarked against Multi-Objective Particle 9 Performance metrics
Swarm Optimization (MOPSO) [15], Multi-Objective Ant
Colony Optimization (MOACO) [24] and Min–Min [40] This study considers four performance metrics to evaluate
task scheduling schemes. The selection of properties for the the efficiency of the developed scheme which are discussed
datacenter host, task and virtual machines is used as in in the following:
[2, 15, 40]. The estimated cost of a unit virtual machine
over a time quantum is adopted as used in [24]. This 9.1 Execution time
estimated cost comprises both the computing cost and
memory cost and varies from one to the other depending on The time a task spent executing on a computing resource
the capacity. The values for the inertia weight and coeffi- (i.e., virtual machine) is significant to cloud customers. To
cient factors ðc1 c2 Þ for MOACO, MOPSO and DMOOTC measure the performance of our proposed DMOOTC
were specified as used in [59]. Table 3 shows properties of scheme, the model in Eq. (10) is used.

123
14830 Neural Computing and Applications (2020) 32:14817–14838

Table 4 Parameter settings for the scheduling schemes 9.4 Quality of Service
Algorithm Parameter Value
The Quality of Service (QoS) represents the fitness of the
MOPSO Particle size 100 proposed DMOOTC scheduling scheme based on any
Self-recognition coefficients ðc1 c2 Þ 2.0 combined objective factors. It is used to reveal the quality
Uniform random number ðR1 Þ [0,1] of standard provided to the customers. The developed
Maximum iteration 1000 algorithm is designed to achieve better QoS [40]. When the
Inertia weight ðW Þ 90–40% execution time and execution cost values are small, QoS
Mixed ratio 2% value must be higher. In this study, Eq. (19) is adopted in
Cat size 100 the evaluation of the QoS of the four scheduling schemes
DMOOTC Count dimension to change 5% [15, 60].
Cat size 100
Self-recognition coefficients ðc1 Þ 2.0
Uniform random number ðR1 Þ [0,1] 10 Simulation results and discussion
Maximum iteration 1000
Inertia weight ðW Þ 90–40% Two task scheduling scenarios are considered in our
Mixed ratio 2% experiments. The simulation results are elaborated in these
MOACO Pheromone persistence a 0.3 sections. For the first scenario, task instances from 20 to
Importance of pheromone (c) 1 100 were used with 20 heterogeneous virtual machines,
Importance of resource innate attribute (b) 1 while for the second scenario, High Performance Com-
Pheromone evaporation value (q) 0.3 puting (HPC2N) Net log [61] containing 527, 371 task
Iteration number 1000 instances was used. The properties for the datacenters, host
Number of ant m 100 and virtual machine settings are similar in configuration to
those of the first scenario but vary in terms of the task sizes.

10.1 First scenario

9.2 Execution cost
For the first scenario, ten independent simulation runs were
conducted in revealing the efficiency of our proposed
The cost of a service is that a customer paid for the cloud
DMOOTC scheme. Tables 5, 6 and 7 show the average
services he/she has consumed [14, 40]. This is derived
value of the simulation runs obtained in terms of execution
from the execution cost. In our study, all virtual machines
time and execution cost. Table 5 shows that the Min–Min
adopted are heterogeneous with varying cost specification
scheduling scheme was able to achieve better results in the
per time quantum. Therefore, virtual machines with high
task scheduling interval from 20 to 50. As the scheduling
speed are said to return high execution cost. Our aim is to
intervals increase over time (see 70 to 100 tasks), perfor-
provide customers with better choice of virtual machines
mance of the Min–Min decreases to an unprecedented
that will minimize the execution cost using the model
level. On the other hand, the MOACO scheme thus per-
developed in Eq. (13).
forms better in the task scheduling interval from 20 to 40.
Its performances also degrades with an increase in task
9.3 Performance improvement rate
sizes above 40. Weakness in performance recorded by the
MOACO scheme is probably caused by the traversing
The performance improvement rate (PIR) is computed in
process of the ant colony approach, which usually leads to
percentage using Eq. (20). It helps to investigate the effi-
its entrapment at the local optimal region during the search
ciency of our proposed DMOOTC scheme toward
process. An improvement over MOACO scheme is seen in
achieving better performance compared to the bench-
MOPSO task scheduling scheme. The MOPSO achieves
marked schemes [2, 22]:
better performance in terms of both the execution time and
PIRð%Þ execution cost under the scheduling interval of 50 to 100
Execution timeðother schemeÞ  Execution timðDMOOTCÞÞ compared to MOACO and Min–Min scheduling schemes.
¼  100
Execution time ðother schemeÞ
The performance shown by our proposed DMOOTC
ð20Þ scheduling scheme is quite remarkable compared to Min–
Min, MOACO and MOPSO schemes. Quality-of-Service
expectations of customers are paramount to a service

123
Neural Computing and Applications (2020) 32:14817–14838 14831

Table 5 Results of the execution time and cost for the scheduling schemes
Task Improved Min–Min MOACO MOPSO DMOOTC
Execution Execution cost Execution Execution cost Execution Execution cost Execution Execution cost
time (s) ($)/h time (s) ($)/h time (s) ($)/h time (s) ($)/h

20 21.46 4.72 27.54 6.05 26.86 5.90 25.40 5.47

30 39.16 8.61 56.32 12.39 70.29 15.45 53.58 11.78
40 60.44 13.29 87.56 19.26 92.29 20.59 53.92 11.86
50 66.62 14.65 163.49 35.96 147.28 32.40 129.19 28.42
60 209.82 46.15 243.89 53.65 206.50 45.43 162.77 35.80
70 318.81 70.13 312.78 68.10 287.21 63.18 240.12 52.82
80 505.02 111.10 407.54 89.65 334.99 73.69 312.28 68.70
90 768.78 169.13 674.89 148.47 538.09 118.37 331.62 72.94
100 1096.40 241.21 754.90 166.07 663.34 145.93 454.52 99.98
Improved Min–Min Improved Minimum Job First Algorithm, MOACO Multi-Objective Ant Colony Optimization, MOPSO Multi-Objective
Particle Swarm Optimization, DMOOTC Dynamic Multi-Objective Orthogonal Taguchi-Cat Algorithm, hr hour

Table 6 Comparison of estimated QoS with cost–time weighted Table 7 Comparison of estimated QoS with time–cost weighted
factor = -0.5 factors = -0.5 and - 0.9
Task Improved Min–Min MOACO MOPSO DMOOTC Task Improved Min–Min MOACO MOPSO DMOOTC

20 - 10.72 - 13.79 - 13.43 - 12.52 20 - 19.31 - 24.80 - 24.17 - 22.54

30 - 19.57 - 28.25 - 35.14 - 26.78 30 - 35.24 - 50.78 - 63.26 - 48.22
40 - 30.22 - 44.01 - 46.80 - 26.96 40 - 54.40 - 79.03 - 84.25 - 48.53
50 - 33.31 - 82.56 - 73.64 - 64.59 50 - 59.96 - 147.95 - 132.55 - 116.27
60 - 104.91 - 123.76 - 103.25 - 81.39 60 - 188.85 - 220.71 - 185.86 - 146.27
70 - 159.53 - 159.34 - 143.61 - 120.06 70 - 286.94 - 284.40 - 258.49 - 216.11
80 - 252.52 - 208.84 - 269.09 - 156.13 80 - 450.93 - 371.86 - 246.30 - 281.06
90 - 468.95 - 351.34 - 347.71 - 162.82 90 - 709.96 - 621.31 - 484.29 - 298.47
100 - 548.25 - 394.86 - 331.68 - 212.10 100 - 986.82 - 696.82 - 597.02 - 409.08
Improved Min–Min Improved Minimum Job First Algorithm, Improved Min–Min Improved Minimum Job First Algorithm,
MOACO Multi-Objective Ant Colony Optimization, MOPSO Multi- MOACO Multi-Objective Ant Colony Optimization, MOPSO Multi-
Objective Particle Swarm Optimization, DMOOTC Dynamic Multi- Objective Particle Swarm Optimization, DMOOTC Dynamic Multi-
Objective Orthogonal Taguchi-Cat Algorithm Objective Orthogonal Taguchi-Cat Algorithm

provider for continuous demand of its services. A service

precisely, the total values for the execution time, execution
provider is expected to meet each customer’s expectations
cost and QoS obtained are reported to provide insight on
within the service-level agreement. In Tables 6 and 7,
the performance of the proposed schemes. We further show
results of the QoS based on the four scheduling schemes
how significant our proposed DMOOTC scheme in terms
are shown. The weighted factors - 0.5 and - 0.9 are
of percentage improvement rate as indicated in Table 9. In
introduced to serve as a control parameter for customers’
the overall performance, the proposed DMOOTC
service selection in terms of virtual machine types to
scheduling scheme is able to achieve minimum execution
guarantee minimum execution time and execution cost.
time with 42.87%, 35.47% and 25.49% improvement
Upon successful simulation runs, the total QoS by Min–
compared to that of the Min–Min, MOACO and MOPSO
Min, MOACO, MOPSO, and that of the proposed
task scheduling schemes. In a similar development,
DMOOTC schemes are - 2792.41, - 2497.66, - 2076.19
DMOOTC scheme is also able achieve 38.62%, 35.32%
and - 1586.55. Comparisons show that the proposed
and 25.56% improvement compared to the benchmarked
DMOOTC scheme has returned the highest QoS value of
schemes. Figures 3 and 4 show the trend of the perfor-
-1586.55 in contrast to those obtained by the Min–Min,
mances of the scheduling schemes. From the trend
MOACO and MOPSO scheduling schemes. In Table 8

123
14832 Neural Computing and Applications (2020) 32:14817–14838

Table 8 Total QoS obtained by the scheduling schemes

Improved Min–Min MOACO MOPSO DMOOTC

Total execution time 3086.51 2732.94 2366.85 1763.40

Total execution cost 631.84 599.60 520.94 387.77
Total estimated QoS based on cost and time weights = -0.5 - 1627.98 - 1406.75 - 1364.35 - 863.35
Total estimated QoS based on cost and time weights = - 0.5 and - 0.9 - 2792.41 - 2497.66 - 2076.19 - 1586.55
Improved Min–Min Improved Minimum Job First Algorithm, MOACO Multi-Objective Ant Colony Optimization, MOPSO Multi-Objective
Particle Swarm Optimization, DMOOTC Dynamic Multi-Objective Orthogonal Taguchi-Cat Algorithm

Table 9 Improvement in
Improved Min–Min MOACO MOPSO DMOOTC
percentage
Total execution time 3086.51 2732.94 2366.85 1763.40
PIR (%) over Improved Min–Min 11.455 23.31 42.87
PIR (%) over MOACO 13.39 35.47
PIR (%) over MOPSO 25.49
Total execution cost 631.84 599.60 520.94 387.77
PIR (%) over Improved Min–Min 5.10 17.55 38.62
PIR (%) over MOACO 13.11 35.32
PIR (%) over MOPSO 25.56
Improved Min–Min Improved Minimum Job First Algorithm, MOACO Multi-Objective Ant Colony
Optimization, MOPSO Multi-Objective Particle Swarm Optimization, DMOOTC Dynamic Multi-Objective
Orthogonal Taguchi-Cat Algorithm, PIR(%) Performance Improvement Rate (in percentage)

Fig. 3 Total execution time

Fig. 4 Total execution cost
illustrated in these figures, it has shown that our proposed
DMOOTC scheme has the ability to provide customers optimum solutions of the multi-objective in terms of exe-
with the best services in terms of virtual machines that can cution time and execution cost.
guarantee them with the minimum execution time and
execution cost compared to the benchmarked schemes. 10.2 Second scenario
The achievements displayed by our customers service
selection scheme are as a result of the incorporation of In this scenario, we used large-workload-containing
orthogonal Taguchi at its local search procedure which HPC2N dataset to determine the performances of the four
helps the scheme in traversing all cats (virtual machines), scheduling schemes. The Parallel Workload Archive—
and the use of Pareto dominance strategy provides the best HPC2N is made available by the High-Performance
Computing Center North (HPC2N). It is a setlog with

123
Neural Computing and Applications (2020) 32:14817–14838 14833

Table 10 Results of the execution time and cost for the scheduling schemes
Task Improved Min–Min MOACO MOPSO DMOOTC
Execution Execution cost Execution Execution cost Execution Execution cost Execution Execution cost
time (s) ($)/h time (s) ($)/h time (s) ($)/h time (s) ($)/h

100 486.68 107.07 967.56 212.86 1193.18 262.50 1134.25 249.53

200 1452.91 319.64 1786.67 393.06 2893.53 632.57 2564.98 564.29
300 3857.67 848.68 7865.24 1730.35 6975.66 1534.64 4587.67 1009.28
400 6231.47 1370.92 9078.56 1997.28 10411.32 2290.49 5089.76 1119.74
500 9918.75 2182.12 11436.78 2516.09 11458.76 2520.92 8986.08 1976.93
600 13810.89 3038.39 15451.23 3399.27 12764.87 2808.27 10679.87 2349.57
700 18380.38 4043.68 16895.87 3717.09 14023.48 3085.16 13867.32 3050.81
800 22984.71 5056.63 19417.56 4271.86 18513.75 4073.02 16894.90 3716.87
900 29688.63 6531.48 24980.23 5495.65 23278.90 5121.35 21856.23 4808.37
1000 35424.75 7793.44 29672.78 6528.01 26867.23 5910.79 25785.45 5672.79
Improved Min–Min Improved Minimum Job First Algorithm, MOACO Multi-Objective Ant Colony Optimization, MOPSO Multi-Objective
Particle Swarm Optimization, DMOOTC Dynamic Multi-Objective Orthogonal Taguchi-Cat Algorithm, hr hour

Table 11 Estimated QoS with

Task Improved Min–Min MOACO MOPSO DMOOTC
time–cost weight factor = - 0.5
100 - 243.35 - 483.80 - 596.62 - 567.15
200 - 726.50 - 893.38 - 1446.85 - 1282.56
300 - 1928.95 - 3932.86 - 3488.04 - 2293.97
400 - 3115.92 - 4539.55 - 5205.97 - 2545.03
500 - 4959.67 - 5718.73 - 5729.73 - 4493.31
600 - 6905.86 - 7726.08 - 6382.82 - 5340.26
700 - 9190.75 - 8448.45 - 7012.16 - 6934.08
800 - 11493.05 - 9709.37 - 10183.12 - 8447.96
900 - 14845.22 - 12490.87 - 11640.16 - 10928.78
1000 - 17713.45 - 14837.29 - 13434.43 - 12893.51
Improved Min–Min Improved Minimum Job First Algorithm, MOACO Multi-Objective Ant Colony
Optimization, MOPSO Multi-Objective Particle Swarm Optimization, DMOOTC Dynamic Multi-Objective
Orthogonal Taguchi-Cat Algorithm

Table 12 Estimated QoS with

Task Improved Min–Min MOACO MOPSO DMOOTC
time–cost weight
factors = -0.5 & - 0.9 100 - 438.03 - 870.83 - 1073.90 - 1020.85
200 - 1307.66 - 1608.05 - 2604.27 - 2308.56
300 - 3472.02 - 7078.95 - 6278.31 - 4129.04
400 - 5608.51 - 8170.98 - 9370.50 - 4580.93
500 - 8927.17 - 10293.45 - 10313.23 - 8087.74
600 - 12930.22 - 13906.57 - 11488.77 - 9612.20
700 - 16542.90 - 15206.79 - 12621.56 - 12481.01
800 - 20686.94 - 17476.39 - 16662.94 - 15205.92
900 - 26720.67 - 22482.97 - 20951.72 - 19671.27
1000 - 31883.35 - 26706.40 - 24181.32 - 23207.69
Improved Min–Min Improved Minimum Job First Algorithm, MOACO Multi-Objective Ant Colony
Optimization, MOPSO Multi-Objective Particle Swarm Optimization, DMOOTC Dynamic Multi-Objective
Orthogonal Taguchi-Cat Algorithm

123
14834 Neural Computing and Applications (2020) 32:14817–14838

Table 13 Total QoS obtained by the scheduling schemes

Improved Min–Min MOACO MOPSO DMOOTC

Total execution time 142,236.84 137,552.48 128,380.68 111,446.51

Total execution cost 31,292.05 30,261.52 28,239.71 24,518.18
Total estimated QoS based on cost and time weights = -0.5 - 71,122.72 - 68,780.38 - 65,119.90 - 48,792.35
Total estimated QoS based on cost and time weights = -0.5 and - 0.9 - 128,517.47 - 123,801.38 - 115,546.52 - 100,305.31
Improved Min–Min Improved Minimum Job First Algorithm, MOACO Multi-Objective Ant Colony Optimization, MOPSO Multi-Objective
Particle Swarm Optimization, DMOOTC Dynamic Multi-Objective Orthogonal Taguchi-Cat Algorithm

Table 14 Improvement in
Improved Min–Min MOACO MOPSO DMOOTC
percentage
Total execution time 142,236.84 137,552.48 128,380.68 111,446.51
PIR (%) over Improved Min–Min 3.29 9.74 21.64
PIR (%) over MOACO 6.66 18.97
PIR (%) over MOPSO 13.19
Total execution cost 31,292.05 30,261.52 28,239.71 24,518.18
PIR (%) over Improved Min–Min 3.29 9.75 21.64
PIR (%) over MOACO 6.66 18.97
PIR (%) over MOPSO 13.19
Improved Min–Min Improved Minimum Job First Algorithm, MOACO Multi-Objective Ant Colony
Optimization, MOPSO Multi-Objective Particle Swarm Optimization, DMOOTC Dynamic Multi-Objective
Orthogonal Taguchi-Cat Algorithm, PIR(%) Performance Improvement Rate (in percentage)

information on about 527, 371 tasks. The HPC2N workload

log was freely provided by Ake Sandgren, who also helped
with background information and interpretation, while
Michael Jack assisted in making sure the log is hosted into
the archive for general usage [2]. We stored the workload
in a sac folder at the CloudSim. The datacenter broker
within the CloudSim is configured to make use of the
scheduling schemes as the main scheduler instead of using
the default scheduler. Tasks are pooled from the sac folder
containing the HPC2N dataset for each run of the simula-
tion based on assigned scheduling intervals (100–1000)
using each of the scheme sets for the simulation.
The results after simulations are indicated in Tables 10,
11, 12, 13 and 14. These results attest to the fact that our
proposed DMOOTC scheme can provide minimum exe- Fig. 5 Total execution time
cution time and execution cost compared to improved
Min–Min, MOACO and MOPSO schemes. More precisely, achieved 21.64%, 18.97% and 13.17% improvement
Table 13 summarizes the whole results in terms of total compared to improved Min–Min, MOACO and MOPSO
execution time and cost. The proposed DMOOTC algorithms. The continuous display of performance by our
scheduling scheme is able to achieve - 48792.35 QoS proposed DMOOTC scheduling scheme is attributed to the
value when a customer selects - 0.5 weight factor for his/ incorporation of orthogonal Taguchi approach and the use
or her cloud service and also achieves - 1000305.31 for of Pareto dominance strategy to provide customers with the
time and cost weight factors of - 0.5 and - 0.9. Table 14 choices of virtual machines that help in retuning an opti-
reports a significant improvement gained by our proposed mum solution of the multi-objective problem. Likewise,
DMOOTC scheme over improved Min–Min, MOACO and figures are used to further show the trend in performance of
MOPSO scheduling schemes in terms of execution time the scheduling scheme under different scheduling intervals
and execution cost. The proposed DMOOTC scheme has as shown in Figs. 5 and 6.

123
Neural Computing and Applications (2020) 32:14817–14838 14835

schemes. Tables 15 and 16 present the computed 95%

confidence intervals for both scenarios according to
Eq. (21) [62]:
s
Confidential Interval ðCIÞ ¼ x t pffiffiffi ð21Þ
n
where x represents the mean; t represents t-distribution that
is derived from the t-distribution table; s represents the
standard deviation of the sample data and n represents the
number of samples. The smaller the value of the confidence
interval, the more precise our estimate. For results shown
in Tables 15 and 16, the 95% confidence intervals obtained
by our proposed DMOOTC scheduling scheme are less
compared to those obtained by the benchmarked schemes.
This means that there is a significant difference in the
results obtained by our proposed DMOOTC scheduling
Fig. 6 Total execution cost schemes compared to the benchmarked schemes. It can be
concluded that our proposed DMOOTC scheduling
11 Statistical analysis on 95% confidential scheme can provide cloud customers with better services
interval that will meet their expectations and adapt the elasticity of
cloud computing environment than the benchmarked
A statistical analysis based on 95% confidence interval is scheduling schemes.
provided to show how significant our obtained results for
both scenarios compared to those of the benchmarked

Table 15 Computed 95%

Improved Min–Min MOACO MOPSO DMOOTC
confidence interval for the first
scenario Degree of freedom 9 9 9 9
Confidence level 0.025 0.025 0.025 0.025
t-distribution 2.262 2.267 2.267 2.262
Mean 342.94 303.21 262.98 195.93
Standard deviation 378.38 264.49 218.04 148.25
Lower bound 72.28 114.02 107.01 90.04
Upper bound 613.60 492.40 418.95 254.14
95% confidential interval (72.28, 613.60) (114.02, 492.40) (107.01, 418.95) (90.04, 254.14)

Table 16 Computed 95% confidence interval for the second scenario

Improved Min–Min MOACO MOPSO DMOOTC

Degree of freedom 9 9 9 9
Confidence level 0.025 0.025 0.025 0.025
t-distribution 2.262 2.267 2.267 2.262
Mean 11,868.01 11986.63 11279.27 9517.90
Standard deviation 10,155.57 7995.00 7048.25 6979.87
Lower bound 3378.74 3412.51 3211.13 2709.68
Upper bound 20,357.28 20,560.75 19,347.41 16,326.12
95% confidential interval (3378.74, 20,357.28) (3412.51, 20,560.75) (3211.13, 19,347.41) (2709.68, 16,326.12)

123
14836 Neural Computing and Applications (2020) 32:14817–14838

12 Conclusion References

Scheduling of cloud service for the purpose of meeting the 1. Gui Z, Yang C, Xia J, Huang Q, Liu K, Li Z, Yu M et al (2016) A
service brokering and recommendation mechanism for better
expectations of each customer in cloud computing is a non- selecting cloud services. PLoS ONE 9(8):e105297. https://doi.
deterministic polynomial times (NP-hard) hard problem. org/10.1371/journal.pone.0105297
Solutions required to facilitate the provisioning of better 2. Abdullahi M, Ngadi MA, Abdulhamid SM (2016) Symbiotic
cloud service are rather too complex to develop. In this paper, Organism Search optimization-based task scheduling in cloud
computing environment. Future Gener Comput Syst
we proposed a cloud customers service selection 56(2016):640–650
scheme known as Dynamic Multi-Objective Orthogonal 3. Thanasias V, Lee C, Hanif M, Kim E, Helal S (2016) VM
Taguchi-Cat (DMOOTC) that served as an ideal solution. The capacity-aware scheduling within budget constraints in IaaS
proposed DMOOTC scheduling scheme not only considers clouds. PLoS ONE 11(8):e0160456. https://doi.org/10.1371/jour
nal.pone.0160456
meeting customers’ QoS expectations, but also facilitates the 4. Gabi D (2014) Surveillance on security issues in cloud comput-
provisioning of several service choices for customers to select ing: a view on forensic perspective. Int J Sci Eng Res
their service preference. Two computing scenarios were 5(5):1246–1252
adopted in evaluation of the efficiency of our proposed 5. Gabi D, Ismail AS, Zainal A (2015) Systematic review on
existing load balancing techniques in Cloud Computing. Int J
DMOOTC scheduling scheme via simulation. The simulation Comput Appl 125(9):16–24
results obtained in both scenarios show our proposed 6. Meena M., Bharadi VA (2016). Performance analysis of cloud
DMOOTC scheduling scheme had returned minimum exe- based software as a service (SaaS) model on public and hybrid
cution time and cost for all scheduled tasks and also provided cloud. In: Proceedings of the 2016 symposium on colossal data
analysis and networking (CDAN). 18–19 March. Indore, Madhya
better QoS compared to the benchmarked schemes. We fur- Pradesh, India, pp 1–6
ther revealed the significance of our proposed DMOOTC 7. Tassabehji R, Hackney R (2016) Hey You? Get Off My Cloud:
scheduling scheme using statistical analysis based on 95% evaluation of cloud Service Models for Business Value within
confidence interval. The statistical results obtained by our Pharma X. J Adv Manag Sci Inf Syst 2:48–52
8. Zhang Y, Qian C, Lv J, Liu Y (2017) Agent and cyber-physical
proposed DMOOTC scheduling scheme are quite significant system based self-organizing and self-adaptive intelligent shop-
than those obtained by the benchmarked schemes. The overall floor. IEEE Trans Ind Inf 13(2):737–747
performances displayed by our proposed DMOOTC 9. Furht B (2010) Cloud computing fundamentals. In: Furht B,
scheme is as a result of the incorporation of an orthogonal Escalante A (eds) Handbook of cloud computing. Springer, New
York, pp 3–19
Taguchi strategy at its local search, which facilitates better 10. Raza HM, Adenola FA, Nafarieh A, Robertson W (2015) The
task mapping on virtual machines and the use of Pareto slow adoption of cloud computing and it workforce. Proc Comput
dominance strategy that provides customers with several Sci 52(2015):1114–1119
service choices to select their preference. Further studies are 11. Zhan Z-H, Liu X-F, Gong Y-J, Zhang J, Chung HS-H, Li Y
(2015) Cloud computing resource scheduling and a survey of its
therefore necessary to investigate the scalability of the pro- evolutionary approaches. J ACM Comput Survey 47(4):1–33
posed DMOOTC scheduling scheme using large workloads. 12. Cui H, Liu X, Yu T, Zhang H, Fang Y, Xia Z (2017) Cloud
service scheduling algorithm research and optimization. Hindawi
Acknowledgements Open access funding provided by Umea Secur Commun Netw 2017:1–7
University. The use of Universiti Teknologi Malaysia facilities and its 13. Gabi D, Ismail AS, Zainal A, Zakaria Z (2018) Quality of Service
services is hereby acknowledged with thanks. (QoS) task scheduling algorithm with taguchi orthogonal
approach for cloud computing environment. In: Saeed F, Gazem
N, Patnaik S, Saed Balaid A, Mohammed F (eds) Recent Trends
Compliance with ethical standards in Information and Communication Technology. IRICT 2017.
Lecture notes on data engineering and communications tech-
Conflict of interest There is no conflict of interest for this manuscript. nologies, vol 5. Springer, Cham, pp 641–649
All authors agreed to its submission. 14. Zuo L, Shu L, Dong S, Zhu C, Hara T (2015) A multi-objective
optimization scheduling method based on the ant colony algo-
Open Access This article is licensed under a Creative Commons rithm in cloud computing. IEEE Access J Rapid Open Access
Attribution 4.0 International License, which permits use, sharing, Publ 3:2687–2699
adaptation, distribution and reproduction in any medium or format, as 15. Ramezani F, Lu J, Taheri J, Hussain FK (2013). Task scheduling
long as you give appropriate credit to the original author(s) and the optimization in cloud computing applying multi-objective parti-
source, provide a link to the Creative Commons licence, and indicate cle swarm. LNCS 8274, 2013: Springer, Berlin, pp 237–251
if changes were made. The images or other third party material in this 16. Zhang F, Cao J, Li K, Khan US, Hwang K (2014) Multi-objective
article are included in the article’s Creative Commons licence, unless scheduling of many tasks in cloud platforms. Futur Gener
indicated otherwise in a credit line to the material. If material is not Comput Syst 37(2014):309–320
included in the article’s Creative Commons licence and your intended 17. Xie Z, Shao X, Xin Y (2016) A scheduling algorithm for cloud
use is not permitted by statutory regulation or exceeds the permitted computing system based on the driver of dynamic essential path.
use, you will need to obtain permission directly from the copyright PLoS ONE 11(8):e0159932. https://doi.org/10.1371/journal.pone.
holder. To view a copy of this licence, visit http://creativecommons. 0159932
org/licenses/by/4.0/.

123
Neural Computing and Applications (2020) 32:14817–14838 14837

18. Deb K (2001) Multi-objective optimization using evolutionary process multiple jobs. PLoS ONE 11(7):e0158491. https://doi.
algorithms. Wiley, New York org/10.1371/journal.pone.0158491
19. Subashini G, Bhuvaneswari MC (2011) Non dominant particle 37. Liu J, Pacitti E, Valduriez P, de Oliveira D, Mattoso M (2016)
swarm optimization for scheduling independent tasks on hetero- Multi-objective scheduling of Scientific Workflows in multisite
geneous distributed environments. Int J Adv Soft Comput Appl clouds. Futur Gener Comput Syst 63(2016):76–95
3(1):1–17 38. Duan R, Prodan R, Li X (2014) Multi-objective game theoretic
20. Gabi D, Ismail AS, Zainal Zakaria Z (2017) Solving task scheduling of bag-of-tasks workflows on hybrid clouds. IEEE
scheduling problem in cloud computing environment using Trans Cloud Comput l 2:29–42
orthogonal Taguchi-cat algorithm. Int J Electr Comput Eng 39. Verma A, Kaushal S (2015) Cost-time efficient scheduling plan
(IJECE) 7(3):1489–1497 for executing workflows in the cloud. J Grid Comput
21. Souza Pardo MH, Centurion AM, Franco Eustáquio PS, Carlucci 2015:495–506. https://doi.org/10.1007/s10723-015-9344-9
Santana RH, Bruschi SM, Santana MJ (2016) Evaluating the 40. Liu G, Li J, Xu J (2013) An improved min-min algorithm in
influence of the client behavior in cloud computing. PLoS ONE cloud computing. Adv Intell Syst Comput. https://doi.org/10.
11(7):e0158291. https://doi.org/10.1371/journal.pone.0158291 1007/978-3-642-33030-8_8
22. Abdulhamid SM, Abd Latiff MS, Abdul-Salaam G, Hussain 41. Beegom ASA, Rajasree MS (2014) A particle swarm optimiza-
Madni SH (2016) Secure scientific applications scheduling tion based pareto optimal task scheduling in cloud computing.
technique for cloud computing environment using global league Lect Notes Comput 8795:79–86
championship algorithm. PLoS ONE 11(7):e0158102. https://doi. 42. Akbari M, Rashidi H (2016) A multi-objectives scheduling
org/10.1371/journal.pone.0158102 algorithm based on cuckoo optimization for task allocation
23. Tran D-H, Cheng M-Y, Cao M-T (2015) Hybrid multiple-ob- problem at compile time in heterogeneous systems. Expert Syst
jective artificial bee colony with differential evolution for the Appl 60(2016):234–248
time–cost–quality tradeoff problem. Knowl Based Syst 43. Voicu C, Pop F, Dobre C, Xhafa F (2014) MOMC: multi-ob-
74(2015):176–186 jective and multi-constrained scheduling algorithm of many tasks
24. Lakra AV, Yadav DK (2015) Multi-objective task scheduling in Hadoop. In: Proceedings of the ninth international conference
algorithm for cloud computing throughput optimization. Proc on P2P, parallel, grid, cloud and internet computing (3PGCIC);
Comput Sci J 48(2015):107–113 2014: IEEE
25. Ramezani F, Lu J, Taheri J, Hussain FK (2015) Evolutionary 44. Bilgaiyan S, Sagnika S, Das M (2015) A multi-objective cat
algorithm-based multi-objective task scheduling optimization swarm optimization algorithm for workflow scheduling in cloud
model in cloud environments. World Wide Web 18:1737–1757. computing environment. Int J Soft Comput 10(1):37–45
https://doi.org/10.1007/s11280-015-0335-3 45. Xu Z, Xu X, Zhao X (2015) Task scheduling based on multi-
26. Rubio JJ, Lughofer E, Meda-Campaña JA, Páramo LA, Novoa objective genetic algorithm in cloud computing. J Inf Comput Sci
JF, Pacheco J (2018) Neural network updating via argument 12(4):1429–1438
Kalman filter for modeling of Takagi-Sugeno fuzzy models. 46. Milani FS, Navin AH (2015) Multi-objective task scheduling in
J Intell Fuzzy Syst 35(2):2585–2596 the cloud computing based on the patrice swarm optimization. Int
27. Soares AM, Fernandes BJT, Bastos-Filho CJA (2018) Pyramidal J Inf Technol Comput Sci 7(5):61–66
neural networks with evolved variable receptive fields. Neural 47. Jena RK (2015) Multi-objective task scheduling in cloud envi-
Comput Appl 29(12):1443–1453 ronment using nested PSO framework. Proc Comput Sci J
28. Rubio JJ (2017) USNFIS: uniform stable neuro fuzzy inference 57(2015):1219–1227
system. Neurocomputing 262:57–66 48. Khajehvand V, Pedram H, Zandieh M (2014) Multi-objective and
29. Liu Y, Wang Z, Yuan Y, Alsaadi FE (2018) Partial-nodes-based scalable heuristic algorithm for workflow task scheduling in
state estimation for complex networks with unbounded dis- utility grids. J Optim Ind Eng 14(2014):27–36
tributed delays. IEEE Trans Neural Netw Learn Syst 49. Pradhan PM, Panda G (2012) Solving Multi-objective problems
29(8):3906–3912 using cat swarm optimization. Int J Expert Syst Appl
30. Rubio JJ (2009) SOFMLS: online self-organizing fuzzy modified 39(2012):2956–2964
least-squares network. IEEE Trans Fuzzy Syst 17(6):1296–1309 50. Shojaee R, Faragardi R. H, Alaee S, Yazdani N (2012) A New cat
31. Li X, Li H, Sun B, Wang F (2018) Assessing information security swarm optimization based algorithm for reliability-oriented task
risk for an evolving smart city based on fuzzy and grey FMEA. allocation in distributed systems. In: symposium on sixth inter-
J Intell Fuzzy Syst 34(4):2491–2501 national telecommunications (IST). IEEE
32. Chu S-C, Tsai P-W (2007) Computational intelligence based on 51. Taguchi G, Chowdhury S, Taguchi S (2000) Robust engineering:
the behavior of cats. Int J Innov Comput Inf Control learn how to boost quality while reducing costs and time to
3(2007):163–173 market. McGraw-Hill, New York
33. Gabi D, Ismail AS, Zainal A, Zakaria Z, Abraham A (2016) 52. Tsai P-W, Pan J-S, Chen S-M, Lio B-Y (2012) Enhanced parallel
Orthogonal Taguchi-based cat algorithm for solving task cat swarm optimization based on taguchi method. Expert Syst
scheduling problem in cloud computing. Neural Comput Appl Appl 39(2012):6309–6319
(2016) 30(6):1845–1863 53. Tsai J-T, Fang J-C, Chou J-H (2013) Optimized tasks scheduling
34. Panda SK, Jana PK (2015) A Multi-objective task scheduling and resource allocation on cloud computing environment using
algorithm for heterogeneous multi-cloud environment. In: Pro- improved differential evolution algorithm. Comput Oper Res
ceedings of the international conference on electronic design, 40(2013):3045–3055
computer networks and automated verification (EDCAV), 2015. 54. Chang H-C, Chen Y-P, Liu T-K, Chou J-H (2015) Solving the
IEEE flexible job shop scheduling problem with makespan optimization
35. Pacini E, Mateos C, Garino CG (2015) Balancing throughput and by using a hybrid Taguchi-genetic algorithm. IEEE J Mag
response time in online scientific Clouds via Ant Colony Opti- 3:1740–1754
mization. Adv Eng Softw 84:31–47 55. Tsai J-T, Liu T-K, Ho W-H, Chou J-H (2008) An improved
36. Wei X, Sun B, Cui J, Xu G (2016) A multi-objective com- genetic algorithm for job-shop scheduling problems using
pounded local mobile cloud architecture using priority queues to Taguchi-based crossover. Int J Adv Manuf Technol
2008(38):987–994

123
14838 Neural Computing and Applications (2020) 32:14817–14838

56. Saule C, Giegerich R (2015) Pareto optimization in algebraic 60. Calheiros RN, Ranjan R, Beloglazov A, De Rose CAF, Buyya R
dynamic programming. Algorithms Mol Biol 10(22):1–20 (2010) CloudSim: a toolkit for modeling and simulation of cloud
57. Kalra M, Singh S (2015) A review of metaheuristic scheduling computing environments and evaluation of resource provisioning
techniques in cloud computing. Egypt Inf J 16(3):275–295 algorithms. Softw Pract Exp 41(1):23–50
58. Farahabady HRM, Lee CY, Zomaya YA (2014) Pareto-optimal 61. HPC2N.TheHPC2NSethlog (2016) http://www.cs.huji.ac.il/labs/
cloud bursting. IEEE Trans Parallel Distrib Syst parallel/workload/l_hpc2n/index.html. Accssed 12 Apr 2018
25(10):2670–2682 62. Hosmer DW, Lemeshow S (1992) Confidence interval estimation
59. Eberhart RC, Shi Y (2000) Comparing inertia weights and con- of interaction. Epidemiology 3(5):452–456
striction factors in particle swarm optimization. In: Proceedings
of the IEEE conference on evolutionary computation (ICEC). Publisher’s Note Springer Nature remains neutral with regard to
IEEE jurisdictional claims in published maps and institutional affiliations.

123