Multi Objective Task Scheduling Using Hybrid Whale Genetic
Multi Objective Task Scheduling Using Hybrid Whale Genetic
Multi Objective Task Scheduling Using Hybrid Whale Genetic
https://doi.org/10.1007/s11277-019-06817-w
Abstract
The system of cloud computing comprises of several servers that are inter-connected in
a datacenter, provisioned dynamically to cater on-demand services through the front-end
interface for the clients. Improvement in virtualization technology has made cloud comput-
ing a viable option for various application services development. Cloud datacenters process
the tasks on the basis of pay as you use manner. Task scheduling is one of the important
research challenges in cloud computing. The formulation of task scheduling probes has
been depicted to be NP-hard hence identifying the solution for a bigger problem is intrac-
table. The dissimilar feature of cloud resources makes task scheduling non-trivial. NP-hard
problem arises due to the dynamic behavior of the dissimilar resources identified in the
cloud computing environment. Task scheduling can be optimized using a meta-heuristic
algorithm. In this paper, we have combined two meta-heuristic techniques, namely Whale
Optimization Algorithm (WOA) and Genetic Algorithm (GA) to devise a new hybridized
algorithm called as Whale Genetic Optimization Algorithm. Our aim is to minimize the
makespan and cost while scheduling the tasks. The simulation is done by using Cloud-
sim toolkit. The results obtained shows significant reduction in the execution time that was
measured in terms of enactment amelioration rate. These results were compared with the
classical WOA and standard GA. The results of the proposed technique provide higher
quality solution for task scheduling.
* Gobalakrishnan Natesan
[email protected]
1
Department of Information Technology, St. Joseph’s College of Engineering, Sathyabama
University, Chennai, Tamil Nadu, India
2
R.M.K College of Engineering and Technology, Chennai, Tamil Nadu, India
13
Vol.:(0123456789)
1888 G. Natesan, A. Chokkalingam
1 Introduction
One of the most trending and the easily recognizable technologies among the academi-
cians, researchers and businessmen is cloud computing as it allows storage of infor-
mation and its utilities and allows it to be used from anywhere [1]. It abides a set of
non-similar processing nodes, virtual systems and services that are provided among the
opponent clients system dynamically on basis of their processing speed, accuracy and
QoS needs. The service provisions of the environment of cloud form three tiers: Soft-
ware as a Service (SaaS), Platform as a Service (PaaS) and Infrastructure as a Service
(IaaS) [2]. These layers stakes administration and monitoring of distinct of the final
stack.
Several industries and forums send their data and services to the cloud everyday due
to its high flexibility and effective structures which enable QoS sanctioned processing
surroundings added with constructive services that can quickly be allowed for use with
optimal effort from management [3]. As several researches concentrate on processes
to provide efficient systems that has high system performance, the active researches
included several attributes like security, storage performance, etc. promoting the system
efficiency and making it avoided many obstacles.
Many achievements have been made in the research of the task scheduling in recent
years. In this research, cloud computing is designed for execution of several applications
of different types at the same time. The virtualization technology and automation manage-
ment in the cloud computing makes the nodes join freely and exit operationally gradually
reducing the cost of data center management. Task scheduling problem is considered as the
toughest problem according to research [4, 5]. Rather than performing the regular routine
of assigning tasks to resources to perform the needed operations, task scheduling works
like a traditional computer OS scheduling. The main aim of a good scheduling algorithm is
to improve the resource utilization rate. In NP-hard problem, the scheduling increases the
efficiency, mapping various tasks of services to a set of resources. Finding solutions to the
NP-hard problem is impractical with known algorithms, which requires an optimum sched-
uler which works with a particular computational speed. In general, NP-hard problems can
be solved by inventory methods, heuristic methods or deterministic method [6].
However, enumeration needs to be used to build an optimum schedules for which
it requires the process of building all scheduler that are possible and the results are
compared one by one in order to select the optimal one which is not applicable for
cloud scheduling due to the large number of tasks. Therefore, the proposed methods
of literature for resolving problem of task scheduling that are either heuristic or meta-
heuristic based [7–9]. The optimal solution can be found by heuristic based methods
basis on some predefined rules and the solution quality was obtained by these methods
on dependence of underlying rules and size of the problem. The solution was hence
acquired by search methods of heuristics are not feasible and are operated at a very high
cost. To solve highly complex problems, the meta-heuristic techniques are applied on
it extensively. These techniques use a set of solutions to tresses the space of solution
unlike the mathematical and heuristic techniques that uses a single candidate solution.
The meta-heuristic algorithm attribute makes them perform better than the heuristic and
mathematical techniques. Some of the well-known techniques for solving the issues of
task scheduling in clouds are Genetic Algorithm (GA) [10], Water drop Algorithm [11],
League championship Algorithm (LCA) [12], Firefly Algorithm [13], Ant Colony Opti-
mization (ACO) [14] and Whale Optimization algorithm (WOA) [15].
13
Multi‑Objective Task Scheduling Using Hybrid Whale Genetic… 1889
In this paper, we have proposed a cloud task scheduling algorithm on WGOA to satisfy
the needs of reducing both time and cost of task on the cloud environment. Hence in need
to check the proposition of the scheduler, we have implemented using cloudsim simulation
environment to execute a set of independent task on data centers with various models of
configuration and cost.
The manuscript is organized as follows: Sect. 2 describes the related work of author’s per-
taining to task scheduling. Section 3 shows the solution framework model of task scheduling
in cloud computing. Problem formulation and proposed hybrid WGOA technique for multi-
objective task scheduling are described in Sects. 4 and 5. Section 6 describes the simulation
results. Finally summary & conclusion of the proposed technique are described in Sect. 7.
2 Related Work
Several researchers have been performed on bag of independent tasks scheduling of applica-
tions on resources that are distributed in cloud computing. In such cases, many algorithms are
used for scheduling and apply a variety of methods and parameters of scheduling [16–19]. In
[20] author presents a multi-objective task scheduling techniques in cloud using resource cost
model. The scheme use improved Ant Colony Optimization (ACO) algorithm to minimize
the budget cost and execution time. Simulation outcome shows that the improved ACO tech-
nique presented is new adept of improving the performance of cloud, decrease the user cost,
increase the utilization of resource and minimize the deadline violation rate.
Thamaraiselvi et al., [21] proposed a framework for scheduling bio-informatics applica-
tion in the cloud environment. The author has used Particle Swarm Optimization (PSO)
algorithm to schedule the bio-informatics tasks to minimize the makespan, complete the
execution time of tasks within the time limit and cost. Self-adaptive learning—PSO based
resource allocation framework in hybrid IaaS cloud environment was proposed by author
[22]. In the hybrid IaaS environment the Cloud Service Provider (CSP) could outsource the
tasks to external CSP since the resources are not enough to accommodate the tasks.
The hybridization of SOS and simulated annealing for task scheduling in cloud environ-
ment was used by the author of [23]. The hybridization algorithm objective is to reduce
the response time, execution time and imbalance degree among VM’s. In [24] author sug-
gested a task scheduling scheme based on Global League Championship optimization to
minimize the makespan and response time in scientific applications.
The authors of [25] proposed Intelligent Water Drops (IWD) algorithm to optimize the
scheduling of workflow on the cloud where it was used and embedded inside the workflow
simulation toolkit and was checked in various environments of cloud with cost models. The
authors of [26] proposed DVFS technique to optimize the energy and cost consumed in
HCS framework by implementation of green service level agreement made between cloud
users. All empty slacks of schedule on processing elements and extended makespan.
Jiachen yang et al. [27] proposed a task scheduling algorithm having game theory
designed for management energy in cloud computing. On basis of game theory algorithm
the reliability of the balanced task is considered. It is implemented using cooperative game
model and game strategy in the task scheduling. The authors of [28] described the pre-
power determination and post-power determination algorithms to perform energy and cost
constrained scheduling on several many core processor of precedence constrained parallel
tasks to be performed on continuous or discrete speed levels by embedding equal speed
method onto our algorithms giving high performance.
13
1890 G. Natesan, A. Chokkalingam
3 Solution Framework
The cloud task scheduling system has three main sections namely the policies module, the
objective function and the scheduler module. The communication and interaction among
the three modules are depicted and implemented in Fig. 1. This is implemented using the
task scheduling technique.
The cloud consumers decide on what task has to be done and come up with a collec-
tion of tasks forming a task pool which all has to be completed on which the scheduling
process has to be applied along with the objective function and then on this data is formed,
the scheduler by making use of the WOA & GA techniques, provides the needed resources
like storage, CPU and memory to the respective servers from the cloud resource pool. The
first module of scheduling the system is the policies module on which contains scheduling
process are applied which are generally set by the cloud providers. These policies of sched-
uling are normally a set of constraints and rules to enable resource allocation and establish
SLA’s for all the applications of cloud. Resource conflicts occurs on a regular basis as the
cloud computing environment provides the illusion of infinite resources but in reality there
are no adequate resources available to satisfy their need/tasks. Hence, this can be settled by
using scheduling policies.
The second module is the objective/fitness function. It is used for determination and
evaluation of the rank and quality of a schedule. It is given for all the tasks and hence a
joint objective function is available for every schedule. It’s main purposes is to test and
make sure ‘n’ tasks are scheduled to ‘m’ cloud resources. So that the goal of minimum
makespan problem is an NP hard problem. The most important component is the sched-
uler which forms the third module of our system and it’s responsible for allocation of
tasks to resources using hybrid WGOA optimized scheduling technique. Initially it col-
lects tasks and needed resource information from the cloud information server and then
secondly it judges whether the resource R j satisfies the requirements of the users task T
i,
if it does then in the end Rj is allocated to T
i.
Consumers or Users
Information
Memory
Server
Cloud
Resource Pool
CPU
Resource Pool
Task Pool
Storage
Resource Pool
Scheduling Policies Scheduler
Cloud
Resource Pool
Objective Function
Fitness Function f(x)
13
Multi‑Objective Task Scheduling Using Hybrid Whale Genetic… 1891
4 Problem Formulation
The Cloud Broker (CB) receives the tasks to be scheduled and it uses the Cloud Infor-
mation Server (CIS) to find the services that are needed to perform the execution of
tasks that are received from the client and then{ scheduling of tasks} is performed on
the services that are discovered. For e.g., tasks T1 , T2 , T3 , … Ty are given to CB at
a particular time interval. The VMs are non-similar having different processing speeds
and memory showing a task { to be executed on } various VMs will end in disparate costs
for execution. Say VMs M1 , M2 , M3 , … Mx are present when CB receives the tasks.
FCFS is used for the execution of tasks to be scheduled present on the VMs. Our main
adage is to perform task scheduling on VMs to achieve higher resource utilization,
minimization of makespan and cost. Therefore expected time to complete (ETC) of the
tasks on which scheduling needs to be performed on each VM is used to make decisions
in scheduling by the proposed method. ETC values are depicted using MIPS ratio of the
task length of VM. ETC values are generally depicted in a matrix format, where the row
represents the no. of tasks to be scheduled and column represent the no. of VM’s that
are available of each matrix.
Every row in ETC matrix determines/depicts the given tasks execution time for each
VM while execution time of each task given on VM represents each column. Because
our aim is to minimize the makespan and cost by identifying the optimal group of tasks
to be performed on VMs. Let Cxy, y = 1, 2, 3 … m , x = 1, 2, 3 … n be consider as execu-
tion time of yth task is executing on the xth VM, where m → no. of VMs and n → no of
tasks. The notations and their descriptions used in our proposed system are represented
in Table 1.
4.1 Performance Metrics
The following metrics are the used to check proposed method performance.
Ty The task y, 1 ≤ y ≤ N
Rx The resource x, 1 ≤ x ≤ K
K, N The amount of resources and tasks
CPUCost (x), StorageCost (x), The cost of CPU, Storage, Memory and Bandwidth of Rx
MemoryCost (x), BandwidthCost (x)
CPUBase The lowest usage of CPU base cost
Storage Case The base cost of storage under the lowest usage
Memory Case The base cost of memory under 1 GB memory
Bandwidth Case The base cost of bandwidth under the lowest usage
Run time cost (x) The duration time of task Ty in resource R
x
CPUTran , StorageTran ,
MemoryTran , BandwidthTran The transmission cost associated with CPU, storage,
memory and bandwidth
13
1892 G. Natesan, A. Chokkalingam
4.1.1 Makespan
Makespan is the time when the execution of the last task is finished. It is one of the
famous metrics for performance of scheduling methods. Lower makespan depicts best
and optimal task scheduling of VMs.
{ }
makespan = minimizeSy =Sch maximumy ∈ Tasks Ty (1)
where Ty denotes the completion time of task ‘y’, the set of all possible schedules are rep-
resented as ‘Sch’ and ‘Tasks’ the set of whole tasks to be scheduled. To make (1) easily
estimable, makespan is expressed in terms of completion time of the virtual machine. The
reason for this assumption is that the time by which the final task is completed. It is equiva-
lent to the time taken to complete its tasks by the last virtual machine. Let completion be
a vector of size nb_VM completion [VM] indicates process finalizing time of VM for the
previously assigned jobs as well as of those already planned for the virtual machine.
Mathematically,
∑ [ ]
completion [VM] = ready_time[VM] + ETC y [VM]
(2)
{y∈Tasks|schedule[y]=VM}
Also, makespan from 1 is as,
{ [ ] }
makespan = maximum completion y |y 𝜖 VM (3)
The EAR is termed as the percentage of amelioration of performance for the technique
y is proposed over the other technique z and is measured using the following equation
( ( ) ( )) 100
EAR(%) = Makespan Cz − Makespan Cy × ( ) (4)
Makespan C y
4.1.3 Cost
Resources and services are of wide range in cloud computing. Say for e.g.: few tasks
require high processing capability resources whereas few other tasks require more stor-
age. Also, different resources have different costs, even the task costs will be vary-
ing. Hence, if we take in the task demand difference for resources, it will reflect in the
relationship among the user budget and resource costs. To talk about this problem, a
resource cost that divides the cost of the resource into parts of memory and CPU can be
proposed.
The CPU cost can be depicted as:
CPU Cost (x) = CPU Base × CPU of VM(x) × Run time cost (x) + CPU Tran cost (5)
where CPU Base
is the cost base i.e. needed for a resource when it is used by the lowest uti-
lization, run time cost is the time that the task is run on the VM and CPU Tran is the cost of
CPU transmission.
MemoryCost (x) = MemoryBase × Memory of VM(x) × Run time cost (x) + MemoryTran cost
(6)
13
Multi‑Objective Task Scheduling Using Hybrid Whale Genetic… 1893
Similarly, MemoryBase is the base cost when 0.5 MB memory used, run time cost is dura-
tion of the task is run on the VM. MemoryTran is the cost of Memory transmission.
The formula of storage cost is defined as,
StorageCost (x) = StorageBase × Storage of VM (x) × Run time cost (x) + StorageTran cost
(7)
StorageBase is the base cost of storage. The cost incurred by the storage transmission is rep-
resented by StorageTran. The run time cost is amount of data is stored during task execution.
The bandwidth cost is defined by the formula as follows:
BandwidthCost (x) = BandwidthBase × Bandwidth of VM (x) × Run time cost (x) + BandwidthTran cost
(8)
The cost incurred by the bandwidth transmission is represented by BandwidthTran. The
runtime cost is amount of bandwidth used during task runtime.BandwidthBase is base cost
of bandwidth.
The depiction of cost function can be represented as follows on basis of the CPU, Band-
width, Storage and Memory.
Cost(x) = CPU cost (x) + Bandwidthcost (x) + Storagecost (x) + Memorycost (x) (9)
n
∑
Total Cost(x) = Cost(x) (10)
x=1
To formulate the objective function, scheduling the tasks are taken into account. The main
intension of the cloud service provider is to maximize the profit, whereas the customers
expect to minimize the execution time and cost cutting. Thus to solve the above problem,
fitness function is calculated to evaluate the quality of feasible solutions. However, for each
optimization problem the fitness function is needed to find near optimal solutions. There-
fore, the fitness value is calculated for our proposed model by using Eq. (11).
( )
Fitness Function xit = 𝜆 ⋅ Makespan + 𝛿 ⋅ Total Cost (11)
( t)
where fitness function is denoted as xi and the values of λ and δ ranges between 1 > λ ≥ 0
and 1 > δ ≥ 0 respectively.
13
1894 G. Natesan, A. Chokkalingam
The best search agent is defined and the positions are updated as the whale gets closer to
the prey. This behavior can be denoted as vector equations given below:
D ⃗ | ⃗ �����������
������ � ⃗ ������⃗ |
dis = ||C ⋅ Xbestpos (t) − Xpos (t)|| (12)
������
⃗ �����������
� ⃗ ⃗ ������⃗ (13)
X pos (t + 1) = Xbestpos (t) − A ⋅ Ddis
where, current iteration is denoted as ‘t’, Coefficient vectors are A and C, best solution
�����������
⃗, X
position vector is X pos is the position vector.
′ ������
⃗
bestpos
To calculate Coefficient vectors A ⃗ and C⃗,
⃗ = 2⃗a ⋅ ⃗r − a⃗
A (14)
⃗ = 2 ⋅ ⃗r
C (15)
Where a⃗ → Linearly decreasing from [2 to 0]. ⃗r → Random vector ranging from [0 to 1].
In this method, the coefficient vector A ⃗ is a arbitrary value in the interval [− a, a], we
decrease the value of a⃗ in subsequent iterations form [2 to 0]. The following Fig. 2 shows
the possible positions of whale from (X, Y) to ( X*, Y*).
13
Multi‑Objective Task Scheduling Using Hybrid Whale Genetic… 1895
Position
The spiral updating position is shown in Fig. 3. In this method, the distance between the
whale and the prey is calculated first and the helix shaped movement of humpback whales
is traced using the given formula.
������
⃗ ������⃗ bl �����������
� ⃗
X pos (t + 1) = Ddis ⋅ e ⋅ cos(2𝜋l) + Xbestpos (t) (16)
| �����������
⃗ |
where Ddis = |Xbestpos (t) − Xpos (t)| and designate the distance of ith whale to the prey. Here,
������
⃗ � ������
⃗
| |
‘b’ → constant for logarithmic spiral and ‘l’ → random no ranging from [− 1 to1].
To elect the finest method, the following probability equation is used,
{
�����������
X � ⃗(t) − A
⃗ ⋅D
������
⃗ if p < 0.5
������
⃗ bestpos dis
X pos (t + 1) = �����������
⃗ (17)
������
⃗
dis
bl
D ⋅ e ⋅ cos(2𝜋l) + X �
(t) if p ≥ 0.5
bestpos
Where ‘p’ is the random number between [0,1], ‘l’ is between [− 1,1] and ‘b’ is a constant
for the spiral shape.
Based on variation of A
⃗ vector the prey can be searched using random variables from [− 1
to 1]. This mechanism emphasizes exploration and allows the WOA algorithm to perform
a global search.
Mathematically,
D ⃗ | ⃗ ������������⃗
������ ������⃗ |
dis = ||C ⋅ Xrandpos (t) − Xpos (t)|| (18)
�����
X ⃗ ������������⃗ ⃗ ������⃗
dis (t + 1) = Xrandpos (t) − A ⋅ Ddis (19)
In the proposed work, genetic algorithm crossover and mutation operations are inte-
grated with WOA. This hybridization is used to produce the near optimal solution for
multi-objective optimization process.
13
1896 G. Natesan, A. Chokkalingam
8.1 Crossover
Crossover is one of the genetic algorithm operators. Cross-overing occurs between pairs of
chromosomes. This operation is a method of replacing a few of the chromosomes in one
worst chromosome by the corresponding chromosome in the other worst chromosome. In
the WO algorithm half of the solution is considered as best case solution and the rest half
is considered as worst case solution. For worst case solution of the WOA, the two point
crossover has been employed in our work.
Consider seven whales, where the first selected whale W1 is assumed to obtain a worst
case solution from the worst population and the second selected whale is W2. Each whale
is coded with an integer array. This integer array is divided into segments. Some of the
integers segments are replaced during the crossover operation. In the below figure seven
tasks are consider and the detailed crossover operation is shown in Fig. 4.
8.2 Mutation
In addition to the two point crossover operation, a mutation process is used usually with
a low probability. A higher probability results in a randomized search. This is done
with the crossover solution. In particular, scramble mutation is deployed here. Scram-
ble mutation is popular with solutions including permutation representations. Scramble
mutation is to perform from the whole chromosome, a part of genes is elected and their
values are jumbled or shuffled randomly as depicted in Fig. 5.
Worst W1,
0 1 1 1 0 1 1 W1 0 1 1 1 0 1 1
Worst W2,
W2 1 0 0 1 1 0 0
1 0 0 1 1 0 0
After two point crossover
Child1 0 1 0 1 0 0 1
Child2 1 0 1 1 1 1 0
After Mutation 0 1 0 0 1 0 1
13
Multi‑Objective Task Scheduling Using Hybrid Whale Genetic… 1897
13
1898 G. Natesan, A. Chokkalingam
The cloudsim toolkit is an event cloud simulation toolkit based on java. It is popularly
named structure for constructing and processing cloud computing architectures and ser-
vices and also enables the processing’s of large cloud computing nodes and environment.
It can supports to schedule data intensive application, computing intensive application
and scientific application. Therefore, to check the proposed task scheduling algorithm, the
cloud simulation tool has been used. It does not support only the checking of scheduling
techniques, but also allows the analysis of several failures that can be caused. So, we used
our WGOA in cloudsim. We verified our experiments by running it on a PC with Intel(R)
Core(TM) i5–457 OS, 4 CPU @ 2.9 GHz, a RAM of 8 GB and a 64-bit windows OS.
The above simulation cases are considered in four sets, where, in each set, the parameters
for the virtual machines are varied—with respect to the number of tasks allotted, number
of virtual machines used, bandwidth values and the completion speed for tasks. The param-
eter setting for standard algorithm experiment is shown in Table 2.
A homogenous configuration implies a collection of systems where the parameters
are fixed for all virtual machines, whereas, a heterogeneous configuration implies varied
parameters for each virtual machine in the network.
SET 1
In the first set, let us assume 5 virtual machines. The number of tasks is 25, with same
task lengths of 15 MI.
1. Homogenous network:
13
Multi‑Objective Task Scheduling Using Hybrid Whale Genetic… 1899
Makespan
300
250
Makespan (s)
200
150
100
50
0
FCFS Min-Min Max-Min WGOA
VM's=10 248.3 119.83 117.25 103.62
Makespan
140
120
Makespan (s)
100
80
60
40
20
0
FCFS Min-Min Max-Min WGOA
VM's = 10 131.19 101.86 101.86 87.13
Fig. 7 Makespan value of heterogeneous environment (MIPS & Bandwidth = 700, 900 & 1100)
algorithm takes the maximum completion time. Other algorithms such as the Min–Min
and Max–Min also obtain closely related results to each other and to that of the proposed
WGOA. The only difference between the homogenous and heterogeneous configurations is
that the ranges for the makespan are less for the heterogeneous configuration.
13
1900 G. Natesan, A. Chokkalingam
2. Heterogeneous network: Here, the values of MIPS and bandwidth are considered to be
700, 900 and 1100 consequently.
SET 2
In the second set, let us consider 10 virtual machines. The experiment has 50 tasks, with
different task lengths and within the range of 10–50 MI.
1. Homogenous network:
Makespan
350
300
250
Makespan (s)
200
150
100
50
0
FCFS Min-Min Max-Min WGOA
VM's =10 320.8 248.3 247.6 189.3
Makespan
300
250
200
Makespan (s)
150
100
50
0
FCFS Min-Min Max-Min WGOA
VM's = 10 293.6 220.6 219.8 167.4
13
Multi‑Objective Task Scheduling Using Hybrid Whale Genetic… 1901
2. Heterogeneous network:
Here, the values of MIPS and bandwidth are considered to be between the ranges of
greater than 500 and less than or equal to 1200.
SET 3
In the third scenario, 10 and 20 virtual machines are considered for this simulation. This
simulation has 75 tasks, with different task lengths for heterogeneous configurations and
equal tasks length for homogeneous configurations. The results obtained have been depicted
in Figs. 10 and 11 for homogeneous and heterogeneous network configurations respectively.
The value of MIPS and the bandwidth are equal here and the task lengths are varied instead.
1. Homogenous network:
2. Heterogeneous network:
Here, all the task lengths are within the range of not more than 20 MI.
Makespan
400
350
Makespan (s)
300
250
200
150
100
50
0
FCFS Min-Min Max-Min WGOA
10 VM's 383.26 339.6 335.1 313.2
20 VM's 326.81 276.3 275.9 246.4
Makespan
400
350
300
Makespan (s)
250
200
150
100
50
0
FCFS Min-Min Max-Min WGOA
VM's = 10 348.4 311.4 311.4 287.3
VM's = 20 297.3 254.1 254.3 238.4
13
1902 G. Natesan, A. Chokkalingam
SET 4
In this set of experiments, 10, 20 and 25 virtual machines are used. This simulation has
100 tasks, with the following five cases:
• All virtual machines are homogenous (MIPS = 1200 and Bandwidth = 1200) and all
tasks have equal length (Task length = 25 MI). The result obtained for this case has
been depicted in Fig. 12.
• All virtual machines are homogenous (MIPS = 1200 and Bandwidth = 1200) and all
tasks are various length (Task length ≤ 25 MI). The result obtained for this case has
been depicted in Fig. 13.
• All virtual machines are heterogeneous (MIPS & Bandwidth range is > 500 to ≤ 1200)
and all tasks are same length (Task length = 25 MI). The result obtained for this case
has been depicted in Fig. 14.
• All virtual machines are heterogeneous (MIPS & Bandwidth range is > 500 to ≤ 1200)
and all tasks are different length (Task length ≤ 25 MI). The result obtained for this case
has been depicted in Fig. 15.
• 50 tasks are of equal length (Task length = 25 MI) and 50 tasks are various length (Task
length ≤ 25 MI). In this case, half of the VMs are homogenous (MIPS = 1200 and Band-
width = 1200) and half of them are heterogeneous (MIPS & Bandwidth range is > 500
Makespan
350
300
Makespan (s)
250
200
150
100
50
0
FCFS Min-Min Max-Min WGOA
VM's = 10 313.7 276.2 274.5 234.8
VM's = 20 255.1 232.8 233.3 203.4
VM's = 25 239.4 214.6 214.6 192.1
Makespan
180
Makespan (s)
150
120
90
60
30
0
FCFS Min-Min Max-Min WGOA
VM's = 10 172.9 142.3 141.8 129.8
VM's = 20 136.1 110.2 111.3 104.6
VM's = 25 112.3 104.01 104.01 101.3
13
Multi‑Objective Task Scheduling Using Hybrid Whale Genetic… 1903
Makespan
300
250
Makespan (s) 200
150
100
50
0
FCFS Min-Min Max-Min WGOA
VM's = 10 277.2 264.1 262.5 231.3
VM's = 20 243.03 229.6 230.1 209.6
VM's = 25 210.4 205.9 206.8 172.3
Makespan
250
Makespan (s)
200
150
100
50
0
FCFS Min-Min Max-Min WGOA
VM's = 10 234.2 167.4 165.3 149.6
VM's = 20 228.3 135.8 134.6 121.7
VM's = 25 218.7 119.1 118.8 101.2
Makespan
300
Makespan (s)
225
150
75
0
FCFS Min-Min Max-Min WGOA
VM's = 10 252.4 183.7 182.6 168.1
VM's = 20 235.6 149.3 147.9 128.3
to ≤ 1200). In this scenario we considered 10 and 20 VMs. The result obtained for this
case has been depicted in Fig. 16.
The proposed Whale Genetic Optimization algorithm is modeled and built to address
the problems of global scheduling of tasks in the cloud computing environment. For
13
1904 G. Natesan, A. Chokkalingam
understand and check the performance of the technique to measure the goodness and effi-
ciency of the WGOA scheme, cost and makespan parameters are kept. Hence, the main
aim of this section is to test and check the performance of proposed method with other
meta-heuristic algorithms such as standard GA and classical WOA. The experiments are
performed continuously up to 25 times and the average is calculated and it is used with
results that are available for comparison. Table 3 represents the setting the parameter of the
selected techniques to perform scheduling. The simulations are then used in four datacent-
ers with 20 hosts. Each host consists, 5 VMs and 10 VMs. If each host has 5 VMs, totally
50 VMS are considering for our simulation. Similarly, each host has 10 VMs the total VMs
are 100 for our experiment. The parameter settings of the cloud simulator are depicted in
Table 4. An assumption of simulation is considered that the tasks are independent and has
no constraint on priority among tasks during non-preemptive and execution. Three various
data sets were considered to check the performance of the proposed method. They are nor-
mal distribution, uniform distribution and HPC2 N [23]. Equal number of tasks of various
sizes namely small, medium and large can be portrayed using uniform distribution. Task of
size medium, few small and large ones can be effectively portrayed using normal distribu-
tion. The performance of distributed systems can be evaluated by HPC2 N set log, which is
one of the standard organized workloads.
1 WOA A [2, 0]
L [− 1, 1]
B 1
R [0, 1]
2 GA Crossover rate 0.4
Mutation rate 0.1
13
Multi‑Objective Task Scheduling Using Hybrid Whale Genetic… 1905
In this section, we present simulation results to explain the efficiency of the proposed
hybrid WGOA task scheduling in terms of makespan. Cloudlet length (task size) was gen-
erated from normal, uniform and HPC2 N distribution. Tasks of varying numbers are cho-
sen as input to portray the scalability of the algorithm. We vary the number of tasks from
100 to 500 and present a comparison between standard GA and classical WOA. In this
simulation, tasks were executed 50 times and an average was calculated. For all three dis-
tributions 50 VMs and 100 VMs were considered for task execution. Figures 17 and 18
shows the performance results of makespan for normal distribution.
Similarly Figs. 19 and 20 depicts the makespan of uniform distribution and Figs. 21
and 22 represent the HPC2 N distribution makespan. From the performance analysis, it
was inferred that the proposed hybrid WGOA optimization method has better capability
to attain near optimal value or gives better makespan for the tested task executions, when
compared with standard GA and classical WOA techniques.
Table 5 show the normal distribution EAR (%) on total makespan of the proposed
hybrid WGOA when compared with the classical WOA and standard GA schedulers. The
proposed hybrid WGOA produces 17.94%, 30.90% makespan improvements when com-
pared to classical WOA and standard GA respectively. A uniform distribution dataset is
tested with our proposed hybrid method with classical WOA, and standard GA. The pro-
posed hybrid WGOA approach makespan is fares 11.51%, 25.83% more in terms of EAR
(%) when compared with standard approaches and the results are shown in the Table 6.
Similarly when HPC2 N dataset is given as test input between our proposed hybrid
techniques and standard approaches, our proposed hybrid WGOA approach resulted in
an improved performance of 7.44% and 22.34%. The obtained results are presented in the
Table 7.
NMakespan of 50 VM's
3000
2500
Makespan (S)
2000
1500
1000
500
0
100 200 300 400 500
GA 587.49 893.12 1241.6 1868.4 2413.4
WOA 481.3 797.9 1114.2 1653.5 2263.7
WGOA 438.42 721.3 1038.7 1219.3 1932.6
Number of Tasks
13
1906 G. Natesan, A. Chokkalingam
1200
Makespan (S)
900
600
300
0
100 200 300 400 500
GA 307.3 467.1 649.4 977.2 1262.2
WOA 251.72 417.30 582.73 864.78 1183.92
WGOA 229.29 377.24 543.24 637.69 1010.75
Number of Tasks
UMakespan of 50 VM's
3000
2500
Makespan (S)
2000
1500
1000
500
0
100 200 300 400 500
GA 528.741 803.808 1117.44 1791.56 2472.06
WOA 433.17 718.11 1002.78 1528.15 2267.33
WGOA 394.578 649.17 934.83 1407.37 1949.34
Number of Tasks
10.2.2 Cost
The second objective cost is simulated and the costs across the three methods are com-
pared. The results of comparison are summarized in the figures provided. In Figs. 23 and
24, the costs of normal distribution for 200 and 400 tasks at varying deadlines are shown.
Costs of uniform distribution for 100 and 300 tasks are depicted in Figs. 25 and 26. Simi-
larly, the costs of HPC2 N of 200 and 500 tasks at different deadlines are shown in Figs. 27
and 28.
13
Multi‑Objective Task Scheduling Using Hybrid Whale Genetic… 1907
1200
Makespan (S)
900
600
300
0
100 200 300 400 500
GA 290.81 442.09 614.59 985.36 1359.63
WOA 238.24 394.96 551.53 840.48 1247.03
WGOA 217.02 357.04 514.16 774.05 1072.14
Number of Tasks
HMakespan of 50 VM's
3500
3000
Makespan (S)
2500
2000
1500
1000
500
0
100 200 300 400 500
GA 939.984 1428.992 1986.56 2439.98 3049.975
WOA 770.08 1276.64 1782.72 2140.825 2676.03125
WGOA 701.472 1154.08 1661.92 2013.335 2516.66875
Number of Tasks
1500
1200
900
600
300
0
100 200 300 400 500
GA 535.79088 814.52544 1132.3392 1390.7886 1738.48575
WOA 438.9456 727.6848 1016.1504 1220.27025 1525.337813
WGOA 399.83904 657.8256 947.2944 1147.60095 1434.501188
Number of Tasks
13
1908 G. Natesan, A. Chokkalingam
For normal distribution we considered 200 and 400 tasks with varied deadlines. From
the simulation, the maximum cost acquired by WGOA varies from 115.92–50.4 and
220.5–95.76. Similarly, for WOA the maximum costs are 128.83–67.67 and 244.77–116.25
and that of GA are 1151.50–77.15 and 287.85–146.59 both of which are higher than the
costs of WGOA.
In uniform distribution, for 100 and 300 tasks on a deadline of 10–100, the maximum
acquired cost of WGOA is in the range of 104.33–45.36 and 187.79–81.65 respectively.
GA WOA WGOA
160
140
120
Cost*10^3 ($)
100
80
60
40
20
0
10 20 30 40 50 60 70 80 90 100
Deadline(s)
13
Multi‑Objective Task Scheduling Using Hybrid Whale Genetic… 1909
GA WOA WGOA
350
300
250
Cost*10^3 ($)
200
150
100
50
0
10 20 30 40 50 60 70 80 90 100
Deadline(s)
GA WOA WGOA
160
140
120
Cost*10^3 ($)
100
80
60
40
20
0
10 20 30 40 50 60 70 80 90 100
Deadline(s)
GA WOA WGOA
300
250
200
Cost*10^3 ($)
150
100
50
0
10 20 30 40 50 60 70 80 90 100
Deadline(s)
13
1910 G. Natesan, A. Chokkalingam
GA WOA WGOA
240
210
180
Cost*10^3 ($)
150
120
90
60
30
0
10 20 30 40 50 60 70 80 90 100
Deadline(s)
GA WOA WGOA
550
500
450
400
Cost*10^3 ($)
350
300
250
200
150
100
50
0
10 20 30 40 50 60 70 80 90 100
Deadline(s)
Likewise, the maximum cost of WOA is in the range of 115.95–60.90 and 208.70–109.62
and that of GA varies from 136.35 to 69.44 and 245.43 to 124.99.
For HPC2 N distribution the cost statistics are quite similar to the above stated
algorithms. For 200 and 500 tasks, the maximum acquired costs are of the scale of
166.92–72.58 and 400.62–174. 18 in the deadline of 10–100, while using WGOA. The
maximum cost for WOA is from 185.51 to 97.44 and 445.23–233.86 and for GA, it lies in
the scale of 218.16–110.10 and 266.64–525.34–266.64
13
Multi‑Objective Task Scheduling Using Hybrid Whale Genetic… 1911
11 Conclusion
In this paper, we propose a new task scheduling algorithm suited for cloud computing envi-
ronment called hybrid WGOA. WGOA is a fusion, or in other words, a hybridization of
two algorithms namely, Genetic Algorithm and Whale Optimization Algorithm. WGOA
ensures the optimization in terms of three conflicting factors, makespan, cost and EAR.
The effectiveness of WGOA has been studied by comparing it with classical WOA and
standard GA and then following it with a simulation of three different datasets (Normal,
Uniform and HPC2 N). By far, WGOA has excelled in performance in all the three distri-
butions. WGOA can produce much better solutions in comparison with the WOA and GA.
From this, it is understood that the proposed method performs scheduling of tasks better
than the other methods. From the simulation, the proposed method has been provides near
optimal solution and significantly reducing the maksepan, cost and EAR. In future work,
we intend to consider other objectives such as energy consumption, security and reliability
in addition to the objectives in this work.
References
1. Buyya, R., Yeo, C. S., Venugopal, S., Broberg, J., & Brandic, I. (2009). Cloud computing and emerg-
ing IT platforms: Vision, hype, and reality for delivering computing as the 5th utility. Future Genera-
tion computer systems, 25(6), 599–616.
2. Zhang, Q., Cheng, L., & Boutaba, R. (2010). Cloud computing: State-of-the-art and research chal-
lenges. Journal of internet services and applications, 1, 7–18.
3. Jennings, B., & Stadler, R. (2015). Resource management in clouds: Survey and research challenges.
Journal of Network and Systems Management, 23(3), 567–619.
4. Mustafa, S., Nazir, B., Hayat, A., & Madani, S. A. (2015). Resource management in cloud computing:
Taxonomy, prospects, and challenges. Computers & Electrical Engineering, 47, 186–203.
5. Kalra, M., & Singh, S. (2015). A review of metaheuristic scheduling techniques in cloud computing.
Egyptian Informatics Journal, 16(3), 275–295.
6. Tsai, J. T., Fang, J. C., & Chou, J. H. (2013). Optimized task scheduling and resource allocation on
cloud computing environment using improved differential evolution algorithm. Computers & Opera-
tions Research, 40(12), 3045–3055.
7. Ali, H. G. E. D. H., Saroit, I. A., & Kotb, A. M. (2017). Grouped tasks scheduling algorithm based on
QoS in cloud computing network. Egyptian Informatics Journal, 18(1), 11–19.
8. Manvi, S. S., & Shyam, G. K. (2014). Resource management for infrastructure as a service (IaaS) in
cloud computing: A survey. Journal of Network and Computer Applications, 41, 424–440.
9. Nzanywayingoma, F., & Yang, Y. (2018). Efficient resource management techniques in cloud com-
puting environment: A review and discussion. International Journal of Computers and Applications,
41(3), 165–188.
10. Gutierrez-Garcia, J. O., & Sim, K. M. (2012). GA-based cloud resource estimation for agent-based
execution of bag-of-tasks applications. Information Systems Frontiers, 14(4), 925–951.
11. Niu, S. H., Ong, S. K., & Nee, A. Y. (2013). An improved intelligent water drops algorithm for solv-
ing multi-objective job shop scheduling. Engineering Applications of Artificial Intelligence, 26(10),
2431–2442.
12. Abdulhamid, S. M., & Latiff, M. S. A. (2014). League championship algorithm based job scheduling
scheme for infrastructure as a service cloud. Preprint arXiv:1410.2208.
13. Karthikeyan, S., Asokan, P., Nickolas, S., & Page, T. (2015). A hybrid discrete firefly algorithm for
solving multi-objective flexible job shop scheduling problems. International Journal of Bio-Inspired
Computation, 7(6), 386–401.
14. Dorigo, M., & Sttzl, T. (2004). Ant colony optimization. Brighton: Bradford Co.
15. Ebrahimi, A., & Khamehchi, E. (2016). Sperm whale algorithm: An effective meta-heuristic algorithm
for production optimization problems. Journal of Natural Gas Science and Engineering, 29, 211–222.
13
1912 G. Natesan, A. Chokkalingam
16. Eswaraprasad, R., & Raja, L. (2017). A review of virtual machine (VM) resource scheduling algo-
rithms in cloud computing environment. Journal of Statistics and Management Systems, 20(4),
703–711.
17. Natesan, G., & Chokkalingam, A. (2017). Opposition learning-based grey wolf optimizer algorithm
for parallel machine scheduling in cloud environment. International Journal of Intelligent Engineering
and Systems, 10(1), 186–195.
18. Pradeep, K., & Jacob, T. P. (2017). CGSA scheduler: A multi-objective-based hybrid approach for task
scheduling in cloud environment. Information Security Journal: Global Perspective, 27(2), 77–91.
19. Ma, T., Chu, Y., Zhao, L., & Ankhbayar, O. (2014). Resource allocation and scheduling in cloud com-
puting: Policy and algorithm. IETE Technical Review, 31(1), 4–16.
20. Zuo, L., Shu, L. E. I., Dong, S., Zhu, C., & Hara, T. (2015). A multi-objective optimization scheduling
method based on the ant colony algorithm in cloud computing. IEEE Access, 3, 2687–2699.
21. Somasundaram, T. S., & Govindarajan, K. (2014). CLOUDRB: A framework for scheduling and man-
aging high-performance computing (HPC) applications in science cloud. Future Generation Computer
Systems, 34, 47–65.
22. Zuo, X., Zhang, G., & Tan, W. (2014). Self-adaptive learning PSO-based deadline constrained task
scheduling for hybrid IaaS cloud. IEEE Transactions on Automation Science and Engineering, 11(2),
564–573.
23. Abdullahi, M., & Ngadi, M. A. (2016). Hybrid symbiotic organisms search optimization algorithm for
scheduling of tasks on cloud computing environment. PLoS ONE, 11(6), e0158229.
24. Latiff, M. S. A., Abdul-Salaam, G., & Madni, S. H. H. (2016). Secure scientific applications schedul-
ing technique for cloud computing environment using global league championship algorithm. PLoS
ONE, 11(7), e0158102.
25. Elsherbiny, S., Eldaydamony, E., Alrahmawy, M., & Reyad, A. E. (2018). An extended intelligent
water drops algorithm for workflow scheduling in cloud computing environment. Egyptian Informatics
Journal, 19(1), 33–55.
26. Kumar, N., & Vidyarthi, D. P. (2017). An energy aware cost effective scheduling framework for hetero-
geneous cluster system. Future Generation computer systems, 71, 73–88.
27. Yang, J., Jiang, B., Lv, Z., & Choo, K. K. R. (2017). A task scheduling algorithm considering game
theory designed for energy management in cloud computing. Future Generation computer systems.
28. Li, K. (2017). Scheduling parallel tasks with energy and time constraints on multiple manycore proces-
sors in a cloud computing environment. Future Generation computer systems.
29. Mirjalili, S., & Lewis, A. (2016). The whale optimization algorithm. Advances in Engineering Soft-
ware, 95, 51–67.
Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.
13
Multi‑Objective Task Scheduling Using Hybrid Whale Genetic… 1913
13