SERBIAN JOURNAL OF ELECTRICAL ENGINEERING

Vol. 21, No. 1, February 2024, 53-71

UDC: 004.76:004.4`416]:620.9 DOI: https://doi.org/10.2298/SJEE2401053S

Intelligent and Metaheuristic Task Scheduling for

Cloud using Black Widow Optimization Algorithm
Sadhana Selvakumar1, Pandiarajan Subramanian1
Abstract: Cloud computing is an internet-based infrastructure for services such
as computations, storage, etc., hosted on physical machines. The machines on
cloud infrastructure scales between a few tens to thousands of machines that are
linked in an unstructured way. In cloud computing, minimizing energy
consumption and its associated costs is the primary goal while preserving
efficiency and performance. It progresses the system’s overall productivity,
reliability, and availability. Furthermore, reducing energy use not only lowers
energy expenses but also helps to safeguard our natural environment by lowering
carbon emissions. The objective of our proposed work is to reduce energy usage
in the cloud environment and enhance its performance. We propose a hybrid
approach that incorporates an energy-aware self-governing task scheduler,
namely, Artificial Neural Network (ANN), and a metaheuristic Black Widow
Optimization (BWO) algorithm to solve the optimization issues. Our suggested
task scheduler focuses on minimizing energy consumption, improving the
makespan, and reducing the operating cost while keeping a low number of active
cloud racks. The cloud environment is highly scalable in this scenario since we
adopt a metaheuristic BWO algorithm. CloudSim simulation framework is used
for implementation and experimental analysis.

Keywords: Black Widow Optimization, Energy efficiency, Bio-inspired, Artificial

Neural Networks, Cloud computing.

1 Introduction
The high-level massive applications in cloud computing have recently been
successful in addressing the expansion of computational needs for enormous
application scales since the cloud provides flexible and durable computing
resources available for usage in the pay-per-use model [1]. Numerous tasks are
usually performed “as a service” model on the cloud environment. Software-as-
a-Service (SaaS) collaboration paradigm provides users with cloud applications
over the Internet, which are accessed on a desktop or workstation using dependent
customer applications, for instance, web browser. It is typically used for cloud-

1
Department of Computer Science and Engineering, KIT-KalaignarKarunanidhi Institute of Technology,
Coimbatore, India; E-mails: sadhana.dsk@gmail.com; evergreenpandi@gmail.com

53
S. Selvakumar, P. Subramanian

based service applications such as webmail, social networking, video-sharing,

and text document editing [2].
Platform-as-a-Service (PaaS) offers a suitable environment for creating
applications to upgrade, test, and accommodate their applications. Moreover, it
offers to the distribution of large-scale applications to scalable and elastic
computing systems. End users can search for the apt use by virtualized computer
equipment known as Virtual Machines (VM) with an advanced CPU processor,
storage zone, memory, and bandwidth, in accordance with Infrastructure-as-a-
Service (IaaS). Numerous VM scenarios are offered to consumers at different
prices, which assist users in managing computer resources at the end of the day.
IaaS offers end-users three key benefits. The primary benefit is that consumers
use cloud resources or services based on their requirements and employ a pay-
per-use model for infrastructure, like how essential services, such as electricity,
gas, and water, might be paid. Based on the needs of its application, users can
utilize this to contract or raise their resource agreement. The second advantage of
IaaS cloud computing is the direct supply of resources that enhances user
application production. The third benefit is that users can acquire rental resources
at any time and from any location, based on the desired service level. The
requirement of resources is still a well-known problem in conducting large-scale
jobs on the IaaS cloud [3].
As demand for high-performance computer resources increases, the Cloud
Data Center’s (CDC) energy usage rate also increases [4]. The energy required
by computing resources and the corresponding cooling system contributes to high
energy and carbon emissions. The data centers use roughly 1.4% of the world’s
electricity usage, which grows at a rate of 12.5% annually [5]. Power
consumption by processing units, cooling facilities and storage facilities, is
42.0%, 15.4%, and 19.3% respectively, and is predicted to increase 12% per
annum. It generates up to 2% of the world’s carbon emissions and is forecast to
increase rapidly [6]. In addition, electricity expenditures represent about 15% of
data center costs [7]. It is a prominent problem to overcome in terms of energy
efficiency. The energy-saving hardware technologies and efficient CDC
management activities can reduce electricity-based costs and carbon emissions
[8]. Poor management of computing resources can lead to excessive energy
consumption and a shorter lifespan.
It is essential to look at the challenges of task planning to reduce energy
usage while enhancing performance (Makespan). There are challenges found in
the task scheduling strategies and resource allocation in the cloud. Metaheuristic
optimization algorithms are employed for solving task scheduling problems in
the cloud. Metaheuristic methods have led to favourable outcomes in the search
region by limiting the search solution area in terms of efficient performance [9].
Several research utilized bio-inspired optimization algorithms [10  13] such as

54
 Intelligent and Metaheuristic Task Scheduling for Cloud using Black Widow…

Genetic Algorithm (GA), Hybrid Particle Swarm Optimization (HPSO),

Intelligent Water Drops (IWD), and Cuckoo Search Algorithm. However, these
meta-heuristic approaches might be still problematic: 1) tendency to decrease
overall performance as the complexity of problems increases; 2) more time
consumed on the same problem instance; 3) the status of the cloud environment
is not taken into consideration, unable to cope with the dynamics in the cloud
environment. Moreover, these techniques might be affected by early convergence
(being stuck in local search), an uneven balance of searching strategies between
local and global, and their exploitation is not as stable as their search capability.
Those concerns are addressed using an optimal Artificial Neural Network [14].
In this study, we suggested an independent task scheduler using ANN to
lessen energy and execution overhead to improve system performance. The
randomly created independent tasks are scheduled in a heterogeneous
environment. The rest of the research work is structured as follows. The related
works based on the usage of optimization algorithms in cloud task scheduling are
reviewed in Section 2. The cloud environment system model and problem
formulation are explained in Sections 3 and 4, respectively. Section 5 presents
the proposed ANN-BWO algorithm and Section 6 discusses the experiment
setup, the results, and their performance analysis. Finally, the paper ends with the
conclusion of the work.

2 Literature Review
Data centers enormous power, and several researchers have recently worked
on cloud task scheduling problems with different optimization targets such as cost
saving, implementation time, dependability, consumption of energy, and so on [9
 11]. Algorithms such as MaxMin, MinMin, and Optimized makespan can
produce solutions with low overhead performance but are not sufficiently
efficient [15]. Meta-heuristic algorithms also build a cost and time efficient
schedule and provide sub-optimal results.
In addition, many earlier studies apply Machine Learning (ML) algorithms
[16  17] to resolve numerous challenges in resource management, such as ANN
[14], and Deep Reinforcement Learning (DRL). The DRL automates the process
of establishing optimum VM settings depending on the present application
requirements. Researchers likewise focus on the automation of the resource
assignment procedure on application demand scales. The applications are not
allocated with the existing computing resources by RL. Instead, they have
dynamically scaled the previously allocated resources, like the processor
capability and memory, according to the application size.
A new method based on the modified PSO algorithm is suggested in [18] to
address task scheduling problems, local optimum and premature convergence
problems. The results demonstrated that the proposed modified PSO can reduce
55
S. Selvakumar, P. Subramanian

the total cost compared to other suggested methods. Garg and Goraya [19]
proposed an energy-efficient heuristic algorithm for task scheduling in a
virtualized cloud with a deadline constraint. This approach comprises two
instances to enhance energy efficiency by executing the maximum workload of
the VM and saving the most energy in the VM’s idle state. Yassa et al. [20]
resolved the problem of scheduling tasks using a variety of computing schemes.
They presented a novel multi-objective optimization method for cloud problems
by introducing the hybrid PSO algorithm for determining the performance of
optimal scheduling.
Safari and Khorsand [21] offered a novel technique for minimizing energy
usage in the job distribution of multiple VMs on heterogeneous cloud systems.
To preserve service quality, they integrated a power-aware and list-based
scheduling technique with a novel method under time constraints. As a result, the
purpose of task scheduling is to enhance performance and energy usage.
Similarly, the authors in [22] researched energy-efficient task scheduling in a
Cloud Data Center environment. Aside from DVFS energy management
techniques and VM reusage, the authors in [21] developed and implemented the
EATS-FFD algorithm, serving as the primary task-scheduling strategy. The goal
is to reduce energy consumption while ensuring QoS.
The research works in [23  25] used ML strategies for estimating the usage
of resources, with a metaheuristic approach, which is consistent for predictions
and scheduling jobs. However, an energy-efficient scheduling decision directly
utilizes a recursive neural network model by them. The authors in [26] presented
a directed search optimization intending to minimize makespan without taking
high energy usage. They iteratively trained an ANN, like any other metaheuristic
method with faster convergence. Yang et al. [27] have utilized ML techniques to
improve the efficiency of large-scale systems. It tries to improve the Quality of
Service (QoS), handle cloud resource requests, and consolidate workloads.

3 System Model
A modern CDC usually includes a big room with multiple rows of racks and
numerous computer resources and supporting appliances. Each rack has a
Cluster-Level switch connected to the other racks, and each computer inside it
has the next set of computer resources connected inside the rack. The other
Information Technology (IT) equipment, including air conditioners, power
distribution components, temperature sensors, and air handling units assists the
primary computing resources. Fig. 1 displays the current CDC, showing how the
computing resources of physical hosts may be grouped into rack unities together
and how each computing resource can represent a Virtual Machine (VM), which
assigns and executes real jobs.

56
 Intelligent and Metaheuristic Task Scheduling for Cloud using Black Widow…

Many users submit several separate tasks, which are then arranged for the
amount of VM resources available. We may observe that job assignments for
certain VM sets schedule tasks automatically through diverse racks. Rack’s
conscious planning adds further drive for energy efficiency and the use of
resources. Rack-conscious scheduling decides the apt rack for a specific activity
or collection of tasks. It can be an active and lightly loaded rack instead of turning
on a completely fresh one because the rack and associated gear, like a fan set, air
blower, interconnection bays, etc., save a considerable computing and cooling
energy. Once the rack is selected for specific tasks, the tasks are scheduled using
the task scheduler. This scheduling method spreads the scheduler over the whole
data center.

Fig. 1 – System Model for the Proposed Work.

4 Problem Formulation
In this section, the problem formulation for the energy consumption and time
estimation for execution are described.
4.1 Estimation of execution time
One of our goals is to enhance the performance, i.e., the time required to
accomplish certain activities using computer resources. To estimate the time
required for a virtual computing resource V0m,n to execute k th task on n th physical
machine of mth rack and calculate it as (1):
Tk
ETkm ,n ,o  , (1)
Com , n

57
S. Selvakumar, P. Subramanian

where, Tk is the working load of the k th task and Com,n refers to the capacity of
virtual computing resource. We presume that each virtual computer resource
works on the host computing resource space shared. So, we are summing up the
time to do tasks allocated to certain virtual computer resources, which are defined
in (2) to calculate the time to finish a virtual computer resource:
FTom , n   k TS m ,n ETkm , n ,o , (2)
o

where FTom,n represents the Virtual Computing Resource and TSom , n is the Virtual
Computing Resource Task Set. The completion time of the high-performance
server machine or multiple computer resource VM ( Pnm ) for all planned activities
is determined as stated in (3): As virtual computing resources function
independently,
FTnm max
oPm
FTom,n , (3)
n

where FTnm denotes a completed time of the Pnm resource for computing.
Likewise, a rack takes time to complete each scheduled work, as set out in (4):
The maximum time among the computing resources for completion:
FTm max m
nRm FTn , (4)

where FTm is the maximum time taken by mth rack to do the allocated tasks.
4.2 Energy consumption
Our subsequent goal is to lower the energy consumption in data centers while
preserving performance. As a result, to carry out specific tasks, we must estimate
the energy consumed using computer resources. Energy consumption in cloud
data centers is controlled primarily by resources such as electricity consumed by
the processing units, storage, and cooling. We use an energy consumption based
on linear regression on the links between CPU utilization and energy
consumption usage that works well even when Dynamic Voltage and Frequency
Scaling (DVFS) is used to simulate energy consumption in the cloud. DVFS is a
technique that adapts input voltage and frequency to present working loads, as its
name suggests. The function of CPU use is electricity consumption. Using this
info, we define Pt m,n energy consumption as follows:
ECnm  Pt m,n  ETm , (5)
where the ECnm table of electricity for the execution of a job during the execution
period is the Pnm calculating resource ETm , and Pt m,n is the current time energy
spent by the CPU for the use of the Pnm computing resource.

58
 Intelligent and Metaheuristic Task Scheduling for Cloud using Black Widow…

ECm   nRm ECnm . (6)

Using the above information, we can compute the total energy consumed by
a data center as follows:
EC   mCDC ECm . (7)
In this study, we have taken the challenge of job planning in a cloud setting
to reduce energy consumption and increase application performance. We aim to
lower the overhead performance of the algorithm at the same time. We plan
individual tasks on a set of computer resources that are accessible and arranged
together into a rack unit. Each computing resource in the CDC can have many
virtual resources to perform tasks autonomously. Each of them has millions of
instructions that can be expressed as a workload. In addition, processing
capability is known in advance together with the energy ingesting rate of each
computing resource. At the cluster and rack levels, every new task queue is large
enough to accommodate all incoming jobs.

5 Proposed Methodology
In this section, we propose a hybrid approach by combining the ability of the
ANN scheduler, which is energy-efficient [14] and the metaheuristic optimization
algorithm called BWO [28] within the CDC to decide on the scheduling
agreement to the present cloud environment.
5.1 Artificial neural network scheduler
In [14], ANN and its applications were briefly reviewed by the authors. ANN
is enthused by human brain functioning and the way information is processed.
The brain consists of thousands of cells known as neurons, formed with the help
of axons, are connected again with many cells. External stimuli are taken as input
by dendrites and generate electrical impulses through the brain network. Each
neuron determines whether or not to transmit the message to another network
neuron. Likewise, ANN is a human brain digital model. It has several nodes in
the human brain akin to actual neurons. These (neurons) nodes are linked to and
can interact with each other via communication linkages (dendrites, axons). A
node collects the input data and then executes straightforward operations of the
data. Each communiqué link is weighted, and they determine which communica-
tions to send and which not. Fig. 2 is a rudimentary digital neural network with a
two-node input, a four-node hidden, and a two-node output (of two neurons).
By adjusting these weights, ANN can learn sophisticated correlations of
effects in the external environment. In our situation, ANN’s external environment
is the task burden and cloud environment from which ANN gets input (X) and
creates bidding details and scheduling choices as outputs via ANN. The ANN
learning process is commonly known as the training process. Two forms of

59
S. Selvakumar, P. Subramanian

training are available, 1) The input data is supervised along with the wanted
outputs, which are fed into ANN, and 2) the input is not supervised. The data is
sent to ANN with a certain power function. The supervised training method is
usually utilized by ANN, making ANN excellent for tasks like prediction and
classification. Therefore, we adopt a scaled gradient back propagation approach
with tan and sígma to train the ANN that the research academy generally uses for
supervised training.

Fig. 2 – Simple Architecture of ANN.

The basic principle of back-propagation is to use one learning input signal to

change weights on the link in the direction of lesser errors using the scaled
conjugate gradient algorithm. The suggested ANN-based task scheduler can
accurately learn the complicated task load correlations and the instantaneous
status of the cloud environment. It significantly minimizes the overhead
execution because it does not work iteratively. It is adaptive in nature by
considering the present state/dynamics of the cloud as an input to scheduling
decisions. Once an ANN-based task scheduler has been trained by sufficient data
(experience) instances, it can forecast the expected outputs of instances, perhaps
not experienced. Several training instances have been established so that a neural
network can understand the complexity associated with these instances. Thus, a
huge network of 5 cached layers is explored, each layer with 20 nodes (total of
100 cached layer nodes), including an input and output layer.
5.2 Black widow optimization algorithm [28  29]
In general, ANN cannot adjust the amount of input and output nodes after
the training because a new neuron with input or output nodes must be introduced.
However, it needs new training with changes in the number of inputs or output
nodes. Researchers employ ANN with a high level of input that takes future
requirements into account to handle the challenge of scalability. But in our
60
 Intelligent and Metaheuristic Task Scheduling for Cloud using Black Widow…

scenario, we cannot take the same strategy as the cloud environment that quickly
grows over time, which is dynamic in nature. To solve this issue, we use the
bioinspired BWO algorithm in our proposed ANN classifier, which is described
as follows. Black widow spiders have unique mating behaviour, including a stage
called cannibalism. In this stage, spiders with unsuitable fitness are ignored from
the circle, leading to rapid convergence and resulting in preventing local optima.
It is realized that BWO can obtain the best global solution by maintaining a
balance between exploration and exploitation. Hence, the BWO algorithm is
considered for this research to solve optimization problems in task scheduling.
The following subsections discuss the different stages of the BWO algorithm.
a) Initial population
The values of the difficulty must be defined as an acceptable construction to
solve the current problem to solve an optimization problem. This construction is
known as “chromosome” or “particle position” in black widow meta-heuristic
optimization language, and “widow” in the NBW method. Black Widow Spiders
are included in the algorithm as potential solutions to any problem. Variable
problems are represented by individual black widow spiders. A test function array
is what this structure is in this article. An initial spider population is used to
produce an entrant widow matrix of size N pop N var, and then perform the
optimization process. To begin the mating process, parent couples are randomly
selected.
b) Procreate
A new generation is born as the sets of each other mate. Each pair mates
individually from the rest of their network simultaneously. Each time they mated,
they lay roughly 1,000 eggs, but only a few offspring make it. To duplicate the
random numbers in the widow's matrix, an Alpha matrix must be produced in this
algorithm. In the following equation (8), x 1 and x 2 are parents’ offspring and
are used to construct them.
 y1  a  x1  (1  a )  x2 ,
 (8)
 y2  (1  a )  x1  a  x2 .
There is no need to duplicate randomly selected integers because this
procedure is repeated N var/2 times. Ultimately, children and mothers join the
ranks and are ranked by their fitness scores. Some of the best people have been
admitted to the newly formed population because of their cannibal rating. All
couples are affected by these rules.
c) Cannibalism
There are three distinct sorts of male predators represented in this picture. To
begin with, a black widow is eating her spouse during or after sexual relations.
We could distinguish women and men based on their fitness levels using
61
S. Selvakumar, P. Subramanian

algorithms. Strong spiders eat their pathetic allies in the second type of
cannibalism. Cannibalism Score (CR) determines the number of survivors in
these algorithms. A third type of ogre is observed occasionally, which feeds on
its mother in the form of little spiders. To distinguish between weak and powerful
spiders, we employ fitness values.
d) Mutation
Randomly chosen Mute Pop members from the population are selected at
this time. Two fundamentals that can be changed in the array are illustrated in
Fig.3 for each solution. Based on the mutation rates, the mute population is
designed.

Fig. 3 – Mutation in BWO.

e) Convergence
There are three possible outcomes for this algorithm's conclusion: (a) a
predetermined number of repeats; (b) For the vast majority of delegates, the best
compliance grade for widows remains unchanged; (c) Definition of precision. As
illustrated in the picture, the key phases of NBW are depicted. Some of the
challenges with optimizing the NBW test utilization are discussed in the next
section. A certain level of precision is considered while evaluating an
experimental algorithm because the best solutions for testing activities are already
well-known. As a result, Section 4 defines the maximum number of repetitions
as stop conditions for the trials.
f) Parameter setting
To get the most out of the proposed plan, several conditions must be met. It
includes the rate of purchases (PP), the rate of consumption (CR), and the rate of
conversion (PM). The settings need to be tweaked for the algorithm to locate the
optimal answers. To optimize the proposed plan, conditions such as enhancing
parameter management, improving navigation on any local User Management
(UM) platform, and improving global location capabilities must be met. Since the
precise number of parameters can control the balance between the absorption
phase and inspection phase, it is possible. PP, CR, and PM all play a role in the
proposed algorithm's control scheme. The ownership ratio (PP) governs how
many people are involved in a company and its products. This parameter provides
a wide range of options for defining a study location because it regulates the
development of distinct offspring. Unqualified people are excluded from the CR
cannibal operator’s population control parameter. The Prime Minister represents
62
 Intelligent and Metaheuristic Task Scheduling for Cloud using Black Widow…

the proportion of voters who have shifted their allegiance. These parameters can
be used to achieve a balance between the use and the search of these parameters.
In the local phase, agents can be controlled by this factor, leading to better
resolution.
The purpose of the ANN scheduler is to program incoming tasks in the rack
with a view to improve performance and reduce energy usage. The functionality
focuses on locating computer resources for the specific activity, so it takes less
time to complete the operation and consumes less energy. The determination of
the best computer resource is based on the usage of ANN correlating with BWO,
namely, the ANN-BWO algorithm.
5.3 ANN-BWO scheduler for task scheduling
In the proposed model, the ANN-based task scheduler predicts the
scheduling of tasks onto the appropriate physical machine of a specified rack
based on the recursive neural network where the weights of it are optimized based
on the given Black Widow Optimization algorithm. The entire procedure is given
in the form of pseudocode in Algorithm 1.

Algorithm 1: ANN-BWO for task scheduling in Cloud environment

Input: Tasks {T}, # Virtual Machine instances, Execution time calculation function,
# input neurons, # Hidden layers, # output neurons
Begin
for each 𝑖 = 1 𝑡𝑜 |𝑛𝑒𝑢𝑟𝑜𝑛𝑠|𝐼𝐿 do
𝑁𝑖 = 𝑇𝑗,𝑖 | 𝑗 = 1 𝑡𝑜 |𝑇|
end for

// Initialize the weight factors and computation of ANN for every axon
for each 𝑖 = 1 𝑡𝑜 |𝑛𝑒𝑢𝑟𝑜𝑛𝑠|𝐼𝐿 do
𝑊𝑖 = 𝑟𝑎𝑛𝑑(0,1)
𝑊𝑁𝑖 = 𝑊𝑖 ∗ 𝑁𝑖
end for

for each 𝑖 = 1 𝑡𝑜 |𝑛𝑒𝑢𝑟𝑜𝑛𝑠|𝐻𝐿 do

for each j= 1 𝑡𝑜 |𝑛𝑒𝑢𝑟𝑜𝑛𝑠|𝐻𝐿 do
𝑊𝑖,𝑗 = 𝑟𝑎𝑛𝑑(0,1)
𝑊𝑘 𝑁𝑖,𝑗 = 𝑊𝑖,𝑗 ∗ 𝑁𝑖 // in sequential order for every subsequent neurons | k
= |HL|
end for
end for
for each 𝑖 = 1 𝑡𝑜 |𝑛𝑒𝑢𝑟𝑜𝑛𝑠|𝑂𝐿 do
𝑊𝑖 = 𝑟𝑎𝑛𝑑(0,1)
𝑊𝑜 𝑁𝑖 = 𝑊𝑖 ∗ 𝑊𝑘 𝑁

63
S. Selvakumar, P. Subramanian

end for
//Error rate computation

for each 𝑖 = 1 𝑡𝑜 |𝑛𝑒𝑢𝑟𝑜𝑛𝑠|𝑂𝐿 do

𝑆𝐹𝑖 = 𝑆𝑜𝑓𝑡𝑚𝑎𝑥(𝑊𝑜 𝑁𝑖 )
𝐸𝑅𝑖 = 𝐴𝑏𝑠(𝐴𝑂𝑖 − 𝑆𝐹𝑖 )
end for
// Weight Optimization using BWO
⃗⃗⃗ // A weights as a single vector
// Input: 𝑊

While (t<=Max) do
𝑌1 , 𝑌2 = Procreate(𝑊𝑖 , 𝑊𝑗 )
𝑌′ = Cannibalism (𝑌1 , 𝑌2 )
𝑌 ′′ = Mutate(𝑌′)
If (𝑓(𝑌 ′′ ) < 𝑓(𝑊𝑖 )) then
𝑊𝑖 = 𝑌′′
Elseif (𝑓(𝑌 ′′ ) < 𝑓(𝑊𝑗 )) then
𝑊𝑗 = 𝑌′′
End if
End while
Repeat for all Tasks.

Output: Scheduled Tasks

6 Experimental Results and Discussion

The experimental simulation environment is implemented using CloudSim
simulator [30]. Table 1 shows the simulation environment details. The experi-
mental analysis of the recommended research is evaluated by performing a
comparative study on various optimization algorithms such as the Hybrid Particle
Swarm Optimization (HPSO), Ant Colony Optimization (ACO), and Intelligent
Water Drops (IWD) algorithm. The user submits task requests to the CloudSim
by varying the number of tasks in a sequence. The total number of instructions to
be performed, file size, length, and more are associated with each cloudlet. Those
cloudlets are later sent to a broker to map the VMs you want to use. The host
creates VMs that are created in the broker. In a data center broker, the task
scheduling algorithm will work. The VMs and data center schedulers can be
shared in time or space. We evaluate the experimental results regarding
Makespan, Energy consumption, and Cost evaluation with variations in the total
number of tasks and total data size.

64
 Intelligent and Metaheuristic Task Scheduling for Cloud using Black Widow…

Table 1
Parameter Settings for CloudSim Simulation.
Type Parameter Values
Number of tasks 100*100
Length of tasks [4000-10000] MI
Tasks
(Cloudlets) Input/output file size [256-1536] MB
Number of processing
1- 4
element (PE) requirement
Number of VMs 50
MIPS or PE 250-2000
Virtual machines
VM memory RAM 512-2048 (MB)
(VM)
Bandwidth 500-1000bits
Type of Manager Time-shared
Number of Data centers 10
Number of hosts 5-10
Data centers
Type of Data centers Heterogenous
Type of Manager Space_shared_time_shared

6.2 Makespan evaluation

It is aimed primarily at reducing the amount of time required, i.e., the whole
duration of the scheduled work, where all work has been completed, or the time
taken for it from the beginning to the end. The following equation demonstrates
the makeup formula.
n
Makespan   Fti , (9)
i 1

where Fti is the finishing time of the task t ( n ) .

The performance of ANN-BWO is correlated with existing optimization
techniques in terms of Makespan (time in seconds), which is presented in Table 2.
The performance evaluation of the Makespan (complete task execution time)
is discussed in this section. It evaluated 100 to 1000 tasks with different arrival
rates such as 20, 40, and 80. The results of the proposed technique ANN-BWO
are compared with HPSO, ACO, and IWD. The CloudSim simulation results are
shown in Table 2. Fig. 4 illustrates the Makespan with an arrival rate of 20. When
the total number of tasks is 700, the Makespan of HPSO, ACO, IWD, and
proposed ANN-BWO are 497, 474, 468, and 399 seconds, respectively. It is noted
that the proposed ANN-BWO outperforms other optimization algorithms
regarding Makespan.

65
S. Selvakumar, P. Subramanian

Table 2
Comparison of Algorithms with Makespan.
Makespan Number of Tasks
(S) / Algorithms 100 200 300 400 500 600 700 800 900 1000
HPSO 227 265 384 436 453 492 497 506 511 516
ACO 236 259 379 434 430 470 474 486 498 511
IWD 215 247 335 387 387 454 468 487 491 524
ANN-BWO 205 220 303 377 372 397 399 448 467 494

600

500

400

300

200

100

0
100 200 300 400 500 600 700 800 900 1000

HPSO ACO IWD ANN-BWO

Fig. 4 – Makespan vs Number of tasks.

6.3 Energy consumption evaluation

In the Table 3, the performance of the proposed ANN-BWO approach is
compared with existing techniques in terms of energy for different numbers of
tasks is presented.
In addition, the performance analysis of energy consumption with tasks
varying between 100-1000 is presented. We can conclude from the experiment
results that the higher the number, the larger the energy consumption. Table 3
and Fig. 5 compare the energy usage of the selected algorithms in this experiment.
The findings show that the energy consumption of the ACO algorithm is
inadequate when compared with other studied algorithms because half of the
scheduling decisions for energy are needed, and half are required to reduce the
map. This experiment also demonstrates that our suggested ANN-BWO
optimization technique saves energy in the fully loaded as well as moderately
loaded cloud environment. The experiment results show that the connections
between task load, processing capacity, and energy consumption rates of
computer resources can be understood using our trained network.
66
 Intelligent and Metaheuristic Task Scheduling for Cloud using Black Widow…

Table 3
Comparison of algorithms with energy consumption.
Energy Consumption Number of Tasks
(S) / Algorithms 100 200 300 400 500 600 700 800 900 1000
HPSO 72 114 132 143 159 166 174 197 228 273
ACO 76 119 144 147 164 174 186 202 233 258
IWD 74 103 129 139 152 159 172 199 226 247
ANN-BWO 70 102 124 136 148 158 167 194 217 239

In addition, the performance analysis of energy consumption with tasks

varying between 100-1000 is presented. We can conclude from the experiment
results that the higher the number, the larger the energy consumption. Table 3
and Fig. 5 compare the energy usage of the selected algorithms in this experiment.
The findings show that the energy consumption of the ACO algorithm is
inadequate when compared with other studied algorithms because half of the
scheduling decisions for energy are needed, and half are required to reduce the
map. This experiment also demonstrates that our suggested ANN-BWO
optimization technique saves energy in the fully loaded as well as moderately
loaded cloud environment. The experiment results show that the connections
between task load, processing capacity, and energy consumption rates of
computer resources can be understood using our trained network.

300
Energy Consumption (KJ)

250

200

150

100

50

0
100 200 300 400 500 600 700 800 900 1000
Number of Tasks

HPSO ACO IWD ANN-BWO

Fig. 5 – Energy consumption vs Number of tasks.

67
S. Selvakumar, P. Subramanian

6.4 Cost evaluation

The cost evaluation analysis is presented in this section. The execution
overhead is another essential indicator that can be used to measure the
performance of the task scheduling. The total size of data processed by the
workflow ranges from 256 to 1536 MB. By differing the total data size, the
variance in the total cost of execution against the distribution of workload on
resources is compared for the various algorithms depicted in Table 4 and Fig. 6,
respectively. The total cost of resource computation is assumed in the range 1-
1.5$/hr for the experiments. The total cost obtained values of the proposed ANN-
BWO are compared with HPSO, ACO, and IWD. From the simulation results, it
is evident that the proposed ANN-BWO can complete tasks of varying sizes with
minimal cost.
Table 4
Comparison of Algorithms with Cost Evaluation.

Cost Evaluation Total Data Size (MB)

($) / Algorithms 256 512 768 1024 1280 1536
HPSO 45.26 45.72 59.62 64.56 71.74 78.92
ACO 38.32 46.41 62.83 73.69 85.95 98.2
IWD 36.77 45.07 57.89 67.69 78.25 88.81
ANN-BWO 34.86 42.56 51.93 63.24 69.03 84.46

120
100
Cost * 103 ($)

80
60
40
20
0
256 512 768 1024 1280 1536
Total Data Size (MB)

HPSO ACO IWD ANN-BWO

Fig. 6 – Cost Evaluation vs Number of Tasks.

68
 Intelligent and Metaheuristic Task Scheduling for Cloud using Black Widow…

7 Conclusion
Cloud computing is a prominent internet paradigm that makes computing a
model of service and creates a pool of virtualized computing resources on
demand. One of the main issues addressed in Cloud Data Centers is energy-
efficient job scheduling. Better Makespan and reduced energy usage indirectly
decrease expenses and enhance cloud performance. The volume of carbon
emissions from tech infrastructure and cloud data centers is also reduced. For the
effective design and deployment of cloud application requirements, adequate
knowledge and understanding of performance and energy consumption is
significant. In this article, we present ANN-BWO, a novel optimization algorithm
for describing energy consumption and performance of cloud-based systems.
Using CloudSim, we performed vast experiments to analyze cloud systems
energy consumption, performance, and execution overhead with various types
and mixes of tasks. We then compared the results of Makespan, energy
consumption and execution overhead for cost estimation with different resource
allocation strategies. The experimental results reveal that the energy consumption
and cost are reduced by approximately 12% and 14%, respectively. We have
investigated the energy cost to achieve the best performance, energy, and cost
balance in cloud computing systems. For future consideration, we anticipate
additionally optimizing the energy consumption and improving the performance
in heterogeneous environments like cloud-fog computing, where the fog node’s
energy consumption and delay description should be examined.

8 References
[1] J. Weinman: The Economics of Pay-per-Use Pricing, IEEE Cloud Computing, Vol. 5, No. 5,
September/October 2018, pp. 99  107.
[2] L. Wu, S. Kumar Garg, R. Buyya: SLA-Based Admission Control for a Software-as-a-Service
Provider in Cloud Computing Environments, Journal of Computer and System Sciences, Vol.
78, No. 5, September 2012, pp. 1280  1299.
[3] M. A. Rodriguez, R. Buyya: A Taxonomy and Survey on Scheduling Algorithms for Scientific
Workflows in IaaS Cloud Computing Environments, Concurrency and Computation: Practice
and Experience, Vol. 29, No. 8, April 2017, p. e4041.
[4] R. Yadav, W. Zhang, K. Li, C. Liu, M. Shafiq, N. Kumar Karn: An Adaptive Heuristic for
Managing Energy Consumption and Overloaded Hosts in a Cloud Data Center, Wireless
Networks, Vol. 26, No. 3, April 2020, pp. 1905  1919.
[5] A. Uchechukwu, K. Li, Y. Shen: Energy Consumption in Cloud Computing Data
Centers, International Journal of Cloud Computing and Services Science (IJ-CLOSER), Vol.
3, No. 3, June 2014, pp. 156  173.
[6] B. Whitehead, D. Andrews, A. Shah, G. Maidment: Assessing the Environmental Impact of
Data Centres Part 1: Background, Energy Use and Metrics, Building and Environment, Vol.
82, December 2014, pp. 151  159.

69
S. Selvakumar, P. Subramanian

[7] A. Greenberg, J. Hamilton, D. A. Maltz, P. Patel: The Cost of a Cloud: Research Problems in
Data Center Networks, ACM SIGCOMM Computer Communication Review, Vol. 39, No. 1,
January 2009, pp. 68  73.
[8] R. Buyya, A. Beloglazov, J. Abawajy: Energy-Efficient Management of Data Center
Resources for Cloud Computing: A Vision, Architectural Elements, and Open
Challenges, arXiv:1006.0308 [cs.DC], June 2010, pp. 1  12.
[9] C.- W. Tsai, J. J. P. C. Rodrigues: Metaheuristic Scheduling for Cloud: A Survey, IEEE
Systems Journal, Vol. 8, No. 1, March 2014, pp. 279  291.
[10] Z. Zhou, J. Chang, Z. Hu, J. Yu, F. Li: A Modified PSO Algorithm for Task Scheduling
Optimization in Cloud Computing, Concurrency and Computation: Practice and
Experience, Vol. 30, No. 24, December 2018, p. e4970.
[11] S. Eswaran, M. Rajakannu: Unevenness Measurement Using the Support Vector Machine and
Dynamic Multiservice Load Balancing with Modified Genetic Algorithm in Cloud-Based
Multimedia System, International Journal of Computer Aided Engineering and Technology,
Vol. 10, No. 6, October 2018, pp. 732  747.
[12] S. Eswaran, M. Rajakannu: Efficient Multimedia Content Storage and Allocation in
Multidimensional Cloud Computing Resources, International Journal of Intelligent Systems
Technologies and Applications, Vol. 18, No. 1-2, February 2019, pp. 20  33.
[13] S. Eswaran, D. Dominic, J. Natarajan, P. B. Honnavalli: Augmented Intelligent Water Drops
Optimisation Model for Virtual Machine Placement in Cloud Environment, IET Networks,
Vol. 9, No. 5, September 2020, pp. 215  222.
[14] M. Sharma, R. Garg: An Artificial Neural Network Based Approach for Energy Efficient Task
Scheduling in Cloud Data Centers, Sustainable Computing: Informatics and Systems, Vol. 26,
June 2020, p. 100373.
[15] S. H. H. Madni, M. S. Abd Latiff, M. Abdullahi, S. M. Abdulhamid, M. J. Usman:
Performance Comparison of Heuristic Algorithms for Task Scheduling in IaaS Cloud
Computing Environment, PLOS ONE, Vol. 12, No. 5, May 2017, p. e0176321.
[16] M. Cheng, J. Li, S. Nazarian: DRL-Cloud: Deep Reinforcement Learning-Based Resource
Provisioning and Task Scheduling for Cloud Service Providers, Proceedings of the 23rd Asia
and South Pacific Design Automation Conference (ASP-DAC), Jeju, South Korea, January
2018, pp. 129  134.
[17] G. Rjoub, J. Bentahar: Cloud Task Scheduling Based on Swarm Intelligence and Machine
Learning, Proceedings of the IEEE 5th International Conference on Future Internet of Things
and Cloud (FiCloud), Prague, Czech Republic, August 2017, pp. 272  279.
[18] S. Eswaran, M. Rajakannu: Multiservice Load Balancing with Hybrid Particle Swarm
Optimization in Cloud-Based Multimedia Storage System with QoS Provision, Mobile
Networks and Applications, Vol. 22, No. 4, August 2017, pp. 760  770.
[19] N. Garg, M. S. Goraya: Task Deadline-Aware Energy-Efficient Scheduling Model for a
Virtualized Cloud, Arabian Journal for Science and Engineering, Vol. 43, No. 2, February
2018, pp. 829  841.
[20] S. Yassa, R. Chelouah, H. Kadima, B. Granado: Multi-Objective Approach for Energy-Aware
Workflow Scheduling in Cloud Computing Environments, The Scientific World Journal,
Vol. 2013, November 2013, p. ID 350934.
[21] M. Safari, R. Khorsand: PL-DVFS: Combining Power-Aware List-Based Scheduling
Algorithm with DVFS Technique for Real-Time Tasks in Cloud Computing, The Journal of
Supercomputing, Vol. 74, No. 10, October 2018, pp. 5578  5600.

70
 Intelligent and Metaheuristic Task Scheduling for Cloud using Black Widow…

[22] A. Alahmadi, D. Che, M. Khaleel, M.M. Zhu, P. Ghodous: An Innovative Energy-Aware

Cloud Task Scheduling Framework, Proceedings of the IEEE 8th International Conference on
Cloud Computing, New York, USA, June 2015, pp. 493  500.
[23] J.L. Berral, Í. Goiri, R. Nou, F. Julià, J. Guitart, R. Gavaldà, J. Torres: Towards Energy-Aware
Scheduling in Data Centers Using Machine Learning, Proceedings of the 1st International
Conference on Energy-Efficient Computing and Networking, Passau, Germany, April 2010,
pp. 215  224.
[24] J.L. Berral, R. Gavalda, J. Torres: Adaptive Scheduling on Power-Aware Managed Data-
Centers Using Machine Learning, Proceedings of the IEEE/ACM 12th International
Conference on Grid Computing, Lyon, France, September 2011, pp. 66  73.
[25] B. Tripathy, S. Dash, S.K. Padhy: Dynamic Task Scheduling Using a Directed Neural Network,
Journal of Parallel and Distributed Computing, Vol. 75, January 2015, pp. 101  106.
[26] A. Agarwal, H. Pirkul, V.S. Jacob: Augmented Neural Networks for Task Scheduling, European
Journal of Operational Research, Vol. 151, No. 3, December 2003, pp. 481  502.
[27] R. Yang, X. Ouyang, Y. Chen, P. Townend, J. Xu: Intelligent Resource Scheduling at Scale:
A Machine Learning Perspective, Proceedings of the IEEE Symposium on Service-Oriented
System Engineering (SOSE), Bamberg, Germany, March 2018, pp. 132  141.
[28] V. Hayyolalam, A.A.P. Kazem: Black Widow Optimization Algorithm: A Novel Meta-
Heuristic Approach for Solving Engineering Optimization Problems, Engineering Applications
of Artificial Intelligence, Vol. 87, January 2020, p. 103249.
[29] B. Pourghebleh, V. Hayyolalam: A Comprehensive and Systematic Review of the Load
Balancing Mechanisms in the Internet of Things, Cluster Computing, Vol. 23, No. 2, June
2020, pp. 641  661.
[30] R.N. Calheiros, R. Ranjan, A. Beloglazov, C.A.F. De Rose, R. Buyya: CloudSim: A Toolkit
for Modeling and Simulation of Cloud Computing Environments and Evaluation of Resource
Provisioning Algorithms, Software: Practice and Experience, Vol. 41, No. 1, January 2011,
pp. 23  50.

71