Volume 8, Issue 8, August 2023 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

A Survey on Task Scheduling based on

Various Meta-Heuristics and Machine
Learning Algorithms in Cloud Computing
B. Suganya Dr.R. Padmapriya
Research Scholar Associate professor & HoD-BCA,
RVS College of Arts and Science School of Computer Studies
Coimbatore, Tamilnadu, RVS College of Arts & Science
India. Coimbatore, Tamilnadu, India

Abstract:- The development of cloud computing in computing architecture that primarily consists of a group of
current decades has led to it serving as the basis for a interlinked and VMs that are provided flexibly and offered
variety of systems. It enables customers to access a list of as 1 or as greater than 1 integrated processing facilities
specified resources, act immediately and adaptably to depending on Service-Level Agreements (SLAs) formed via
customer preferences, and only be charged for actual negotiations between the service providers of clouds and
utilization. One of the most important problems in cloud customers” [2]. A large-scale dispersed processing
computing is Task Scheduling (TS). The issue is how to architecture called cloud computing is abstract, virtualized
equitably distribute and organize the user-provided and dynamically operated dependent on the monetary scale
tasks for Virtual Machine (VM) execution. Also, user of the operator. The primary function of cloud computing is
experience is directly impacted by the effectiveness of the management of computer resources, storage, multiple
scheduling efficiency. As a result, the TS issue in cloud platforms and applications that are rented out to outside
computing has to be more precisely addressed. In cloud customers over the internet [3].
computing, the TS is essential such that the optimal
scheduling of task requests may boost network Cloud computing is a quickly developing model for
efficiency. The main objective of TS is to assign tasks to processing that aims to alleviate cloud clients from the
appropriate processors to create the shortest deadline maintenance of hardware, software, networks and
achievable without compromising on priority criteria. information resources, as well as, transfer such obligations
Numerous research has been conducted to design TS to cloud service providers [4]. The essential features of
schemes based on various metaheuristic and machine cloud computing are distribution, virtualization and
learning algorithms that satisfy several criteria such as flexibility. Clouds offer a huge variety of resources, such as
minimization of the makespan, execution cost and computing platforms, data centers, storage, networks,
energy. They have demonstrated that conventional TS is firewalls, and applications delivered as services. In addition,
effective only to satisfy certain criteria and have devised it offers strategies for controlling those services ensuring
an optimum solution using multi-objectives in cloud that cloud clients may utilize them without experiencing any
computing. This paper presents a systematic and performance-related issues. The 3 categories of cloud
extensive analysis of TS algorithms in cloud computing computing services are Infrastructure as a Service (IaaS),
depending on the different optimization and machine Platform as a Service (PaaS) and Software as a Service
learning algorithms. Also, it addresses the challenges in (SaaS). Such categories are based on the degree of
those algorithms and recommends a few possible abstraction and the communication pattern of the providers
solutions for improving the utilization of cloud [5].
computing.
A. Architecture of Cloud Computing
Keywords:- Cloud computing, Task scheduling, Virtual Different types of enterprises use cloud computing
machine, Makespan, Metaheuristic, Machine learning, platforms to preserve information in the clouds therefore
Optimization. they may retrieve it anytime they need it. The 2 types of
cloud infrastructure as shown in Fig. 1 are a front end and a
I. INTRODUCTION back end, which are linked by the internet [6].

Internet-connected supercomputing is known as cloud

computing. It is a sort of global technology that merely joins
enormous computer groups utilizing a variety of techniques,
including remote computing, virtualization, etc. Clients may
transfer many data into the cloud systems and utilize a great
processing ability with the aid of their local computer [1]. It
provides clients with a range of storage, networking, and
processing capabilities over the Internet. The cloud, as
defined by R. Buyya, is “a concurrent and dispersed

IJISRT23AUG382 www.ijisrt.com 936

Volume 8, Issue 8, August 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

Fig. 1: Architecture of Cloud Computing

The back end is in the role of providing cloud B. Task Scheduling and Its Categories
applications with data protection. The back end is used by The cloud comprises a variety of resources, which are
the network operators. It oversees the management of every distinct from one another in terms of various resources, and
resource required to provide operations. It includes a since the expense of executing jobs in the cloud with those
security system, a huge amount of information storage, resources is distinct, therefore TS in the cloud differs from
hosts, VMs, traffic management systems, deployment conventional strategies of TS. As a result, TS in the cloud
models, etc. requires more emphasis since cloud operations rely on it. TS
is crucial for increasing the adaptability and dependability of
Indeed, individuals interact with the front end. cloud-based applications. The primary aim of allocating jobs
Programs and user interactions are required for front-end to resources in line with scheduling constraints is to
access to cloud computing. Computers, web browsers, and determine the optimal schedule in which to perform multiple
smartphones are included. The access methods for cloud jobs such that to provide the client with the optimum
storage are distinct from those for traditional storage because outcome [7].
the cloud holds a large quantity of information from a wide
range of individuals. The majority of operators implement In cloud computing, various resources, including
several access methods. The following includes a few cloud containers, firewalls, and networks, are often dynamically
computing architectural components: assigned by the order and specifications of the job and its
 User infrastructure: It is regarded as a front-end subtasks. As a result, work scheduling in the cloud becomes
component. It offers a Graphical User Interface (GUI) for a flexible issue because no previously established schedule
communicating with the cloud. can be helpful while executing a job. Since the workflow is
 Internet: It serves as a channel for 2 ends to interact with unpredictable, processing methods are also unpredictable,
one another. and resources are also unpredictable when several workloads
 Application: The client might seek access to any are using resources concurrently, the scheduling is
application or network. unpredictable due to these factors.
 Service: It offers IaaS, SaaS and PaaS.
TS in the cloud refers to selecting the optimal
 Runtime cloud: It provides the VMs with an operational
resources provided for workload completion or allocating
and dynamic platform.
system resources to workloads in a way that minimizes the
 Storage: It is one of the key components of cloud workload execution period. In scheduling strategies, a
computing infrastructure. It provides a lot of storage collection of workloads is formed by assigning a weight to
capacity in the cloud for handling and storing all jobs, with the significance of individual workload
information. depending on a variety of factors. After that, workloads are
 Infrastructure: It provides functions at the application, selected based on their importance and given to the
host and network levels. It includes both hardware and processing systems that can meet a predetermined target
software components. function [8].
 Management: It is employed to handle every component
of the back end. As well, it creates cooperation among Two major categories of TS are:
them.  Fixed scheduling: It schedules workloads in a well-
 Security: It executes a privacy method in the back end. known setting, i.e. it contains the data regarding the
overall arrangement of workloads, resource allocation
before processing and prediction of the workload
processing period [9].

IJISRT23AUG382 www.ijisrt.com 937

Volume 8, Issue 8, August 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
 Dynamic scheduling: It should rely not solely on the The cloud computing TS process has 3 stages [14] as
allocated workloads to the cloud system, yet also on the illustrated in Fig. 2:
present conditions of systems to create scheduling  Initial stage: It comprises a collection of workloads
choices [9]. (cloudlets), which are transmitted by the cloud clients for
 Direct scheduling: If new workloads exist, then they are processing.
allocated to VMs immediately [10].  Second stage: It translates workloads to appropriate
 Batch scheduling: Workloads are clustered into a batch resources to obtain the maximum resource usage and a
before transmission. It is also known as mapping services less makespan.
[11].  Third stage: It comprises a collection of VMs that are
 Preemptive scheduling: All workloads are disturbed utilized to process the workloads.
while processing and may be shifted to the other resource
to finish processing [12].
 Non-preemptive scheduling: VMs are not rescheduled to
new workloads until completing the processing of the
allocated workloads [13].

Fig. 2: Overview of Task Scheduling in Cloud Computing

C. Necessity of Task Scheduling in Cloud Computing D. Classification of Meta-Heuristic Task Scheduling

The primary goal of scheduling is to respond to arriving Algorithms in Cloud Computing
requests from end clients by identifying the optimum cloud The purpose of TS differs from every system to others,
resources that must increase both the system usage rates and under the QoS criteria. As a result, several studies were
essential quality metrics. Different efficiency measures for established that focus on TS using meta-heuristics
cloud computing exist, including makespan, financial cost, algorithms. A new, robust classification is presented in Fig.
processing cost, reaction time, power usage, dependability, 3 to help individuals comprehend metaheuristic TS
etc. To meet the needs of end clients and service providers algorithms in cloud computing more extensively and
while maintaining the SLA, an effective TS strategy should effectively [16-17]. These algorithms may be grouped into 4
be employed to assess and enhance such factors. Owing to distinct categories depending on the specific kind of
challenges including resource distribution, dynamism, and scheduling issue, the major goal of scheduling, the task-
heterogeneity, traditional scheduling mechanisms are unable resource mapping strategy and the scheduling restriction.
to tackle these issues [15-16]. Considering the primary goal Depending on the relationship between the entering
of solving the possible issues of overloading and workloads, these algorithms are further divided into
underloading in cloud TS, a scheduling algorithm is dependent and independent workloads. Depending on the
therefore required for fair and appropriate allocation of kind of scheduling algorithms (schedulers), they are
diverse tasks among VMs depending on the availability of classified as classical/heuristics or meta-heuristics.
resources.  Type of scheduling issue: It is necessary to develop an
optimized strategy that satisfies the goals by selecting the
The benefits of TS approaches [16] are (i) controlling most optimum result because there is often a balance
the Quality-of-Service (QoS) efficiency of cloud computing, among optimization goals. It is feasible to evaluate the
(ii) controlling the processor and storage, (iii) increasing optimality of a specific strategy in contrast to another one
resource usage when reducing the overall workload that already exists in a single objective optimization.
processing period, (iv) ensuring fairness for every workload, Whilst this cannot be done effectively in Multi-objective
(v) enhancing the number of workloads that are properly Optimization Problems (MOPs), it may be done
finished, (vi) allocating workloads on real-time applications, indirectly [18].
(vii) obtaining a maximum network throughput and (viii)  Major goal of scheduling: Whenever a task scheduling
enhancing load distribution. procedure is carried out, a minimum single goal value is
required for obtaining higher performance. The most
often used goals may be stated as follows: throughput,

IJISRT23AUG382 www.ijisrt.com 938

Volume 8, Issue 8, August 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
makespan, economic cost, processing expense (i.e., usage having ambiguous features and also being possible to be
of CPU, memory and so on), dependability and altered. Thus, to address QoS demands and reduce SLA
accessibility, flexibility or scalability, privacy and power breaches, these strategies are created and implemented
usage [19]. [20].
 Task-resource mapping method: To effectively exploit  Scheduling restriction: Due to the potential impact on the
the allocated resources depending on the state of the SLA when a huge variety of services are failing to satisfy
cloud system and the given tasks, static, dynamic, deadline, priority, budget, and fault tolerance
Artificial Intelligence (AI)-based and prediction-based requirements, such variables are important in the sector
translation of cloud resources to arriving workloads is of cloud scheduling [19].
conducted. Resources and tasks are well-known for

Fig. 3: Taxonomy of Task Scheduling Algorithms in Cloud Computing

Various algorithms have been developed over the past II. LITERATURE SURVEY
decades for TS in cloud applications. The primary purpose
of this paper is to give a comprehensive overview of TS A. Survey on Task Scheduling Based on Optimization
algorithms in cloud computing using various optimization Algorithms in Cloud Computing
and machine learning techniques. Also, a comparative An Improved Particle Swarm Optimization (IPSO)
analysis is presented to highlight the benefits and drawbacks algorithm [21] are developed to achieve the best distribution
of those algorithms in a tabular form, which supports us to for a huge amount of tasks. This was performed by
suggest possible future directions. partitioning the allocated tasks into batches dynamically.
Also, the resource usage condition was taken in the
The following sections have been prepared as follows: generation of all batches. Once obtaining a sub-optimal
Section II studies and analyzes the TS based on various result for all batches, every sub-optimal result for batches
optimization algorithms, whereas Section III studies and was added to the absolute distribution map. Moreover, the
analyzes the TS based on machine learning algorithms in loads over the absolute distribution map were balanced by
cloud systems. Section IV summarizes the complete study the IPSO.
and offers suggestions for future enhancement.

IJISRT23AUG382 www.ijisrt.com 939

Volume 8, Issue 8, August 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
An Immune-based PSO (IMPSO) algorithm [22] were AEO in the procedure of discovering the best decision for
presented to allocate workflow in the cloud paradigm. The the issue under concern.
aim was to reduce the execution cost and makespan under
user-defined deadline restraints. Li & Han [23] presented a A 3-level scheduling framework depending on the
flexible TS based on the hybrid discrete Artificial Bee whale-Gaussian cloud called GCWOAS2 [29], which were
Colony (ABC) algorithm. Initially, the TS issue was client task level, TS level and data center level to define the
formulated as a Hybrid Flowshop Scheduling (HFS) issue. whole procedure of TS. Initially, an opposition-based
In multi-objective HFS, reduction of the maximum end training scheme was adopted to initialize the scheduling
period, maximum system workload and overall workloads of plans and find the best scheduling strategy. After that, a
each system were measured concurrently. Many kinds of dynamic fine-tuning factor was applied to adaptively fine-
perturbation patterns were considered to improve hunting tune the search region. To improve the arbitrariness of
capabilities. Also, an enhanced adaptive perturbation pattern exploration, a whale optimization algorithm was designed
was included to balance the exploitation and exploration depending on the Gaussian cloud scheme. Further, a multi-
capability. Moreover, a deep-exploitation operator was objective TS scheme using the Gaussian whale-cloud
applied to enhance the exploitation capabilities for effective optimization was introduced to find the global best
TS. scheduling plan.

To designed a multi-objective TS optimization Introducing a framework to estimate the present

depending on the fuzzy defense algorithm [24]. The main condition of the active tasks based on the outcomes of the
aim was to choose the shortest period, the degree of resource QoS forecast allocated by an Auto-regressive Integrated
load balance and the cost of multi-objective task execution Moving Average (ARIMA) [30] framework optimized by
by creating a mathematical framework, which provides the the Kalman filter. After that, a scheduling strategy was
objective factor and determines the impact of multi- determined by the joint PSO and Gravitational Search
objective TS. Those objective values were resolved by the Algorithm (GSA) based on the QoS conditions to ensure the
fuzzy self-defense scheme to get the global best result of the client’s QoS via allocating the workflow.
objective factor.
A TS technique [31] were presented to jointly reduce
An Improved Whale Optimization (IWC) algorithm energy cost and mean task loss probability of clouds. In this
[25] are developed to enhance the TS performance in cloud technique, the issue was modeled and solved by an adaptive
computing. Initially, a cloud computing TS and allocation bi-objective differential growth depending on simulated
system with a period, cost and VMs was built. Then, a viable annealing to compute a real-time and near-optimum group of
strategy for all whale individuals related to the cloud results. Moreover, an absolute knee result was selected
computing TS was applied to obtain the optimal whale based on the minimum Manhattan distance to characterize
individual using the inertial weight mechanism, which appropriate servers in clouds and task distribution amid
enhances the local hunting capability and avoids early online sites.
convergence. Also, add and delete functions were used to
monitor individuals after all iterations, which were ended TS method [32] were depending on the Advanced
and modified to choose individuals with greater efficiency. Phasmatodea Population Evolution (APPE) algorithm in a
heterogeneous cloud setting. This algorithm was used to
A metaheuristic model termed MDVMA [26] were minimize the time needed to obtain solutions by enhancing
developed for dynamic VM distribution with optimized TS the convergent progress of the closest best solutions. Also, a
in a cloud computing paradigm. In this model, a multi- restart mechanism was included to avoid the algorithm from
objective scheduling scheme was adopted by the Non- entering the local optimization and balance its search and
dominated Sorting Genetic Algorithm (NSGA)-II algorithm exploitation abilities. Moreover, the valuation function was
to optimize TS, which reduces energy utilization, makespan applied to discover the optimal solutions according to the
and cost concurrently to achieve tradeoff to the cloud service makespan, resource cost and load balancing level.
providers according to their demands.
A semi-adaptive real-time TS scheme named the
A multi-objective restricted optimization issue [27] Improved Genetic Algorithm is designed for Permutation-
were analyzed to recognize the best scheduling strategies for based Optimization Problems (IGA-POP) [33] for bag-of-
systematic tasks to be employed in unreliable cloud tasks in the cloud-fog paradigm. In this scheme, the TS issue
scenarios. The main aim was to reduce the estimated task was modeled as a POP. First, the IGA was applied to find
execution period and monetary expense under probabilistic various permutations for arrived tasks at all scheduling
restraints on deadline and budget. This issue was resolved by cycles. After that, the tasks were allocated to the VM based
the combined Monte Carlo method and Genetic Algorithm on the optimal permutation to accomplish a better tradeoff
(MCGA), the cloud clients were permitted to select the between the makespan and the overall performance cost
strategy of the Pareto optimum group ensuring their when satisfying deadline restraints.
demands and interests. An alternated TS method [28] were
designed for IoT requests in a cloud-fog paradigm
depending on the modified Artificial Ecosystem-based
Optimization (AEO) by the operators of the Salp Swarm
Algorithm (SSA) to improve the exploitation capability of

IJISRT23AUG382 www.ijisrt.com 940

Volume 8, Issue 8, August 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
B. Survey on Task Scheduling Based on Machine Learning A novel framework depending on the multi-agent
Algorithms in Cloud Computing system called DRL for resource distribution and TS [40] to
A smart QoS-aware TS model based on Deep minimize the cost and power in cloud computing. In this
Reinforcement Learning (DRL) [34] was developed for framework, a Quantile Regression DQN (QR-DQN) scheme
applications in clouds. It can learn to create suitable online was developed to create a suitable strategy and the best long-
task-to-VM solutions for constant task requests immediately term solutions to assign resources and schedule tasks to
from its experiences with no previous data. Based on this corresponding VMs.
process, the tasks were scheduled by the service providers to
constrained resources under QoS demand limits. TS scheme in a cloud paradigm depending on the
multi-criteria decision-making approach [41]. In this
fat-tree structure-based method named Large-scale scheme, the TS was modeled as a non-linear restricted
Tasks processing using Deep Reinforcement (LTDR) [35] optimization dilemma and solved by the Technique for
training to find the best TS policy. This was achieved by Order of Preference by Similarity to Ideal Solution
using a virtual network mapping scheme depending on a (TOPSIS), which incorporates an Entropy Weight Method
deep Convolutional Neural Network (CNN) and Q-learning (EWM) to reduce makespan, cost and power usage, as well
algorithms. Also, a policy network was applied to create as, improve the consistency.
node mapping decisions and the link mapping method was
performed using the distributed value factor. Then, tasks Self-adapting TS scheme called ADATSA [42] based
were scheduled to the appropriate physical nodes and on the learning automata for container cloud. Initially, a
processed effectively. learning automata scheme and the objective factor were
designed for the system on the TS issue. After that, an
A novel scheduling model called Spear [36] is efficient incentive-penalty strategy was performed to
developed to reduce the makespan of complicated tasks schedule tasks combined with the idle condition of resources
when considering task dependencies and heterogeneous and the operating condition of tasks in the present
resource needs. In this model, a Monte Carlo Tree Search atmosphere. Additionally, the atmosphere was designed by
(MCTS) was applied in the TS phase and the DRL was cluster, node and task, as well as the chance of task chosen,
trained to direct the processes in the MCTS. Using this was optimized by scheduling implementation to improve the
DRL, exploration ability was enhanced by concentrating adaptability to the cloud scenario of the allocation.
favorable branches of the search tree. Moreover, a model of task load monitoring with a buffer
queue was created to perform dynamic scheduling
2-phase TS and resource distribution model [37] were depending on priority.
designed, which utilizes many smart schedulers to resolve
the cooperative scheduling issue between TS and resource A multi-objective TS scheme depending on the
distribution. A Heterogeneous Distributed Deep Learning Decision Tree (DT) [43] in a heterogeneous cloud scenario.
(HDDL) framework was applied in the TS phase to allocate A novel TS-DT algorithm was adopted to assign and
various tasks to several cloud data centers. Also, a Deep Q- implement the applications’ tasks. The major aim was to
Network (DQN) framework was applied as a resource resolve the multi-objective TS challenge by reducing
scheduler to arrange VMs for tasks to physical servers for makespan, ensuring load balancing amid VMs and
implementation. increasing resource usage.

TS scheme depending on the DRL model [38] such as An improved training-enabled TS model depending on
DQN to adaptively schedule tasks with precedence the task Criticality and Collapse-Aware Scheduling (CCAS)
connection to cloud servers to reduce the task scheme [44]. In this scheme, 2 distinct strategies were
implementation period. To achieve this, the aspects of designed such as the TS strategy depending on task CCAS
servers and tasks were considered as state inputs and server and an ensemble forecast strategy such as Gradient Boosting
numbers were considered as activities. To reduce the DT (GBDT) to proactively estimate the system usage and
execution period, the negative change value of makespan task implementation status by capturing the high-level
from a particular state to the other state was described as the attributes via training the task variables. Also, a smart
incentive. Also, the task precedence connection restraint was scheduling scheme was adopted for best resource
accomplished during the state shift phase. The issue of TS of distribution.
cloud-based systems and intended to reduce the
computational cost under resource and deadline restraints An independent TS method in cloud computing based
[39]. To solve this issue, a clipped double deep Q-learning on the utilization of the Multi-Objective Artificial Bee
method was introduced based on the target network and Colony with Q-learning (MOABCQ) algorithm [45]. This
experience relay schemes, which allocates the tasks to their algorithm was used to compute the order of tasks for
corresponding VMs. appropriate resources and schedule the most suitable tasks
according to the execution time, cost and usage of resources.
Also, it was integrated with the First Come First Serve
(FCFS) and the Largest Job First (LJF) heuristic TS schemes
to achieve load balancing among VMs.

IJISRT23AUG382 www.ijisrt.com 941

Volume 8, Issue 8, August 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
III. COMPARATIVE ASSESSMENT

This part compares the merits and demerits of the different metaheuristic-based TS algorithms for cloud applications in
Table 1.

Table 1: Assessment of Various Metaheuristic-based TS Algorithms in Cloud Systems

Ref. No. Algorithm Merits Demerits Outcomes
[21] IPSO It can avoid imbalanced task It did not consider the cost and No. of workloads=3000:
scenarios because of the rebalancing energy use in the objective Makespan=540sec;
procedure after obtaining the final functions. Standard variance=15;
scheduling maps. Degree of imbalance=0.1
[22] IMPSO It could achieve the optimum result It needs to group the Genome tasks:
at a rapid convergence time. workloads before scheduling Cost=7.2$/hr;
them to cloud resources due to Makespan=40000sec;
the dependency among Cybershake tasks:
workloads and tradeoffs Cost=1.5$/hr;
among groups. Makespan=700sec;
Sipht tasks:
Cost=1.14$/hr;
Makespan=2500sec
[23] Hybrid It improves the convergence ability The local exploration Mean makespan=23sec;
discrete ABC to find the best solution during TS. capability was not efficient. Computational
Also, the tradeoff between time=1.25sec
exploitation and exploration
was not effective.
[24] Fuzzy self- The deadline violation rate was It did not consider the No. of workloads=180:
defense reduced so that the resource usage makespan and energy usage as Highest execution
rate on the VM was comparatively the objective function. period=790sec;
satisfactory. Deadline rate=0.02%;
VM number=80:
VM resource
usage=0.98%
[25] IWC It can reduce the cost and time for The impact on the memory No. of workloads=800:
TS. load value was not clear, Economic cost=0.75$;
which needs to enhance the Time utilization=1.1msec;
memory usage. Memory load=0.41;
[26] MDVMA It was helpful to achieve optimal TS It restricts the convergence Total makespan=7338sec;
using NSGA- with less energy use, makespan and speed and a few results of the Total energy
II cost. Pareto front were not obtained, utilization=31.93kWh;
which may be best compared Total cost=7338$
to the best results.
[27] MCGA It can enable the clients to select the The optimum results were not No. of workloads:24:
strategy of the Pareto optimum achieved under tight deadlines Runtime=146sec;
group guaranteeing their demands and costs, because the Processing
and interests. variability raises. period=3789sec
[28] Modified It can achieve better mean It needs to consider additional No. of workloads:800:
AEO using makespan and throughput under objective values like energy Mean makespan=35.1sec;
SSA both synthetic and real tasks. usage and economic costs. Mean throughput
period=3000sec
[29] GCWOAS2 It may decrease the workload It did not function efficiently No. of workloads=100:
execution period and balance the in terms of operating costs. No. of iterations=100:
load of VMs. Overall cost=0.263$;
Time cost=0.15sec;
Load cost=0.331;
[30] Combined It can decrease the SLA violation It did not consider the No. of workloads=800:
PSO and rate efficiently while increasing the periodicity of client workload SLA violation
GSA amount of workload. information and the client’s rate=0.03%;
QoS factors. Also, it needs Workload completion
more objective functions to cost=850$;
increase the efficiency of TS.
[31] Simulated It can minimize the energy cost and The convergence speed and Mean energy

IJISRT23AUG382 www.ijisrt.com 942

Volume 8, Issue 8, August 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
annealing mean error probability of diversity of acquired results cost=1.335×104$;
workloads. were not improved. Convergence
speed=18.09sec
[32] APPE It has a rapid convergence period It was solely appropriate to No. of workloads=500:
and better makespan. solve the fixed TS challenges Performance index
and not distribute resources evaluation function
based on the arriving period of value=450
workloads.
[33] IGA-POP It can achieve a good tradeoff It considers only the static No. of workloads=400:
between the makespan and the cloud computing platform so Elapsed period=615sec;
overall execution cost. the resource usage was not Makespan=113.7sec;
enhanced. Also, it needs Execution cost=2122.4G$
advanced meta-heuristics and
machine learning algorithms
to solve the dynamic TS issue.

Table 1 addresses that many researchers focused on environments. From this viewpoint, a few researchers design
optimized TS in cloud applications using various machine learning algorithms with and without meta-
metaheuristic algorithms like PSO, GA, NSGA-II, ABC, etc. heuristics to accomplish dynamic TS in cloud systems,
Each algorithm has disadvantages regarding exploration, which are studied in Section III. Here, the merits and
exploitation and convergence abilities. To combat these demerits of those machine learning-based TS algorithms are
issues, more advanced and machine learning algorithms listed in Table 2.
must be developed to achieve TS in dynamic cloud

Table 2: Assessment of Various Machine Learning-based TS Algorithms in Cloud Systems

Ref. No. Algorithm Merits Demerits Outcomes
[34] QoS-aware TS It can effectively decrease It needs to extend to a more Success rate=98.3%;
using DRL the mean task response sophisticated cloud platform Response period=158ms;
period and ensure the QoS using multiple objective
at a high level of client values.
experience.
[35] LTDR using It can distribute the It needs to improve the policy Time=2500sec:
CNN and Q- workloads on the model by raising the number Throughput=3.7requests/sec;
learning appropriate physical nodes. of neural levels. Long-term
revenue/cost=0.44$
[36] MCTS and It can decrease the It considers only a single Makespan=820.1sec;
DRL makespan efficiently by objective, whereas more Runtime=500sec
enhancing the exploration objective functions were
ability. needed to enhance the
efficiency of TS.
[37] HDDL and It has better scalability and It needs to achieve near- Energy usage=10.48kWh;
DQN computation time in real- global optimization via Latency rate=0.37%;
time cloud TS. enhancing the cooperative Task delay=33.21sec
capability of many training
frameworks.
[38] DQN It has a less mean It needs to consider multiple No. of workloads=100:
execution period as objectives such as cost, Mean
increasing the number of deadline of workloads, etc., to makespan=45.3467sec;
workloads. enhance TS efficiency. Standard variance=1.7932;
Mean CPU
period=0.0267sec
[39] Clipped It can achieve a better It did not reduce the load on No. of workloads=9:
double deep balance between the cloud data center and it Execution period=130sec
Q-learning exploration and should select proper value
exploitation. for achieving effective TS.
[40] QR-DQN It efficiently reduces both Increasing the number of No. of workloads=1000:
energy and time cost. workloads discards more Normalized energy
workloads because of cost=0.03;
exceeding intervals or Time cost=0.1sec
resources.
[41] TOPSIS with It has less cost and energy It did not consider the client’s Mean communication to

IJISRT23AUG382 www.ijisrt.com 943

Volume 8, Issue 8, August 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
EWM usage within significant interest-based factors like computation ratio=10:
restraints. cost and deadline restraints Mean cost=1.5$;
for TS. Mean energy
usage=4×109kWh
[42] ADATSA It achieves good It did not consider the Time=20min:
using learning environment adaptability, heterogeneity of cloud Resource imbalance
automata resource optimization resources. Also, the degree=0.13;
efficacy and QoS environment system trained Resource residual
efficiency. from the constant incentive- degree=0.65;
penalty variables was not Response delay=440msec;
ideal. Throughput=150req/sec
[43] TS-DT It can decrease the mean The energy usage was high. No. of VMs=40:
makespan and enhance the Mean
mean resource usage. makespan=153.65msec;
Resource usage=99.297%;
Deviation of load
balance=0.37
[44] CCAS and It enhances resource usage, It did not guarantee error-free No. of workloads=45:
GBDT flexibility and dynamism. workloads processing. Execution period=630msec;
Latency=5msec;
Mean system usage=90%
[45] MOABCQ It has less time complexity It did not ensure that this Synthetic task database:
and makespan. Also, it has algorithm was ideal and the No. of workloads=800:
a good resource usage rate. efficiency of the network was Makespan=24sec;
not optimized in each test Mean throughput=35tasks/s;
database. Cost=150G$;
Mean resource usage
rate=0.801%;
Degree of imbalance=0.117

Table 2 states that some researchers have concentrated REFERENCES

on TS based on machine learning algorithms, i.e. DQN,
DRL, etc., with a few meta-heuristic algorithms. Even [1.] L. Bohu, Z. Lin and C. Xudong, “Introduction to
though those algorithms outperform single-objective cloud manufacturing,” Zte Communications, vol. 8,
optimization algorithms, TS in multi-cloud, fog-cloud, or no. 4, pp. 6-9, 2020.
edge-cloud platforms is problematic. It is vital to apply the [2.] R. Buyya, C. S. Yeo, S. Venugopal, J. Broberg and I.
other sophisticated machine learning algorithms to enhance Brandic, “Cloud computing and emerging IT
TS in different kinds of cloud environments. Also, it must be platforms: vision, hype, and reality for delivering
tested in a real-time scenario to analyze the efficiency of TS computing as the 5th utility,” Future Generation
algorithms. Computer Systems, vol. 25, no. 6, pp. 599-616, 2009.
[3.] S. K. Sahana, “Emerging computing platforms for
IV. CONCLUSION solving complex engineering problems,” In Methods,
Implementation, and Application of Cyber Security
This study presents a broad review of various TS Intelligence and Analytics, IGI Global, pp. 165-180,
algorithms in cloud computing based on a variety of meta- 2022.
heuristics and machine learning algorithms. According to the [4.] M. Taghipour, M. E. Soofi, M. Mahboobi and J.
findings of this study, many academics have been Abdi, “Application of cloud computing in system
experienced in designing TS algorithms that schedule the management in order to control the
best workloads to the proper VMs in the cloud paradigm. process,” Management, vol. 3, no. 3, pp. 34-55, 2020.
Amongst, the MOABCQ algorithm can reduce the makespan [5.] G. Bhatta and M. Pandey, “A case study on hybrid
and enhance resource usage rate with less computational cloud approach to automate the cloud services based
time complexity. Conversely, its performance cannot be on decision support system,” Review of International
guaranteed to be perfect. Also, not all test databases can Geographical Education Online, vol. 11, no. 8, pp.
facilitate optimizing network efficiency. So, advanced 1669-1683, 2021.
machine learning and meta-heuristics algorithms can be [6.] Odun-Ayo, M. Ananya, F. Agono and R. Goddy-
incorporated to achieve optimized TS and analyze the Worlu, “Cloud computing architecture: a critical
effectiveness of those algorithms in both static and dynamic analysis,” In 18th IEEE International Conference on
cloud computing applications in the future. Computational Science and Applications, pp. 1-7,
2018.

IJISRT23AUG382 www.ijisrt.com 944

Volume 8, Issue 8, August 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
[7.] M. Ibrahim, “Task scheduling algorithms in cloud [20.] S. Subbaraj and R. Thiyagarajan, “Performance
computing: a review,” Turkish Journal of Computer oriented task-resource mapping and scheduling in fog
and Mathematics Education, vol. 12, no. 4, pp. 1041- computing environment,” Cognitive Systems
1053, 2021. Research, vol. 70, pp. 40-50, 2021.
[8.] F. Ebadifard and S. M. Babamir, “Autonomic task [21.] H. Saleh, H. Nashaat, W. Saber and H. M. Harb,
scheduling algorithm for dynamic workloads through “IPSO task scheduling algorithm for large scale data
a load balancing technique for the cloud-computing in cloud computing environment,” IEEE Access, vol.
environment,” Cluster Computing, vol. 24, no. 2, 7, pp. 5412-5420, 2018.
1075-1101, 2021. [22.] P. Wang, Y. Lei, P. R. Agbedanu and Z. Zhang,
[9.] T. McSweeney, N. Walton and M. Zounon, “An “Makespan-driven workflow scheduling in clouds
efficient new static scheduling heuristic for using immune-based PSO algorithm,” IEEE
accelerated architectures,” In International Access, vol. 8, pp. 29281-29290, 2020.
Conference on Computational Science, Springer, [23.] J. Q. Li and Y. Q. Han, “A hybrid multi-objective
Cham, pp. 3-16, 2020. artificial bee colony algorithm for flexible task
[10.] J. Yao and N. Ansari, “Fog resource provisioning in scheduling problems in cloud computing
reliability-aware IoT networks,” IEEE Internet of system,” Cluster Computing, vol. 23, no. 4, pp. 2483-
Things Journal, vol. 6, no. 5, pp. 8262-8269, 2019. 2499, 2020.
[11.] Z. Zhao, S. Liu, M. Zhou, D. You and X. Guo, [24.] X. Guo, “Multi-objective task scheduling
“Heuristic scheduling of batch production processes optimization in cloud computing based on fuzzy self-
based on petri nets and iterated greedy defense algorithm,” Alexandria Engineering
algorithms,” IEEE Transactions on Automation Journal, vol. 60, no. 6, pp. 5603-5609, 2021.
Science and Engineering, vol. 19, no. 1, pp. 251-261, [25.] L. Jia, K. Li and X. Shi, “Cloud computing task
2020. scheduling model based on improved whale
[12.] W. Chen, X. Zhou and J. Rao, “Preemptive and low optimization algorithm,” Wireless Communications
latency datacenter scheduling via lightweight and Mobile Computing, pp. 1-13, 2021.
containers. IEEE Transactions on Parallel and [26.] D. Alsadie, “A metaheuristic framework for dynamic
Distributed Systems, vol. 31, no. 12, pp. 2749-2762, virtual machine allocation with optimized task
2019. scheduling in cloud data centers,” IEEE Access, vol.
[13.] N. Panwar, S. Negi, M. M. S. Rauthan and K. S. 9, pp. 74218-74233, 2021.
Vaisla, “TOPSIS–PSO inspired non-preemptive tasks [27.] M. C. Calzarossa, M. L. Della Vedova, L. Massari, G.
scheduling algorithm in cloud environment,” Cluster Nebbione and D. Tessera, “Multi-objective
Computing, vol. 22, no. 4, pp. 1379-1396, 2019. optimization of deadline and budget-aware workflow
[14.] T. Aladwani, “Types of task scheduling algorithms in scheduling in uncertain clouds,” IEEE Access, vol. 9,
cloud computing environment,” Scheduling pp. 89891-89905, 2021.
Problems-New Applications and Trends, pp. 1-12, [28.] M. Abd Elaziz, L. Abualigah and I. Attiya,
2020. “Advanced optimization technique for scheduling IoT
[15.] M. Kumar, S. C. Sharma, A. Goel and S. P. Singh, “A tasks in cloud-fog computing environments,” Future
comprehensive survey for scheduling techniques in Generation Computer Systems, vol. 124, pp. 142-154,
cloud computing,” Journal of Network and Computer 2021.
Applications, vol. 143, pp. 1-33, 2019. [29.] L. Ni, X. Sun, X. Li and J. Zhang, “GCWOAS2:
[16.] E. H. Houssein, A. G. Gad, Y. M. Wazery and P. N. multiobjective task scheduling strategy based on
Suganthan, “Task scheduling in cloud computing Gaussian cloud-whale optimization in cloud
based on meta-heuristics: review, taxonomy, open computing,” Computational Intelligence and
challenges, and future trends,” Swarm and Neuroscience, pp. 1-17, 2021.
Evolutionary Computation, vol. 62, pp. 1-41, 2021. [30.] M. A. Rakrouki and N. Alharbe, “QoS-aware
[17.] N. Kaur, A. Kumar and R. Kumar, “A systematic algorithm based on task flow scheduling in cloud
review on task scheduling in fog computing: computing environment,” Sensors, vol. 22, no. 7, pp.
taxonomy, tools, challenges, and future 1-20, 2022.
directions,” Concurrency and Computation: Practice [31.] H. Yuan, J. Bi and M. Zhou, “Energy-efficient and
and Experience, vol. 33, no. 21, pp. 1-25, 2021. QoS-optimized adaptive task scheduling and
[18.] M. Abdullahi, M. A. Ngadi, S. I. Dishing and B. I. E. management in clouds,” IEEE Transactions on
Ahmad, “An efficient symbiotic organisms search Automation Science and Engineering, vol. 19, no. 2,
algorithm with chaotic optimization strategy for pp. 1233-1244, 2022.
multi-objective task scheduling problems in cloud [32.] N. Zhang, S. C. Chu, P. C. Song, H. Wang and J. S.
computing environment,” Journal of Network and Pan, “Task scheduling in cloud computing
Computer Applications, vol. 133, pp. 60-74, 2019. environment using advanced phasmatodea population
[19.] R. Ghafari, F. H. Kabutarkhani and N. Mansouri, evolution algorithms,” Electronics, vol. 11, no. 9, pp.
“Task scheduling algorithms for energy optimization 1-16, 2022.
in cloud environment: a comprehensive [33.] S. Abohamama, A. El-Ghamry and E. Hamouda,
review,” Cluster Computing, pp. 1-59, 2022. “Real-time task scheduling algorithm for IoT-based
applications in the cloud–fog environment,” Journal

IJISRT23AUG382 www.ijisrt.com 945

Volume 8, Issue 8, August 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
of Network and Systems Management, vol. 30, no. 4,
pp. 1-35, 2022.
[34.] Y. Wei, L. Pan, S. Liu, L. Wu and X. Meng, “DRL-
scheduling: An intelligent QoS-aware job scheduling
framework for applications in clouds,” IEEE
Access, vol. 6, pp. 55112-55125, 2018.
[35.] Wu, G. Xu, Y. Ding and J. Zhao, “Explore deep
neural network and reinforcement learning to large-
scale tasks processing in big data,” International
Journal of Pattern Recognition and Artificial
Intelligence, vol. 33, no. 13, pp. 1-29, 2019.
[36.] Z. Hu, J. Tu and B. Li, “Spear: Optimized
dependency-aware task scheduling with deep
reinforcement learning,” In IEEE 39th International
Conference on Distributed Computing Systems, pp.
2037-2046, 2019.
[37.] J. Lin, D. Cui, Z. Peng, Q. Li and J. He, “A two-stage
framework for the multi-user multi-data center job
scheduling and resource allocation,” IEEE
Access, vol. 8, pp. 197863-197874, 2020.
[38.] T. Dong, F. Xue, C. Xiao and J. Li, “Task scheduling
based on deep reinforcement learning in a cloud
manufacturing environment,” Concurrency and
Computation: Practice and Experience, vol.32, no. 11,
pp. 1-12, 2020.
[39.] S. Swarup, E. M. Shakshuki and A. Yasar, “Task
scheduling in cloud using deep reinforcement
learning,” Procedia Computer Science, vol. 184, pp.
42-51, 2021.
[40.] T. Oudaa, H. Gharsellaoui and S. B. Ahmed, “An
agent-based model for resource provisioning and task
scheduling in cloud computing using DRL,” Procedia
Computer Science, vol. 192, pp. 3795-3804, 2021.
[41.] M. S. Kumar, A. Tomar and P. K. Jana, “Multi-
objective workflow scheduling scheme: a multi-
criteria decision making approach,” Journal of
Ambient Intelligence and Humanized
Computing, vol. 12, no. 12, 10789-10808, 2021.
[42.] L. Zhu, K. Huang, Y. Hu and X. Tai, “A self-adapting
task scheduling algorithm for container cloud using
learning automata,” IEEE Access, vol. 9, pp. 81236-
81252, 2021.
[43.] H. Mahmoud, M. Thabet, M. H. Khafagy and F. A.
Omara, Multiobjective task scheduling in cloud
environment using decision tree algorithm,” IEEE
Access, vol. 10, pp. 36140-36151, 2022.
[44.] N. Khan, N. Iqbal, A. Rizwan, S. Malik, R. Ahmad
and D. H. Kim, “A criticality-aware dynamic task
scheduling mechanism for efficient resource load
balancing in constrained smart manufacturing
environment,” IEEE Access, vol. 10, pp. 50933-
50946, 2022.
[45.] Kruekaew and W. Kimpan, “Multi-objective task
scheduling optimization for load balancing in cloud
computing environment using hybrid artificial bee
colony algorithm with reinforcement learning,” IEEE
Access, vol. 10, pp. 17803-17818, 2022.

IJISRT23AUG382 www.ijisrt.com 946