Wireless Networks (2024) 30:1085–1105

https://doi.org/10.1007/s11276-023-03532-0 (0123456789().,-volV)(0123456789().
,- volV)

ORIGINAL PAPER

Hybrid optimization enabled VM scaling based load distribution

and optimal switching strategy in cloud data center
A. Poobalan1 • P. Shanthakumar2 • M. Robinson Joel3

Accepted: 28 July 2023 / Published online: 3 November 2023

 The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2023

Abstract
Cloud data centers provide incredible services to their customers ubiquitously based on demand and pay-per-use strategy.
In virtualized data centers (DC), CPU, RAM, and bandwidth are assigned to a virtual machine (VM) from a group of
pooled resources. One of the key issues for virtualized DC is VMs consolidation as it achieves better performance and also
reduces the cost. In recent years, researchers have paid more attention to developing the global best solution in DCs.
Meanwhile, the global best solution results in a number of redundant migrations and is not desirable for large-scale cloud
computing environments. While designing modern software, care must be taken for scaling processes based on consumer
demands to slash down the costs of the system. This research proposes an effective strategy based on optimization-enabled
VM scaling-based load distribution and optimal switching strategy in the cloud data center. Here, horizontal scaling or
vertical scaling is employed to address the overloading complications. The network structure of the cloud is defined
depending upon the fat tree model and load distribution of cloud DCs is carried out using pelican Taylor manta ray foraging
optimization (P-Taylor MRFO) algorithm by considering multiple objectives, such as power, load, latency, and bandwidth.
Based on the load distribution, the switching of the cloud DC to the desired mode is carried out utilizing actor critic neural
network. If the system is overloaded, either vertical scaling or horizontal scaling is done based on a predefined threshold
elastic scaling using the proposed pelican Adam optimization algorithm (PAOA) based on the horizontal cost that is based
on CPU, memory, and hard disk. The PAOA is devised by integrating Pelican optimization algorithm and Adam opti-
mization. However, the proposed model has attained superior results with a minimum load of 0.329, power of 0.532,
energy consumption of 0.357, and latency of 0.315.

Keywords Cloud computing  Cloud data center  Virtual machine (VM)  Manta ray foraging optimization (MRFO) 
Pelican optimization algorithm (POA)

1 Introduction

Cloud computing has emanated as a significant computing

platform across the world [1]. It has allowed us to distribute
& A. Poobalan IT utilities between various users from everywhere. It has
[email protected] been employed in large-scale applications, such as medi-
& P. Shanthakumar cine, academics, research community, agriculture [2],
[email protected] business, image processing, government sectors, and so on.
1
Department of Computer Science and Engineering, The scalability of resources is two eminent characteristic of
University College of Engineering, Dindigul, cloud computing that performs on-demand computing and
Tamil Nadu 624709, India storage services accessible to the consumers. Users can
2
Department of Computer Science and Engineering, Nehru obtain the on-demand service by the cloud at anywhere
Institute of Technology, Coimbatore, Tamil Nadu 641105, using any terminal device hitched to the Internet [3, 4].
India Nevertheless, it has been analyzed that there is a number of
3
Department of Information Technology, Kings Engineering differences in consumer resource requirement and original
College, Sriperumpudur, Chennai, Tamil Nadu 602117, India

123
1086 Wireless Networks (2024) 30:1085–1105

utilizations that needs to be encountered to minimize the a smoother and quicker control of the allotted resour-
resource wastage at the DC, which is the main constituent ces [17]. The drawback is that the limit of possible control
of information technology have increased rapidly at an actions is constrained. For example, one cannot insignifi-
unusual level as IT establishers including, Microsoft, IBM, cantly grow a VM to be larger than the PM it is deployed
Google, and various large corporations have proliferated on [18]. Resource allotment is one of the huge issues in
DC to assist their cloud computing services. Such DCs are cloud DC because of resource heterogeneity. Reinforce-
connected to numerous servers and switches that utilize ment learning is used for load balancing, handover Control,
huge amounts of energy, thus elevating the operating and resource allocation [19] [20, 21]. With the tremendous
expenses and maximizing the carbon dioxide emission into growth of applications over DC, optimization of such
the environment [5]. The cloud DCs should be connected utilities over cloud DC is a very challenging task. One such
with an efficient resource provisioning scheme to elevate fundamental advancement is the cloud virtualization
utility consumption and to guarantee the non-violation of advancement in which VMs are represented to PMs. Var-
service level agreement (SLA) terms. The prior data on ious optimization methodologies are developed by many
resource requirements can enhance resource utilization and researchers so far depending upon the diverse techniques,
minimize SLA violations [6]. some of them utilize heuristics whereas others utilize
Virtualization allows cloud DCs to share physical metaheuristics in which VMs are represented by PMs [22].
machines (PM’s) resources among various tenants and this The core aim of this research is to establish an efficient
technology can consolidate different VMs into the same model based on optimization-enabled VM scaling-based
PM by means of server consolidation and then slash down load distribution and optimal switching strategy in cloud
the inactive PMs to save energy. Virtual Machine Place- DC using the proposed PTaylor-MRFO _ ACNN?PAOA.
ment (VMP) is arranged on PMs in a DC and the two main In this research, either horizontal scaling or vertical scaling
fundamental goals of VMP are to reduce power con- can be used to resolve the overloading issue. Initially, the
sumption as well as resource wastage [7]. By performing network structure of the cloud is described based on the fat
virtualization [8] to the physical servers, VMs are formed tree model and the load distribution of cloud DC is per-
to serve cloud user’s requests [9]. VM consolidation formed utilizing P-Taylor MRFO by considering the mul-
[10, 11] is a mechanism to reduce the count of active tiple objectives, like power, load, latency, and bandwidth in
physical servers by transferring VMs from underutilized to which Taylor MRFO is already designed by inheriting the
one server to another and then slashing down the features of Taylor series concept with Manta Ray Foraging
underutilized users. If the physical server is in an Optimization (MRFO). Accordingly, the switching of DC
overutilization condition, then migration is needed. A to the desired mode is carried out utilizing ACNN. If the
physical server is overutilized if the host has no free space. system is overloaded, either horizontal scaling or vertical
Owing to this reason, certain VMs must be migrated into scaling is performed depending upon predefined threshold
loaded physical servers [12, 13]. Since virtualization elastic scaling. The vertical scaling is accomplished by
enables many VMs to run over on one hardware model, it including CPU, and RAM in the model. On the other hand,
protects against the issue of server spreading. In certain horizontal scaling is carried out by adding machines to the
cases, the utilities needed by the applications may not be PM using the proposed PAOA based on the horizontal cost
enough to perform all information. In order to overcome that relies on CPU, memory, and hard disk.
such issues of offering demand resources, VM scaling The highlight of this research is shown as follows:
comes forward to offer resources to the VMs when there is
• Proposed PTaylor-MRFO _ ACNN?PAOA: An effica-
a sudden increase in the workload [14]. This scaling can
cious model for VM scaling-based load distribution and
enhance the resource utilization, but it is sparse by inactive
optimal switching strategy in cloud DC is proposed
resources of the physical server.
using PTaylor-MRFO _ ACNN?PAOA. Here, load
Scaling of allotted resources in a cloud model can be
distribution is accomplished through PTaylor MRFO
performed in two ways, namely horizontal scaling and
and ACNN is employed to carry out the optimal
vertical scaling. Horizontal scaling is one that allots or de-
switching strategy in cloud DC. Moreover, both vertical
allots VMs in which the resource allocation varies in dis-
and horizontal scaling is performed using thedesigned
crete steps. The major benefit is that it can add or remove
PAOA, which is derived from the unification of POA
more and more VMs. On the other hand, it is very slow
and Adam optimization.
while deploying a new VM and it consumes more time.
Other than this disadvantage, the time delay is typically The residual section of this investigation article is arranged
varying [15, 16]. Vertical scaling is to adjust the size of as follows: Sect. 2 enumerates the motivation part by
already allotted resources by varying the count of memory analyzing the review of existing models corresponding to
or CPU allotted to previously deployed VM. This provides VM scaling techniques along with their pros and cons that

123
Wireless Networks (2024) 30:1085–1105 1087

stimulate the researchers to model an efficient framework. convergence. Moreover, an adaptive plant intelligent
The system model of fat tree-based cloud storage is enu- behavior optimization (APIBO) was employed to reallo-
merated in Sect. 3 and the proposed strategy for VM cate the unsuccessful workload to obtain fault tolerance.
scaling-based load distribution and optimal switching The simulation results proved the designed model was very
strategy in cloud DC is explained in Sect. 4. The results effective with respect to the response time, resource con-
part is discussed in Sect. 5 and the research conclusion is tention, throughput, execution time, and resource utiliza-
revealed in Sect. 6. tion. To minimize the energy as well as the response
duration of the data center is still considered as a major
challenging issue.
2 Motivation Li et al. [26] designed an effective model by exploiting
the massive data in the data center (DC) to stabilize the
This part frames the major motivation behind this proposed control policy. In order to perform such process, the
model of VM scaling-based load distribution and optimal cooling control policy was impelled into an energy cost
switching strategy in cloud DC by reviewing the pros and minimization problem with temperature issues and com-
cons of the existing models associated with VM scaling in bined with the Deep Reinforcement Learning (DRL)
cloud DC. paradigm. More particularly, the method introduced an
end-to-end cooling control algorithm (CCA), which was
2.1 Literature review mainly depend upon actor critic framework. Here, an
evaluation structure was trained to recognize the energy
This section enumerates the overall review of diverse expense counter. The results provided by this method were
techniques related to the best switching strategy in cloud superior and it successfully learned the model dynamics
DC. Zhou, et al. [3] presented an energy-saving approach from the monitoring information, thereby enhancing the
depending on Artificial Bee Colony (ABC) for a Fat tree. cooling effectiveness. Azizi et al. [7] presented an effective
Besides, the linear programming model was introduced heuristic algorithm called MinPR that concentrates more
depending upon multi-commodity network flow. The on power consumption and resource wastage optimization
optimal fitness parameter for energy utilization was to overcome the issues of energy consumption and resource
achieved using the optimization model ABC. The experi- utilization in the cloud DC. This model reduced the overall
mental outcomes proved that the optimal path was attained power utilization by decreasing the number of active PMs
by the developed ABC algorithm, thereby providing and preferred the high-powered machine. In addition, it
superior results in the optimization of energy consumption. minimized resource wastage by increasing and stabilizing
Priya and Gnanasekaran [23] designed an effective strategy resource utilization within the PMs. In order to attain such
named Efficient Load Optimization and Resource Mini- fundamentals, Resource Usage Factor model was devel-
mization (ELORM) algorithm in order to optimize the oped that balances the allocation of VMs on PMs utilizing
tasks at various hybrid P2P cloud data center areas and reward and penalty models. The designed model consid-
diverse users in the cloud environment. Here, the resources erably reduced the overall power utilization and resource
were well distributed depending upon the load forecasting wastage of DC. However, it failed to consider the similarity
in DCs. This model improved load optimization by regu- among virtual machines and data center network topology.
lating reliability and stability between the consumer base Huang et al. [17] devised an allocation technique for
and DC at the time of data transmission. Wei et al. [24] optimization depending on user requirements in a cloud
developed an improved Ant Colony Optimization (ACO) DC. Initially, the method allotted the VM’s to the respec-
with an adaptive parameter setting, named AP-ACO to tive PM’s based on the user requests, utilization of hard-
maintain its fast convergence and search capability. The ware resources and throughput of PM’s. Depending upon
performance of the designed model was validated in the workload, the PM’s load that hosts the VM’s was very
comparison with the First Fit Decreasing (FFD) and Ran- high. It is necessary to calculate the CPU utilization
dom algorithm. The developed scheme was highly desir- threshold in order to take the migration decision. The
able to various connections and highly effective for diverse developed approach delivered better performance in
congestion patterns. Mishra and Manjula [25] designed an accordance with the overall energy consumption. It failed
effective load balancing algorithm named Meta-heuristic to address different topologies, like B-cube and hypercube
based multi-objective optimization in a wide-scale plat- topologies. Soltanshahi et al. [5] modeled a novel approach
form of cloud computing. The chore of this research was to for the allotment of VM’s to physical hosts in cloud DCs
allocate the works to groups utilizing a multi-objective utilizing the Krill Herd algorithm. This strategy effectively
memetic algorithm (MOMA). Here, a local search minimized energy utilization by stabilizing the consolida-
approach was utilized to minimize the possibility of earlier tion of VM’s and slashed down the inactive servers without

123
1088 Wireless Networks (2024) 30:1085–1105

degrading the service quality. The major issue faced by this offers better communication among external paradigms
method was that it was inefficient in diagnosing the host and DC and then the information is redirected in the
overload utilizing deep learning algorithms. aggregation layer. The application-oriented requests from
DC are forwarded to the servers from the core layer.
2.2 Research gaps However, the switches present in the core layer are costly
in nature and consist of low upstream ports hitched to the
• In [24], the devised model achieved remarkable results outside world and downstream ports related to the switches
in both power consumption and communication cost. of the aggregation layer. Moreover, the aggregation
However, it did not slash down the computational time. switches are highly reliable for interaction and the edge
In the devised scheme, the best solution is finding based switches are highly meant for authorizing the servers.
on load, running time, and energy. So that, the com-
putational time of this method is low.
• The ELORM algorithm modeled in [23] considerably 4 Proposed Ptaylor-MRFO_ACNN1PAOA
reduced the resource utilization and response time but, model for VM scaling-based load
it failed to satisfy the user demands accurately. In the distribution and optimal switching
devised scheme, the optimal switching is done by strategy in cloud DC
ACNN, which offered superior accuracy with maxi-
mum processing speed. In recent days, research has concentrated more on the
• It is very significant to highlight the specific issues of globally optimal solution in virtualized data centers and
both the operators and parameters in the designed this makes a lot of insignificant migration and cannot be
model introduced in [25]. In addition, the method also liable for large-scale cloud computing environments. While
failed to minimize the cost. In the devised scheme, the constructing modern software, care must be taken to enable
inherent features of MRFO are incorporated into the applications to scale them depending upon the user’s
Taylor series and POA for offering a better solution in demands. Many organizations select to host their
load distribution, thereby minimizing the cost and it microservices architectures in cloud DCs to cut down the
also enhances the efficiency of the model. expenses. To counterpart this issue, this research proposes
• Owing to the tremendous evolution of cloud computing, an effective strategy based on optimization-enabled VM
DCs face increasing power utilization and electricity scaling-based load distribution and optimal switching
bills. This significant rise in power consumption has strategy in cloud DC. In this research, either horizontal
confined the viable establishment of cloud services and scaling or vertical scaling is utilized to encounter the
it complicated the DC operators. Hence, energy overloading issue. The network structure of the cloud is
efficiency has become a significant parameter in high- defined based on the fat tree cloud model and the load
scale DC. In the devised scheme, the energy is the one allocation of cloud DCs is carried out depending upon
of the parameter for evaluating the best solution. P-Taylor MRFO by considering the multi-objectives, such
as load, power, latency, and bandwidth. Depending upon
load allocation, the switching of DC to the desired mode is
done by employing ACNN. If the system is overloaded
3 System architecture of fat tree-based
either vertical scaling or horizontal scaling is done based
cloud storage
on a predefined threshold elastic scaling in which vertical
scaling is carried out by including CPU, and RAM to the
This section expounds on the fat tree-based cloud storage
model. In the contrary, horizontal scaling is done by
model depicted in Fig. 1, which is comprised of three
including nodes to VMs using the proposed PAOA
layers, namely core, aggregation, and edge layer. The core
depending upon horizontal cost that depends on the CPU,
layer is constructed with a number of PMs and VMs are
memory, and hard disk. However, the PAOA is newly
impelled in the aggregation layer, wherein switches are
devised by consolidating POA and Adam optimiza-
constructed within the edge layer. The fat tree has a
tion [27]. Figure 2 shows the schematic diagram of the
number of pods and each includes switches of double
designed model for VM scaling and optimal switching
layers, namely aggregation and edge layer. In the latter
strategy in cloud DC.
one, switches of each port are hitched to the servers and the
residuals are linked with switches in the aggregation. In the
4.1 Model parameter
core layer, every single port is connected with all switches
utilizing switches of the aggregation layer. Accordingly,
The fat tree-based cloud model is constructed for assigning
data broadcasting is performed using switches and this
resources in a cloud environment and it is depicted in

123
Wireless Networks (2024) 30:1085–1105 1089

Fig. 1 Fat tree-based cloud storage model

Fig. 1. The main chore is to find the best resources that are Em ¼ Ech þ numl  El þ RnumEconfigm o  Econfigmo
very efficient and significant in allocating the resources to
ð3Þ
each VM thus; it reduces energy utilization and increases
energy efficiency. Let us assume the cloud model with a set Here the energy consumption of chassis and line cards are
of PM and it is denoted as P ¼ fP1 ; P2 ; . . .Pa ; . . .Px g, specified as Ech and El , numl shows the amount of line
where Pa denotes the ath PM, such that 1  a  x and each cards opened by switches and Econfigm o implies the energy
PM consists of number of VM’s and the VM contained in consumption in the port at speed config o.
the ath PM is represented as In order to save maximum energy, the congestion is
V ¼ fV1a ; V2a ; . . .Vba ; . . .Vya g; 1  b  y. Here, implies the concentrated more on switches and then the ports are tuned
total number of VM’s present in ath PM. Each VM consists finely. The inactive state of the switches in the active mode
of various switches and it is modeled using the below is turned into either off or sleep mode with energy uti-
expression, lization of switches given by,
8
Sab ¼ fS1 ; S2 ; . . .Sn ; . . .Sm g ð1Þ < Eint þ numact  Eport ; ACTIVE MODE
>
Em ¼ Esleeping  0; ; SLEEPING MODE
Here Sn implies the nth switch present in the bth VM of ath >
:
0 ; OFF MODE
PM and Sm shows the overall count of switches present in
bth VM. ð4Þ
Each switch is selected for optimizing the data center by The overall ports opened by switches are signified as
considering some parameters, like power, bandwidth, numact and the stable energy consumption of ports is
latency, and load and the expression is given as follows, implied as Eport . The consolidated energy consumption of
A ¼ fBn ; Wn ; Ln ; Hn g ð2Þ chassis and line cards is represented as Eint and the energy
utilization of the switches at the sleep model is signified as
Here Bn and Wn symbolizes the bandwidth and power of nth Esleeping .
switch, the load and latency of nth switch is indicated as Ln
and Hn , respectively.
4.1.2 Load model
4.1.1 Energy model
The load of the switch is estimated depending upon the
proportion of resources consumed by the switch to the
Based on the working condition, the cloud model consumes
resource strength and it is computed as follows,
a certain amount of energy [3] and the working mode of  
switches can be categorized into three modes, namely Rutilize ðnÞ
LSn ¼ ð5Þ
active mode, off mode, and sleep mode. The static energy Rstrength ðnÞ
consumption generated by the switches is expressed as,

123
1090 Wireless Networks (2024) 30:1085–1105

Fig. 2 Block diagram of the

proposed model for VM
scaling-based load distribution
and optimal switching strategy
in the cloud datacenter

Here, the resource exploited by nth switch and the resource 4.2 Load distribution model using proposed
strength of the individual switch is signified as Rutilize ðnÞ P-Taylor MRFO
and Rstrength ðnÞ, respectively.
The expression for resource utilized and resource Typically, the load allocation in the cloud is performed
between the physical hosts and VM. This technique allows
strength of nth switch is formulated as follows,
the workloads evenly between all the nodes. The significant
Rutilize ðnÞ ¼Butilize
n þ Wnutilize þ Lutilize
n þ Hnutilize ð6Þ limitation in the cloud model is the load distribution, which
is highly useful for better service quality, limited service
Rstrength ðnÞ ¼Bstrength
n þ Wnstrength þ Lwarm þ Hnstrength ð7Þ
time, and reduces expense. The schematic view of the load
allocation system wielding designed P-Taylor MRFO and

123
Wireless Networks (2024) 30:1085–1105 1091

the optimal switching strategy utilizing ACNN is illus-

trated in Fig. 3. Let us consider that the user assigns the
task to the cloud, where the monitoring and job scheduling
are carried out. Due to the insufficient processing strength
of VM, then the calling of the load scheduler is performed.
The scheduling and task distribution is done at the load
scheduler using devised P-Taylor MRFO. After that, the
switching scheduler is employed to vary a single process to
another by preserving all conditions of implementing tasks. Fig. 4 Solution encoding
Finally, the switching state is achieved utilizing ACNN and
it is then subjected to cloud model monitoring and illustrates the solution encoding for load distribution using
scheduling. The designed P-Taylor MRFO enhances the the proposed P-Taylor MRFO.
load distribution in the cloud platform by managing the
stability between the DC and user base. In addition, it 4.2.2 Fitness function
divides the resources based on the load in the data centre.
Nevertheless, the P-Taylor MRFO is newly achieved by the The fitness factor is exploited to find the best solution for
combination of the Taylor concept with POA and MRFO. load distribution and it is comprised of some parameters
that include load, running time, and energy. The fitness
4.2.1 Solution encoding function is computed as follows,
1
The solution representation is a significant factor to esti- = ¼ Rek¼1 Tk ½LSnk þ Ekn  ð8Þ
mate the optimal solution for encountering the optimization e
problems. In order to select the optimal solution for load Here the running time of the task k is expressed as Tk , LSnk
distribution and to process the tasks by selecting the opti- symbolizes the load of nth switch and the energy of nth
mal switches, P-Taylor MRFO is proposed. For optimal switch is denoted as Ekn .
load distribution, every job is assumed and its respective
switches are adopted. Depending upon the fitness, the 4.2.3 Algorithmic procedure of P-Taylor MRFO
solution set achieves the best jobs. Suppose the count of
jobs be implied as Je and overall count of switches for MRFO [28] is a bio-inspired optimization algorithm, which
performing each job ranging from 1 to Q. Figure 4 is inspired mainly based on the intelligent characteristics of

Fig. 3 Load distribution system

using devised P-Taylor MRFO
and optimal switching condition
utilizing ACNN

123
1092 Wireless Networks (2024) 30:1085–1105

Manta rays. It is one of the well-known marine organisms. 1

Mi ðt þ 1Þ ¼
These organisms do not have teeth but, they feed on ½1  2ð1  q  aÞ
2 2 33
plankton, which is a microscopic organism. Here, three 1:3591Mi ðt  1Þ þ 1:359Mi ðt  2Þ  0:6795Mi ðt  3Þ
foraging strategies are followed by the Manta ray. On the 6 26 77
6 4 þ0:2259Mi ðt  4Þ  0:0555Mi ðt  5Þ þ 0:0104Mi ðt  6Þ 5 7
6 7
other hand, POA [29] describes the hunting behavior of 4 1:38e3 Mi ðt  7Þ þ 9:92e5 Mi ðt  8Þ 5
pelicans, which are the search agents searching for food ½1  q  a þ Mbest ðq þ aÞ
sources. By incorporating the inherent features of MRFO ð12Þ
into the Taylor series [30] and POA can offer a better
solution in load distribution, thereby minimizing the cost By subtracting Mi ðtÞ on both sides, the equation becomes
and it also enhances the efficiency of the model. The 1
M i ðt þ 1Þ  Mi ðtÞ ¼
algorithmic steps included in P-Taylor MRFO are descri- ½1  2ð1  q  aÞ
2 2 33
bed as below: 1:3591Mi ðt  1Þ þ 1:359Mi ðt  2Þ  0:6795Mi ðt  3Þ
6 26 77
Step 1: Initialization of Manta ray population 6 4 þ0:2259Mi ðt  4Þ  0:0555Mi ðt  5Þ þ 0:0104Mi ðt  6Þ 5 7
6 7  Mi ðtÞ
4 3 5
1:38e Mi ðt  7Þ þ 9:92e Mi ðt  8Þ 5
Let us initiate the population of manta ray in g dimen-
½1  q  a þ Mbest ðq þ aÞ
sional search area with j count of members and it is given
below, ð13Þ

M ¼ fM1 ; M2 ; . . .Mi ; . . .Mj g ð9Þ From POA, the standard expression of POA is derived
depending upon the case =pe \=i ,
where Mi denotes ith manta ray in jth population and the
population is randomly distributed in a g dimensional area. Mi ðt þ 1Þ ¼Mi ðtÞ þ rand:ðPei  r:Mi ðtÞÞ ð14Þ
Step 2: Evaluate the fitness factor
Mi ðt þ 1Þ ¼Mi ðtÞ þ randPei  rand:r:Mi ðtÞ ð15Þ
It is imperative to derive the best solution for load dis-
tribution and it is computed utilizing Eq. (8). Mi ðt þ 1Þ ¼Mi ðtÞ½1  randr þ randPei ð16Þ
Step 3: Chain foraging strategy
Mi ðtÞ½1  randr ¼Mi ðt þ 1Þ  randPei ð17Þ
The very first step is the chain foraging step. If fifty or
more manta rays begin to search for a food source, they Mi ðt þ 1Þ  randPei
form a line. On the other hand, smaller male manta rays are Mi ðtÞ ¼ ð18Þ
ð1  randrÞ
sticked on the back of the female manta ray to catch the
plankton. Usually, manta rays have the capability to notice Substituting Eq. (18) in Eq. (13), the expression becomes,
the location of plankton. If the concentration of plankton is Mi ðt þ 1Þ  Mi ðtÞ ¼
1
maximum, then the location of the manta ray must be the ½1  2ð1  q  aÞ
2 2 33
best solution. Here, the manta ray considers the best 1:3591Mi ðt  1Þ þ 1:359Mi ðt  2Þ  0:6795Mi ðt  3Þ
6 26 77
solution determined so far and in every iteration each 6 4 þ0:2259Mi ðt  4Þ  0:0555Mi ðt  5Þ þ 0:0104Mi ðt  6Þ 5 7
6 7
4 3 5
1:38e Mi ðt  7Þ þ 9:92e Mi ðt  8Þ 5
manta ray is upgraded using the optimal solution deter-
½1  q  a þ Mbest ðq þ aÞ
mined. The chain foraging strategy is mathematically Mi ðt þ 1Þ  randPei
modeled as follows, 
ð1  randrÞ
Mig ðt þ 1Þ ð19Þ

Mig ðtÞ þ q:ðMbest
g
ðtÞ  Mig ðtÞÞ þ a:ðMbest
g
ðtÞ  Mig ðtÞÞ; i ¼ 1 Mi ðt þ 1Þ 1
¼ Mi ðt þ 1Þ þ ¼
Mi ðtÞ þ q:ðMi1 ðtÞ  Mi ðtÞÞ þ a:ððMbest ðtÞ  Mig ðtÞÞ; i ¼ 2
g g g g
1  randr ½1  2ð1  q  aÞ
2 2 33
ð10Þ 1:3591Mi ðt  1Þ þ 1:359Mi ðt  2Þ  0:6795Mi ðt  3Þ
6 26 77
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 6 4 þ0:2259Mi ðt  4Þ  0:0555Mi ðt  5Þ þ 0:0104Mi ðt  6Þ 5 7
6 7
a ¼ 2  q  j logðqÞ j ð11Þ 4 3 5
1:38e Mi ðt  7Þ þ 9:92e Mi ðt  8Þ 5
½1  q  a þ Mbest ðq þ aÞ
Here Mig ðtÞsignifies the position of ith individual at time randPei
þ Mi ðtÞ þ
instant t in gth area and q represents the uneven number ð1  randrÞ
that lies in the range of [0, 1], the weight coefficient is ð20Þ
indicated as a and the plankton with maximum concen-
g
tration is denoted as Mbest ðtÞ. By considering the above two
conditions, the updated solution of Taylor MRFO is
derived.
The update expression for Taylor MRFO is given by,

123
Wireless Networks (2024) 30:1085–1105 1093

Mi ðt þ 1Þð1  randrÞ þ Mi ðt þ 1Þ 1
¼ Here Mi ðt þ 1Þ shows the location of ith solution at itera-
1  randr ½1  2ð1  q  aÞ
2 2 33 tion ðt þ 1Þ, r is a parameter equal to 1 0r 2. In addition,
1:3591Mi ðt  1Þ þ 1:359Mi ðt  2Þ  0:6795Mi ðt  3Þ
6 24 þ0:2259M ðt  4Þ  0:0555M ðt  5Þ þ 0:0104M ðt  6Þ 7
6
6
i i i 57
7
Mi ðt  1Þ, Mi ðt  2Þ, Mi ðt  3Þ, and Mi ðt  4Þ represents
6 7 the location of ith solution at iteration t  1; t  2; t  3;
4 1:38e3 Mi ðt  7Þ þ 9:92e5 Mi ðt  8Þ 5
½1  q  a þ Mbest ðq þ aÞ and t  4.
þ Mi ðtÞ þ
randPei Step 4: Cyclone foraging mechanism
ð1  randrÞ The second step is the cyclone foraging strategy in
ð21Þ which if the plankton concentration is maximum, pairs of
Mi ðt þ 1Þ½1  randr þ 1 1 manta rays accumulate. Their ends connect up with the
¼ heads in a spiral shape to create a spiraling end in the
1  randr ½1  2ð1  q  aÞ
2 2 33 center of the cyclone and it refined the water sweeps up
1:3591Mi ðt  1Þ þ 1:359Mi ðt  2Þ  0:6795Mi ðt  3Þ
6 24 þ0:2259M ðt  4Þ  0:0555M ðt  5Þ þ 0:0104M ðt  6Þ 7
6
57
towards the water surface level. This makes the plankton to
6 i i i 7
6 7 get into their wide mouths. If a group of manta rays
4 3 5
1:38e Mi ðt  7Þ þ 9:92e Mi ðt  8Þ 5
identifies a food source, they will create a lengthy foraging
½1  q  a þ Mbest ðq þ aÞ
randPei
chain and move toward the food source in a spiral shape.
þ Mi ðtÞ þ The spiral shaped-movement is mathematically framed as
ð1  randrÞ
follows,
ð22Þ 
Mi ðt þ 1Þ ¼ Mbest þ q:ðMi1 ðtÞ  Mi ðtÞÞ þ ecw :Cosð2pwÞ:ðMbest  Mi ðtÞÞ
1
Mi ðt þ 1Þ ¼ ð Ni ðt þ 1Þ ¼ Nbest þ q:ðNi1 ðtÞ  Ni ðtÞÞ þ ecw :Cosð2pwÞ:ðNbest  Ni ðtÞÞ
½1  2ð1  q  aÞ
2 2 33 ð26Þ
1:3591Mi ðt  1Þ þ 1:359Mi ðt  2Þ  0:6795Mi ðt  3Þ
6 26 77
6 4 þ0:2259Mi ðt  4Þ  0:0555Mi ðt  5Þ þ 0:0104Mi ðt  6Þ 5 7 Here w is an uneven value that lies in the limit of [0, 1].
6 7
4 1:38e3 Mi ðt  7Þ þ 9:92e5 Mi ðt  8Þ 5 The mathematical formula of cyclone foraging can be
½1  q  a þ Mbest ðq þ aÞ shown as,
randPei 1  radnr
þ Mi ðtÞ þ Þð Þ Mig ðt þ 1Þ ¼ fMbest
g g
þ q:ðMbest ðtÞ
ð1  randrÞ 2  randr ð27Þ
 Mig ðtÞÞ þ b:ðMbest
g
ðtÞ  Mig ðtÞÞ; i ¼ 1
ð23Þ
stþ1
b ¼ 2eq1 s :Sinð2pq1 Þ ð28Þ

2 2 3 3
1:3591Mi ðt  1Þ þ 1:359Mi ðt  2Þ  0:6795Mi ðt  3Þ
6 6 þ0:2259M ðt  4Þ  0:0555M ðt  5Þ þ 0:0104M ðt  6Þ 7 7
6 6 i i i 7 7
6 26
6 3 5
7
7 7
½ð2  randrÞð1  randrÞ6
6 4 1:38e M i ðt  7Þ þ 9:92e M i ðt  8Þ 5 7
7
6 7
4 ½1  q  a þ Mbest ðq þ aÞ 5
þmi ðtÞð1  2ð1  q  aÞÞð2  randrÞð1  randrÞ þ randpei ð1  2ð1  q  aÞÞ
Mi ðt þ 1Þ ¼ ð1  randrÞ
1  2ð1  q  aÞð1  radnrÞð2  randrÞ
ð24Þ

Here s and b represents the maximum count of iterations

The update solution of P-Taylor MRFO is mathematically and the weight coefficient. Additionally, implies the
given as follows, uneven number that exists in the limit of [[0, 1].
Mi ðt þ 1Þ ¼
1 Overall individuals unevenly do the search process in
ð1  2ð1  q  aÞÞð2  randrÞ
2 2 2 333 accordance with the source as their reference location.
1:3591Mi ðt  1Þ þ 1:359Mi ðt  2Þ  0:6795Mi ðt  3Þ
6 6 26
6 ð2  randrÞð1  randrÞ6 4 þ0:2259M i ðt  4Þ  0:0555M i ðt  5Þ þ 0:0104Mi ðt  6Þ
777
577
Therefore, cyclone foraging exhibits better exploitation for
6 6 77
6
6
4 1:38e3 Mi ðt  7Þ þ 9:92e5 Mi ðt  8Þ 57
7
considered space with the optimal solution determined.
4 ½1  q  a þ Mbest ðq þ aÞ 5
This is exploited to enhance the exploration phase. It is
þmi ðtÞð1  2ð1  q  aÞÞð2  randrÞð1  randrÞ þ randpei ð1  2ð1  q  aÞÞ
imperative to seek a new location apart from the current
ð25Þ best solution by allocating a new uneven location in the

123
1094 Wireless Networks (2024) 30:1085–1105

whole search space as their reference location. The math- SL. No Pseudo-code of proposed P-Taylor MRFO
ematical expression for the exploration phase is given as
follows, 19 Update solution using Eq. (26)
g 20 Then Mbest ¼ Mi ðt þ 1Þ
Mrand ¼ LBg þ q:ðUBg  LBg Þ ð29Þ
21 end for
Mig ðt þ 1Þ ¼ g
fMrand g
þ q:ðMrand ðtÞ 22 end while
ð30Þ
 Mig ðtÞÞ þ g
b:ðMrand  Mig ðtÞÞ; i ¼1 23 Return optimal solution

Here the uneven location created in the search area is

g
implied as Mrand , the upper and lower bounds of the gth
dimension are expressed as UBg and LBg , respectively.
Step 5: Somersault foraging strategy 4.3 Optimal switching strategy based on ACNN
This somersault foraging strategy is fascinating in nat-
ure. If manta rays determine a plankton, they will perform a Once the load distribution of cloud DC is performed using
sequence of backward somersaults encircling the plankton the proposed P-Taylor MRFO, the optimal switching
to attract it towards each individual. Due to the cyclic strategy is carried out using ACNN. If the load found is
movement of somersault, food intake can be optimized underloaded, ACNN is exploited to perform the optimal
easily. The location of the food is considered as the centre. switching strategy. If the load is overloaded, check it with
By keeping the centre as a major point, all the individuals the threshold value and then demand either horizontal
begin to swim into and from motion encircling the centre scaling or vertical scaling accordingly based on the pro-
and somersault to a current location. This nature is math- posed PAOA.
ematically defined as given below,
Mig ðt þ 1Þ ¼ Mig ðtÞ þ d:ðq2 :Mbest
g
 q3 :Mig ðtÞÞ; i ¼ 1 4.3.1 Architecture of ACNN

ð31Þ ACNN [31] is a supervised classifier, which mainly con-

where d is the somersault parameter, q2 and q3 are the sists of two operating modules, namely actor and critic
uneven numbers that exist in the limit of [0, 1] modules. The fundamental aim of this ACNN is used to
Step 6: Termination predict the switching state and it facilitates the processing
The steps are continued and iterated until the optimal speed delivered with superior accuracy. On the other hand,
solution is obtained so far. Algorithm 1 elucidates the the critic module offers a control action that provides
pseudo-code of devised P-Taylor MRFO. effective forecasting. Figure 5 illustrates the structure of
ACNN.
(i) Actor module: Let us consider the input of the actor
SL. No Pseudo-code of proposed P-Taylor MRFO mode as, which is a system state in order to approximate
1 Input: Mi ðtÞ; s the optimal action and the output derived is notated as. The
2 Output: Mi ðt þ 1Þ actor module is performed perfectly using the parameter-
3 Begin
ized technique, like neural network. In actor module, a
4 Initialize the population
simple feed-forward neural structure with only one hidden
layer is considered for operation. The output of the actor is
5 Formulate objective factor using Eq. (8)
expressed as follows,
6 for i ¼ 1 to Mm do
7 if rand\0:5 then // cyclone foraging OAa ðsÞ ¼ WeTt Tt Tt
Aa ðsÞgðIAa ðsÞXAa ðsÞÞ ¼ WeAa ðsÞrAa ðsÞ ð32Þ
t
8 if smax \rand then
Here XAa ðsÞ is the input, the weight matrix among input
9 Update solution utilizing Eq. (25)
and hidden layer is signified as IAa ðsÞ, whereas WeAa ðsÞ
10 Else
denotes the weight amongst hidden and output layer.
11 Update solution using Eq. (23)
Moreover, the activation parameter rðÞ is represented as
12 end if follows, which is a hyperbolic tangent function.
13 else // chain foraging
Update solution using Eq. (21)
ez  ez
14 rðzÞ ¼ ð33Þ
15 end if ez þ ez
16 Then Mbest ¼ Mi ðt þ 1Þ (ii) Critic module: Critic module imitates the control action
17 //Somersault foraging considered under the latest condition by formulating the
18 for i ¼ 1 to Mm do

123
Wireless Networks (2024) 30:1085–1105 1095

Fig. 5 Architecture of ACNN

further accumulative reward-to-go measure expressed as to do a number of tasks simultaneously. This can be
follows, accomplished to implement the strategy of scaling, which
is a remarkable solution that increases the overall
RewardðsÞ ¼ RTt d
d¼0 v Rðs þ d þ 1Þ ð34Þ
throughput. If the predicted load is overloaded one, either
where 0\v\1 is a discount factor, denotes the time and Tt horizontal or vertical scaling will be demanded and this
indicates the terminal step. Hence, a same three layer feed- process is carried out using the proposed PAOA.
forward neural structure is employed to approximate the
Reward(s) in Eq. (34). 4.4.1 Demand for horizontal scaling
The output of critic module is an approximation of
Reward(s), which is defined by, The horizontal scaling [32] is performed by maximizing
the hardware in the machinery. If the load is higher than the
CðsÞ ¼ WeTt Tt
Cc ðsÞrðICc ðsÞXCc ðsÞÞ ¼ WeTt
Cc ðsÞrCc ðsÞ ð35Þ
threshold value, the system is demanded for horizontal
Here XCc is the input of the critic, the weight matrix scaling by adding VM to PM. This is mainly the process of
Tt
amongst input and hidden layer is implied as ICc . On the including more and more nodes in the current system.
other hand, the weight matrix among hidden and output
layers is specified as WeTt 4.4.2 Solution encoding
Cc ðsÞ.

4.4 Horizontal and vertical scaling using The solution encoding represents the solution vector or the
proposed PAOA optimal solution for a given optimization issue. When the
load is greater than the threshold measure, optimal VMs
There is a necessity to improve the performance of cloud are included to PM that facilitates horizontal scaling. Fig-
storage in both the domain of software and hardware in ure 6 portrays the solution encoding for horizontal scaling.
order to maximize the throughput. Hence, there is a high
demand for throughput profit when the system is requested

123
1096 Wireless Networks (2024) 30:1085–1105

Step 1: Initialization of pelican population

Each population member is considered as a candidate
3 2 4 2 solution in all population-based algorithms. Let us assume
the pelican population in a—search area and the population
Index of PM members are determined based on a matrix known as
Fig. 6 Solution encoding for horizontal scaling population matrix.
Z ¼ fZ1 ; Z2 ; . . .Zu ; . . .ZNn g ð42Þ
4.4.3 Fitness function
Here Zu refers the uth candidate solution among Nn number
The imperative function of the fitness factor is to estimate of total pelican population. At first, the members are evenly
the best VM added to PM while demanding the horizontal initiated depending upon the lower and upper bound of the
scaling and it is expressed as follows, problem represented using the below equation,

= ¼ Ryb¼1 ðDVM
b
1
þ DVM
b Þ
0
ð36Þ Zu;v ¼ LBv þ rand:ðUBv þ LBv Þ; u ¼ 1; 2; . . .; Nn ð43Þ

The cost of VMs [33] includes the memory cost, CPU, and where Zu;v refers to the measure of vth variable represented
hard disk. by uth candidate solution. rand specifies the uneven mea-
sure that lies in the value of [0, 1]. LBv and UBv is the vth
1 ;Mem
DVM
b
1
¼DVM
b þ DbVM1 ;CPU þ DVM
b
1 ;hd
ð37Þ lower bound and upper bound, respectively.
0 ;Mem Step 2: Determine the objective factor
DVM
b
0
¼DVM
b þ DbVM0 ;CPU þ DVM
b
0 ;hd
ð38Þ
The objective factor is employed to determine the best
The constraint expressions of integer programming are VM included to PM when the system is overloaded and it is
defined as follows: already formulated in Eq. (31).
Step 3: Exploration phase
Ryb¼1 UbMem Mem
Uneed ð39Þ
The exploration phase is simply said as the movement of
Ryb¼1 UbCPU CPU
Uneed ð40Þ pelicans toward their food source. At this phase, the pelican
finds the location of the food source and then gradually
Ryb¼1 Ubhd hd
Uneed ð41Þ moves close to the determined area. The chore part in POA
Here y0 and y1 depict the number of VM before and after is that the location of the prey is created evenly in the
search area and this nature is defined as,
resource expansion, respectively. Similarly, DVM
b
0
and DVM
b
1

represents the VM cost before and after the expansion Zu ðt þ 1Þ ¼Zu ðtÞ þ rand:ðPeu  r:Zu ðtÞÞ; =Pe \=u ð44Þ
change. The memory cost of bth VM is indicated as DMem b Zu ðt þ 1Þ ¼Zu ðtÞ þ randPet  rand:r:Zu ðtÞ ð45Þ
and UbMem is the memory size of bth VM. The CPU cost and
the CPU size of bth VM are signified as DCPU and UbCPU , Zu ðt þ 1Þ ¼Zu ðtÞ½1  randr þ randPet ð46Þ
b
respectively. The hard disk size of bth VM is notated as
The standard expression for Adam optimization is given
Ubhd and Dhd
b specifies the hard disk cost of bth VM. by,
qffiffiffiffiffiffiffiffiffi 
4.4.4 Algorithmic steps of PAOA b
Zu ðtÞ ¼ Zu ðt  1Þ  k GðtÞ= YbðtÞ þ n ð47Þ

POA is a swarm-based optimization algorithm that imitates Substituting Eq. (47) in Eq. (46),
the nature of the hunting behavior of pelicans. The pelican
is the largest creature with a long beak and large bag sit- Zu ðt þ 1Þ ¼ ½Zu ðt  1Þ
qffiffiffiffiffiffiffiffiffi 
uated in its throat that helps the pelican to catch and ð48Þ
b
k GðtÞ= b
Y ðtÞ þ n ð1  randrÞ þ randPet
swallow the prey easily. Generally, the pelican work in
groups and at instant of determining the prey location, it
dives toward the food source from a height of 10–20 m. Now, by considering the else condition from POA, the
The pelicans are skilled hunters and this has been consid- update solution becomes,
ered as an inspirational principle while designing POA. By Zu ðt þ 1Þ ¼Zu ðtÞ þ randðZu ðtÞ  Pet Þ ð49Þ
inheriting the features of Adam optimization [26] into POA
can deliver better performance to accomplish the task of Zu ðt þ 1Þ ¼Zu ðtÞ þ randZu ðtÞ  randPet ð50Þ
both horizontal and vertical scaling. Both scaling types are
Zu ðt þ 1Þ ¼Zu ðtÞð1 þ randÞ  randPet ð51Þ
effectively carried out using the proposed PAOA. The steps
involved in PAOA are expounded below, Substituting Eq. (42) in Eq. (46),

123
Wireless Networks (2024) 30:1085–1105 1097

Fig. 7 Solution encoding for

vertical scaling

Table 1 Experimental parameters Step 5: Termination

The aforementioned procedure is continued until it sat-
Parameter Value
isfies the determined criteria. The pseudo code of devel-
Learning rate 0.001 oped PAOA is explained in Algorithm 2.
Epsilon decay 0.995
Activation function Relu
SL. Pseudo code of developed PAOA
Loss MSE
No
Optimizer Adam
Lower bound 1 1 Input: Zu ðtÞ, population size Nn, number of iterations s
Upper bound 25 2 Output: Zu ðt þ 1Þ
Total number of iteration 10 3 Initialize the pelican population
4 Estimate the objective function using Eq. (31)
5 For t ¼ 1 : s
6 Generate the location of the prey randomly
7 For r ¼ 1 : Nn
Zu ðt þ 1Þ ¼ ½Zu ðt  1Þ
qffiffiffiffiffiffiffiffiffi  8 //Exploration phase
ð52Þ
b
k GðtÞ= YbðtÞ þ n ð1 þ randÞ  randPet 9 For v ¼ 1
10 Determine new solution of vth dimension using Eq. (39)
where Zu ðt þ 1Þ represents the position of uth solution at 11 End
iteration t þ 1, the position of uth solution at iteration t  1 12 Update the uth population member using Eq. (43) and
is indicated as Zu ðt  1Þ. The location of the prey and the Eq. (47)
random number is depicted as Pei and r, respectively. 13 //Exploitation phase
Step 4: Exploitation phase 14 For v ¼ 1
After reaching the water surface, the pelicans spread 15 Determine new solution of vth dimension using Eq. (48)
their wings to move the food source to swim in a shallow 16 End
water surface so that they can easily catch prey in their 17 Update the uth population member using Eq. (49)
throat bag. This behavior can be modeled to attain a better 18 End
solution and it is expressed as follows, 19 Update best solution
 t 20 End
Phase2
Zu;v ¼ Zuv þ P 1  ð2rand  1Þ:Zu;v ð53Þ
s 21 Return best solution
Phase2
Here Zu;v refers the new solution of uth pelican in vth
dimension depending upon phase 2, p is a fixed number,
and the maximum count of iterations is denoted as s.
During this phase, efficient upgrading has been utilized
4.4.5 Demand for vertical scaling
to either accept or reject the location of the pelican and it is
modeled using the below expression,
8 Vertical scaling is done by adding sufficient resources to
Phase2
>
< Zu ; =Phase2
u \=u ; the available physical machines. This can be achieved by
Zu ¼ ð54Þ adding memory, elevating the storage capacity and maxi-
>
:
Zu ; else mizing the virtual memory. If the load is below to the
threshold value then, the model is demanded for vertical
Here ZuPhase2 represents the new solution of uth pelican. scaling and it is performed using the proposed PAOA and

123
1098 Wireless Networks (2024) 30:1085–1105

Fig. 8 Analysis based on task size = 100, a Load, b power, c energy consumption, d latency

its algorithmic steps are already elaborated under load is below to the threshold measure, optimal resources
Sect. 4.4.4. are added to the PM that facilitates the vertical scaling.
Figure 7 portrays the solution encoding for vertical scaling.
4.4.6 Solution encoding

The solution encoding represents the solution vector or

optimal solution for a given optimization issue. When the

123
Wireless Networks (2024) 30:1085–1105 1099

Fig. 9 Assessment based on task size = 200, a load, b power, c energy consumption, d latency

4.4.7 Fitness function Here the cost of scaled CPU, memory, and hard disk is
depicted as DbVM1 ;CPU , DVM
b
1 ;Mem
, and DbVM1 ;hd , respectively.
The fitness factor is to estimate the best resources added to Moreover, DLlb specifies the license cost.
the VM while the system demands the vertical scaling and
it is defined as follows,
= ¼ Ryb¼1 DVM
b
1
þ DLl
b ð55Þ
i ;mem i ;CPU i ;hd
where DVM
b
i
¼ DVM
b þ DVM
b þ DVM
b

123
1100 Wireless Networks (2024) 30:1085–1105

Fig. 10 Assessment based on task size = 300, a load, b power, c energy consumption, d latency

5 Results and discussion 5.1 Experimental setup

The outcomes of devised PTaylor-MRFO_ACNN?PAOA The demonstration of the proposed PTaylor-

in regard to the measures are described in this part. Apart MRFO_ACNN?PAOA is executed in PYTHON with
from this, the comparative assessment of the proposed cloudsim tool utilizing PC having 10 OS, 4GB RAM, and
approach based on various task sizes is also elaborated in intel core-i3 processor. The parameters are provided in
this section. Table 1.

123
Wireless Networks (2024) 30:1085–1105 1101

Table 2 Comparative analysis

Task Methods/metrics ABC ELORM AP- MOMA- CCA MinPR Taylor-MRFO Proposed PTaylor-MRFO
size ACO APIBO ?ACNN _ACNN?PAOA

100 Load 0.839 0.783 0.720 0.643 0.584 0.430 0.348 0.298
Power 0.891 0.862 0.804 0.782 0.758 0.721 0.646 0.644
Energy 0.895 0.878 0.839 0.819 0.754 0.693 0.641 0.414
consumption
Latency 0.752 0.682 0.621 0.585 0.444 0.370 0.349 0.331
200 Load 0.837 0.792 0.760 0.705 0.652 0.379 0.336 0.329
Power 0.873 0.834 0.795 0.674 0.667 0.604 0.541 0.532
Energy 0.901 0.872 0.853 0.827 0.698 0.659 0.516 0.357
consumption
Latency 0.885 0.849 0.807 0.755 0.404 0.351 0.320 0.315
300 Load 0.872 0.841 0.724 0.641 0.557 0.404 0.362 0.352
Power 0.931 0.913 0.899 0.835 0.747 0.660 0.598 0.569
Energy 0.925 0.901 0.894 0.872 0.809 0.580 0.579 0.405
consumption
Latency 0.787 0.741 0.696 0.674 0.667 0.604 0.541 0.532

5.2 Evaluation metrics competing schemes, like ABC [3], ELORM [23], AP-ACO
[24], MOMA-APIBO [25], CCA [26], MinPR [7], and
The evaluation measures considered by the proposed Taylor-MRFO?ACNN.
scheme for estimation are described as follows:
5.4 Comparative assessment
5.2.1 Power
This section enumerates the comparative analysis of the
Power is metric that is used to define the total amount of proposed PTaylor-MRFO_ACNN?PAOA in terms of
power utilized by the system to do a particular task. evaluation metrics based on different task sizes by varying
the swarm size.
5.2.2 Load
5.4.1 Assessment based on task size=100
The load is computed depending upon the proportion of
resources utilized to the resource strength and it is com- Figure 8 depicts a comparative evaluation of the proposed
puted using Eq. (5). PTaylor-MRFO_ACNN?PAOA based on task size=100
by elevating the swarm sizes from 5 to 20. Figure 8a, b
5.2.3 Latency shows the assessment of PTaylor-MRFO_ACNN?PAOA
with respect to load and power, respectively. If the swarm
Latency is also referred as the response time, which is size is 20, the load obtained by the proposed scheme is
employed to calculate the amount of time required by the 0.298, whereas the power obtained by the proposed
system to process a user request. PTaylor-MRFO_ACNN?PAOA is 0.644. However, the
existing techniques yielded the power as 0.891 for ABC,
5.2.4 Energy consumption 0.862 for ELORM, 0.804 for AP-ACO, 0.782 for MOMA-
APIBO, 0.758 for CCA, 0.721 for MinPR, and 0.646 for
The total amount of energy utilized by the system model to Taylor-MRFO?ACNN. Figure 8c displays the evaluation
process the tasks is computed using Eq. (3). of the proposed model in accordance with energy con-
sumption. The energy consumed by the proposed PTaylor-
5.3 Competing techniques MRFO_ACNN?PAOA for swarm size 20 is 0.414J. The
evaluation done by the proposed PTaylor-
The performance improvement of the proposed PTaylor- MRFO_ACNN?PAOA in regard to latency is represented
MRFO_ACNN?PAOA is analyzed and compared with the in Fig. 8d. When the swarm size is 20, latency acquired by

123
1102 Wireless Networks (2024) 30:1085–1105

Table 3 Statistical analysis

Methods Load Power Energy consumption Latency
Best Mean Variance Best Mean Variance Best Mean Variance Best Mean Variance
Task size-100

ABC 0.839 0.835 0.004 0.891 0.888 0.003 0.895 0.890 0.005 0.752 0.750 0.002
ELORM 0.783 0.780 0.003 0.862 0.860 0.002 0.878 0.875 0.008 0.682 0.680 0.002
AP-ACO 0.720 0.717 0.003 0.804 0.800 0.004 0.839 0.835 0.004 0.621 0.618 0.003
MOMA-APIBO 0.643 0.640 0.003 0.782 0.780 0.002 0.819 0.813 0.005 0.585 0.580 0.005
CCA 0.584 0.580 0.004 0.758 0.755 0.003 0.754 0.750 0.004 0.444 0.440 0.004
MinPR 0.430 0.426 0.004 0.721 0.718 0.003 0.693 0.690 0.003 0.370 0.366 0.004
Taylor-MRFO?ACNN 0.348 0.345 0.003 0.646 0.643 0.003 0.641 0.638 0.003 0.349 0.345 0.004
Proposed PTaylor-MRFO 0.298 0.296 0.002 0.644 0.643 0.001 0.414 0.412 0.002 0.331 0.330 0.001
_ACNN?PAOA
Task size-200
ABC 0.837 0.834 0.003 0.873 0.870 0.003 0.901 0.898 0.004 0.885 0.880 0.005
ELORM 0.792 0.790 0.002 0.834 0.830 0.004 0.872 0.870 0.002 0.849 0.845 0.004
AP-ACO 0.760 0.756 0.004 0.795 0.790 0.005 0.853 0.850 0.003 0.807 0.805 0.002
MOMA-APIBO 0.705 0.700 0.005 0.674 0.670 0.004 0.827 0.824 0.003 0.755 0.751 0.004
CCA 0.652 0.650 0.002 0.667 0.662 0.005 0.698 0.694 0.004 0.404 0.400 0.004
MinPR 0.379 0.375 0.004 0.604 0.600 0.004 0.659 0.657 0.002 0.351 0.347 0.004
Taylor-MRFO?ACNN 0.336 0.332 0.004 0.541 0.538 0.003 0.516 0.513 0.003 0.320 0.318 0.002
Proposed PTaylor-MRFO 0.329 0.327 0.002 0.532 0.530 0.002 0.357 0.356 0.002 0.315 0.313 0.002
_ACNN?PAOA
Task size-300
ABC 0.872 0.869 0.003 0.931 0.928 0.003 0.925 0.920 0.005 0.787 0.783 0.004
ELORM 0.841 0.837 0.004 0.913 0.910 0.003 0.901 0.899 0.002 0.741 0.738 0.003
AP-ACO 0.724 0.720 0.004 0.899 0.895 0.004 0.894 0.890 0.004 0.696 0.693 0.003
MOMA-APIBO 0.641 0.638 0.003 0.835 0.830 0.005 0.872 0.870 0.002 0.674 0.670 0.004
CCA 0.557 0.553 0.004 0.747 0.743 0.004 0.809 0.805 0.004 0.667 0.664 0.003
MinPR 0.404 0.400 0.004 0.660 0.655 0.005 0.580 0.575 0.005 0.604 0.600 0.004
Taylor-MRFO?ACNN 0.362 0.360 0.002 0.598 0.595 0.003 0.579 0.576 0.003 0.541 0.538 0.003
Proposed PTaylor-MRFO 0.352 0.351 0.001 0.569 0.567 0.002 0.405 0.403 0.002 0.532 0.530 0.002
_ACNN?PAOA

the proposed PTaylor-MRFO_ACNN?PAOA, existing as 0.315. The load obtained by the conventional models,
models ABC, ELORM, AP-ACO, MOMA-APIBO, CCA, like ABC is 0.837, ELORM is 0.792, AP-ACO is 0.760,
MinPR, and Taylor-MRFO?ACNN is 0.331, 0.752, 0.682, MOMA-APIBO is 0.705, CCA is 0.652, MinPR is 0.379,
0.621, 0.585, 0.444, 0.370, and 0.349, respectively. and Taylor-MRFO?ACNN is 0.336. The existing models
consumed the energy consumption value as 0.901J for
5.4.2 Assessment based on task size=200 ABC, 0.872J for ELORM, 0.853J for AP-ACO, 0.827J for
MOMA-APIBO, 0.698J for CCA, 0.659J for MinPR, and
Figure 9 illustrates the comparative evaluation of the pro- 0.516J for Taylor-MRFO?ACNN.
posed PTaylor-MRFO_ACNN?PAOA with respect to
performance measures based on task size=200. Figure 9a– 5.4.3 Assessment based on task size=300
d specifies the analysis of the proposed PTaylor-
MRFO_ACNN?PAOA in terms of load, power, energy Figure 10 portrays the analysis of the proposed PTaylor-
consumption, and latency. By considering the swarm size MRFO_ACNN?PAOA in accordance with the evaluation
as 20, the proposed technique attained the load as 0.329, metrics depending upon task size=300. Figure 10a displays
power as 0.532, energy consumption as 0.357J, and latency the evaluation of the proposed approach using load. If the

123
Wireless Networks (2024) 30:1085–1105 1103

swarm size is 20, load computed by the proposed PTaylor- offering an effective solution is difficult owing to the
MRFO_ACNN?PAOA is 0.352, whereas the power complexities like machine heterogeneity, multi-dimen-
achieved by the proposed model is 0.569 and it is portrayed sional resources, and high cloud DCs. The tremendous
in Fig. 10b. Meanwhile, the load computed by the con- demand for computational power has resulted to the
ventional approaches are 0.841, 0.724, 0.641, 0.557, 0.404, breakthrough of high-scale DCs that need a high amount of
and 0.362 for swarm size=20. On the other hand, the power energy, thus producing high operating expenses. It is
attained by the existing models, such as ABC is 0.931, essential to eliminate wasting resources and avoid lengthy
ELORM is 0.913, AP-ACO is 0.899, MOMA-APIBO is response time for an efficient solution. To encounter this
0.835, CCA is 0.747, MinPR is 0.660, and Taylor- problem, an efficacious model based on optimization
MRFO?ACNN is 0.598. Figure 10c represents the analy- enabled VM scaling-based load distribution and optimal
sis of the proposed PTaylor-MRFO_ACNN?PAOA with switching strategy in the cloud data center. Initially, net-
respect to energy consumption and assessment of the work structure of the cloud is described depending upon
designed model based on latency is depicted in Fig. 10d. the fat tree model and the load allocation of DCs is done
By increasing the swarm size to 20, energy consumed by using P-Taylor MRFO by considering the multi-objectives
the devised approach is 0.405J, whereas latency obtained like power, load, latency, and bandwidth. Accordingly, the
by the traditional approaches, such as ABC is 0.787, switching of the cloud data center to the desired mode is
ELORM is 0.741, AP-ACO is 0.696, MOMA-APIBO is performed using ACNN. If the system is overloaded, either
0.674, CCA is 0.667, MinPR is 0.604, and Taylor- horizontal or vertical scaling is carried out based on pre-
MRFO?ACNN is 0.541. defined threshold elastic scaling using devised PAOA.
Here, vertical scaling is performed by including CPU, and
5.5 Comparative discussion RAM to the model, wherein horizontal scaling is done by
including nodes to the machines. Nevertheless, the devised
Table 2 elucidates the comparative discussion of the PAOA is obtained by the integration of POA with Adam
introduced PTaylor-MRFO_ACNN?PAOA. From the optimization. Meanwhile, the proposed approach has
discussion, it is clearly understood that the devised model achieved a minimum load of 0.329, power of 0.532, min-
has obtained a minimum load of 0.329, power of 0.532, imum energy consumption of 0.357, and minimum latency
minimum energy consumption of 0.357, and latency of of 0.315.
0.315 while considering the task size as 200. In the devised
approach, the inherent features of MRFO are incorporated
into the Taylor series and POA for offering a better solu-
tion in load distribution, thereby minimizing the cost and it References
also enhances the efficiency of the model. Also, the ACNN
1. Ashok Kumar, C., & Vimala, R. (2020). Load balancing in cloud
is used to predict the switching state and it facilitates the environment exploiting hybridization of chicken swarm and
processing speed delivered with superior accuracy. enhanced raven roosting optimization algorithm. Multimedia
Research, 3(1), 45–55.
5.6 Statistical analysis 2. Indira, C. K., Tiwari, T., Tiwari, M., Saxena, Gaurav, D.,
Chauhan, A., & Bajaj, S. H. (2023). Eco-friendly green cloud
structure with internet of things for astute agriculture. Journal of
The statistical analysis of the model is depicted in Table 3. Pharmaceutical Negative Results, 14(2), 2973–2986.
Here, the best, mean, and variance of the models are 3. Zhou, Q., Lou, J., & Jiang, Y. (2019). Optimization of energy
evaluated for load, power, energy consumption, and consumption of green data center in e-commerce. Sustainable
Computing: Informatics and Systems, 23, 103–110.
latency metrics. 4. Sharma, R., & Singh, R. (2022). A highly reliable and cost-ef-
fective service model for finite population clouds: Analysis and
implementation. Arabian Journal for Science and Engineering,
6 Conclusion 47, 1181–1196.
5. Soltanshahi, M., Asemi, R., & Shafiei, N. (2019). Energy-aware
virtual machines allocation by krill herd algorithm in cloud data
Cloud is the most fascinating advancement utilized in centers. Heliyon, 5(7), e02066.
every side of the business and in every domain and it is 6. Kumar, J., Saxena, D., Singh, A. K., & Mohan, A. (2020).
easily achieved through certain parameters, such as Biphase adaptive learning-based neural network model for cloud
datacenter workload forecasting. Soft Computing, 24(19),
deployment method, resource, load optimization, and so 14593–14610.
on. Nowadays, load distribution in DCs is a significant part 7. Azizi, S., Zandsalimi, M. H., & Li, D. (2020). An energy-efficient
of the cloud with respect to system achievement. The key algorithm for virtual machine placement optimization in cloud
role is to satisfy the user demands effectively with mini- data centers. Cluster Computing, 23(4), 3421–3434.
mum resource utilization and processing time. However,

123
1104 Wireless Networks (2024) 30:1085–1105

8. Yavari, M., Ghaffarpour Rahbar, A., & Fathi, M. H. (2019). center under varying workloads. Cluster Computing, 23(4),
Temperature and energy-aware consolidation algorithms in cloud 3079–3093.
computing. Journal of Cloud Computing, 8(1), 1–16. 26. Li, Y., Wen, Y., Tao, D., & Guan, K. (2019). Transforming
9. Devagnanam, J., & Elango, N. M. (2020). Optimal resource cooling optimization for green data center via deep reinforcement
allocation of cluster using hybrid grey Wolf and Cuckoo search learning. IEEE Transactions on Cybernetics, 50(5), 2002–2013.
algorithm in cloud computing. Journal of Networking and 27. Kingma, D. P., & Ba, J. (2014). Adam: A method for stochastic
Communication Systems, 3(1), 31–40. optimization. arXiv:1412.6980.
10. Xiao, H., Hu, Z., & Li, K. (2019). Multi-objective VM consoli- 28. Zhao, W., Zhang, Z., & Wang, L. (2020). Manta ray foraging
dation based on thresholds and ant colony system in cloud optimization: An effective bio-inspired optimizer for engineering
computing. IEEE Access, 7, 53441–53453. applications. Engineering Applications of Artificial Intelligence,
11. Ranjbari, M., & Torkestani, J. A. (2018). A learning automata- 87, 103300.
based algorithm for energy and SLA efficient consolidation of 29. Trojovsky, P., & Dehghani, M. (2022). Pelican optimization
virtual machines in cloud data centers. Journal of Parallel and algorithm: A novel nature-inspired algorithm for engineering
Distributed Computing, 113, 55–62. applications. Sensors, 22(3), 855.
12. Mandal, R., Mondal, M. K., Banerjee, S., & Biswas, U. (2020). 30. Yang, Q., Zhang, H. H., & Zhang, H. (2001). Taylor series pre-
An approach toward design and development of an energy-aware diction: A cache replacement policy based on second-order trend
VM selection policy with improved SLA violation in the domain analysis. In Proceedings of the 34th Annual Hawaii International
of green cloud computing. The Journal of Supercomputing, 76(9), Conference on System Sciences, pp. 7.
7374–7393. 31. Zhao, D., Wang, B., & Liu, D. (2013). A supervised actor-critic
13. Wang, J. (2021). Grey wolf optimization and crow search algo- approach for adaptive cruise control. Soft Computing, 17(11),
rithm for resource allocation scheme in cloud computing. Mul- 2089–2099.
timedia Research, 4(3). 32. Patil, R., & Singh, R. (2013). Scaling in cloud computing.
14. Choudhury, T., Dewangan, B. K., Tomar, R., Singh, B. K., Toe, International Journal of Advance Research, IJOAR, 1, 21–27.
T. T., & Nhu, N. G. (Eds.). (2021). Autonomic computing in 33. Li, C., Tang, J., & Luo, Y. (2020). Elastic edge cloud resource
cloud resource management in industry 4.0. Springer. management based on horizontal and vertical scaling. The Jour-
15. Leitner, P., & Cito, J. (2016). Patterns in the chaos-a study of nal of Supercomputing, 76(10), 7707–7732.
performance variation and predictability in public IaaS clouds.
ACM Transactions on Internet Technology (TOIT), 16(3), 1–23. Publisher’s Note Springer Nature remains neutral with regard to
16. Netaji, V. K., & Bhole, G. P. (2020). Optimal container resource jurisdictional claims in published maps and institutional affiliations.
allocation using hybrid SA-MFO algorithm in cloud architecture.
Multimedia Research, 3(1), 11–20.
Springer Nature or its licensor (e.g. a society or other partner) holds
17. Huang, Y., Xu, H., Gao, H., Ma, X., & Hussain, W. (2021).
exclusive rights to this article under a publishing agreement with the
SSUR: An approach to optimizing virtual machine allocation
author(s) or other rightsholder(s); author self-archiving of the
strategy based on user requirements for cloud data center. IEEE
accepted manuscript version of this article is solely governed by the
Transactions on Green Communications and Networking, 5(2),
terms of such publishing agreement and applicable law.
670–681.
18. Millnert, V. & Eker, J. (2020). HoloScale: Horizontal and vertical
scaling of cloud resources., In IEEE/ACM 13th International
Conference on Utility and Cloud Computing (UCC), Dr. A. Poobalan Poobalan A. has
pp. 196–205. graduated from Anna Univer-
19. Mohajer, A., Bavaghar, M., & Hamid, F. (2020). Mobility-aware sity, Chennai in B.E Computer
load balancing for reliable self-organization networks: Multi- Science and Engineering from
agent deep reinforcement learning. Reliability Engineering & RVS College of Engineering
System Safety, 202, 107056. and Technology, Dindigul.
20. Dong, S., Zhan, J., Hu, W., Mohajer, A., Bavaghar, M., & Mir- Masters in Engineering (M.E)
zaei, A. (2023). Energy-efficient hierarchical resource allocation Degree from Anna University of
in uplink-downlink decoupled NOMA HetNets. IEEE Transac- Technology, Trichy in Com-
tions on Network and Service Management. puter Science and Engineering
21. Mohajer, A., Sorouri, F., Mirzaei, A., Ziaeddini, A., Jalali Rad, from RVS College of Engi-
K., & Maryam, B. (2022). Energy-aware hierarchical resource neering and Technology, Din-
management and backhaul traffic optimization in heterogeneous digul and Doctor of Philosophy
cellular networks. IEEE Systems Journal, 16(4), 5188–5199. (Ph.D ) degree in Information
22. Madhumala, R. B., Tiwari, H., & Devarajaverma, C. (2021). A and Communication Engineer-
reliable frame work for virtual machine selection in cloud Dat- ing from Anna University, Chennai. He is currently working as an
acenter using particle swarm optimization. International Journal Assistant Professor in the Department of Computer Science &
of Mathematics and Computer Science, 16(2), 677–685. Engineering at University College of Engineering, Dindigul, Tamil
23. Priya, B., & Gnanasekaran, T. (2020). To optimize load of hybrid Nadu, India. His research interest includes Cloud computing, Net-
P2P cloud data-center using efficient load optimization and work security, Image Processing. He is a Life member of ISTE.
resource minimization algorithm. Peer-to-Peer Networking and
Applications, 13(2), 717–728.
24. Wei, W., Gu, H., Lu, W., Zhou, T., & Liu, X. (2019). Energy
efficient virtual machine placement with an improved ant colony
optimization over data center networks. IEEE Access, 7,
60617–60625.
25. Mishra, S. K., & Manjula, R. (2020). A meta-heuristic based
multi objective optimization for load distribution in cloud data

123
Wireless Networks (2024) 30:1085–1105 1105

Dr. P. ShanthaKumar received Dr. M. Robinson Joel received

his Bachelor’s degree in Science the Doctorate Ph.D degree in
and his Master’s degree in Computer and Information
Computer Applications from Technology from the Manon-
Madurai Kamaraj University in maniam Sundaranar University,
1997 and 2001, respectively. He Tirunelveli on March 2019. He
obtained his M.E in Computer completed his M.Tech and
Science and Engineering and MCA from the same university
Ph.D degree’s from Anna year 2008 and 2002 respec-
University in 2010 and 2016.. tively. He is currently working
Currently, he is working as a as an Associate Professor in the
Professor in the Department of Department of Information
Computer Science & Engineer- Technology at Kings Engineer-
ing, Nehru Institute of Tech- ing College, Chennai. He has 19
nology, Coimbatore, Tamil years of teaching experience.
Nadu, India. He was worked in different capacities in various engi-
neering college for past two decades. His research interests include
medical image processing.

123