Планування ресурсів у мережах IoT edge computing з використанням гібридного алгоритму глибокого навчання
The proliferation of the Internet of Things (IoT) and wireless sensor networks enhances data communication. The demand for data communication rapidly increases, which calls the emerging edge computing paradigm. Edge computing plays a major role in IoT networks and provides computing resources close...
Saved in:
| Date: | 2022 |
|---|---|
| Main Authors: | , |
| Format: | Article |
| Language: | English |
| Published: |
The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute"
2022
|
| Subjects: | |
| Online Access: | https://journal.iasa.kpi.ua/article/view/261572 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Journal Title: | System research and information technologies |
| Download file: |  |
Institution
System research and information technologies| _version_ | 1867334426490306560 |
|---|---|
| author | Vijayasekaran, G. Duraipandian, M. |
| author_facet | Vijayasekaran, G. Duraipandian, M. |
| author_institution_txt_mv | [
{
"author": "G. Vijayasekaran",
"institution": "Department of Computer Science and Engineering of Sir Issac Newton College of Engineering and Technology, Nagapattinam"
},
{
"author": "M. Duraipandian",
"institution": "Department of Computer Science and Engineering of Hindusthan Institute of Technology, Coimbatore"
}
] |
| author_sort | Vijayasekaran, G. |
| baseUrl_str | http://journal.iasa.kpi.ua/oai |
| collection | OJS |
| datestamp_date | 2022-12-21T22:15:21Z |
| description | The proliferation of the Internet of Things (IoT) and wireless sensor networks enhances data communication. The demand for data communication rapidly increases, which calls the emerging edge computing paradigm. Edge computing plays a major role in IoT networks and provides computing resources close to the users. Moving the services from the cloud to users increases the communication, storage, and network features of the users. However, massive IoT networks require a large spectrum of resources for their computations. In order to attain this, resource scheduling algorithms are employed in edge computing. Statistical and machine learning-based resource scheduling algorithms have evolved in the past decade, but the performance can be improved if resource requirements are analyzed further. A deep learning-based resource scheduling in edge computing IoT networks is presented in this research work using deep bidirectional recurrent neural network (BRNN) and convolutional neural network algorithms. Before scheduling, the IoT users are categorized into clusters using a spectral clustering algorithm. The proposed model simulation analysis verifies the performance in terms of delay, response time, execution time, and resource utilization. Existing resource scheduling algorithms like a genetic algorithm (GA), Improved Particle Swarm Optimization (IPSO), and LSTM-based models are compared with the proposed model to validate the superior performances. |
| doi_str_mv | 10.20535/SRIT.2308-8893.2022.3.06 |
| first_indexed | 2025-07-17T10:27:55Z |
| format | Article |
| fulltext |
G. Vijayasekaran, M. Duraipandian, 2022
86 SN 1681–6048 System Research & Information Technologies, 2022, № 3
TIДC
ТЕОРЕТИЧНІ ТА ПРИКЛАДНІ ПРОБЛЕМИ
ІНТЕЛЕКТУАЛЬНИХ СИСТЕМ ПІДТРИМАННЯ
ПРИЙНЯТТЯ РІШЕНЬ
UDC 519-62
DOI: 10.20535/SRIT.2308-8893.2022.3.06
RESOURCE SCHEDULING IN EDGE COMPUTING IOT
NETWORKS USING HYBRID DEEP LEARNING ALGORITHM
G. VIJAYASEKARAN, M. DURAIPANDIAN
Abstract. The proliferation of the Internet of Things (IoT) and wireless sensor net-
works enhances data communication. The demand for data communication rapidly
increases, which calls the emerging edge computing paradigm. Edge computing
plays a major role in IoT networks and provides computing resources close to the
users. Moving the services from the cloud to users increases the communication,
storage, and network features of the users. However, massive IoT networks require a
large spectrum of resources for their computations. In order to attain this, resource
scheduling algorithms are employed in edge computing. Statistical and machine
learning-based resource scheduling algorithms have evolved in the past decade, but
the performance can be improved if resource requirements are analyzed further. A
deep learning-based resource scheduling in edge computing IoT networks is pre-
sented in this research work using deep bidirectional recurrent neural network
(BRNN) and convolutional neural network algorithms. Before scheduling, the IoT
users are categorized into clusters using a spectral clustering algorithm. The pro-
posed model simulation analysis verifies the performance in terms of delay, re-
sponse time, execution time, and resource utilization. Existing resource scheduling
algorithms like a genetic algorithm (GA), Improved Particle Swarm Optimization
(IPSO), and LSTM-based models are compared with the proposed model to validate
the superior performances.
Keywords: edge computing, cloud computing, Internet of Things (IoT), resource
scheduling, deep learning.
INTRODUCTION
The Internet and smart devices have become indispensable elements in daily life.
People depend on their smart devices for daily activities like payment, healthcare,
virtual reality, games, etc. These different applications increase the resource re-
quirements of smart devices. Cloud computing has been adopted to meet resource
demands. The cost-effective cloud solutions offer numerous advantages in infor-
mation technology and ensure that users receive essential computing, storage, and
communication services based on their needs [1]. The tremendous applications
based on IoT networks enhance the quality of life (Fig. 1). However, the high
bandwidth requirement for IoT applications increases energy consumption,
transmission bandwidth, and delay. Moreover, providing all the user-requested
services using cloud computing is difficult since more than ten billion edge de-
vices are deployed every day, and the rate is increasing [2]. To overcome these
issues, edge computing paradigms have been introduced.
Resource scheduling in edge computing IoT networks using hybrid deep learning algorithm
Системні дослідження та інформаційні технології, 2022, № 3 87
Edge computing IoT networks allow the process to be performed near the
device or node. The user-requested functions and services from the edge cloud are
moved near to the user in edge computing to provide better storage and computing
facilities [3]. Figure 1 depicts a simple illustration of IoT edge networks.
Though edge computing provides a wide range of services in various do-
mains, it is essential to look into resource management. The increased number of
user services increases the bandwidth demand on IoT networks. The resource
scarcity problem and the computational complexities of IoT networks reduce the
overall quality of services. Edge computing provides user-requested resources
incorporating multiple techniques like clustering algorithms, and scheduling algo-
rithms to define the user demand. Since the resource requirements to compute
data types in IoT systems are different, the data processing is generally performed
at regular intervals. In order to process diverse data, the system requires different
computation resources, which should be provided by edge computing by switch-
ing resources from one to the other. This process will increase the computation
time. To avoid this, clustering techniques are used in edge computing before allo-
cating the resources [4].
Traditional clustering techniques like k-means, fuzzy c-means, hierarchical
clustering, etc. are used in various research models. However, the conventional
methods lag in performance while handling large data volumes. Moreover, con-
ventional clustering techniques require additional dimensionality reduction tech-
niques, which increase the overall computation cost [5]. Considering this limita-
tion, in our previous work, we have employed an improved spectral clustering
algorithm that clusters the resource requirements based on the data similarity [25].
The process flow of the clustering model is simply illustrated in Fig. 2. Compared
to conventional cloud computing services, edge computing will provide minimum
latency and support a wide range of IoT applications. As discussed, to satisfy the
computation requirements of IoT and improve the quality of services, a hybrid
deep learning-based resource allocation procedure is presented in this research
work. Summarized research contributions are presented as follows.
Presented a hybrid deep learning model for resource allocation in edge
computing using deep bidirectional recurrent neural network and convolutional
neural network technique.
Presented an intense experimental analysis of the proposed model in
terms of different metrics like resource utilization, response time, execution time,
delay, and efficiency.
Fig. 1. IoT edge networks
G. Vijayasekaran, M. Duraipandian
ISSN 1681–6048 System Research & Information Technologies, 2022, № 3 88
Presented a comparative analysis of the proposed model to validate the
superior performance with conventional techniques like Improved PSO (IPSO),
Genetic Algorithm (GA), and LSTM based resource scheduling procedures.
The remaining part of the article is arranged in the following order: Litera-
ture analysis of existing scheduling approaches are presented in section 2. Sec-
tion 3 presents the hybrid deep learning based resource scheduling model. Section
4 presents the details of simulation results and discussion and finally, the features
are concluded in section 5.
RELATED WORKS
Resource scheduling strategies that have evolved in the past few years are
considered for literature analysis, and the observations are summarized in this
section based on the methodology, feature merits, and demerits. The recent trends
in resource scheduling in edge computing are analyzed in [6] based on resource
allocation, computation offloading, and resource provisioning. The techniques
that have evolved for scheduling are categorized into centralized and distributed
approaches. Applications related to these centralized and distributed approaches
are analyzed and discussed in detail, which provides a basic ideology about the
resource scheduling procedures. A hybrid resource scheduling procedure in edge
computing was reported in [7] as a four-layer computing system that supports in-
telligent operations in a smart manufacturing environment. The presented two-
phase hybrid algorithm incorporates greedy and threshold strategies for resource
scheduling to minimize energy consumption and maximize efficiency in a manu-
facturing environment.
A dynamic scheduling approach for edge computing was reported in [8] in-
corporates deep reinforcement learning and deterministic policy gradient methods
to minimize delay, energy consumption, and cache fetching costs. The presented
learning models schedule the resources based on cache, offload status for un-
cached tasks, offloading transmission power, and edge computing resource status.
The combined approach minimizes the cost function and performs better than
conventional deep Q networks. The major objective of edge computing is to pro-
vide suitable computing resources for user requests in a static and dynamic envi-
ronment. The issues in resource allocation are formulated as a nonlinear optimiza-
tion problem in [9] and presented with a regularization-based agnostic online
algorithm. The presented approach split the major issues into subcategories and
formulated an objective function for each subcategory. Convex programming is
Normalized
Matrix
Cluster
nodes
n
Number
of nodes
Graph
Similarity
Matrix
Adjacency
Matrix
Laplacian
Matrix
Fig. 2. Improved spectral clustering algorithm
Resource scheduling in edge computing IoT networks using hybrid deep learning algorithm
Системні дослідження та інформаційні технології, 2022, № 3 89
used to solve the objectives, and the experimental results validate the better per-
formance over the online greedy one-shot solution.
A stochastic optimization was formulated in [10] considering the resource
scheduling and offloading in local devices, back-end cloud and base station. To
attain the objective of minimum energy consumption, and meet the QoS require-
ments, the stochastic optimization problem is converted into a dynamic optimiza-
tion problem. To obtain an optimal solution to the dynamic optimization problem,
a Lyapunov optimization theory based offloading and scheduling procedure was
presented, which improves the overall performance and minimizes energy con-
sumption. Similarly, a delay based Lyapunov function was utilized in [11], [12] to
minimize the scheduling delay in multi-server edge computing systems. Without
traffic statistics, the computation and communication delay can be formulated
using the function and directly minimize the latency compared to traditional ap-
proaches.
Task offloading and resource scheduling in hybrid edge cloud computing re-
ported in [13] considers the applications as graphs and analyzes the resource re-
quirements to minimize the rent cost, time, and energy. Semidefinite and dual de-
composition methods are employed to obtain offloading and scheduling decisions.
Furthermore, the deep reinforcement learning model is employed to obtain dis-
crete offloading decisions and computation frequencies. Better convergence and
scheduling performance are the observed features of the presented work. The task
offloading and resource scheduling procedure reported in [14] employs a decom-
position model to reduce the computation complexity of the system. Logic-based
bender decomposition is employed to obtain the optimal solution for master and
subcategorized issues in edge computing resource management. The presented
model attains better performance in delay-sensitive applications compared to con-
ventional approaches.
An optimal task offloading and scheduling process reported in [15] considers
the completion latency and energy consumption to frame the research objective.
Based on the Markov decision procedure, a reinforcement learning model is em-
ployed in the presented work. Initially, it converts the problems of dynamic net-
work conditions and task generations into a decision process. Then a double Q
network is utilized along with a neural network to define the rewards attained by
the system. The presented approach additionally includes a context-aware atten-
tion mechanism that assigns different weights to each action to validate better per-
formances. Mobile edge computing, proximity-aware task offloading and sched-
uling were presented in [16]. The research model initially considers the
distributed resources, user mobility, energy requirements, and task properties as a
mixed integer non-linear programming problem. The optimal solution for the given
problem is obtained using a genetic algorithm and a heuristic mobility-aware
scheduling scheme was presented for effective task assignment with minimum
delay and energy constraints compared to traditional methodologies. A deadline-
aware task dispatching and scheduling model presented in [17] is used to schedule
the new tasks and take the decision to replace the existing tasks to meet the dead-
line constraints. The non-trivial analysis provides better scalability compared to
centralized algorithms. The presented approach minimizes the latency in sensitive
applications and reduces the deadline miss ratio compared to traditional ap-
proaches.
G. Vijayasekaran, M. Duraipandian
ISSN 1681–6048 System Research & Information Technologies, 2022, № 3 90
Collaborative offloading and resource allocation algorithms reported in [18]
improve the overall performance of edge computing systems and make sure that
the response time limits are met. Presented migrating birds optimization algorithm
identifies the optimal solution for the resource allocation problem considering the
memory, CPU, energy, task queue, and servers. Maximizing the service rate with
better load balancing and minimizing energy requirements are the observed fea-
tures of the presented research model. A similar collaborative task scheduling
model for edge computing IoT networks was reported in [19], which defines the
offload state based on energy consumption and execution time. The presented
approach defines when to execute the offload tasks based on the local task execu-
tion, which improves the overall throughput and deadline satisfaction ratio for
critical tasks.
The resource allocation procedure for vehicle-mounted edge computing re-
ported in [20] employed a piecewise linear approximation and relaxation proce-
dure to obtain the optimal solution. Further, a gap-adjusted branch and bound al-
gorithm are presented that includes a lookahead branch scheme to improve the
scheduling performance over conventional scheduling models. A similar vehicu-
lar edge computing model reported in [21] incorporated the Markov decision pro-
cess into the deep reinforcement learning model to obtain better training effi-
ciency. The presented deep reinforcement learning model implementation is
defined based on the proximal policy optimization algorithm. A convolutional
neural network was also incorporated to approximate the value and policy func-
tions to extract the essential features. From the literature analysis, it can be ob-
served that the features of deep learning algorithms are not effectively utilized in
resource scheduling. Most of the scheduling procedures still follow the linear and
non-linear optimization problem, and solution practices. Moreover, the perform-
ance of such optimization models is also not up to the mark. Reinforcement learn-
ing is widely used in resource scheduling, but the architecture can lead to an over-
load state and diminish the results. Considering these limitations, a hybrid deep
learning-based resource scheduling model is presented in this research work in the
following section.
PROPOSED WORK
The proposed resource scheduling model is developed using a bidirectional recur-
rent neural network model which includes a 1D-convolutional neural network
model, recurrent network, and fully connected neural network block. Instead of a
conventional recurrent neural network, the presented model is combined with a
convolutional neural network to obtain better performances. The included convo-
lution block is trained to learn the features from input data. The resource require-
ments are considered as input data and details of resources are formulated as
complete information using a one-hot encoding procedure. Then the time series
resource requests are processed using a bidirectional recurrent neural network
model which includes a long short-term memory unit. Finally, the fully connected
neural network is used to sample the output which improves the scheduling per-
formances.
Initially, the convolutional neural network receives the requested resource
details as input which includes the features like resource category, duration, sub-
Resource scheduling in edge computing IoT networks using hybrid deep learning algorithm
Системні дослідження та інформаційні технології, 2022, № 3 91
class, etc., Improved spectral clustering algorithm which is utilized in our previ-
ous work is incorporated in this model also to cluster the resource requirements.
These clustered information features are converted into quantified information
using the encoding procedure. Instead of an integer encoding procedure, one-hot
encoding is used in the proposed work. Since in integer encoding if the require-
ments are encoded in natural order there may be a chance for featuresimposed
which affect the performances. We consider the time duration between resource
requests as a distance factor in the encoding procedure and in the proposed work
distance between each request is considered as same. i.e., the time duration
between one request to another request is considered as same to reduce the ex-
perimental computation complexity.
One-hot encoding present the information as a one-dimensional array func-
tion with equal length where the value 1 in the array represents the specific re-
source request. Though the dimension of one hot encoding array is high due to the
zero elements in the array which is used to fill the complete data. To manage this
dimensionality issue, 1D-Convolutional neural networks are incorporated in the
proposed architecture. CNN reduces the computational complexity by extracting
the features using feature weights and activation functions. In the presented CNN
model, rectified linear unit (ReLU) activation function is used and the procedure
is formulated using filters },,,{ 21 wwww and features },,,{ )()2()1( nffff
as follows.
wfwfc pnmpnm )2(1)2( ,
where 1
2
s
pn
m ;
),0(max)( xxxf ,
where mc is the convoluted value and the input size is reduced into m from n in
one layer. if more number of layers are included then the size can be reduced fur-
ther. Though employing CNN in the scheduling process reduces the computation
cost significantly it reduces the complexities of recurrent block which includes
LSTM and fully connected neural network models.
Following the 1D-CNN model, a bidirectional recurrent neural network is
employed in the proposed architecture. The presented deep RNN model is the ma-
jor element in the scheduling process that analyzes the resource requests effec-
tively. Since the conventional deep neural network doesn’t have the ability to pre-
serve the learned information, it is not able to schedule the resource in the future.
The process must be repeated again in order to satisfy the resource requests.
These limitations are overcome by the presented RNN model, which learns the
information in both directions, i.e., forward and backward, so that the system
needs not be trained again for new requests. The RNN architecture is able to ad-
dress the long-term dependencies. However, if the request gap is too large for the
resources, then the RNN will exhibit poor performance. So, a type of RNN that
performs better than conventional RNNs is employed in the proposed work.
Long short-term memory (LSTM) is a type of RNN model that is utilized as
a bidirectional model to predict the user resource requests in the scheduling
process. To attain better performance, the cells in the model obtain the input from
G. Vijayasekaran, M. Duraipandian
ISSN 1681–6048 System Research & Information Technologies, 2022, № 3 92
the previous layers and time step as a two directional propagation. The proposed
network model is bidirectional, so two activation functions are employed, which
are represented as t
and t
. The only difference between the activation func-
tions is the input and output direction. Several gates like forget, update and output
gates are used in the LSTM cells to controls the information flow over time steps.
Two functions like sigmoid and tanh functions are used in the system design
which is mathematically formulated as
xe
x
1
1
)( ,
xx
xx
ee
ee
x
)(tanh .
The inputs for the LSTM cell are obtained from the previous layers along
with time step, activation and direction steps. The cell calculates the output
through the activation function along with gates. Initially the forget gate hold or
drop an information. If a resource request is set then LSTM is used to track the
resource schedule over time slot. If the identified resource is need to be changed
due to unavailability, then its associated previous stored values will also be re-
moved as a update process. The forget gate is the main authority which decides
which information has to be stored or removed from the memory. Consider the
previous activation function as )1( t and its current time step is )(t , then the
forget gate is represented as a function
))](,[( 1
f
t
f
t
f tw ,
where the weights are represented as fw , previous activation function is repre-
sented as 1t )(t , the current time step input is represented as )(t and forget
bias term is represented as f . The above function results into a vector function
in the range [0,1]. Further this forget gate values are multiplied element wise with
the previous cell 1t
f . If the value of the multiplication process is obtained as
zero or nearer to zero then the selected resources are removed from the schedule.
If the value is near to or absolutely one then the resource will be held for further
process. The tanh function will create new candidate values if an information is
removed by the forget gate. New candidate selection is based on the following
formulation
))](,[(tan 1
twh tt
f ,
where w represents the weight function and represents the bias term for
tanh function. The sigmoid function in the network updates the information about
new candidate. The update gate has full control over the candidate values and it
will add the values to the cell state. The major process of update gate is to relate
the candidate to the respective previous files. The update operation is mathemati-
cally expressed as
))](,[( 1
u
t
u
t
u tw ,
Resource scheduling in edge computing IoT networks using hybrid deep learning algorithm
Системні дослідження та інформаційні технології, 2022, № 3 93
where t
u is the vector and the values are present between [0,1]. Similar to previ-
ous operation an element wise multiplication is performed with t̂ to compute
t . uw represents the update gate weight and u represents the update term bias
function. The information flow is decided by the forget gate and update gate de-
cides which candidate values are need to be updated. based on this process, the
new cell state is updated as follows.
tt
u
t
f
t
f
t ˆ1 .
Finally, the outputs are obtained by the output gate which considers the cell
state as follows:
]))](,[( 1
tw tt ;
) (tan ttt h ,
where the output weight is represented as w and bias term for output is repre-
sented as . In the proposed method, the output can be the requested resource
for schedule and subsequent resource requests which are need to be scheduled.
Fig. 3 depicts the overall architecture of proposed resource scheduling model.
The final block in the proposed architecture is fully connected neural net-
work block which is normal neural network. without neural network, SoftMax
function can be used to predict the schedule as a probability function. The re-
source with low probability values are removed and high probability values are
held in the scheduling process. while using fully connected network, instead of
Fig. 3. Overall architecture of the proposed model
G. Vijayasekaran, M. Duraipandian
ISSN 1681–6048 System Research & Information Technologies, 2022, № 3 94
static threshold-based result prediction, the network model learns the information
from output of previous blocks. Leveraging all the outputs from bidirectional
RNN the performance of proposed model is improved due to the neural network
learning procedure. The process is similar to image classification from the outputs
of CNN blocks but here it needs to learn from the RNN block. Mathematically
the network model is formulated for one layer with activation function is given as
)()( ][]1[][][][ w ,
where the activation function is represented as ][ , weight function is repre-
sented as ][w and layer is represented as . The bias term for the layer is
represented as ][ . The maximum probability obtained for the weight are for-
mulated as
jf
j
we
we
fP
1
)|( ,
where the probability of prediction from SoftMax output is represented as
)|( fP , represents the set input and jw represents the weights. To train
the fully connected neural network, cross entropy function is employed in the
proposed work. The output of the cross-entropy function measures the error
between predicted results and actual results. The major objective of the training
process is to reduce the prediction error. The function is mathematically formulated as
)1(log)1())(log(
1 )]([)(][
1
tiii
i
,
where the desired output function is represented as )(i , training samples are rep-
resented as and activation function for layer is represented as )]([ t . The
layer parameters are updated using Adam optimizer. The optimization function is
mathematically expressed as
w
w
ww
t
t
;
where
'
1
1
)(1
][
][
w
w
t
t ,
and
][
)1(][][ 1
1
1
1
w
ww tt
;
'
1
1
)(1
][
][
w
w
t
t ,
and
][
)1(][][ 2
1
2
1
w
ww tt
,
where 1 and 2 represents the weighted average functions, the learning rate is
represented as , the squares of gradients before bias correction is represented as
Resource scheduling in edge computing IoT networks using hybrid deep learning algorithm
Системні дослідження та інформаційні технології, 2022, № 3 95
1t , t represents the after bias correction. In order to avoid zero is added in
the denominator. Similarly, the CNN blocks are also optimized in the sameman-
ner as neural network block. Summarized pseudocode for the proposed resource
scheduling model is presented as follows.
Pseudocode for the proposed deep bidirectional RNN based resource sched-
uling
Initialize learning procedure
Input: encoded data, network parameters
Output: optimal resource
While do
Forward model
Load network parameters and data
For t time slots do
Initialize forward propagation from CNN
Block LSTM to neural network block
Store the prediction results
Store the back propagation values
perform predicted results update
Model backward
For t time slots do
Obtain parameter gradients in neural network block
Obtain LSTM block parameters
Obtain CNN block parameters
Compute the gradient and store
Perform update of parameters
Update optimizer function
End
End
End
RESULTS AND DISCUSSION
The proposed resource scheduling algorithm using deep recurrent neural network
and convolutional neural network performance is verified through simulation
analysis and compared with existing resource scheduling techniques. Intel Berke-
ley research laboratory benchmark dataset has been used for experimentation. The
dataset includes sensor readings from 54 sensors acquired from light, voltage,
humidity and temperature sensors. The time duration for data collection is about 2
months and measurements are performed for every 31 seconds. Simulation analy-
sis is performed in NetBeans version 8.1 installed in an Intel i5 processor with
16GB memory. Performance metrics like response time, execution time, resource
G. Vijayasekaran, M. Duraipandian
ISSN 1681–6048 System Research & Information Technologies, 2022, № 3 96
utilization, efficiency is considered for scheduling model. Clustering accuracy and
convergence rate are used for improved spectral clustering model performance
analysis. As detailed analysis of spectral clustering performances are explored in
the previous research work this experimental analysis mainly focused the per-
formance of deep learning-based scheduling techniques. Table 1 depicts the de-
tails of clustering model performance over existing k-means and fuzzy c-means
clustering performances.
T a b l e 1 . Performance analysis of clustering models
Algorithms Convergence Rate, % Clustering Accuracy, %
k-means 94.30 92.73
FCM 95.90 95.09
Improved spectral
clustering algorithm 99.00 99.15
Fig. 4 depicts the performance comparative analysis of proposed model and
existing scheduling methods like improved particle swarm optimization (IPSO),
genetic algorithm (GA), LSTM based scheduling methods. The resource
utilization is measured based on number of clusters and maximum resource
utilization is obtained by the presented model on contrary to existing
methodologies. The average resource utilization attained by the IPSO and GA
models are 95.92% and 95.62% which is 4% lesser than the proposed scheduling
model. Resource utilization attained by the LSTM based scheduling model is
98.68% which is lesser than the proposed deep BRNN model.
The response time analysis is comparatively presented in Fig. 5 for the
proposed model and existing models. The minimum time taken by the algorithm
to schedule a resource is measured as response time. It is observed from the
results if the number of clusters are minimum the response time of all the models
is less whereas it gradually increases as the number of clusters increases. The
average response acquired by the hybrid deep learning BRNN model is 1.54
seconds which is much better than the existing resource scheduling approaches.
Further the performance of all the models are measured in terms of overall
execution time. The process of requesting resources, time taken to check the
Fig. 4. Resource Utilization
Resource scheduling in edge computing IoT networks using hybrid deep learning algorithm
Системні дослідження та інформаційні технології, 2022, № 3 97
resource availability, time acquired to schedule the resource are collectively
measured as overall execution time.
Fig. 6 depicts the comparative analysis of proposed model and existing mod-
els execution times with respect to number of iterations. The maximum iteration
is selected into 50 and the execution time gradually increases for all the methods
when iteration increases. The overall execution time attained by the proposed
model is 14.82 seconds. Though the performance of existing methods are varied
in seconds, the small time difference will introduce huge impact in service viola-
tions and affect the quality of services. Presented model schedules the resources
effectively based on the prediction characteristics of bidirectional model which
reduces the further analysis if same resource is requested in future. Whereas exist-
ing methods analyze the request again and confirms the resource status and
schedule to the respective request increases the overall execution time.
The average delay exhibited by the existing models and proposed model is
analyzed and depicted in Fig. 7. The time taken by a user to acquire an optimal
resource is generally termed as waiting time. Whereas delay describes about the
Fig. 5. Response Time Analysis
Fig. 6. Execution Time Analysis
G. Vijayasekaran, M. Duraipandian
ISSN 1681–6048 System Research & Information Technologies, 2022, № 3 98
time which exceeds the stipulated predefined time period. Since all the requests
are fixed with a minimum waiting time and meanwhile the system need to search
for the resource and schedule them for further process. if the minimum waiting
time is exceeding over a period and the resources are scheduled after that then it is
measured as delay. Here the proposed model exhibits minimum delay compared
to existing scheduling procedures. The prediction performance and memory utili-
zation of proposed model not only reduces the execution time also it reduces the
delay by schedule the resource based on existing utilization. Whereas there is no
process followed in the existing scheduling procedures increases the delay.
The overall efficiency of all the approaches are comparatively analyzed and
depicted in Fig. 8. Based on the response time, resource utilization, execution
time and delay the efficiency is measured. If the algorithm exhibits maximum de-
lay definitely it will introduce an impact in the efficiency. Similarly, if the re-
source utilization is low then that system could not be considered as an efficient
one. It is essential for a system to complete the required process with minimum
response time and execution time. Considering all these factors, the proposed
model attains maximum efficiency score compared to other methods. Since the
proposed scheduling procedure reduces the response time, execution time and
improves the resource utilization and minimizes the delay which indicates the
maximum efficiency.
T a b l e 2 . Performance Comparative Analysis
Methods
Resource
Utilization
(%)
Response
Time
(s)
Execution
Time
(s)
Average
Delay
(s)
Efficiency
(%)
IPSO 95.92 2.20 26.09 5.42 94.50
GA 95.62 1.98 20.73 4.30 96.00
LSTM 98.68 1.66 17.55 3.14 98.00
Proposed BRNN 99.12 1.54 14.82 2.86 98.90
Table 2 depicts the summary of proposed model and existing resource
scheduling models performances in terms of resource utilization, response time,
Fig. 7. Average Delay Analysis
Resource scheduling in edge computing IoT networks using hybrid deep learning algorithm
Системні дослідження та інформаційні технології, 2022, № 3 99
execution time, average delay and efficiency. It can be observed from the results
that the proposed model attains better performance than existing approaches. im-
proved scheduling performance will increase the overall performance and quality
of services in cloud integrated IoT networks.
CONCLUSION
A hybrid deep learning model for resource scheduling in edge computing Internet
of Things (IoT) network is presented in this research work. The presented sched-
uling algorithm includes the deep bidirectional recurrent neural network with
convolutional neural network block to improve the scheduling performance in
edge computing. Initially the resource requests are clustered using improved spec-
tral clustering algorithm and converted into encoded image. The encoded infor-
mation features are processed by one-dimensional convolutional neural network
model followed by bidirectional long-short term memory which is type of RNN
model. The final results select the optimal resources and schedule to respective
requests to reduce the computational complexity of IoT network. Simulation
analysis of proposed model demonstrates the better performances compared to
existing scheduling models which is based on improved particle swarm optimiza-
tion algorithm, genetic algorithm and LSTM based model. Further, this research
work can be extended by introducing concatenated deep learning techniques to
avoid initial clustering process and improve the overall performances.
REFERENCES
1. Vishnu Kumar Kaliappan, S. Gnanamurthy, and K. Mohanasundaram, “Reduced
power consumption by resource scheduling in mobile cloud using optimized neural
network,” Materials Today: Proceedings, vol. 46, no.15, pp.6453–6458, 2021.
2. Yanling Shao, Chunlin Li, and Youlong Luo, “Cost-effective replication manage-
ment and scheduling in edge computing,” Journal of Network and Computer Appli-
cations, vol.129, pp. 46–61, 2019.
Fig. 8. Efficiency Analysis
G. Vijayasekaran, M. Duraipandian
ISSN 1681–6048 System Research & Information Technologies, 2022, № 3 100
3. Xiao Ma, Shangguang Wang, Shan Zhang, Peng Yang, Chuang Lin, and Xuemin
Shen, “Cost-Efficient Resource Provisioning for Dynamic Requests in Cloud As-
sisted Mobile Edge Computing,” IEEE Transactions on Cloud Computing, vol. 9,
no. 3, pp. 968–980, 2021.
4. Yongsheng Hao, Jie Cao, and Jinglin Du, “Energy-aware scheduling in edge com-
puting with a clustering method,” Future Generation Computer Systems, vol.117,
pp.259–272, 2021.
5. Jun Liu, Tianfu Yang, and Bo Sun, “Resource allocation and scheduling in the intel-
ligent edge computing context,” Future Generation Computer Systems, vol.121, pp.
48–53, 2021.
6. Quyuan Luo, Shihong Hu, Changle Li, Guanghui Li, and Weisong Shi, “Resource
Scheduling in Edge Computing: A Survey,” IEEE Communications Surveys & Tuto-
rials, vol. 23, no. 4, pp. 2131–2165, 2021.
7. Xiaomin Li, Jiafu Wan, Hong-Ning Dai, Muhammad Imran, Min Xia, and Antonio
Celesti, “A Hybrid Computing Solution and Resource Scheduling Strategy for Edge
Computing in Smart Manufacturing,” IEEE Transactions on Industrial Informatics,
vol. 15, no. 7, pp. 4225–4234, 2019.
8. Samrat Nath and Jingxian Wu, “Deep reinforcement learning for dynamic computa-
tion offloading and resource allocation in cache-assisted mobile edge computing sys-
tems,” Intelligent and Converged Networks, vol. 1, no. 2, pp. 181–198, 2020.
9. Lin Wang, Lei Jiao, Jun Li, Julien Gedeon, and Max Mühlhäuser, “MOERA: Mobil-
ity-Agnostic Online Resource Allocation for Edge Computing,” IEEE Transactions
on Mobile Computing, vol. 18, no. 8, pp. 1843–1856, 2019.
10. Fengjun Zhao, Ying Chen, Yongchao Zhang, Zhiyong Liu, and Xin Chen, “Dynamic
Offloading and Resource Scheduling for Mobile-Edge Computing With Energy Har-
vesting Devices,” IEEE Transactions on Network and Service Management, vol. 18,
no. 2, pp. 2154–2165, 2021.
11. Yuan Zhang, Peng Du, Jiang Wang, Teer Ba, Rui Ding, and Ning Xin, “Resource
Scheduling for Delay Minimization in Multi-Server Cellular Edge Computing Sys-
tems,” IEEE Access, vol. 7, pp. 86265–86273, 2019.
12. Bin Zhang and Dexu Chen, “Resource scheduling of green communication network
for large sports events based on edge computing,” Computer Communications,
vol.159, pp. 299–309, 2020.
13. Qi Zhang, Lin Gui, Shichao Zhu, and Xiupu Lang, “Task Offloading and Resource
Scheduling in Hybrid Edge-Cloud Networks,” IEEE Access, vol. 9, pp. 85350–
85366, 2021.
14. Hyame Assem Alameddine, Sanaa Sharafeddine, Samir Sebbah, Sara Ayoubi, and
Chadi Assi, “Dynamic Task Offloading and Scheduling for Low-Latency IoT Ser-
vices in Multi-Access Edge Computing,” IEEE Journal on Selected Areas in Com-
munications, vol. 37, no. 3, pp. 668–682, 2019.
15. Tong Liu, Yameng Zhang, Yanmin Zhu, Weiqin Tong, and Yuanyuan Yang,
“Online Computation Offloading and Resource Scheduling in Mobile-Edge Comput-
ing,” IEEE Internet of Things Journal, vol. 8, no. 8, pp. 6649–6664, 2021.
16. Umber Saleem, Yu Liu, Sobia Jangsher, Yong Li, and Tao Jiang “Mobility-Aware
Joint Task Scheduling and Resource Allocation for Cooperative Mobile Edge Computing,”
IEEE Transactions on Wireless Communications, vol. 20, no. 1, pp. 360–374, 2021.
17. Jiaying Meng, Haisheng Tan, Xiang-Yang Li, Zhenhua Han, and Bojie Li, “Online
Deadline-Aware Task Dispatching and Scheduling in Edge Computing,” IEEE
Transactions on Parallel and Distributed Systems, vol. 31, no. 6, pp. 1270–1286, 2020.
18. Haitao Yuan and MengChu Zhou, “Profit-Maximized Collaborative Computation
Offloading and Resource Allocation in Distributed Cloud and Edge Computing Sys-
tems,” IEEE Transactions on Automation Science and Engineering, vol. 18, no. 3,
pp. 1277–1287, 2021.
Resource scheduling in edge computing IoT networks using hybrid deep learning algorithm
Системні дослідження та інформаційні технології, 2022, № 3 101
19. Youngjin Kim, Chiwon Song, Hyuck Han, Hyungsoo Jung, and Sooyong Kang,
“Collaborative Task Scheduling for IoT-Assisted Edge Computing,” IEEE Access,
vol. 8, pp. 216593–216606, 2020.
20. Yu Liu, Yong Li, Yong Niu, and Depeng Jin, “Joint Optimization of Path Planning
and Resource Allocation in Mobile Edge Computing,” IEEE Transactions on Mobile
Computing, vol. 19, no. 9, pp. 2129–2144, 2020.
21. Wenhan Zhan et al., “Deep-Reinforcement-Learning-Based Offloading Scheduling
for Vehicular Edge Computing,” IEEE Internet of Things Journal, vol. 7, no. 6,
pp. 5449–5465, 2020.
22. Yi-Han Chiang, Tianyu Zhang, and Yusheng Ji, “Joint Cotask-Aware Offloading
and Scheduling in Mobile Edge Computing Systems,” IEEE Access, vol. 7,
pp. 105008–105018, 2019.
23. Shuaishuai Guo, Dalei Wu, Haixia Zhang, and Dongfeng Yuan, “Resource Modeling
and Scheduling for Mobile Edge Computing: A Service Provider’s Perspective,”
IEEE Access, vol. 6, pp. 35611–35623, 2018.
24. Tien Van Do, N.H. Do, and Peter Hegyi, “Comparison of scheduling algorithms for
multiple mobile computing edge clouds,” Simulation Modelling Practice and The-
ory, vol. 93, pp. 104–118, 2019.
25. G. Vijayasekaran and M. Duraipandian, An Efficient Clustering and Deep Learning
Based Resource Scheduling for Edge Computing to Integrate Cloud-IoT. Wireless
Pers Commun. 2022. Available: https://doi.org/10.1007/s11277-021-09442-8.
Resieved 00.00.2022
INFORMATION ON THE ARTICLE
G. Vijayasekaran, Department of Computer Science and Engineering, Sir Issac Newton
College of Engineering and Technology, Nagapattinam, India, e-mail:
gvijayasekaran2@gmail.com
M. Duraipandian, Department of Computer Science and Engineering, Hindusthan Insti-
tute of Technology, Coimbatore, India, e-mail: durainithi@gmail.com
ПЛАНУВАННЯ РЕСУРСІВ У МЕРЕЖАХ IOT EDGE COMPUTING З
ВИКОРИСТАННЯМ ГІБРИДНОГО АЛГОРИТМУ ГЛИБОКОГО НАВЧАННЯ /
Г. Віджаясекаран, М. Дурайпандіан
Анотація. Поширення Інтернету речей (IoT) і бездротових сенсорних мереж
покращує передачу даних. Попит на передачу даних швидко зростає, що ви-
кликає появу парадигми периферійних обчислень. Граничні обчислення віді-
грають важливу роль у мережах IoT і надають обчислювальні ресурси поблизу
користувачів. Перенесення служб із хмари до користувачів розширює комуні-
каційні, сховища та мережеві функції користувачів. Однак масивні мережі IoT
потребують великого обсягу ресурсів для своїх обчислень. Щоб досягти цього,
у граничних обчисленнях використовуються алгоритми планування ресурсів.
Алгоритми планування ресурсів, засновані на статистиці та машинному на-
вчанні, розвинулися протягом останнього десятиліття, але їх продуктивність
можна покращити, якщо додатково проаналізувати вимоги до ресурсів. У ро-
боті подано глибоке планування ресурсів на основі навчання в периферійних
обчислювальних мережах IoT з використанням глибокої двонаправленої реку-
рентної нейронної мережі (BRNN) і алгоритмів згорткової нейронної мережі.
Перед плануванням користувачі IoT класифікуються в різні кластери за допо-
могою спектрального алгоритму кластеризації. Пропонований аналіз моделю-
вання перевіряє продуктивність з точки зору затримки, часу відгуку, часу ви-
конання та використання ресурсів. Існуючі алгоритми планування ресурсів, як-
от генетичний алгоритм (GA), покращена оптимізація роїв частинок (IPSO) і
моделі на основі LSTM, порівнюються із запропонованою моделлю для під-
твердження кращої продуктивності.
Ключові слова: периферійні обчислення, хмарні обчислення, інтернет речей
(IoT), планування ресурсів, глибоке навчання.
|
| id | journaliasakpiua-article-261572 |
| institution | System research and information technologies |
| keywords_txt_mv | keywords |
| language | English |
| last_indexed | 2025-07-17T10:27:55Z |
| publishDate | 2022 |
| publisher | The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" |
| record_format | ojs |
| resource_txt_mv | journaliasakpiua/42/56732311dfd93bc718a87b5b8da26b42.pdf |
| spelling | journaliasakpiua-article-2615722022-12-21T22:15:21Z Resource scheduling in edge computing IoT networks using hybrid deep learning algorithm Планування ресурсів у мережах IoT edge computing з використанням гібридного алгоритму глибокого навчання Vijayasekaran, G. Duraipandian, M. периферійні обчислення хмарні обчислення інтернет речей IoT планування ресурсів глибоке навчання edge computing cloud computing Internet of Things IoT resource scheduling deep learning The proliferation of the Internet of Things (IoT) and wireless sensor networks enhances data communication. The demand for data communication rapidly increases, which calls the emerging edge computing paradigm. Edge computing plays a major role in IoT networks and provides computing resources close to the users. Moving the services from the cloud to users increases the communication, storage, and network features of the users. However, massive IoT networks require a large spectrum of resources for their computations. In order to attain this, resource scheduling algorithms are employed in edge computing. Statistical and machine learning-based resource scheduling algorithms have evolved in the past decade, but the performance can be improved if resource requirements are analyzed further. A deep learning-based resource scheduling in edge computing IoT networks is presented in this research work using deep bidirectional recurrent neural network (BRNN) and convolutional neural network algorithms. Before scheduling, the IoT users are categorized into clusters using a spectral clustering algorithm. The proposed model simulation analysis verifies the performance in terms of delay, response time, execution time, and resource utilization. Existing resource scheduling algorithms like a genetic algorithm (GA), Improved Particle Swarm Optimization (IPSO), and LSTM-based models are compared with the proposed model to validate the superior performances. Поширення Інтернету речей (IoT) і бездротових сенсорних мереж покращує передачу даних. Попит на передачу даних швидко зростає, що викликає появу парадигми периферійних обчислень. Граничні обчислення відіграють важливу роль у мережах IoT і надають обчислювальні ресурси поблизу користувачів. Перенесення служб із хмари до користувачів розширює комунікаційні, сховища та мережеві функції користувачів. Однак масивні мережі IoT потребують великого обсягу ресурсів для своїх обчислень. Щоб досягти цього, у граничних обчисленнях використовуються алгоритми планування ресурсів. Алгоритми планування ресурсів, засновані на статистиці та машинному навчанні, розвинулися протягом останнього десятиліття, але їх продуктивність можна покращити, якщо додатково проаналізувати вимоги до ресурсів. У роботі подано глибоке планування ресурсів на основі навчання в периферійних обчислювальних мережах IoT з використанням глибокої двонаправленої рекурентної нейронної мережі (BRNN) і алгоритмів згорткової нейронної мережі. Перед плануванням користувачі IoT класифікуються в різні кластери за допомогою спектрального алгоритму кластеризації. Пропонований аналіз моделювання перевіряє продуктивність з точки зору затримки, часу відгуку, часу виконання та використання ресурсів. Існуючі алгоритми планування ресурсів, як-от генетичний алгоритм (GA), покращена оптимізація роїв частинок (IPSO) і моделі на основі LSTM, порівнюються із запропонованою моделлю для підтвердження кращої продуктивності. The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" 2022-10-30 Article Article application/pdf https://journal.iasa.kpi.ua/article/view/261572 10.20535/SRIT.2308-8893.2022.3.06 System research and information technologies; No. 3 (2022); 86-101 Системные исследования и информационные технологии; № 3 (2022); 86-101 Системні дослідження та інформаційні технології; № 3 (2022); 86-101 2308-8893 1681-6048 en https://journal.iasa.kpi.ua/article/view/261572/264994 |
| spellingShingle | периферійні обчислення хмарні обчислення інтернет речей IoT планування ресурсів глибоке навчання Vijayasekaran, G. Duraipandian, M. Планування ресурсів у мережах IoT edge computing з використанням гібридного алгоритму глибокого навчання |
| title | Планування ресурсів у мережах IoT edge computing з використанням гібридного алгоритму глибокого навчання |
| title_alt | Resource scheduling in edge computing IoT networks using hybrid deep learning algorithm |
| title_full | Планування ресурсів у мережах IoT edge computing з використанням гібридного алгоритму глибокого навчання |
| title_fullStr | Планування ресурсів у мережах IoT edge computing з використанням гібридного алгоритму глибокого навчання |
| title_full_unstemmed | Планування ресурсів у мережах IoT edge computing з використанням гібридного алгоритму глибокого навчання |
| title_short | Планування ресурсів у мережах IoT edge computing з використанням гібридного алгоритму глибокого навчання |
| title_sort | планування ресурсів у мережах iot edge computing з використанням гібридного алгоритму глибокого навчання |
| topic | периферійні обчислення хмарні обчислення інтернет речей IoT планування ресурсів глибоке навчання |
| topic_facet | периферійні обчислення хмарні обчислення інтернет речей IoT планування ресурсів глибоке навчання edge computing cloud computing Internet of Things IoT resource scheduling deep learning |
| url | https://journal.iasa.kpi.ua/article/view/261572 |
| work_keys_str_mv | AT vijayasekarang resourceschedulinginedgecomputingiotnetworksusinghybriddeeplearningalgorithm AT duraipandianm resourceschedulinginedgecomputingiotnetworksusinghybriddeeplearningalgorithm AT vijayasekarang planuvannâresursívumerežahiotedgecomputingzvikoristannâmgíbridnogoalgoritmuglibokogonavčannâ AT duraipandianm planuvannâresursívumerežahiotedgecomputingzvikoristannâmgíbridnogoalgoritmuglibokogonavčannâ |