5224 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO.

4, APRIL 2023

An Edge Server Placement Algorithm Based on

Graph Convolution Network
Chen Ling , Zebang Feng, Liyan Xu , Qian Huang, Yinsheng Zhou, Weizhe Zhang , Senior Member, IEEE,
and Rahul Yadav , Senior Member, IEEE

Abstract—Efficient edge server placement techniques select op-

timal locations for edge servers to improve network and energy
performance. However, without prior knowledge of users’ resource
demands, the possibilities of optimal location for these edge servers
within a network are vast, which is a challenging problem. On
the other hand, most existing techniques ignore the influence of
user mobility on user’s resources demand, violation of low-latency,
and high energy consumption of 5G networks. Therefore, this
article addresses the edge server placement problem using net-
work traffic to estimate the user resource demands. We first use Fig. 1. The Network Topology of Edge Computing Environment.
a network traffic prediction model based on Graph Convolution
Network to generate network traffic distribution. Second, the I. INTRODUCTION
problem of edge server placement is formulated as a constraint
N RECENT years, the number of mobile and Internet-of-
optimization problem that places edge servers strategically to
balance energy and latency. Searching randomly through many
possible solutions and selecting the most descriptive optimal so-
I Things (IoT) devices, such as mobile phones, wearable
gadgets, tablets, virtual reality, etc., has exploded. In addition,
lutions can be time-consuming. Therefore, we used the particle
swarm optimization (PSO) algorithm to optimize network delay
it was expected that these devices would rise to more than 75
and energy consumption, especially for high mobility areas. Ex- billion by 2025. Traditionally, due to the physical size constraint,
perimental results are obtained to compare the performance of these devices are generally offloaded their computational
the proposed algorithm with existing methods. We evaluate the incentive task to cloud servers. Unfortunately, cloud servers are
algorithm based on the real dataset from Shenzhen, Futian Dis- placed far away from users, resulting in a significant delay in
trict. The results show our proposed algorithm averagely reduces
edge servers’ total cost and overloaded numbers by 23.98% and
transmission [1]. However, Edge computing has recently been
52.71%, respectively. presented to overcome this problem.
Edge computing emerges as a potential paradigm to directly
Index Terms—Edge computing, graph convolution network, deliver computing, network, and storage resources services to
edge server placement, heuristic algorithm, 5G network.
mobile or IoT devices at the network edge, which can alleviate
the pressure over the cloud, reduce latency, and break the physi-
cal limitations of devices simultaneously [2], [3]. A fundamental
and critical issue in Edge computing is server deployment, i.e.,
Manuscript received 7 April 2022; revised 24 September 2022 and 11
November 2022; accepted 1 December 2022. Date of publication 5 Decem- determining where to place the servers to maximize performance
ber 2022; date of current version 18 April 2023. This work was supported and minimize service latency.
in part by the National Key Research and Development Program of China In deploying edge computing architecture, edge server place-
under Grant 2021YFB3101102, in part by the Joint Funds of the National
Natural Science Foundation of China under Grant U22A2036, in part by ment is the first step, deciding the critical energy and network
the Key-Area Research and Development Program of Guangdong Province performance of edge computing services. Edge server placement
under Grant 2020B0101360001, in part by the Fundamental Research Funds aims to choose edge servers’ locations and satisfy users’ resource
for the Central Universities under Grant [Link].2021007, and in part
by Shenzhen Colleges and Universities Stable Support Program under Grant demands, as shown in Fig. 1. The ideal edge server placement
GXWD20220817124251002. The review of this article was coordinated by Prof. method is placing edge servers as more as possible. However, too
Xu Chen. (Corresponding authors: Liyan Xu; Weizhe Zhang. Professor Weizhe many edge servers consume enormous energy and maintenance
Zhang is the first co-corresponding author.)
Chen Ling and Weizhe Zhang are with the School of Computer Science and costs. Furthermore, small-scale placement of edge servers may
Technology, Harbin Institute of Technology, Harbin 150001, China (e-mail: create coverage holes [4] of edge computing services, increasing
18b903060@[Link]; wzzhang@[Link]). the network delay. Therefore, an effective edge server placement
Zebang Feng and Liyan Xu are with the College of Architecture and
Landscape Architecture, Peking University, Beijing 100871, China (e-mail: method is needed to trade off the network delay and energy
1801210364@[Link]; xuliyan@[Link]). consumption.
Qian Huang and Yinsheng Zhou are with the Service Lab, Huawei
Technologies, Shenzhen 518129, China (e-mail: huangqian16@[Link];
[Link]@[Link]). A. Motivations and Problems
Rahul Yadav is with the Cyberspace Security Research Center, Peng Cheng
Laboratory, Shenzhen 518055, China (e-mail: rahulyd@[Link]). Some researchers have proposed edge server placement meth-
Digital Object Identifier 10.1109/TVT.2022.3226681 ods to optimize energy consumption [5], network delay [6],
0018-9545 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See [Link] for more information.

Authorized licensed use limited to: C. V. Raman Global University - Bhubaneswar. Downloaded on February 04,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
 LING et al.: EDGE SERVER PLACEMENT ALGORITHM BASED ON GRAPH CONVOLUTION NETWORK 5225

and other performances [7], [8]. These algorithms formalize minimal cost of workload information collection. Considering
the problem model based on the working state of base stations that most existing prediction methods require sufficient training
and users’ locations. Through the solvation of the problem sets, the other easily collected datasets are needed to remedy the
model, these algorithms generate the edge server placement lack of training sets. Therefore, we apply the Graph Convolution
plans. However, current edge server placement methods are far Network (GCN) [11] in the demand prediction process, using
from real-world circumstances. the interaction information to increase the prediction accuracy.
The drawbacks of the current edge server placement algo- Thirdly, for the resource demands in high mobility areas, net-
rithm: Firstly, most edge server placement methods use the work delay should remain low, and resources should be sufficient
communication time of users and base stations to estimate the for the demands changes. As a result, an additional placement
resource demands [5], but it is inaccurate. For example, the com- strategy is needed to reserve more resources in these areas.
munication time of voice and video calls is often similar, but their However, traditional base stations and edge servers suffer from
resource demands are different. Secondly, most research [5], [6], heavy workloads in high mobility areas, causing the overload
[7], [8] analyzes the workloads of base stations and places edge problem. With the development of 5G network, the 5G base
servers to process these workloads. However, these workloads’ stations can process more workload and provide much lower
information collecting is extremely expensive, even unable to network delay. Therefore, we introduced the 5G edge servers in
be realized in some rural areas [9]. Thirdly, 5G networks require our edge server placement method to provide sufficient resources
low network delay and high reliability, but some areas where and low network delay in high mobility areas. Although 5G edge
many people visit and leave often violate these requirements [6]. servers incur higher energy consumption than traditional edge
We call these areas high mobility areas, which possess high servers, their performance in high mobility areas is crucial for
mobility features in urban areas. The network delay in these the solution of the overload problem.
areas is often increased because of the users’ mobility. Moreover, In addition, the edge server placement problem is essentially
edge servers’ workload changes intensely in high mobility areas a location selection problem, an NP-hard problem [12]. Simply
due to people’s mobility [6]. Changed workload may exceed searching all possible edge servers’ potential locations cannot
edge servers’ resource capacities and overload them, further get the optimal placement plan in a short time. Therefore,
influencing 5G networks’ reliability. we need a heuristic algorithm to search for the optimal edge
These drawbacks motivate us to develop a new edge server server placement plan quickly. Considering that particle swarm
placement method. In this new method, there are three main optimization (PSO) algorithm [13] can quickly search for the
challenges. Firstly, edge server placement needs to estimate optimal solution, we develop our edge server placement method
the users’ resource demands. However, the current estimation based on PSO algorithm. The contributions of our work are
method is inaccurate. So that the first challenge of our work summarized as follows:
is CH1: how to estimate the resource demands accurately. r Targeting CH1, we propose a network-traffic-based re-
Secondly, edge server placement should predict base stations’ source demands estimation model. This model uses net-
workload to save the cost of workload information collection. work traffic to estimate the resource demands of each area,
So that the second challenge of our work is CH2: how to predict which reflects the resource demands more precisely.
base stations’ workloads based on a small range of workload r Targeting CH2, we predict the network traffic distribution
information and other easily collected datasets. Finally, edge based on GCN. The prediction of network traffic distri-
servers should satisfy the low network delay and high reliability bution uses the easily collected datasets in datacenters.
requirements of 5G networks, especially in high mobility areas. Therefore, the proposed edge server placement method
Edge servers should reserve more resources for high mobility significantly minimizes the cost of workload information
areas to meet the significantly increased resource demands. collection. Moreover, its sufficient accuracy of prediction
Therefore, the last challenge is CH3: how to place the edge results can further maximize the performance of edge
servers in high mobility areas and meet the intense resource servers. The effect of network traffic prediction is also
demands changes. verified in simulation environments.
r Targeting CH3, we propose two edge server placement
algorithms based on PSO, named minimum Energy and
B. Contributions Delay in 5G (mEaD-5G) and minimum Energy and Delay
This paper proposes an edge server placement method to (mEaD). By placing edge servers in appropriate places,
optimize the edge servers’ network delay and energy consump- these two algorithms effectively optimize edge computing
tion simultaneously. Firstly, for the estimation method of users’ environments’ network delay and energy consumption.
resource demands, we investigate that resource demands of Furthermore, mEaD-5G places 5G edge servers and adds
most existing edge computing applications are often related separate constraints in high mobility areas, which is the
to network traffic [10]. Therefore, network traffic reflects the bottleneck of the 5G networks’ performance. Finally, it
resource demands better than the connection time, which can retains low network delay and enough resources in high
estimate users’ resource demand. mobility areas.
Secondly, for the prediction method of base station workload, The simulation experiments on a real-world dataset show
we assume that the edge servers are placed in an area with that our proposed algorithms, mEaD and mEaD-5G, averagely
minimal base stations’ workload information, meaning a reduce the energy consumption and network delay by 23.98%

Authorized licensed use limited to: C. V. Raman Global University - Bhubaneswar. Downloaded on February 04,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
 5226 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 4, APRIL 2023

TABLE I smart cities. However, this framework did not define areas with
ABBREVATIONS USED IN THIS PAPER
low network delay and high stability requirements. Kun et al. [7]
proposed a heterogeneous edge server placement algorithm to
optimize the service response time and fairness. It has two
stages, offline and online. The offline stage generates the optimal
layout strategy, and the online stage dynamically adjusts the
edge servers according to the dynamic characteristics of each
edge server. Nevertheless, the offline stage ignores future users’
demands in the edge server placement, which may cause the
online stage to be triggered frequently.
Souza et al. [8] proposed a cloud-edge service placement
(CESP) model with parallel data flow to place the parallel
tasks. By applying parallel data streams, CESP can effectively
reduce the tasks’ network delay. In addition, they did not con-
sider the task operation cost in the service placement model,
which cannot meet some users’ demands. Zhong et al. [14]
compared with an existing method, named Energy-aware Edge proposed a containerized task co-location (CTCL) scheduler that
Server Placement (EESP) [5]. Moreover, the overloaded edge avoids severe Quality of Service(QoS) degradation effectively.
servers’ numbers are averagely reduced by 52.71% in high However, CTCL did not consider the mobility problem in the
mobility areas. task scheduler, which may cause the QoS degradation problem.
This paper is organized as follows: Section II describes related Benamer et al. [15] aimed to optimize edge computing appli-
work on edge server placement and place attribute prediction cations’ network delay and proposed a heuristic algorithm to
methods. Section III describes the model of edge computing find the optimal placement position. Nevertheless, the heuristic
environment and the performance metrics of edge servers. In algorithm did not compare with the existing methods. Brogi
Section IV, we introduce the structure of the proposed method. and Forti [16] proposed a placement model to optimize IoT
Then, in Section V, we describe the design of the network applications’ network delay and bandwidth. In addition, this
traffic prediction model. Sections VI and VII introduce the model focuses more on implementing real edge computing IoT
problem model of edge server placement and the algorithm applications, which ignores the performance comparison with
procedures of mEaD and mEaD-5G. In Section VIII, we evaluate existing research. Maio and Brandic [17] proposed an edge
the performance of mEaD and mEaD-5G in terms of total cost computing offloading architecture based on the directed acyclic
and high mobility area. Finally, Section IX summarizes our work graph (DAG) to offload delay-tolerant applications and extend
and draws some conclusions. the mobile devices’ battery time. In addition, this offloading
architecture did not consider the offloading cost, which may
II. RELATED WORK increase the energy consumption in the offloading process.
In this section, we summarize the research on edge server
placement. These researches include two categories: energy
consumption optimization and network delay optimization. The B. Energy Consumption Optimization on Edge Server
research on delay optimization mainly focuses on service qual- Placement
ity. In contrast, the research on energy consumption optimization
mainly focuses on the edge network’s load balance and resource The energy consumption of the edge computing environment
allocation. In addition, we also summarized the research on is mainly related to two factors: the number and the workload
location attribute prediction in the urban environment, which is of edge servers. The idle edge servers consume a lot of wasted
related to the network traffic prediction problem. Table I shows energy. We need to avoid idle edge servers and reduce the number
the abbreviations used in this paper. of edge servers to reduce network energy consumption. Wang
and Li [5] proposed an energy-aware edge server placement
algorithm based on PSO. It maximizes edge servers’ utilization
A. Network Delay Optimization on Edge Server Placement
and further minimizes energy consumption based on an edge
Edge servers use wireless communication to connect with the servers’ energy consumption model. However, this algorithm
mobile terminal, so the network delay of the edge computing ignored the network delay optimization, which may degrade
services is related to the distance between the two commu- the QoS of edge computing services. Gao et al. [18] proposed
nicating parties. Currently, most researchers use the distance a placement algorithm for edge computing tasks. Dividing the
between the edge servers and the mobile terminals to measure edge computing tasks and introducing Named Data Network
the network delay and simplify the problem model. Gonzálezand (NDN) effectively decreases the energy consumption of edge
and Pastor [6] proposed an edge server placement framework in a computing environments. Nevertheless, they did not conduct a
5G network with network delay and stability constraints. It uses performance evaluation.
a hybrid simulated annealing algorithm to find the placement Huang et al. [19] studied the placement process of edge
location to ensure the performance of edge computing services in geographic information system (GIS) services and proposed

Authorized licensed use limited to: C. V. Raman Global University - Bhubaneswar. Downloaded on February 04,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
 LING et al.: EDGE SERVER PLACEMENT ALGORITHM BASED ON GRAPH CONVOLUTION NETWORK 5227

a heuristic algorithm to optimize energy consumption. How- trajectories and street view images of different locations in
ever, this service placement algorithm ignores the task integra- urban areas to predict the functional classification of locations.
tion problem, which is important in map services. Althamary Compared with the traditional location classification model, this
et al. [20] proposed a cache placement algorithm in edge com- model further considers the interaction information and image
puting environment, which reduces the processing of duplicate features between locations and improves prediction accuracy.
content by caching popular content, thereby reducing network The experimental results show that attribute features of urban
energy consumption. However, this algorithm did not consider areas are related to the geographical features of the locations and
the cache location. Ching-Hsien Hsu et al. [21] proposed a multi- the interaction between them. Nevertheless, they ignore these
objective edge sever placement algorithm to optimize network additional attributes and limit the prediction accuracy.
delay and energy consumption based on a mixed-integer linear In conclusion, these related works have some critical
programming method. Nevertheless, they ignored the users’ limitations:
mobility problem. r The edge server placement method needs an effective
Yang et al. [22] proposed a sorting algorithm to find the workload estimation method. Most existing methods use
appropriate access points and optimize the energy consumption. the simulated datasets to verify the methods’ performance,
However, this model only focuses on the connection number which is different from the real-world edge computing
of the edge servers, which ignores the resource capacities of environment.
the edge servers. Bhatta and Mashayekhy [23] proposed a het- r The edge server placement method should balance net-
erogeneous edge server placement algorithm based on Genetic work delay and energy consumption. The research on the
Algorithm (GA) [24] for the Internet of Vehicles environment balance between network delay and energy consumption
to optimize the cost of edge servers. However, they ignored has limited progress.
the resource utilization, which may overload the edge servers r The edge server placement should consider the workload
when minimizing their cost. Rodrigues et al. [25] proposed an changes caused by the users’ mobility. Moreover, the edge
energy-aware task allocation strategy for the healthcare IoT server placement should preserve the resources to process
environment to search for suitable task placement locations. the significantly changed workload.
However, this algorithm needs the energy consumption results
for each task, which is difficult in some scenarios. Yi et al. [26] III. PRELIMINARIES
proposed a video application placement system based on the
This section introduces the model of the edge computing en-
edge-first strategy to provide low-latency video services. In
vironment and the performance metrics of edge servers. Table II
addition, this system might offload the applications with high
shows the notations used in this paper.
network delay, which degrades the QoS of these applications.
Aryal and Altmann [27] proposed a virtual machine placement
algorithm based on GA [24] in a cloud environment to adjust the A. Network Topology of Edge Computing Environment
resource allocation and optimize energy consumption. However, The edge computing environment is a wireless network. It
this algorithm did not consider the mobility problem. can be modeled as an undirected graph G = (V, E), and V =
B ∪ S. B = {b1 , b2 , . . . , bn } represents the base station set, in
which bi represents a base station and n is the total number
C. Research on Prediction of Location Attributes in the Urban of base stations. S = {s1 , s2 , . . . , sk } is the edge server set, in
Environment
which sj represents an edge server, and the total number of
Location attributes in urban environments often have a time edge servers is k. E = {(bi , sj )|bi ∈ B, sj ∈ S, bi ⇔ sj } rep-
and space correlation. Therefore, most research on location resents the communication correlations between base stations
attribute prediction introduces space and time information into and edge servers, where bi ⇔ sj represents that bi communi-
the prediction model to improve the model’s accuracy. Wang cates with sj . Based on E, we can get the base station sets
et al. [28] proposed a network traffic prediction model. This managed by each edge server, which can be represented by
model uses a GCN with a graph recurrent unit to extract the A = {a(s1 ), a(s2 ), . . . , a(sk )}, where a(st ) represents the base
interaction features and then generate network traffic’s pre- station set managed by edge server st . Fig. 1 shows the detailed
dicted value. Experiment results showed that the model has edge computing environment.
better accuracy in the coarse-grained network traffic predic- To avoid topology confusion and resource redundancy, we
tion, such as city-to-city. However, this model cannot predict assume that each base station only connects with one edge
the network traffic for a specific area inside the city. Ji and server. Therefore, there are two constraints for the edge server
Meng [29] proposed a GCN-based network traffic classification placement problem.
model. The model extracts the characteristics of the association r Each edge server must be placed at the location of a base
relationship between different network elements and applies station and manage this base station.
the representation learning method to obtain higher prediction r The resource demand of each base station should be satis-
accuracy when the training set is small. In addition, this model fied by edge servers.
ignores the influence difference of the interactions, which limits Above all, each base station generally possesses one edge
the classification accuracy. Zhu et al. [30] proposed a location server. We call the edge server that manages the base station as
classification model based on GCN, which uses pedestrian the managing edge server of that base station.

Authorized licensed use limited to: C. V. Raman Global University - Bhubaneswar. Downloaded on February 04,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
 5228 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 4, APRIL 2023

TABLE II server sj is defined as

NOTATIONS USED IN THIS PAPER
P (sj ) = Pidle + (Pf ull − Pidle ) × u(sj ), (2)
where P (sj ) is the energy consumption of edge server sj . Pf ull
is the fully loaded energy consumption of edge server sj , and
Pidle is the idle energy consumption. u(sj ) is the utilization
of edge server sj , depending on the resource demands of base
stations. In the edge computing environment, network traffic
reflects the resource demands. For most edge computing appli-
cations, such as VR, AR, and data mining applications, higher
network traffic means more data need to be processed, requiring
more resources [10]. Therefore, we use network traffic to mea-
sure the base stations’ resource demands, and the utilization of
edge server sj is defined as

bi ∈a(sj ) f (bi )
u= , (3)
fm
where f (bi ) is the network traffic of base station bi . fm is the
theoretical max network traffic of edge servers, defined by their
hardware conditions.
Based on equations 2 and 3, the energy consumption of the
all edge servers is

k
Psum = P (sj ) . (4)
j=1

Moreover, considering that energy consumption’s value is

much larger than the network delay’s, we need to normalize
these two values. The normalized energy consumption is Pnorm ,
shown as
Psum
Pnorm = , (5)
Pideal

f (bi )
Pideal = bi × Pf ull , (6)
fm
where Pideal is the ideal total energy consumption when each
edge server is fully loaded.
For the normalization of average network delay Znorm , we set
the ideal average network delay as half of edge servers’ coverage,
B. Edge Servers’ Network Delay and Energy Consumption which means all base stations are covered evenly. The detailed
definition is shown as follows,
This paper aims to balance edge servers’ network delay and n
i=1 D (bi , sj )
1
energy consumption. To simplify the calculation of network Znorm = n
. (7)
delay, we define it as the distance between the base station and 0.5Cm
its managing edge server [5]. The locations of base station bi and
edge server sj are L(bi ) and L(sj ), respectively. The network IV. EDGE SERVER PLACEMENT METHOD BASED ON NETWORK
delay is shown as follows TRAFFIC PREDICTION
 Our proposed edge server placement method consists of two
D(bi , sj ) = (L(bi ) − L(sj ))2 . (1) stages: network traffic-prediction stage and placement-decision
stage. In the network traffic-prediction stage, we use a semi-
In addition, edge servers’ wireless communication ability supervised model to predict the network traffic distribution. In
is finite. Therefore, the edge server placement method should detail, we input the users’ movement traces, street view images,
control the network delay in edge servers’ coverage Cm [6]. and functional building records into this model and predict
The network delay between each base station and its managing the network traffic distribution, which provides the basis for
edge server must be lower than Cm . the placement-decision stage. Then, in the placement-decision
For the energy consumption of edge servers, we introduce the stage, we propose two edge server placement algorithms, mEaD
utilization of edge servers. The energy consumption of the edge and mEaD-5G, to place edge servers and minimize network

Authorized licensed use limited to: C. V. Raman Global University - Bhubaneswar. Downloaded on February 04,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
 LING et al.: EDGE SERVER PLACEMENT ALGORITHM BASED ON GRAPH CONVOLUTION NETWORK 5229

TABLE III
Algorithm 1: Edge Server Placement Method. FUNCTIONAL BUILDING ENCODING
input : T race, users’ movement traces; Image, street
view images; F unction, functional building records.
output: Edge server placement plan.
1: //Network traffic-prediction stage;
2: generate InteractionN etwork based on T race;
3: use resnet18 generate V isualF eature based on Image;
4: generate F unctionF eature based on F unction;
5: P redictionResult = GCN(InteractionN etwork,
V isualF eature, F unctionalF eature);
6: //Placement-decision stage;
7: use mEaD, mEaD-5G search for edge server placement
plan;
8: return edge server placement plan classification network. It transforms these street view images
into 512-dimensional visual feature vectors. Finally, We gener-
ate the functional features from the functional building records.
delay and energy consumption. In detail, we select each edge The functional building records are encoded based on Table III
servers’ placement location and define the communication cor- and transformed into 20-dimensional functional feature vectors.
relations between edge servers and base stations based on the Moreover, these datasets also need to be normalized before
base station’s location and network traffic distribution. More- being inputted into GCN model. Firstly, we normalize the in-
over, mobility features are considered in mEaD-5G to avoid the put datasets. We calculate the symmetric normalized Laplacian
overload problem caused by users’ mobility. Algorithm 1 shows matrix based on the interaction network, which is
the pseudocode of our proposed method.
Lsym = In − D− 2 N D− 2 ,
1 1
The details of these two stages are described in the following (8)
sections. Moreover, the time complexity of our proposed edge
server placement method depends on the time complexity of where In is the identity matrix, and D is the degree matrix of
these two stages. Because the network traffic prediction model the interaction network. For visual feature vectors and functional
we use is semi-supervised, its operation time is a constant feature vectors, we calculate the average of each place’s feature
value. Therefore, the time complexity of our proposed method is vectors. Then, we concatenate these vectors and generate the
the time complexity of the placement-decision stage, including feature matrix H 0 .
mEaD and mEaD-5G. We will analyze the time complexities of After the dataset normalization, we input symmetric normal-
mEaD and mEaD-5G in the following sections. ized Laplacian matrix and feature matrix into GCN model. GCN
model uses multiple graph convolution layers to extract the
V. NETWORK TRAFFIC DISTRIBUTION PREDICTION places’ information. Equation (9) shows the detailed operation
BASED ON GCN of each layer,
 
Firstly, the edge server placement method predicts the net- H l+1 = σ Lsym H l W l , (9)
work traffic distribution based on easily collected datasets in
the network traffic-prediction stage. Network traffic distribution where H l represents the input feature matrix of layer l, and W l
prediction is essentially a spatial prediction problem. In spatial is the weight matrix of layer l, which is learning parameters.
prediction problems, attribution of a location correlates with the σ(·) is the activation function, which adds non-linear fitness
related attributions and the interactions with other areas [30]. ability for GCN model. In GConv1 and GConv2, the activation
GCN model is an effective method to solve the spatial prediction function is relu, and the activation function in GConv3 is
problem. Therefore, we use GCN model to predict network logSof tmax. With the symmetric normalized Laplacian matrix
traffic using both interaction dataset and areas’ attributions. In and operation of each layer, one area’s features influence the
detail, we input the users’ movement traces, street view images, other areas’ output through the interaction network. In detail, by
and functional building records into a GCN model to predict multiplying the symmetric normalized Laplacian matrix and the
the network traffic distribution. Because the GCN model is feature matrix, each area’s feature vector multiplies with others’
semi-supervised, we can use the users’ movement traces in feature vectors whose number of users’ movement traces is more
both training and validating areas but the network traffic of than 1. Moreover, the more users’ movement traces in the areas,
only training areas. Hence, the proposed edge server placement the more they influence the other areas’ output. Based on this,
method need fewer network traffic dataset. GCN model can extract features from the interaction dataset and
The users’ movement traces formulates the interaction net- thereby increases the prediction accuracy.
work N . Nij represents the number of people moving from To simplify the network traffic prediction problem, we per-
L(bi ) to L(bj ). Then, we generate the visual features from the form logarithm processing on the network traffic and classify the
street view images. In detail, we input the street view images network traffic into 8 categories, transforming the network traffic
of four directions, east, south, west, and north, to resnet18 [31] prediction problem into a classification problem. The detailed

Authorized licensed use limited to: C. V. Raman Global University - Bhubaneswar. Downloaded on February 04,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
 5230 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 4, APRIL 2023


n
ysj (bi ) = 1 ∀1  j  k (12-c2)
i=1

D(bi , sj )  Cm (12-c3)
f (sj )  fm , (12-c4)

where xbi (sj ) and ysj (bi ) are the decision variables. xbi (sj )
represents the communication correlation between base station
bi and edge server sj . Suppose xbi (sj ) = 1, edge server sj will
be the managing edge server of base station bi . Otherwise, they
do not have any communication correlations. ysj (bi ) represents
the placement decision of edge server sj . Suppose ysj (bi ) = 1,
the edge server sj will be placed on base station bi . Otherwise, it
will not be placed there. Constraint 12-c1 constraints that each
base station only connects with one edge server. In constraint
12-c2, each edge server must be placed at one base station and
manage this base station. Moreover, constraint 12-c3 represents
Fig. 2. GCN Model Structure. that the network delay between base stations and their managing
edge servers should be lower than the edge server’s coverage Cm .
network traffic classification process is shown as follows, Constraint 12-c4 means that the network traffic of edge servers
traf f icClass = log3 traf f ic − 9. (10) should be lower than fm .

Finally, Fig. 2 shows the detailed structure of GCN model.

B. Particle State Encoding Method
VI. MINIMUM ENERGY AND DELAY EDGE SERVER Edge server placement is an NP-hard problem, so we need
PLACEMENT ALGORITHM a heuristic algorithm to search for the optimal edge server
placement plan. As a heuristic algorithm, PSO [13] algorithm
In this section, we introduce the problem model of edge server generates the initial particle swarm and updates each particle’s
placement. Then, we propose our minimum Energy and Delay position and speed based on the local and global optimal. It
(mEaD) edge server placement algorithm. makes each particle move forward to the optimal solution,
accelerating the search process and increasing the possibility of
A. Problem Model of Edge Server Placement finding the optimal solution. Compared with the other heuristic
Our edge server placement method aims to minimize both algorithms, PSO does not require the continuity and convexity of
the average network delay and total energy consumption of the objective function. Since the edge server placement problem
edge servers. However, there are some network traffic hot spot is discrete, PSO algorithm is suitable for edge server placement
areas close to base stations with high network traffic, and the problem. Moreover, by combining global and local optimal, PSO
edge servers there have low network delay but high energy algorithm is more likely to find the global optimal edge server
consumption. On the contrary, edge servers placed far from hot placement than the existing algorithms. Therefore, we develop
spot areas have low energy consumption but high network delay. mEaD algorithm based on PSO algorithm.
Therefore, these two factors cannot be minimum simultaneously, In mEaD, we encode the edge server placement plan as a
and the edge server placement method should balance these two matrix M with n × n size. In matrix M , each row represents
factors. Here we define the total cost of edge server placement a base station, and each column represents the edge server. In
Calsum as detail, if Mij = 1, edge server sj placed at base station bj will
manage the base station bi . So that we can get the constraints of
Calsum = [α × Pnorm + (1 − α) × Znorm ] , (11) M : (1) In each row, there must be one element equal to 1 and the
where α is a weight coefficient parameter representing energy other elements is 0. (2) If there is an element in column j is equal
consumption’s importance. If α = 1, edge server placement to 1, then Mjj = 1. Then, we define the particle position vector
algorithm will only focus on energy consumption and ignore X as the position of edge servers. If Xi = 1, an edge server will
the network delay. Furthermore, if α = 0, it will only focus on be placed at L(bi ).
network delay and ignore the other. To better illustrate the method, we take Fig. 1 as an example,
Then, we give the problem model of edge server placement including 4 base stations and 3 edge servers. As shown in Fig. 1,
based on definitions mentioned before, shown as each edge server manages 1 or 2 base stations. Moreover, the
edge servers are placed at the first base station they manage. In
min Calsum (12) detail, these three edge servers are placed at the first, second and

k fourth based stations. Then, the first edge server manages the first
s.t. xbi (sj ) = 1 ∀1  i  n (12-c1) and third base stations. And the second and fourth base stations
j=1 are managed by the second and third edge servers, respectively.

Authorized licensed use limited to: C. V. Raman Global University - Bhubaneswar. Downloaded on February 04,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
 LING et al.: EDGE SERVER PLACEMENT ALGORITHM BASED ON GRAPH CONVOLUTION NETWORK 5231

Therefore, the position vector X and matrix M of this edge

Algorithm 2: mEaD(minimum Energy and Delay)
server placement plan are
Algorithm.
X = [1, 1, 0, 1] (13)
⎡ ⎤
1 0 0 0
⎢ 0 1 0 0 ⎥
M =⎢⎣ 1 0 0
⎥. (14)
0 ⎦
0 0 0 1

C. Particle Velocity and Update Methods

Then, mEaD also needs the update methods of position vector
X and matrix M to realize the search process. We define the
particle’s velocity as an n-dimensional vector V , in which Vi =
{0, 1}. In addition, we define the position vector update method
as bitwise exclusive OR, which is
Xit+1 = Xit ⊗ Vit+1 . (15)
Based on this, the particle moves to a new position Xit+1 .
Then, the velocity vector should be updated based on the lo-
cal and global optimal positions. Therefore, we calculate the
probability distribution of the global optimal, local optimal, and
present positions, shown as
i
1/Calsum
p1i = (16)
i
1/Calsum ilbest + 1/Calgbest
+ 1/Calsum sum
ilbest
1/Calsum
p2i = (17)
i
1/Calsum ilbest + 1/Calgbest
+ 1/Calsum sum
gbest
1/Calsum
p3i = . (18)
1/Calsumi ilbest + 1/Calgbest
+ 1/Calsum sum

The probability distribution satisfies that l=1,2,3 pli = 1.
With the probability distribution, the update method of the
velocity vector is defined as
 t   t 
Vit+1 = p1i Vit ⊕ p2i Xilbest ⊗ Xit ⊕ p3i Xgbest ⊗ Xit ,
(19)
where ⊕ is the addition operation of velocity vectors. In
detail, we defined that (1, ∗, 0) = p1i (1, a, 0) ⊕ p2i (1, b, 0) ⊕
p3i (1, c, 0). ∗ is the uncertain bit, which is
⎧
⎨a 0 < rand  p1
∗= b p1 < rand  p1 + p2 , (20)
⎩
c p1 + p2 < rand  1
where rand is a random number between 0 and 1. With these
update methods, each particle is more likely to move to the
position with a lower total cost. of edge servers, making the matrix M violate the constraints.
Therefore, in Line (19), mEaD sets the rows and columns of
D. The Procedure of mEaD Algorithm M to 0 that violate the constraints, generating some unassigned
Finally, based on the above definitions, we propose the edge base stations. In Line (20)–(32), mEaD assigns the unassigned
server placement algorithm, mEaD. Algorithm 2 shows the base stations to the edge servers with enough resources. In detail,
detailed procedure of mEaD. mEaD calculates the average cost-profit avgCost and the total
In Line (1)–(14), mEaD randomly generates the initial particle cost increment appendCost for the unassigned base stations,
swarm with the corresponding position vector X, velocity vector shown as
V and matrix M . Then, in Line (16)–(18), mEaD updates the par-
ticles’ states based on the update method of velocity and position Calsum (lbest) − tmpCost
avgCost = , (21)
vectors. However, the update process will change the positions number of unallocated base stations

Authorized licensed use limited to: C. V. Raman Global University - Bhubaneswar. Downloaded on February 04,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
 5232 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 4, APRIL 2023

D (sj , bi ) Therefore, we finally used the total user number to measure the
appendCost = α × P (sj ) + (1 − α) × . (22) area mobility feature.
n
With the measuring of area mobility features, we define high
Through avgCost, mEaD can easily infer if the current posi- mobility areas: If the total user number of an area is higher than
tion could have a better total cost than local optimal. Moreover, the user limit, the area will be a high mobility area, represented
with the calculation of appendCost, mEaD compares the total by
cost increment of all unassigned base stations. Therefore, mEaD 
can prune part of the base station assignments and decrease the 0 if user < user limit
zi = . (23)
time consumption. In detail, when avgCost < 0, there is no 1 if user  user limit
placement plan better than the local optimal, so mEaD randomly
assigns the base stations to the edge servers (Line (25)–(27)). In Then, we update the definition of energy consumption and
contrast, when avgCost > 0, mEaD assigns the base stations network delay. For the energy consumption, we improve the
to the edge server with the minimum appendCost (Line (29)– definition of edge server utilization, shown as
(30)). Finally, if there is no edge server with enough resources 
bi ∈a(sj ) (1 + zi × β)f (bi )
to assign the base stations, mEaD places the new edge servers u (sj ) = . (24)
in the places of the unassigned base stations (Line (31)–(32)). fm
In (24), the edge servers must reserve the resource to process
VII. MINIMUM ENERGY AND DELAY IN 5G (MEAD-5G) EDGE the intense increased network traffic. β represents the resource
SERVER PLACEMENT ALGORITHM reservation proportion of high mobility areas. Its value reflects
the network traffic difference between high mobility and other
Although mEaD optimizes energy consumption and network areas. We will set the value of β in the following experiment
delay of edge servers, the influence of users’ mobility on user’s section. Then, the updated energy consumption definition is
resources demand is ignored. In edge computing environments, shown as
some high mobility areas have high mobility features. Users
in high mobility areas often require edge computing servers to 
k 
k

respond to them before leaving edge servers’ coverage, meaning Psum = P (sj ) = [Pidle + (Pf ull − Pidle ) × u (sj )].
j=1 j=1
low network delay. Moreover, high mobility areas often have in-
(25)
tense changes in the number of users and resources demand [32].
Moreover, we update the definition of average network delay,
Therefore, edge server placement method should focus more on
shown as
high mobility areas and satisfy low network delay requirements
and high resources demands. 1
k
With the development of a 5G network, the real edge com- Zavg = [D(bi , sj ) + zi × L], (26)
k j=1
puting environment must be heterogeneous, including some 5G
edge devices [6]. Therefore, we introduce 5G edge servers to
where L represents the delay penalty, which is set based on
edge server placement problem to provide resources for high
the communication technology parameters. We will also set the
mobility areas. Compared with regular edge servers used by
value of it in the following experiment section. Based on this
mEaD, 5G edge servers have more resources and lower network
network delay penalty, the high mobility base station can be
delay but smaller coverage and higher energy consumption.
served in optimal network delay.
Then, we improve the problem model of mEaD and propose
our minimum Energy and Delay in 5G (mEaD-5G) edge server
placement algorithm. The detailed model definition and proce- B. Improvement of Problem Model
dure of mEaD-5G are shown below. The placement strategies of 5G edge servers and regular edge
servers in mEaD-5G are different. Therefore, we need to use
A. Energy Consumption and Network Delay in mEaD-5G different decision variables for the regular and 5G edge servers.
In detail, we introduce xbi (sj ) and ysj (bi ). xbi (sj ) represents
We assume that the total user number within 24 hours corre- whether 5G edge server sj manages base station bi , and ysj (bi )
lates with the area mobility features. The more users appear in an
represents whether 5G edge server sj locates at the location of
area, the higher mobility feature this area is, and then the network
base station bi . Moreover, the coverage of 5G edge servers is
traffic changes more significantly. To verify this assumption, We
Cm , and fm is the max network traffic of 5G edge servers. For
calculated the Pearson’s correlation coefficient [33] of total user
the network delay and energy consumption, we use the same
number, average velocity and trace number with the network
normalization method with mEaD, shown as
traffic variance. Experimental results show that the total user
number has the strongest positive correlation among these three Psum
values, of which the Pearson’s correlation coefficient is 0.606. Pnorm = , (27)
Pideal
Moreover, the Pearson’s correlation coefficients of average ve-
locity and trace number are below 0.00064 and −0.00463, indi- Zavg
Znorm = . (28)
cating a weak relationship with the variance of network traffic. 0.5Cm

Authorized licensed use limited to: C. V. Raman Global University - Bhubaneswar. Downloaded on February 04,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
 LING et al.: EDGE SERVER PLACEMENT ALGORITHM BASED ON GRAPH CONVOLUTION NETWORK 5233

Based on the above definitions, the problem model of mEaD- Moreover, the regular and 5G edge servers can be transformed
5G show as into each other. Thirdly, the update method of the velocity vector
is
min [α × Pnorm + (1 − α) × Znorm ] (29)
 t   t 

k
Vit+1 = p1i Vit ⊕ p2i Xilbest Xit ⊕ p3i XgbestXit .
s.t. xbi (sj ) + xbi (sj ) = 1 ∀i, 1  i  n (29-c1) (32)
j=1
We also take an example in Fig. 1. Assuming the velocity

n vector is [0,1,0,2], the new particle position vector is
ysj (bi ) = 1 ∀j, 1  j  k (29-c2)
i=1
Xit+1 = [1, 1, 0, 1] [0, 1, 0, 2] = [1, 0, 0, 2]. (33)

n
ysj (bi ) = 1 ∀j, 1  j  k (29-c3)
i=1 As shown in (33), the first edge server is kept as is, the second
edge server is removed and the last edge server is transformed
zi × L + D(bi , sj )  Cm (29-c4)
from the regular edge server to the 5G edge server.
zi × L + D(bi , sj )  Cm (29-c5) With the newly defined position vector, velocity vector, and
their update methods, we propose mEaD-5G edge server place-

k
ment algorithm, shown as Algorithm 3.
(1 + zi × β)f (bi ) × xbi (sj )  fm (29-c6)
Different from mEaD, mEaD-5G calculates the total network
j=1
traffic of 10 closest base stations in the particle swarm initializa-

k tion (Line (11)–(17)). If the total network traffic is larger than the
(1 + zi × β)f (bi ) × xbi (sj )  fm . (29-c7) max network traffic of the regular edge server, this area should
j=1 be a network traffic hot spot area. mEaD-5G places a 5G edge
Equation (29) shows the objection function of mEaD-5G. server here to provide more resources and better performance.
In constraints 29-c1, each base station must have only one Then, in the particle state update process, mEaD-5G assigns the
managing edge server. Constraints 29–c2 and 29–c3 represent base stations to regular or 5G edge servers. In Line(40)–(42),
that each 5G and regular edge server can only locate at one base mEaD-5G places 5G edge servers at the unassigned base stations
station. Constraints 29–c4 and 29–c5 represent network delay with network traffic higher than LargeT raf f ic. If the base sta-
constraints of 5G and regular edge servers. The network delay tions’ network traffic is lower than LargeT raf f ic, mEaD-5G
of edge servers cannot be higher than its coverage. In constraints will place a regular one (Line (43)–(45)).
29–c6 and 29–c7, network traffic cannot be larger than fm and
fm .
D. Analysis of Time Complexity
C. The Procedure of mEaD-5G The time complexity of mEaD and mEaD-5G depends on
the iteration number T , swarm size P , and base station number
This section updates the particle state encoding method and
n. In each iteration, mEaD and mEaD-5G update the position
introduces the detailed procedure of mEaD-5G. Firstly, we up-
and velocity vectors of P number of particles, and this process
date the definition of matrix M and position vector X. The new
repeats for T times. For each particle, the update of the position
defined matrix M represents the service correlations between
vector will generate unassigned base stations, which should be
edge servers and base stations, and its element is 0, 1 or 2.
then assigned to new edge servers. We consider the worst-case
If Mij = 1, regular edge server sj manages base station bi .
scenario, the update of the position vector removes all edge
If Mij = 2, the base station bi will be served by a 5G one.
servers and all base stations need to be assigned. Therefore,
Otherwise, there will not be any communication correlations
this assignment process can be seen as a bin package problem,
between bi and sj . Moreover, the element value of defined
of which the time complexity is o(n2 ). In conclusion, the time
position vector X is 0, 1, or 2. It represents that no edge server,
complexities of mEaD and mEaD-5G are o(T P n2 ).
regular edge server, or 5G edge server is placed at this location,
respectively.
Secondly, we define the position vector update method of VIII. EXPERIMENT
mEaD-5G as
A. Dataset
Xit+1 = Xit Vit+1 , (30)
We test mEaD, mEaD-5G, and other existing algorithms
where is defined as in datasets of Shenzhen, Futian District. In detail, the dataset
 includes the network traffic records, the users’ movement traces,
[2] if a = 1 and b = 2
[a] [b] = . (31) the street view images, and the functional building records.
[a − b] else
Tables IV, V, VI, and Fig. 3 are examples of network traffic
Based on this position vector update method, the position of records, users’ movement traces, functional building records,
the edge server will be changed based on the particle velocity. and street view images.

Authorized licensed use limited to: C. V. Raman Global University - Bhubaneswar. Downloaded on February 04,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
 5234 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 4, APRIL 2023

TABLE IV
Algorithm 3: mDaE-5G(minimum Energy and Delay in 5G) AN EXAMPLE OF BASE STATION NETWORK TRAFFIC RECORDS
Algorithm.

TABLE V
AN EXAMPLE OF USERS’ MOVEMENT TRACES

TABLE VI
AN EXAMPLE OF FUNCTIONAL BUILDING RECORDS

Fig. 3. An example of street view images.

points of the users’ movement traces. The dwell points are where
users stay for at least 1800 seconds. We keep the dwell points
in the users’ movement trace dataset and delete the other points
and abnormal high-velocity records. Thirdly, apply logarithm
processing on the network traffic records, shown in (10). Finally,
we calculate the number of each functional building type for
functional building records.

C. Experiment Environment
We input all the processed datasets into SuperMap iDesktopX
10.1 to set up a simulation environment. SuperMap iDesktopX
10.1 is a geographic information processing platform that can
set up the geographic simulation environment. Then, we use
Python3.8 to realize mEaD, mEaD-5G and other baseline al-
gorithms and calculate edge servers’ energy consumption and
network delay. The experiment platform’s CPU is Intel Xeon
E5-2690, the memory is 64 GB, and its operating system is
WindowsServer 2012 R2.
B. Data Preprocessing
For the parameters of edge servers, we investigate the per-
We preprocess the whole dataset before inputting the datasets formance and energy consumption of 4G and 5G base station
into the edge server placement methods. Firstly, we divide Futian devices [32], [34]. The power of 4G single system base station
District into 50 × 50 meters grids and replace the records’ is 2000 W, seen as the energy consumption of regular edge
locations with these grids. Secondly, we calculate the dwell servers. Moreover, that of 5G edge servers is 3500 W. We set

Authorized licensed use limited to: C. V. Raman Global University - Bhubaneswar. Downloaded on February 04,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
 LING et al.: EDGE SERVER PLACEMENT ALGORITHM BASED ON GRAPH CONVOLUTION NETWORK 5235

the idle power of edge servers account for 60% of the full state
power. In addition, the regular and 5G edge servers’ max network
traffic are 2.2 × 108 KBytes and 2.2 × 109 KBytes, calculated
based on the peak network traffic data. We also set the value of
largeT raf f ic as 2.2 × 108 KBytes. For the high mobility base
station definition, we set the user limit to 70 percentile of all base
station user numbers, which is 41. In the following experiments,
we set regular edge servers’ coverage as 500–750 meters, which
increased by 50 meters. Based on the investigations, we set that
the coverage of 5G edge servers is 40% of that of regular edge
servers.

D. Baseline Algorithms
In our experiment, we use the following baseline algorithms
to compare with our algorithms.
Fig. 4. Network Traffic Distribution Prediction Results. (a) Prediction results
Top-K: Top-K sorts all base stations based on their network of GCN model. (b) Real-world network traffic distribution.
traffic. Then, Top-K places edge servers in order and assigns
base stations to edge servers with minimum network delay. It
repeats this process until all base stations are assigned.
Random: Random places edge servers randomly and assigns many users’ movement traces. Therefore, the interaction, visual
base stations to edge servers with minimum network delay. It and functional features here all indicate high network traffic.
also repeats until all base stations are assigned. Moreover, its strong interaction information also influences the
EESP [5]: EESP generates the initial particle swarm based neighboring areas, resulting in higher network traffic than other
on Random. Then, it uses PSO algorithm to place edge servers, areas. In contrast, the black rectangle area is Lotus Mountain
which tries to place the least amount of edge servers to decrease Park. Its visual and functional features result in low network
the energy consumption of edge servers. traffic. In addition, fewer users’ movement traces significantly
Genetic Algorithm (GA): GA randomly generates the place- decrease the network traffic, also decreasing the neighborhood’s
ment plans as the population. Moreover, the base stations are as- network traffic.
signed to the closest edge servers. Then, it iteratively crossovers On the other hand, the network traffic hot spot areas have more
and mutates each two placement plans to find the optimal place- users and network traffic, resulting in high energy consumption.
ment plan. In detail, crossover keeps placement plans’ same edge Moreover, many users’ movement traces there also means high
servers and exchanges the different ones, and mutation randomly mobility features, requiring a low network delay. The network
changes the position of one edge server. traffic hot spot areas limit edge servers’ performance. So that
in edge server placement method, the network traffic prediction
E. Result of Network Traffic Prediction accuracy in network traffic hot spot areas is most important. With
accurate prediction, our algorithm can place more edge servers
The first stage of our proposed edge server placement method near these network traffic hot spot areas with a short distance
is network traffic-prediction. We use GCN model to predict the to optimize energy consumption and network delay. As shown
network traffic distribution of the simulation area. In detail, we in Fig. 4, the locations of the network traffic hot spot areas are
use the average network traffic in Internet rush hour as the output similar, and their network traffic is also similar. Therefore, GCN
of GCN model, which is 9:00–11:00 and 21:00–23:00. We use model’s prediction accuracy is enough for edge server placement
2% of the network traffic records as a training set and 10% for method.
validation. Moreover, some grids have no street view pictures nor
functional building records, and thus they cannot be inputted into
GCN model. To address this problem, we need to use the other F. Effects of Network Traffic Distribution
grids’ attributes to interpolate the missing attributes. In detail, the This section explains the effect of network traffic and in-
visual and functional features of the areas are generally similar teraction information in our edge placement method. First, the
to their neighbors. Therefore, for each grid missing features, we network traffic datasets determine the upper bound of the edge
calculated the average feature vector of 8 neighbor grids as its servers’ performance. Ideally, all base stations’ network traffic is
feature vector. Based on this, all grids have complete feature collected to support the location optimization of the edge servers
vectors and can be input into GCN model. Finally, the mean for approaching the optimal performance. However, in most
absolute error (MAE) of prediction results is 1.5395, and Fig. 4 cases, only a small amount of base stations’ network traffic is col-
shows the detailed results. lected, and the edge servers suffer the performance degradation
As shown in Fig. 4, only a few areas are network traffic hot caused by the workload from uncollected base stations. Because
spot areas, and most areas have low network traffic. For example, a considerate placement of edge servers needs the support of
the red rectangle area is the Shenzhen Convention & Exhibition network traffic, we use GCN model to predict the areas’ network
Center, a business center and a network traffic hot spot area with traffic which cannot be collected. In addition, GCN model makes

Authorized licensed use limited to: C. V. Raman Global University - Bhubaneswar. Downloaded on February 04,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
 5236 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 4, APRIL 2023

Fig. 5. Differences between Total Costs.

Fig. 6. Edge Servers’ Overloaded Edge Servers Number.

errors in network traffic prediction and causes the edge servers’
location deviation and performance degradation. Therefore, the
integrity and prediction accuracy of network traffic directly 700 meters, meaning the most appropriate communication dis-
influences the edge servers’ performance. tance is 350 meters. Therefore, we use different values of delay
Second, interaction information is necessary for the prediction penalty L to make sure the network delay is below 350 meters,
of network traffic. Actually, the training set input into GCN shown as
model is too small to extract enough network traffic features from ⎧
⎪
⎪ 150 Cm = 500
it. To address this problem, we transformed the users’ movement ⎪
⎪
⎪
⎪ 200 Cm = 550
traces into the symmetric normalized Laplacian matrix and ⎨
250 Cm = 600
input it into GCN model to extract interaction information and L= . (34)
⎪
⎪ 300 Cm = 650
improve the prediction accuracy of network traffic. To clarify the ⎪
⎪
⎪
⎪ 350 Cm = 700
effect of interaction information in GCN model, we evaluated ⎩
400 Cm = 750
the accuracy in GCN model with two representative cases. The
first case inputs the symmetric normalized Laplacian matrix into Finally, we compare the base stations’ network traffic in high
GCN model, which is described in the previous content. The mobility and other areas. We find that the network traffic peak
second case, in contrast, replaces the symmetric normalized in high mobility areas is 1.6 times of the others’ network traffic,
Laplacian matrix with an identity matrix. Same as traditional so we set the reservation proportion β = 0.6.
deep learning models, the GCN model’s output in the second
case is generated by the multiplication of feature matrix and H. Total Cost Comparison
weights. Experimental results show that the MAE of the first
With the network traffic distribution prediction and the values
case is 1.5395 and the second one is 2.5145. As the second
of parameters, we compare the performance of the edge server
case’s MAE is obviously larger, the interaction information does
placement algorithms in our experiment. For the baseline algo-
improve the accuracy of network traffic distribution prediction.
rithms, we use the training set of GCN model as their input.
For base stations outside the training set, we set their network
G. Parameters of mEaD and mEaD-5G
traffic as that of the closest base stations in the training set.
For mEaD and mEaD-5G, the weight coefficient α is vital for Then, we get the edge server placement plan from all algorithms
edge server placement method, effectively balancing network and calculate the edge servers’ total cost each hour. For the
delay and energy consumption. Therefore, we need to find the overload edge servers, we place additional edge servers to assist
appropriate value of α. The differences between the randomly in processing the overload part of network traffic. Fig. 6 shows
generated edge server placement plans’ total costs should be the maximum total costs and maximum overload edge servers’
minor. Here we use different values of α (from 0.0 to 1.0 in steps number of the edge servers placed by mEaD, mEaD-5G and
of 0.1) to calculate the total costs of 100 edge server placement baseline algorithms.
plans generated by Random. To obtain an appropriate value of α, As shown in Fig. 6, mEaD and mEaD-5G averagely reduce
we calculate the differences between maximum and minimum edge servers’ total cost and overloaded numbers by 23.98% and
total costs, as shown in Fig. 5. In Fig. 5, when alpha = 0.4, 52.71%, respectively. In detail, mEaD has the minimum total
the difference between the maximum and minimum total costs cost, and the total cost of mEaD-5G is just high than mEaD.
is the lowest, 0.02975. Therefore, in the following simulation It means that mEaD and mEaD-5G have better optimization
experiment, we set α = 0.4. on network delay and energy consumption. Because mEaD and
In addition, we calculate the ideal energy consumption based mEaD-5G consider network delay, they place edge servers close
on the method described in section 4, which is Pideal = to the base stations. Moreover, the particles’ states update pro-
242488 W. Moreover, the interval of base stations usually is cess considers the cost profit, increasing the possibility of finding

Authorized licensed use limited to: C. V. Raman Global University - Bhubaneswar. Downloaded on February 04,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
 LING et al.: EDGE SERVER PLACEMENT ALGORITHM BASED ON GRAPH CONVOLUTION NETWORK 5237

Fig. 8. Edge servers’ average network delay in high mobility area.

Fig. 7. Edge servers’ total cost.

the optimal edge server placement plan. However, mEaD-5G

uses 5G edge servers in the edge server placement, which
increases the energy consumption of edge servers. Therefore,
the total cost of mEaD-5G is slightly higher than that of mEaD.
However, EESP only optimizes energy consumption but ig-
nores network delay. Therefore, the total cost of EESP is higher
than that of mEaD and mEaD-5G. Conversely, GA focuses
on energy consumption and network delay, but its simple as-
signment mechanism places many unnecessary edge servers.
Therefore, too many edge servers increase the total cost of GA,
resulting in a higher total cost than EESP. Fig. 9. Edge servers’ maximum utilization in high mobility area.
Finally, Random and Top-K have the highest total costs among
these edge server placement algorithms. It means that simply
placing the edge servers randomly or to the network traffic hot maximum utilization of edge servers in 24 hours. Figs. 8 and 9
spot areas is not efficient for edge server placement problem. show the detailed results.
Moreover, we find that mEaD and mEaD-5G has fewer over- As shown in Fig. 8, mEaD has a lower network delay, which
load edge servers. That is because mEaD and mEaD-5G use the means that mEaD effectively controls the network delay in the
prediction results of network traffic distribution as the placement high mobility areas. In addition, the average network delay of
basis, effectively avoiding the overload problem. In addition, mEaD-5G is higher than mEaD, but it is still in an appropriate
mEaD-5G considers the network traffic change in high mobility coverage. It means that mEaD-5G focuses more on resource
base stations and reserves the resource for significant network reservation than network delay in high mobility areas. The net-
traffic changes. It makes mEaD-5G has the fewest overload edge work delay of EESP is high than the others. Without considering
servers. network delay, EESP also has a high network delay, like the
Because of the lack of the prediction of network traffic distri- other areas. Moreover, because of the consideration of network
bution, EESP has many overload edge servers. In addition, GA delay, GA has a lower network delay than EESP. However, GA’s
also lacks the predictions of network traffic distribution, result- inefficient search mechanism makes the edge servers far from
ing in a high number of overload edge servers. Moreover, GA the edge servers in high mobility areas. Therefore, its network
assigns the base stations to their closest edge servers. Therefore, delay is still higher than mEaD and mEaD-5G. Because of the
the edge servers in the network traffic hot spot areas manage all high density of edge servers, Random and Top-k have the lowest
surrounding base stations, which have too much workload and network delay.
create the maximum number of overload edge servers. At last, In addition, as shown in Fig. 9, the average maximum utiliza-
Random and Top-k place excessive edge servers. Although the tion of mEaD and mEaD is in an appropriate range. The average
overload edge serves’ number of them is the minimum, the total maximum utilization of mEaD is about 80%, representing that
costs of edge servers placed by them are too high. the resource is enough for the present network traffic. However,
for the future network traffic peak, mEaD may cause many edge
servers to be overloaded. The average maximum utilization of
I. Edge Server’s Performance in High Mobility Area
mEaD-5G is about 60%, indicating that its reserved resource
To compare the edge server performance in high mobility is more than that of mEaD. So that mEaD-5G can process the
areas, we calculate the average network delay and average future network traffic peak perfectly.

Authorized licensed use limited to: C. V. Raman Global University - Bhubaneswar. Downloaded on February 04,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
 5238 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 4, APRIL 2023

measurement by considering the channel interference problem

in wireless communication.

REFERENCES
[1] I. A. Elgendy and R. Yadav, “Survey on mobile edge-cloud computing: A
taxonomy on computation offloading approaches,” in Security and Privacy
Preserving for IoT and 5G Networks, Berlin, Germany: Springer, 2022,
pp. 117–158.
[2] R. Yadav et al., “Smart healthcare: Rl-based task offloading scheme for
edge-enable sensor networks,” IEEE Sensors J., vol. 21, no. 22, pp. 24
910–24 918, Nov. 2021.
[3] R. Yadav, W. Zhang, O. Kaiwartya, H. Song, and S. Yu, “Energy-latency
tradeoff for dynamic computation offloading in vehicular fog computing,”
IEEE Trans. Veh. Technol., vol. 69, no. 12, pp. 14 198–14 211, Dec. 2020.
Fig. 10. Total cost in 24 hours. [4] X. Deng, Z. Tang, L. T. Yang, M. Lin, and B. Wang, “Confident information
coverage hole healing in hybrid industrial wireless sensor networks,” IEEE
Trans. Ind. Inform., vol. 14, no. 5, pp. 2220–2229, May 2018.
[5] Y. Li and S. Wang, “An energy-aware edge server placement algorithm in
Moreover, EESP’s average maximum utilization is the highest mobile edge computing,” in Proc. IEEE Int. Conf. Edge Comput., 2018,
among these algorithms, above 100%. It indicates that EESP pp. 66–73.
cannot provide enough resources for the high mobility area. [6] A. Santoyo González and C. Cervelló Pastor, “Edge computing node place-
ment in 5G networks: A latency and reliability constrained framework,”
In addition, GA has lower average maximum utilization than in Proc. 6th IEEE Int. Conf. Cyber Secur. Cloud Comput. 5th IEEE Int.
mEaD, mEaD-5G and EESP. It is because GA places more Conf. Edge Comput. Scalable Cloud, 2019, pp. 183–189.
edge servers than those three algorithms. Random and Top-k [7] K. Cao, L. Li, Y. Cui, T. Wei, and S. Hu, “Exploring placement of het-
erogeneous edge servers for response time minimization in mobile edge-
have the minimum average maximum utilization, 25%. It is also cloud computing,” IEEE Trans. Ind. Inform., vol. 17, no. 1, pp. 494–503,
because they place too many edge servers, and each edge server Jan. 2021.
is underloaded. [8] F. R. de Souza, M. D. de Assunçao, E. Caron, and A. d. S. Veith, “An
optimal model for optimizing the placement and parallelism of data stream
processing applications on cloud-edge computing,” in Proc. IEEE 32nd Int.
Symp. Comput. Architecture High Perform. Comput., 2020, pp. 59–66.
J. Edge Server’s Performance in Each Hour [9] L. Nie et al., “Network traffic prediction in industrial Internet of Things
backbone networks: A multitask learning mechanism,” IEEE Trans. Ind.
To further analyze the edge servers’ performance, we also Inform., vol. 17, no. 10, pp. 7123–7132, Oct. 2021.
compare the total cost performance of our algorithms and base- [10] M. C. Calzarossa, M. L. Della Vedova, L. Massari, D. Petcu, M. I. M.
line algorithms in each hour. Fig. 10 shows the detailed total Tabash, and D. Tessera, “Workloads in the Clouds,” in Principles of
Performance and Reliability Modeling and Evaluation: Essays in Honor
cost of each algorithm, of which the edge servers’ coverage is of Kishor Trivedi on His 70th Birthday, ser. Springer Series in Reliability
550 m. Engineering, L. Fiondella and A. Puliafito, Eds. Cham, Switzerland:
In Fig. 10, network traffic varies with time and further influ- Springer International Publishing, 2016, pp. 525–550.
[11] T. N. Kipf and M. Welling, “Semi-supervised classification with graph
ences the performance of edge servers. In detail, in 0:00–7:00, convolutional networks,” in Proc. 5th Int. Conf. Learn. Representations,
most edge servers have low network traffic, resulting in low total 2017.
costs. Then, from 8:00, most people start to work, so the network [12] S. Guha and S. Khuller, “Greedy strikes back: Improved facility location
algorithms,” J. Algorithms, vol. 31, no. 1, pp. 228–248, Apr. 1999.
traffic begins to increase and causes an increment in total cost. [13] J. Kennedy and R. Eberhart, “Particle swarm optimization,” in Proc. Int.
Moreover, the total cost reaches the highest at 12:00 and 21:00, Conf. Neural Netw., vol. 4, 1995, pp. 1942–1948.
meaning the highest workload on edge servers. Finally, the [14] Z. Zhong, J. He, M. A. Rodriguez, S. Erfani, R. Kotagiri, and R. Buyya,
“Heterogeneous task co-location in containerized cloud computing envi-
performance of all the edge server placement algorithms follows ronments,” in Proc. IEEE 23 rd Int. Symp. Real-Time Distrib. Comput.,
the above rules. 2020, pp. 79–88.
[15] A. R. Benamer, H. Teyeb, and N. Ben Hadj-Alouane, “Latency-Aware
Placement Heuristic in Fog Computing Environment,” in On the Move
IX. CONCLUSION to Meaningful Internet Systems. OTM 2018 Conferences, ser. Lecture
Notes in Computer Science, H. Panetto, C. Debruyne, H. A. Proper, C. A.
This paper defines a cost minimization in edge computing Ardagna, D. Roman, and R. Meersman, Eds. Cham, Switzerland: Springer
International Publishing, 2018, pp. 241–257.
servers placement problem to address the trade-off between [16] A. Brogi and S. Forti, “QoS-Aware deployment of IoT applications
energy cost and users’ delay tolerance. Firstly, we propose through the fog,” IEEE Internet Things J., vol. 4, no. 5, pp. 1185–1192,
an edge server placement method based on a network traffic Oct. 2017.
[17] V. De Maio and I. Brandic, “First hop mobile offloading of DAG computa-
prediction model. Then, we propose two edge server placement tions,” in Proc. 18th IEEE/ACM Int. Symp. Cluster, Cloud Grid Comput.,
algorithms, mEaD and mEaD-5G, focusing on network delay 2018, pp. 83–92.
and energy consumption in high mobility areas. Extensive exper- [18] K. Xiao, Z. Gao, Q. Wang, and Y. Yang, “A heuristic algorithm based on re-
source requirements forecasting for server placement in edge computing,”
iments are conducted on a large real-world dataset. Our proposed in Proc. IEEE/ACM Symp. Edge Comput., 2018, pp. 354–355.
solutions show the total cost of edge servers and overloaded [19] J. Zhang, B. Ma, and J. Huang, “Deploying GIS services into the edge:
edge servers by 23.98% and 52.71% less than the existing solu- A study from performance evaluation and optimization viewpoint,” Secur.
Commun. Netw., vol. 2020, Oct. 2020, Art. no. e8822990.
tions. Moreover, results show significant advantages on the total [20] I. Althamary, C.-W. Huang, P. Lin, S.-R. Yang, and C.-W. Cheng,
performance of edge servers, especially in high mobility areas. “Popularity-based cache placement for fog networks,” in Proc. 14th Int.
One promising future direction is to improve network delay Wireless Commun. Mobile Comput. Conf., 2018, pp. 800–804.

Authorized licensed use limited to: C. V. Raman Global University - Bhubaneswar. Downloaded on February 04,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
 LING et al.: EDGE SERVER PLACEMENT ALGORITHM BASED ON GRAPH CONVOLUTION NETWORK 5239

[21] S. Wang, Y. Zhao, J. Xu, J. Yuan, and C.-H. Hsu, “Edge server place- [28] Y. Gui, D. Wang, L. Guan, and M. Zhang, “Optical network traffic
ment in mobile edge computing,” J. Parallel Distrib. Comput., vol. 127, prediction based on graph convolutional neural networks,” in Proc. Opto-
pp. 160–168, May 2019. Electron. Commun. Conf., 2020, pp. 1–3.
[22] G. Yang, Q. Sun, A. Zhou, S. Wang, and J. Li, “Poster abstract: Access [29] X. Ji and Q. Meng, “Traffic classification based on graph convolutional
point ranking for cloudlet placement in edge computing environment,” in network,” in Proc. IEEE Int. Conf. Adv. Elect. Eng. Comput. Appl., 2020,
Proc. IEEE/ACM Symp. Edge Comput., 2016, pp. 85–86. pp. 596–601.
[23] D. Bhatta and L. Mashayekhy, “Generalized cost-aware cloudlet place- [30] D. Zhu et al., “Understanding place characteristics in geo-
ment for vehicular edge computing systems,” in Proc. IEEE Int. Conf. graphic contexts through graph convolutional neural networks,”
Cloud Comput. Technol. Sci., 2019, pp. 159–166. Ann. Amer. Assoc. Geographers, vol. 110, no. 2, pp. 408–420,
[24] D. E. Goldberg, B. Korb, and K. Deb, “Messy genetic algorithms: Motiva- Mar. 2020.
tion, analysis, and first results,” Complex Syst., vol. 3, no. 5, pp. 493–530, [31] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
1989. recognition,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2016,
[25] M. M. E. Mahmoud, J. J. P. C. Rodrigues, K. Saleem, J. Al-Muhtadi, pp. 770–778.
N. Kumar, and V. Korotaev, “Towards energy-aware fog-enabled cloud of [32] F. Xu, Y. Li, H. Wang, P. Zhang, and D. Jin, “Understanding mobile traffic
things for healthcare,” Comput. Elect. Eng., vol. 67, pp. 58–69, Apr. 2018. patterns of large scale cellular towers in urban environment,” IEEE/ACM
[26] S. Yi, Z. Hao, Q. Zhang, Q. Zhang, W. Shi, and Q. Li, “LAVEA: Latency- Trans. Netw., vol. 25, no. 2, pp. 1147–1161, Apr. 2017.
aware video analytics on edge computing platform,” in Proc. IEEE 37th [33] W. Kirch, Pearson’s Correlation Coefficient. The Netherlands, Springer,
Int. Conf. Distrib. Comput. Syst., 2017, pp. 2573–2574. 2008, pp. 1090–1091.
[27] R. G. Aryal and J. Altmann, “Dynamic application deployment in feder- [34] M. H. Alsharif, A. H. Kelechi, J. Kim, and J. H. Kim, “Energy
ations of clouds and edge resources using a multiobjective optimization efficiency and coverage trade-off in 5G for eco-friendly and sus-
AI algorithm,” in Proc. 3rd Int. Conf. Fog Mobile Edge Comput., 2018, tainable cellular networks,” Symmetry, vol. 11, no. 3, Mar. 2019,
pp. 147–154. Art. no. 408.

Authorized licensed use limited to: C. V. Raman Global University - Bhubaneswar. Downloaded on February 04,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.