Computer Communications 169 (2021) 71–80

Contents lists available at ScienceDirect

Computer Communications
journal homepage: www.elsevier.com/locate/comcom

Intelligent workload allocation in IoT–Fog–cloud architecture towards

mobile edge computing
M. Abbasi a ,∗, E. Mohammadi-Pasand a , M.R. Khosravi b,c
a Department of Computer Engineering, Engineering Faculty, Bu-Ali Sina University, Hamedan, Iran
b
Department of Computer Engineering, Persian Gulf University, Bushehr, Iran
c
Telecommunications Group, Department of Electrical and Electronic Engineering, Shiraz University of Technology, Shiraz, Iran

ARTICLE INFO ABSTRACT

Keywords: Because of the tremendous growth in the number of smart vehicular devices and 5G mobile technologies, the
Internet of Things (IoT) Internet of Things (IoT) has experienced rapid expansion. This has led to a considerable increase in the volume
Mobile edge computing (MEC) of sensory data produced from, but not limited to, monitoring devices, traffic congestion in cities, safety, and
Multi-objective genetic algorithm
pollution control. Cloud computing can deal with the corresponding workload by providing virtually unlimited
Workload allocation
computational resources. But, given the importance of the quality of service and security in delay-sensitive
requests, other solutions like fog computing have also been introduced to speed up processing and management
of sensory data in real scenarios like smart grid and IoT. Processing workloads at the network edge reduces
the delay in mobile edge computing, but it highly increases the consuming power. Therefore, there is an
urgent need for the improvement of the energy model of fog devices at the network edge. This paper is an
attempt to modify this model using the green energy concept and reduce both delay and power consumption in
multi-sensorial frameworks in secure IoT systems. In the proposed method, a Genetic Algorithm (GA) is used
for handling a large number of requests and the corresponding quality and security limitations. Simulation
results show that the proposed method can simultaneously reduce the delay and the power consumption of
edge devices compared to a baseline strategy.

1. Introduction between the end-users and the cloud will limit the processing of these
requests in the cloud [14,15].
Internet of Things (IoT) is overgrowing as a new paradigm of
wireless communication. The main idea behind it is the connection of 1.2. Communication bandwidth
all objects such as sensors in smart cities [1,2], mobile edge devices [3],
intelligent vehicles [4,5], wearable gadgets [6], and e-healthcare items
It is predicted that the data produced each year by smart mobile
to the secure internet [7,8]. Due to the limited battery life and com-
networks in the United States will reach 1000 petabytes in the near
puting capability of the terminal devices [9], the cloud server usually
future [16]. The traffic produced in Google servers has already reached
displays the role that receives and processes the data offloaded from
all kinds of terminals in traditional IoT networks [10,11]. Due to the 1 petabyte each month. Also, AT&T networks use 200 petabytes of
abundance of energy sources in the cloud, cloud-based networks have bandwidth each year. Sending all the data to the cloud requires a
a higher processing capacity than terminal nodes [12]. However, the large amount of bandwidth used for sophisticated methods of high-
development of the IoT has challenged cloud computing in several performance flow processing [13,17,18].
ways [13], as follows.
1.3. Resource limitations
1.1. Delay
Many IoT devices have limited resources and are unable to fulfill
Many manufacturing systems, smart networks, and petroleum pro- their own computation needs [8,15]. Communication between these
duction systems are sensitive to delays greater than dozens of mil- devices and the cloud imposes high costs because it entails the use of
liseconds between sensors and control nodes. Thus, the long distance complicated protocols that need substantial resources [5,8,14].

∗ Corresponding author.
E-mail address: abbasi@basu.ac.ir (M. Abbasi).

https://doi.org/10.1016/j.comcom.2021.01.022
Received 19 July 2020; Received in revised form 27 December 2020; Accepted 19 January 2021
Available online 23 January 2021
0140-3664/© 2021 Elsevier B.V. All rights reserved.
M. Abbasi, E. Mohammadi-Pasand and M.R. Khosravi Computer Communications 169 (2021) 71–80

1.4. New security challenges The use of fog computing in the architecture of the IoT with the aim
of improving power consumption and QoS has turned into a significant
The cyber-security paradigm which has been designed to address field of research [33]. In this paper, we focus on the smaller distance
the current organizational networks and data centers is incapable of between client and server nodes and consider the increased flexibility
coping with the following security challenges [19–22]: of network topology in the architecture of fog computing systems to
⋅ Updating numerous applications and security credentials in a large decrease their power consumption [34,35]. At the same time, given a
number of mobile devices: large number of fog devices and terminal nodes with limited power
By increasing the number of devices that are connected to the supply, decision-making about workload allocation is the focus of our
Internet, managing their security credential and updating their security attention. The next section is a review of studies previously conducted
applications becomes a challenge. about workload allocation in terms of delay and power consumption.
⋅ Protecting devices with limited resources:
Many IoT devices do not have sufficient resources to protect them- 2. Related work
selves. The lifespan of these devices can be very long, and upgrading
their hardware and software is impractical. However, these devices In 2016, Ruilong Deng et al. [36] studied the interaction between
must remain secure for a long time. So, a fundamental question arises: fog and cloud. Workload allocation in the cloud and the fog was for-
how can a large number of such devices with limited resources be mulated as an optimization problem. The aim was to minimize power
protected from security attacks? consumption in the face of delay limitations. Their study proposed
⋅ Assessing the security status of distributed systems reliably: an approximate algorithm that divides the problem into three sub-
The IoT supports a large number of large distributed systems. There- problems: total processing delay and total power consumption of fog
fore, the ability to detect the reliability of these distributed systems devices, total processing delay and total power consumption of the
reliably is essential. However, common approaches to issues of scal- cloud, and delay of transmission between fog devices and the cloud.
ability support and reliable regulatory needs are problematic. Many The processing delay of fog devices at the network edge is less than that
IoT devices with limited resources are not able to support remote of the cloud, but their energy model is not more optimum, affecting the
authentication. Even if possible, forcing devices to perform remote efficiency of computation loads on the network.
authentication is costly and complex. Mubarak et al. [37] investigated workload allocation using a genetic
⋅ Responding to security compromises without disrupting the com- algorithm and BIP. Their criteria for the processing performance of the
munication: fog include power consumption and QoS (delay). They developed a
Today’s disaster response solutions rely more on brutal mechanisms model for each criterion. The power consumption model is divided into
such as shutting down a potentially damaged system, reinstalling and three sub-problems: total power consumption during the transmission
restarting its software, or replacing components and their subsystems. of a user’s request, total power consumption during the processing
These reactions ignore the danger of the devices and can cause irre- and storing of the request, and total power consumption during the
versible disruption to important systems. However, uninterrupted and processing of the user’s request in the cloud. The delay model is divided
safe operation is especially important in most cases, even if the system into two sub-problems: the sum of delays in the transmission of the
is compromised. user’s request from the target device to the data center, and the cloud’s
To overcome the above challenges in cloud-based mobile networks, delay in the processing and storing of data. Finally, the fog architecture
the fog computing model has been exploited. This model aims to is modeled as an undirected graph and two optimization methods based
process mobile workloads near the end-users and support the wide on genetic algorithm and BIP are used. The issue with this method is
variety of IoT requests [23]. Fog computing reduces the traffic im- that it could not improve power consumption and delay on small and
posed on the core network and provides higher security due to the medium scales on the IoT.
processing of requests on a more limited scale [24]. Fog devices can In 2017, D. Kanapram et al. [38] introduced a framework for exam-
be efficiently used by mobile network operators in the form of small ining the difference between efficiency and energy consumption. They
stations [10]. They can use both power supplies and solar energy created this framework for optimizing resource and energy consump-
to recharge their batteries [25]. Using technologies such as network tion in cloud management software. Also, they presented a multilayer
function virtualization (NFV) and software-defined networking (SDN), architecture and implemented it with the OpenStack tool to have a
we can implement services such as flow processing [26,27], network high level of abstraction of energy consumption in cloud and fog
trust [28], cashing [29], and control on fog devices using a central computing. In this architecture, they modeled the effect of different
controller and develop the central cloud at the network edge [30,31]. management methods in terms of workload allocation and examined
Given that fog devices at the network edge are heterogeneous, the the effect of these methods on energy consumption with different levels
packets received in terminal nodes are processed in the nearest fog of granulation (for example, a broker or the whole cloud). Despite their
device, thereby preventing them from being sent to cloud servers. If acceptable performance, their proposed method was still slow.
the packets are processed in fog devices (instead of cloud servers), the Yousef-poor et al. [39] proposed a general framework for workload
end-user is provided with a higher quality of service (QoS) [13,32]. On allocation on the IoT. This framework is aimed at minimizing the
the other hand, due to the layers’ difference in computation power, the response time. For this purpose, the decisions are made based on
allocation of workloads to the layers has become an important issue. whether the requests are easy or difficult to process. If the duration
There are various strategies for a balanced distribution of workloads of response to the requests is less than a specific threshold, the fog
among the layers, e.g., filling each layer as much as possible. To de- node accepts the request; otherwise, the requests are sent to one of the
velop an optimum allocation strategy, several factors, including delay, neighbors of the fog device or to the cloud. If the number of requests
bandwidth, energy, and cost, should be taken into account [13]. is small, this method assigns all the processes to the fog device or
Mobile IoT devices are carried by their owners, such as wear- to a close neighbor. One of the drawbacks of this method is that it
able devices (e.g., fitness trackers, wearable cameras, smart clothing, examines different scenarios within a distributed environment. Another
and sports bracelets) and mobile smart devices (smartphones, smart- problem is that the allocation of requests to the fog or the cloud cannot
watches, smart glasses, vehicles). All the devices of an owner can be determined by looking at the processing delay. To allocate the
form a group and communicate with each other through wireless requests, we should consider the power consumption of fog devices,
networks [23]. On wearable devices, a mobile phone can be turned cloud devices, and routers that mediate between the fog and the cloud.
into a fog device to provide local control and analysis programs for Zhang et al. [10] proposed a method, named FEMTO, for balanced
the wearable device [12]. workload allocation and minimizing power consumption on fog-based

72
M. Abbasi, E. Mohammadi-Pasand and M.R. Khosravi Computer Communications 169 (2021) 71–80

IoT networks. Their proposed method is based on the minimization Table 1

Summary of key notations.
of energy consumption and the fair distribution of loads among fog
Symbol Unit Description
devices. Increased distance between the terminal node and the central
controller can increase power consumption and delay. Therefore, this 𝑙 bit Overall task size of the terminal node
𝑝𝑖 – The probability that fog node 𝐢 is idle
distance should be taken into account for workload allocation. In the 𝑙𝑚𝑖𝑛 bit Minimum overall task size of the terminal node
mentioned study, fog devices with different configurations may vary in 𝑙𝑚𝑎𝑥 bit Maximum overall task size of the terminal node
their power consumption. As a result, workloads are processed locally 𝑙𝑖 bit Subtask size offloaded to fog node 𝐢
𝑙𝑡 bit Subtask size processed locally at the terminal node
if the power consumption of fog devices exceeds a certain threshold.
W Hz Spectrum bandwidth for task offloading
The local processing of loads will lead to a rise in the processing of Bi – The spectral efficiency of the wireless link from the
requests. On the other hand, modification of the energy model of fog terminal node to fog node 𝐢
devices may affect the power consumption of other fog devices and the 𝛾i – Path loss
ki – Shadowing factors of this wireless link
QoS on the end-user’s side.
Ii dBm Interference power
In 2018, Hang Wu et al. [40] introduced an algorithm called GLOBE. N0 dBm/Hz Noise power spectral density
In this algorithm, they combine two geographical load balancing mech- Pi J/s Transmission power of the terminal node to fog node 𝐢
nt cycle/bit CPU cycles for processing 1-bit data at the terminal node
anisms and workload input control to enhance the performance of
ni cycle/bit CPU cycles for processing 1-bit data at fog node 𝐢
network stations. The mentioned algorithm operates online without fi cycle/s CPU frequency of the terminal node
the need for future system information and responds to significant ft cycle/s CPU frequency of the fog node 𝐢
challenges due to battery status and energy constraints. GLOBE, in 𝜃i J/cycle Energy consumption per CPU cycle of fog node 𝐢
𝜃t J/cycle Energy consumption per CPU cycle of fog node 𝐢
comparison with other offline algorithms, reduces costs by 50% and B J Maximum capacity of battery for fog nodes
balances system usage and system performance. But this algorithm Overall energy consumption of fog nodes
does not optimally distribute computational workload between network Esum J Upper bound of the transmission power of the terminal
nodes. node
Pmax J/s Maximum power
In 2018, Jiafu Wan et al. [10] proposed a workload balancing and
planning method called ELBS to improve energy consumption based
on fog calculations. In their method, formulas are proposed for the
model of energy consumption related to user volume on fog nodes, 3. System modeling
then an optimization function is obtained with the aim of balancing the
workload in clusters. This function is improved by a batch optimization In this model, workloads are distributed according to delay and
algorithm. Finally, a multipurpose system is formed to achieve a cluster power consumption. The overall delay includes delays in the processing
distribution schedule. Experiments with the method on robots show of requests in the terminal node, the processing of requests in fog de-
that although the performance of the robots in the created environment vices, and the transmission of packets on the wireless network between
has improved compared to other scheduling methods and the workload the terminal node and the central controller. The power consumption
has been distributed equally among them, a solution to create balance of the terminal node and fog devices and the energy needed by the
in energy consumption and delay is not raised, and only the energy terminal node to transmit the requests over the wireless network should
standard is optimized. be considered for decision-making. Fog devices are available with
In 2019, Forough Shirin Abkenar et al. [33] proposed a new a probability of 𝑝𝑖 and regularly inform the controller about their
workload distribution algorithm, called the Energy Balance Algorithm availability.
(EBA), for a Fog–IoT three-layer network. EBA uses two optimization As can be seen in Fig. 1, the fog devices are connected to a central
models to reduce energy consumption and bandwidth in the network controller, and the requests are distributed among the devices based
and ensures energy balance between all fog nodes. The first optimiza- on the probability of their availability. The workload comes from the
tion model, called Best Transfer Power and Transfer Speed (BTPR), user’s side to the terminal node, and the node either processes them
finds the optimal transmission power and transmission speed of the locally or distributes them among the available fog devices. As can be
end nodes so that no request is lost. In the second optimization model, seen in Fig. 1, the fog devices are connected to a central controller and
a topology is presented between the end nodes and the fog node, the requests are distributed among the devices based on the probability
which is called the best model of important nodes. The purpose of this of their availability. The workload comes from the user’s side to the
topology is to find the best fog node to serve the final node so that terminal node, and the node either process them locally or distributes
energy balance is established between all fog nodes. The simulation them among the available fog devices. Table 1 summarizes the main
results show that the proposed EBA can achieve a reduction in energy parameters. We shall examine our proposed systems in terms of four
consumption across the network and fog nodes. This study, like other models: workload, delay, power consumption, and battery status. For
studies, attempts to optimize only one criterion (energy consumption) each model, a set of equations are presented. Next, the models are
and does not provide a solution to balance energy consumption and simulated using a genetic algorithm.
latency.
Among the factors that have been largely neglected in previous 3.1. Workload
studies are decision-making based on multiple indexes, lack of attention
to the power consumption of certain network nodes, QoS on different As this scenario is based on a discrete-time model, the simulation
scales of the IoT, and varying transmission delay for devices and cloud time is divided into smaller intervals, and the number of active fog
servers. Therefore, if the energy model of cloud devices is improved, devices is specified in each interval. 𝑙 denotes the rate at which the
they will be able to receive and process more workload, and the QoS entire workload arrives at the terminal node ([𝑙𝑚𝑖𝑛 .𝑙𝑚𝑎𝑥 ] ∈ 𝑙). Next, the
will be enhanced for the end-user. Also, the use of bandwidth will terminal node decides how much of the workload should be processed
considerably improve. locally, which is indicated by 𝑙𝑡 . Finally, the workload 𝑙 − 𝑙𝑡 , which is
In addition to modeling the system, the next section will explain denoted by 𝑙𝑖 , is allocated to the available fog nodes. The number of
the problem of terminal nodes and fog devices concerning power fog devices available in each allocation may vary with the probability
consumption and delay. of 𝑝𝑖 .

73
M. Abbasi, E. Mohammadi-Pasand and M.R. Khosravi Computer Communications 169 (2021) 71–80

Fig. 1. Task offloading model in a secure IoT–Fog paradigm in MEC.

3.2. Delay 𝑑𝑖 . If 𝑙𝑖 bits are allocated to the fog device 𝑖, 𝑙𝑖 𝑛𝑖 CPU cycles are
required. 𝑙𝑖 is sent to the device 𝑖 and, after processing, the result
The proposed model recognizes three types of delay: is returned to the terminal node. The result of the processing is
often in form of a small packet and a control signal [41]. As
• Delay in the transmission of workloads over the wireless network: in [41,42], and [43], the delay in sending the final result to the
The delay in the transmission of workloads between the ter- terminal node is not taken into account. 𝑓𝑖 is the CPU frequency
minal node and the central controller is denoted by 𝑑𝑡𝑟𝑎𝑛𝑠 and of the fog device 𝑖 [10].
calculated by the following equation [10]. 𝑙𝑖 𝑛𝑖
𝑑𝑖 = (4)
𝑙𝑖 𝑓𝑖
𝑑𝑡𝑟𝑎𝑛𝑠 = (1)
𝑊 𝐵𝑖
Thus, the overall delay is:
In this equation, 𝑊 is the bandwidth between the terminal 𝑙𝑡 𝑛𝑡 𝑙 𝑙𝑛
node and the controller, and 𝐵𝑖 is the spectral efficiency of the 𝐷𝑡𝑜𝑡𝑎𝑙 = 𝑑𝑡 + 𝑑𝑡𝑟𝑎𝑛𝑠 + 𝑑𝑖 = + 𝑖 + 𝑖 𝑖 (5)
𝑓𝑡 𝑊 𝐵𝑖 𝑓𝑖
wireless link of the terminal node and the fog device 𝑖. Given the
transmission power of the sender of the signal, 𝐵𝑖 is calculated 3.3. Power consumption
by the following equation [10].
( )
pi 𝛾i ki This system contains three types of energy:
Bi = log2 1 + (2)
Ii + WN0
• The energy consumed for the processing of packets in the terminal
where 𝛾𝑖 and 𝑘𝑖 are the channel loss rate and the shadowing
node: The energy consumed for the processing of packets in the
factor in the wireless link, respectively. In this equation, 𝐼𝑖 and
terminal node is denoted by 𝐸𝑡 . The energy consumed in each
𝑁0 denote the interference power and the noise power spectral
CPU cycle of the terminal node is 𝜃𝑡 . The energy consumed for
density respectively.
the processing of one bit of data in the terminal node is 𝑛𝑡 𝜃𝑡 . If 𝑙𝑡
• Delay in the processing of workloads in the terminal node: The delay bits are allocated to the terminal node, the power consumption
in the local processing of workloads in the terminal node is will be calculated by the following equation, see more in [10].
denoted by 𝑑𝑡 . If 𝑙𝑡 bits are allocated to the terminal node, 𝑙𝑡 𝑛𝑡 𝐸𝑡 = 𝑙𝑡 𝑛𝑡 𝜃𝑡 (6)
CPU cycles are required. Given the frequency of the terminal
node (𝑓𝑡 ), the delay in local processing is obtained by the • The energy consumed for the transmission of 𝑙𝑖 bits to the central
following equation. controller through the wireless channel: The energy consumed
𝑙𝑡 𝑛𝑡 for the transmission of 𝑙𝑖 bits to the central controller through
𝑑𝑡 = (3)
𝑓𝑡 the wireless channel is denoted by 𝐸𝑡𝑟𝑎𝑛𝑠 . 𝐵𝑖 represents the
spectral efficiency of the wireless link and 𝑝𝑖 is the transmission
• Delay in the processing of workloads in the fog device 𝑖: Delay in power of the signal transmitter. If 𝑙𝑖 bits are allocated to the fog
the processing of workloads in the fog device 𝑖 is denoted by device 𝑖 with the current bandwidth 𝑊 , the transmission energy

74
M. Abbasi, E. Mohammadi-Pasand and M.R. Khosravi Computer Communications 169 (2021) 71–80

will be calculated according to Eq. (7) [10]. Based on [41,42], and also crossover and mutation operators. Subsequently, the two pop-
and [43], the amount of consumed energy is not taken into ulations combine, and a new population (𝑅𝑡 ) will be formed. In the next
account in transmitting the final result to the terminal node. step, the new population is divided into dominated and non-dominated
𝑙𝑖 𝑝𝑖 categories. Then it is filled with various parts of non-dominated Pareto
𝐸𝑡𝑟𝑎𝑛𝑠 = (7) fronts. Filling the new population initiates with the first non-dominated
𝑊 𝐵𝑖
Pareto front and carries on with the second non-dominated Pareto
• The energy consumed for the processing of packets in the fog device front. Consequently, this process goes on until the capacity of the
𝑖: The energy consumed for the processing of packets in the population is completed. Pareto fronts that cannot be substituted in
fog device 𝑖 is denoted by 𝐸𝑖 . The energy consumed in each the new population are removed. If only a specific number of single
CPU cycle of the fog device is 𝜃𝑖 . The energy consumed for the Pareto fronts are permitted to enter the new population, instead of
processing of one bit of data in the terminal node is 𝑛𝑖 𝜃𝑖 . If 𝑙𝑖 bits using random selection, a certain procedure called crowding distance
are allocated to the fog device 𝑖, the power consumption will be sorting is used [50]. NSGA II algorithm is altered to resolve the problem
calculated by the following equation: of workload allocation in fog computing. Our suggested multi-objective
optimization algorithm is introduced as Algorithm 1.
𝐸𝑖 = 𝑙𝑖 𝑛𝑖 𝜃𝑖 (8)

Therefore, the overall power consumption for 𝑙 is as follow- 4.1. Defining chromosome
ing [10].
𝑙𝑖 𝑝𝑖 According to the nature of the problem under study, which contains
𝐸𝑡𝑜𝑡𝑎𝑙 = 𝐸𝑡 + 𝐸𝑡𝑟𝑎𝑛𝑠 + 𝐸𝑖 = 𝑙𝑡 𝑛𝑡 𝜃𝑡 + + 𝑙𝑖 𝑛𝑖 𝜃𝑖 (9) both latency and energy consumption of user requests, each workload is
𝑊 𝐵𝑖
allocated to the terminal node or one of the accessible fog devices. Next,
3.4. Battery status their latency and energy consumption are calculated. The allocation
of each request is shown by the index of the corresponding terminal
The maximum total capacity of the batteries at the network edge node or one of the fog devices. The first step in implementing the
in the interval 𝑡 is 𝑏(𝑡) ∈ [0, 𝐵]. These batteries are recharged by solar algorithm is to develop the random workload allocation of requests to
and wind energy. The battery level is initially taken as zero and will the terminal node or one of the fog devices (lines 3–5 of Algorithm
be recharged after 𝑡 intervals. If the battery used at the network edge 1). Next, crossover and mutation will occur and, based on the fitness
is empty, the power from the electrical grid will be used. If the sum of function (lines 6–11 and 25–27 of Algorithm 1) of delay and energy
the energy consumed by the fog devices 𝐸𝑠𝑢𝑚 is less than the level of consumption of each solution, the best answer will be chosen.
the batteries at the edge, the entire energy required for the processing
of the packets will be supplied by the batteries. Otherwise, the amount
of battery consumption will be 𝑏(𝑡) and the amount of power consumed
from the electric grid will be 𝐸𝑠𝑢𝑚 − 𝑏(𝑡). If battery energy is used for
the processing of requests, the power consumption model of fog devices
will improve, more requests will be sent to these devices, and the QoS
for the end-user will increase.

𝑃1 ∶ min 𝐷𝑡𝑜𝑡𝑎𝑙 (10)

s.t. 𝐸𝑡𝑟𝑎𝑛𝑠 < 𝐸𝑚𝑎𝑥
0 ≤ 𝑙𝑖 ≤ 𝑙, 0 ≤ 𝑃𝑖 ≤ 𝑃𝑚𝑎𝑥

In the problem 𝑃1 , the aim is to minimize delay for the end-

user depending on the energy limitations and the number of packets
transmitted to fog devices. 𝐸𝑚𝑎𝑥 is the maximum power consumption
of the terminal node. The first limitation indicates that during the
transmission of requests to the central controller, the energy consumed
by the terminal node to transmit the requests must not exceed the
maximum energy of this node. According to the second limitation, the
number of requests sent to the central controller by the terminal node
must not exceed the load transmitted to this node. The transmission
power of the terminal node ranges between 0 and 𝑃𝑚𝑎𝑥 .

4. Proposed method

In this section, a model of one terminal node with 10 fog devices

is presented. Requests are processed at the terminal node or on the
available fog device. Due to the existence of multiple variables, the
solution to the problem of the workload allocation of terminal nodes
and fog devices is not a straightforward one. Therefore, NSGA II multi-
objective algorithm was used in this study. NSGA II is a well-known
multi-objective algorithm [44]. This algorithm has efficiently solved a
large number of multi-objective optimization problems [45,46]. With
the emergence of multi-objective optimization problems in the IoT,
this meta-heuristic algorithm has been widely used and its high perfor-
mance in solving myriad problems has been proved [47–49]. NSGA II
algorithm was presented by [50]. In each generation of the algorithm,
the child population (𝑃𝑡 ) is developed by the parent population (𝑄𝑡 )

75
M. Abbasi, E. Mohammadi-Pasand and M.R. Khosravi Computer Communications 169 (2021) 71–80

4.2. Non-dominated sorting Table 2

Simulation parameters setting.

When a new population which is the result of the combination Parameter Quantity

of parent and child populations is created, non-dominated sorting is 𝐖 The fog radius (10–90) m
𝐥𝐦𝐢𝐧 2 MB
utilized for dividing solutions into various categories (lines 12 and 30
𝐥𝐦𝐚𝐱 8 MB
of Algorithm 1) [50]. 𝐏𝐦𝐚𝐱 1 W
nt 1000 cycle/bit
4.3. Crowding distance 𝐧𝐢 200–2000 cycle/bit
fi 1–15 GHz (cycle/s)
𝐟𝐭 2 GHz (cycle/s)
For estimating the density of answers, crowding distance is em- 𝜽𝐢 1–10 × 10−10 J/cycle
ployed (lines 13 and 31 of Algorithm 1). The following steps are 𝜽𝐭 5 × 10−10 J/cycle
followed to calculate the crowding distance for each non-dominated pi 0.1–1
solution [50], as follows. 𝑬 𝑔𝑟𝑒𝑒𝑛 10 W
N 10
(1) The value of 𝑙 is equal to the number of non-dominated solution.
(2) For each fitness function 𝑚, 𝑓𝑚 collection is incrementally sorted
and called 𝐼 𝑚 . We consider a fog-enabled IoT network with N fog nodes which are
(3) For each fitness function, the distance between the beginning randomly distributed in the fog cluster. The terminal nodes located
and end-points is equal to 𝑑𝐼𝑚 = ∞, and the crowding distance in this cluster transmit the packet to the selected fog node through
𝑗
for the midpoints is calculated by Eq. (11) as the wireless link with a bandwidth of 10 MHz at each fog node, the
( ) ( )
Im Im interference power 𝐼𝑖 and the noise power spectral density 𝑁0 are
j+1 j−1
fm − fm −43 dBm and −173 dBm/Hz, respectively. The path loss factor 𝛾𝑖
dm
I = (11) ( )
j fmmax − fmmin (in dB) is obtained through −(38.46 + 20 log10 𝐷𝑖 ), where 𝐷𝑖 (in m)
is the distance between the terminal node and fog node 𝑖. Besides,
4.4. Selection operator a shadowing factor of −5 dB is adopted for each fog node. Among
the N FNs in this cluster, one half is passive, and the other half is
In order to generate a child population in each generation, parents active. Below, we shall arrange two different scenarios considering the
must be chosen from the current population. The NSGA II algorithm allocation of workloads to the terminal node and accessible fog nodes
uses a specific mechanism called binary tournament selection for choos- with two different values of transmission power. In the first scenario,
ing a parent. First, two random answers are selected from the popula- the terminal node transmits the packets to the central controller with
tion. For selecting a parent, one of the answers should be better than its maximum power, and in the other one, the terminal node transmits
the others. Nevertheless, if both solutions have the same non-dominated
the packets to the central controller with half of its capacity. In the
level, they are compared in terms of crowding distance and, under those
latter case, the total energy consumption will increase, but there is no
circumstances, the solution with a greater crowding distance will be
change in the energy consumption of fog devices. In these scenarios, the
selected (line 29 of Algorithm 1) [50].
change of overall power consumption over 𝜃i is examined. The power
The computational complexity of the proposed NSGA II algorithm
consumption per CPU cycle is plotted, where ni = 800 cycle/bit and
for the current two-objective problem of workload allocation is
( ) fi = 5 GHz. The average service delay and energy consumption in these
𝑂 2𝐺𝑁 2 , Where G and N denote the number of generations and size
scenarios are analyzed.
of the population, respectively [48,50–52]. To decrease the computa-
tion time of the algorithm, especially in real-time implementation of the
proposed method, the size of the population should be held as small as 5.1. Workload allocation for maximum transmission power
possible. This key criterion is carefully preserved in the implementation
of the algorithm. In this scenario, the terminal node transmits the packets to the
central controller with its maximum power. Fig. 2 shows two different
4.5. Crossover operator modes of workload distribution. In the first mode, workload distribu-
tion is implemented using a genetic algorithm without green energy
In order to use the crossover operator, two parents are chosen by and based on the settings presented in [10]. In the second mode,
the selection operator. In this study, a single-point crossover operator the workload is distributed among fog devices using green energy
[ ]
is utilized (lines 16–21 of Algorithm 1) [50]. with the values 0 ∶ 𝐸𝑔𝑟𝑒𝑒𝑛 [10]. In Fig. 2, the vertical axis shows the
overall delay for 2, 4, 6, and 8 MB workloads, and the horizontal axis
4.6. Mutation operator represents the values of 𝜃𝑖 in the fog device 𝑖. As it can be seen in the
plot, when the workload transmitted to the terminal node is 2 MB, using
To utilize the mutation operator, a solution is selected by using green energy for 𝜃𝑖 values less than 0.6 does not reduce the overall
Binary Tournament Selection. One of the genes is altered according to delay. Therefore, the overall delay will remain at 3.5 s. The reason
a random number (line 23 of Algorithm1) [50]. why the delay is not reduced is the processing of all requests in fog
devices. For 𝜃𝑖 values greater than 0.6, however, more requests are sent
5. Implementation and evaluation to fog devices and the delay is reduced by 0.9 and 0.79 s for 𝜃𝑖 values
of 0.8 and 1, respectively. When the received workload reaches 4 MB
The proposed algorithm was implemented on a system with a 5-core for 𝜃𝑖 values less than 0.6, no reduction is observed in the delay. Once
3.5 GHz CPU and 8 GB RAM. In the following, the results of implement- again, as all the requests are processed in fog devices, the overall delay
ing the proposed algorithm in MATLAB R2013a are evaluated. Table 2 remains at 7 s. But with an increase in 𝜃𝑖 and use of green energy, the
shows the setting used for the simulation of scenarios. delay is reduced by more than 1 s. As the input workload is increased
In this implementation, we vary the workload 𝑙 allocated to the to 6 MB, the delay is reduced by 0.74 and 1.33 s for 𝜃𝑖 values of 0.8
terminal node from 𝑙𝑚𝑖𝑛 to 𝑙𝑚𝑎𝑥 MBs to see how it affects the energy con- and 1, respectively. In this case, more requests are sent to fog devices.
sumption 𝐸𝑖 and delay 𝐷𝑡𝑜𝑡𝑎𝑙 . Similar to the implementation, requests When the transmitted workload reaches 8 MB, the delay is reduced by
are received by the terminal node and can be processed by 1.02 s if 𝜃𝑖 is increased by 1.

76
M. Abbasi, E. Mohammadi-Pasand and M.R. Khosravi Computer Communications 169 (2021) 71–80

Fig. 2. Overall delay with the maximum transmission of the terminal node for 2 MB, 4 MB, 6 MB, and 8 MB (GE: with Green Energy, NGE: without Green Energy).

Fig. 3. Overall energy consumption of fog devices with the maximum transmission of the terminal node for 2 Mbyte, 4 Mbyte.

Fig. 3 represents the power consumption of fog devices for workload 5.2. Workload allocation for half-capacity transmission power
sizes of 2 and 4 MB. The vertical axis shows the power consumption of
fog devices and the horizontal axis represents the variations in 𝜃𝑖 for In this scenario, the terminal node transmits the packets to the
the fog device 𝑖. The power consumption of fog devices is examined in central controller with half of its capacity. The energy consumed by the
two modes. The first mode is based on the FEMTO method as proposed terminal node to transmit packets to the central controller is more than
in [10], and the second mode is based on obtaining optimum power when it sends them with full power. In Fig. 5, the vertical axis shows the
consumption in fog devices using genetic algorithm and green energy. overall delay, and the horizontal axis represents the variations in 𝜃𝑖 for
As shown in Fig. 3, for 2 and 4 MB of workload, the power consumption the fog device 𝑖. As in Fig. 2, the overall delay for 2, 4, 6, and 8 MB of
of fog devices using green energy indicates a significant difference as workload is examined in two modes. When the workload transmitted to
long as 𝜃𝑖 is 0.6. If 𝜃𝑖 exceeds 0.6, the difference begins to decrease. the terminal node is 2 MB, using green energy for 𝜃𝑖 values less than 0.6
The reason behind this difference is that, for 𝜃𝑖 values greater than will not reduce the delay, and the overall delay will remain at 4.03 s.
0.6, more workload is transmitted to fog devices, resulting in increased
The reason why the delay is not reduced is the processing of all requests
consumption of green energy for processing more packets. For 𝜃𝑖 values
in fog devices. For 𝜃𝑖 values greater than 0.6; however, more requests
less than 0.6, however, the number of requests sent to fog devices does
are sent to fog devices, and the delay is reduced by 0.83 and 0.69 s
not change. As a result, the green energy applied will not reduce the
for 𝜃𝑖 values of 0.8, and 1, respectively. When the received workload
power consumption of fog devices.
reaches 4 MB for 𝜃𝑖 values less than 0.6, no reduction is observed in
Fig. 4 represents the power consumption of fog devices for 6 and
the delay. Once again, as all the requests are processed in fog devices,
8 MB of workload. The vertical axis shows the power consumption of
fog devices, and the horizontal axis represents the variations in 𝜃𝑖 for the overall delay remains 8.06 s. But with the increase in 𝜃𝑖 , and the
the fog device 𝑖. As in Fig. 3, the power consumption of fog devices is use of green energy, the delay is reduced by more than 0.9 s. As the
examined in two modes. According to the results in Fig. 4, the increased received workload is increased to 6 MB, the delay is reduced by 0.71,
workload of the terminal node and the use of green energy have and 1.29 s for 𝜃𝑖 values of 0.8 and 1, respectively. The reason is that
reduced the power consumption of fog devices. Due to the increase in more requests are sent to fog devices. When the transmitted workload
the number of packets sent to the terminal node, part of the packets reaches 8 MB, the delay is reduced by 0.89 s if 𝜃𝑖 is increased by 1.
is processed locally and the rest are sent to fog devices. The green As a result, the green energy applied will not reduce the power
energy applied to fog devices increases proportionally to the increase in consumption of fog devices. If the terminal node transmits the workload
workload, and to the maximum value 𝐸𝑔𝑟𝑒𝑒𝑛 . Therefore, more reduction to the central controller with half of its capacity, the overall power
can be observed in the power consumption of fog devices with 6 and 8 consumption will decrease. However, the power consumption of fog
MB of workload. devices will not change.

77
M. Abbasi, E. Mohammadi-Pasand and M.R. Khosravi Computer Communications 169 (2021) 71–80

Fig. 4. Overall energy consumption of fog devices with the maximum transmission of the terminal node for 6 Mbyte, 8 Mbyte.

Fig. 5. Overall delay with a half-capacity transmission power of the terminal node for 2 MB, 4 MB, 6 MB, and 8 MB (GE: with Green Energy, NGE: without Green Energy).

Fig. 6. Overall energy consumption of fog devices with the maximum transmission of the terminal node for 6 Mbyte, 8 Mbyte.

78
M. Abbasi, E. Mohammadi-Pasand and M.R. Khosravi Computer Communications 169 (2021) 71–80

Fig. 7. Overall energy consumption of fog devices with a half capacity transmission power of the terminal node for 6 Mbyte, 8 Mbyte.

Fig. 6 represents the power consumption of fog devices for workload consumption of fog devices but also significantly decreases the delay in
sizes of 2 and 4 MB. The vertical axis shows the power consumption of the processing of packets on the IoT. In comparison with the FEMTO
fog devices and the horizontal axis represents the variations in 𝜃𝑖 for method, the NSGA II algorithm resulted in a more significant reduction
the fog device 𝑖. As in Fig. 3, the power consumption of fog devices in both power consumption and delay for different workloads. In the
is examined in two modes. For 2 and 4 MB of workload, the power two scenarios simulated using a genetic algorithm, the IoT requests
consumption of fog devices using green energy shows a significant were sent to the controller with full and half capacity. In both scenarios,
difference as long as the 𝜃𝑖 value is 0.6, but the difference begins to power consumption, and delay were improved. Therefore, this model
decrease when 𝜃𝑖 exceeds 0.6. This difference is due to the fact that could be used in networks where green energy is not available to the
more workload is sent to fog devices for 𝜃𝑖 values greater than 0.6. processing resources. In these circumstances, solar or wind energy or
As a result, the provided green energy is used for processing a larger any other renewable energies can be utilized as a complementary power
number of packets. For 𝜃𝑖 values less than 0.6; however, the number of supply to reduce the power consumption of fog devices.
requests sent to fog devices does not change. Deep learning methods, including deep neural networks, have proven
Fig. 7 represents the power consumption of fog devices for packets to be powerful methods in many complicated multi-objective optimiza-
with sizes of 6 and 8 MB. The vertical axis shows the power consump- tion problems [53,54]. Adapting these methods to the multi-objective
tion of the fog device 𝑖, and the horizontal axis represents the variations workload allocation in secure MEC will be our future study.
in 𝜃𝑖 for the fog device 𝑖. As in Fig. 3, the power consumption of fog
devices is examined in two modes. According to the results in Fig. 7, the CRediT authorship contribution statement
increased workload of the terminal node, and the use of green energy
has reduced the power consumption of fog devices. Due to the increase M. Abbasi: Supervision, Conceptualization, Methodology, Formal
in the number of packets sent to the terminal node, part of the packets is analysis, Visualization, Writing - review & editing. E. Mohammadi-
processed locally, and the rest are sent to fog devices. The green energy Pasand: Data curation, Software, Investigation, Writing - original draft.
applied to fog devices increases proportionally to the increase in the M.R. Khosravi: Conceptualization, Validation, Writing - review & edit-
number of packets and to the maximum value 𝐸𝑔𝑟𝑒𝑒𝑛 . Therefore, more ing.
reduction can be observed in the power consumption of fog devices
with 6 and 8 MB packets. Declaration of competing interest

6. Conclusions The authors declare that they have no known competing finan-
cial interests or personal relationships that could have appeared to
The transmission of large volumes of data from sensors to the cloud influence the work reported in this paper.
imposes long delays on the workloads in the 5G empowered IoT. In
the meanwhile, processing these workloads at the mobile network edge References
will significantly increase power consumption. Therefore, workload
allocation with an emphasis on delay to provide QoS as well as on [1] X. Xu, Q. Huang, X. Yin, et al., Intelligent offloading for collaborative smart city
energy consumption has become an important issue. Fog devices have services in edge computing, IEEE Internet Things J. (2020).
[2] L.U. Khan, I. Yaqoob, N.H. Tran, et al., Edge computing enabled smart cities: A
relatively small delays, but their high power-consumption at the net-
comprehensive survey, IEEE Internet Things J. (2020).
work edge leads to less workload being allocated to them. In previous [3] Z. Ullah, F. Al-Turjman, L. Mostarda, et al., Applications of artificial intelligence
studies, workload allocation has been addressed without modifying the and machine learning in smart cities, Comput. Commun. 154 (2020) 313–323.
energy model of fog devices, which led to an increased delay for the [4] M. Abbasi, M. Yaghoobikia, M. Rafiee, et al., Energy-efficient workload allocation
end-user. In this study, we attempted to compromise between power in fog-cloud based services of intelligent transportation systems using a learning
classifier system, IET Intell. Transp. Syst. (2020).
consumption and delay in fog devices through the modification of the [5] M. Abbasi, M. Yaghoobikia, M. Rafiee, et al., Efficient resource management
energy model of these devices using the NSGA II algorithm. The simu- and workload allocation in fog–cloud computing paradigm in IoT using learning
lation results suggest that this modification not only reduces the power classifier systems, Comput. Commun. 153 (2020) 217–228.

79
M. Abbasi, E. Mohammadi-Pasand and M.R. Khosravi Computer Communications 169 (2021) 71–80

[6] M. Al-Zinati, T. Almasri, M. Alsmirat, et al., Enabling multiple health security [32] M. Chiang, S. Ha, I. Chih-Lin, et al., Clarifying fog computing and networking:
threats detection using mobile edge computing, Simul. Model. Pract. Theory 101 10 questions and answers, IEEE Commun. Mag. 55 (2017) 18–20.
(2020) 101957. [33] F.S. Abkenar, A. Jamalipour, EBA: Energy balancing algorithm for fog-IoT
[7] L. Atzori, A. Iera, G. Morabito, The internet of things: A survey, Comput. Netw. networks, IEEE Internet Things J. 6 (2019) 6843–6849.
54 (2010) 2787–2805. [34] S. Yi, C. Li, Q. Li, A survey of fog computing: concepts, applications and issues,
[8] A.A. Abdellatif, A. Mohamed, C.F. Chiasserini, et al., Edge computing for in: Proceedings of the 2015 Workshop on Mobile Big Data, 2015, pp. 37–42.
energy-efficient smart health systems: Data and application-specific approaches, [35] F. Jalali, K. Hinton, R. Ayre, et al., Fog computing may help to save energy in
in: A. Mohamed (Ed.), Energy Efficiency of Medical Devices and Healthcare cloud computing, IEEE J. Sel. Areas Commun. 34 (2016) 1728–1739.
Applications, Academic Press, 2020, pp. 53–67, (Chapter 3). [36] R. Deng, R. Lu, C. Lai, et al., Optimal workload allocation in fog-cloud computing
[9] J. Kwak, Y. Kim, J. Lee, et al., DREAM: Dynamic resource and task allocation toward balanced delay and power consumption, IEEE Internet Things J. 3 (2016)
for energy minimization in mobile cloud systems, IEEE J. Sel. Areas Commun. 1171–1181.
33 (2015) 2510–2523. [37] A. Mebrek, L. Merghem-Boulahia, M. Esseghir, Efficient green solution for a bal-
[10] G. Zhang, F. Shen, Z. Liu, et al., FEMTO: Fair and energy-minimized task anced energy consumption and delay in the IoT-Fog-Cloud computing, in: 2017
offloading for fog-enabled IoT networks, IEEE Internet Things J. 6 (2018) IEEE 16th International Symposium on Network Computing and Applications
4388–4400. (NCA), 2017, pp. 1–4.
[11] X. Xu, X. Zhang, X. Liu, et al., Adaptive computation offloading with edge for [38] D. Kanapram, G. Lamanna, M. Repetto, Exploring the trade-off between per-
5G-envisioned internet of connected vehicles, IEEE Trans. Intell. Transp. Syst. formance and energy consumption in cloud infrastructures, in: 2017 Second
(2020). International Conference on Fog and Mobile Edge Computing (FMEC), 2017,
[12] A. Botta, W. De Donato, V. Persico, et al., Integration of cloud computing and pp. 121–126.
internet of things: a survey, Future Gener. Comput. Syst. 56 (2016) 684–700. [39] A. Yousefpour, G. Ishigaki, J.P. Jue, Fog computing: Towards minimizing delay in
[13] M. Chiang, T. Zhang, Fog and IoT: An overview of research opportunities, IEEE the internet of things, in: 2017 IEEE International Conference on Edge Computing
Internet Things J. 3 (2016) 854–864. (EDGE), 2017, pp. 17–24.
[14] M. Abbasi, E. Mohammadi Pasand, M.R. Khosravi, Workload allocation in IoT- [40] H. Wu, L. Chen, C. Shen, et al., Online geographical load balancing for energy-
fog-cloud architecture using a multi-objective genetic algorithm, J. Grid Comput. harvesting mobile edge computing, in: 2018 IEEE International Conference on
18 (2020) 43–56. Communications (ICC), 2018, pp. 1–6.
[15] Z. Zhou, S. Yu, W. Chen, et al., CE-IoT: Cost-effective cloud-edge resource [41] X. Chen, L. Jiao, W. Li, et al., Efficient multi-user computation offloading for
provisioning for heterogeneous IoT applications, IEEE Internet Things J. (2020). mobile-edge cloud computing, IEEE/ACM Trans. Netw. 24 (2015) 2795–2808.
[16] M. Lesk, How Much Information is There in the World?, 1997. [42] D. Huang, P. Wang, D. Niyato, A dynamic offloading algorithm for mobile
[17] M. Abbasi, A. Shokrollahi, M.R. Khosravi, et al., High-performance flow classifi- computing, IEEE Trans. Wireless Commun. 11 (2012) 1991–1995.
cation using hybrid clusters in software defined mobile edge computing, Comput. [43] C. Xian, Y.-H. Lu, Z. Li, Adaptive computation offloading for energy conservation
Commun. 160 (2020) 643–660. on battery-powered systems, in: 2007 International Conference on Parallel and
[18] M. Abbasi, A. Shokrollahi, Enhancing the performance of decision tree-based Distributed Systems, 2007, pp. 1–8.
packet classification algorithms using CPU cluster, Cluster Comput. (2020). [44] S.-A.N. Alexandropoulos, C.K. Aridas, S.B. Kotsiantis, et al., Multi-objective
[19] L.A. Tawalbeh, F. Muheidat, M. Tawalbeh, et al., IoT privacy and security: evolutionary optimization algorithms for machine learning: A recent survey, in:
Challenges and solutions, Appl. Sci. 10 (2020) 4102. Approximation and Optimization, Springer, 2019, pp. 35–55.
[20] J. Mineraud, O. Mazhelis, X. Su, et al., A gap analysis of Internet-of-Things [45] S. Wang, D. Zhao, J. Yuan, et al., Application of NSGA-II algorithm for fault
platforms, Comput. Commun. 89 (2016) 5–16. diagnosis in power system, Electr. Power Syst. Res. 175 (2019) 105893.
[21] E. Grande, M. Beltrán, Edge-centric delegation of authorization for constrained [46] A. Ranjbar, S. Talatahari, F. Hakimpour, The application of multi-objective
devices in the Internet of Things, Comput. Commun. (2020). charged system search algorithm for optimization problems, Sci. Iran. 26 (2019)
[22] X. Xu, X. Liu, X. Yin, et al., Privacy-aware offloading for training tasks of 1249–1265.
generative adversarial network in edge computing, Inform. Sci. (2020). [47] P. Wang, F. Xue, H. Li, et al., A multi-objective DV-Hop localization algorithm
[23] J. Ni, K. Zhang, X. Lin, et al., Securing fog computing for internet of things based on NSGA-II in internet of things, Mathematics 7 (2019) 184.
applications: Challenges and solutions, IEEE Commun. Surv. Tutor. 20 (2017) [48] H. Ma, A.S. da Silva, W. Kuang, NSGA-II with local search for multi-objective
601–628. application deployment in multi-cloud, in: 2019 IEEE Congress on Evolutionary
[24] K. Yang, K. Zhang, J. Ren, et al., Security and privacy in mobile crowdsourcing Computation (CEC), 2019, pp. 2800–2807.
networks: challenges and opportunities, IEEE Commun. Mag. 53 (2015) 75–81. [49] S. Tavakoli-Someh, M.H. Rezvani, Multi-objective virtual network function
[25] J. Xu, L. Chen, S. Ren, Online learning for offloading and autoscaling in energy placement using NSGA-II meta-heuristic approach, J. Supercomput. 75 (2019)
harvesting mobile edge computing, IEEE Trans. Cogn. Commun. Netw. 3 (2017) 6451–6487.
361–373. [50] K. Deb, A. Pratap, S. Agarwal, et al., A fast and elitist multiobjective genetic
[26] M. Abbasi, M. Rafiee, A calibrated asymptotic framework for analyzing packet algorithm: NSGA-II, IEEE Trans. Evol. Comput. 6 (2002) 182–197.
classification algorithms on GPUs, J. Supercomput. 75 (2019) 6574–6611. [51] K. Deb, S. Agrawal, A. Pratap, et al., A fast elitist non-dominated sorting genetic
[27] H. Tahaei, F. Afifi, A. Asemi, et al., The rise of traffic classification in IoT algorithm for multi-objective optimization: NSGA-II, in: Parallel Problem Solving
networks: A survey, J. Netw. Comput. Appl. 154 (2020) 102538. from Nature PPSN VI, Berlin, Heidelberg, 2000, pp. 849–858.
[28] S.N. Matheu, A. Robles Enciso, A. Molina Zarca, et al., Security architecture for [52] M.T. Jensen, Reducing the run-time complexity of multiobjective EAs: The
defining and enforcing security profiles in dlt/sdn-based iot systems, Sensors 20 NSGA-II and other algorithms, IEEE Trans. Evol. Comput. 7 (2003) 503–515.
(2020) 1882. [53] M. Loni, S. Sinaei, A. Zoljodi, et al., DeepMaker: A multi-objective optimiza-
[29] S. Khodaparas, A. Benslimane, S. Yousefi, A software-defined caching scheme for tion framework for deep neural networks in embedded systems, Microprocess.
the Internet of Things, Comput. Commun. (2020). Microsyst. 73 (2020) 102989.
[30] X. Peng, J. Ren, L. She, et al., BOAT: A block-streaming app execution scheme [54] M. Caramia, P. Dell’Olmo, Multi-objective optimization, in: M. Caramia, P.
for lightweight IoT devices, IEEE Internet Things J. 5 (2018) 1816–1829. Dell’Olmo (Eds.), Multi-Objective Management in Freight Logistics: Increasing
[31] J. Ren, H. Guo, C. Xu, et al., Serving at the edge: A scalable IoT architecture Capacity, Service Level, Sustainability, and Safety with Optimization Algorithms,
based on transparent computing, IEEE Netw. 31 (2017) 96–105. Springer International Publishing, Cham, 2020, pp. 21–51.

80