This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2874484, IEEE
Transactions on Cloud Computing
1

On the Planning and Design Problem of

Fog Computing Networks
Faisal Haider, Decheng Zhang, Marc St-Hilaire and Christian Makaya

Abstract—This paper proposes an exact model for the planning and design problem of fog networks. More precisely, a mathematical
model is proposed to simultaneously determine the optimal location, the capacity and the number of fog node(s) as well as the
interconnection between the installed fog nodes and the cloud. The goal of the model is to minimize the delay in the network and the
amount of traffic sent to the cloud data center. To address this multi-objective optimization problem, three optimization techniques are
used: the weighted sum, the hierarchical and the trade-off methods. The weighted sum method aggregates all the lone objective
functions into a single objective by applying a weighted vector. The hierarchical method takes a sequential approach by tightly
constraining the more important objective function. The trade-off method solves a single objective function and translates all other
objective functions into constraints. These methods are then compared in terms of average delay, amount of traffic sent to the cloud
and amount of CPU time required to find optimal solution(s). Since we are dealing with a multi-objective optimization problem and that
multiple optimal solutions can be found, the fuzzy-based mechanism and the hypervolume indicator have been used. Computational
results show that as the problem size increases, the delay and the traffic also increase in a linear form; whereas, the solution time
increases in non-polynomial time. The weighted sum method was able to achieve the best trade-off results for the delay and the traffic,
whereas the hierarchical method was able to return minimum delay but with worse traffic going to the cloud. As the model considers
realistic edge device traffic parameters, constraints, and various topology aspects, it can be helpful for the planning and deployment of
fog networks and how they operate within a cloud infrastructure.

Index Terms—Fog computing, edge computing, cloud computing, multi-objective optimization, mathematical model, network planning

1 I NTRODUCTION

T HE Internet of Things (IoT) has become a well known

term today. The objective of IoT is to connect billions
of smart devices to the Internet, generating a huge volume
an essential role. Hence, with the shift of handling traffic
from the cloud to the fog and improving Quality-of-Service
(QoS), fog computing will be an added paradigm in the
of data. Cloud computing, with its extensive computation cloud architecture.
capacity and storage, has been a propulsive force for various Fog networks will be a composition of distributed enti-
applications such as in IoT context. With the integration of ties called fog nodes which will enable the deployment of
big data, machine learning applications and an upsurge of fog services. For a typical fog network, there will be a lot of
connected devices in the IoT, the cloud infrastructure alone fog nodes to be installed in order to serve the large number
is not suited for latency-sensitive applications. Issues of high of edge devices over the entirety of the geographical region
bandwidth utilization, reliability, performance, security are of interest. Consequently, there are many factors such as the
just a few reasons to promote the notion of fog and edge number and capacity of nodes, locations, traffic distribution
computing [1], [2], [3], [4]. and so on that need to be considered. With the complexity of
Fog computing can be defined as a non-trivial extension fog networks and the huge number of fog nodes and edge
of the cloud located near the edge devices. The main idea devices, service providers need efficient planning tools to
of fog computing is front-end cloud services, i.e., provid- design optimal fog networks.
ing the same services as the cloud in a limited or full The purpose of the fog network is to improve the end
service from the edge of the network. Fog computing is user experience by reducing parameters like delay in the
an immensely virtualized architecture focused on broadly network, traffic going to the cloud and so on. Hence, it
distributed services and applications. It is a Micro Data is fundamental for the service providers to consider delay
Center (MDC), enabling services like storage, computation and traffic parameters as part of the planning problem.
and networking between the edge devices in the IoT. The Optimizing the two parameters convert the problem into a
distributed directory system in fog application enables in multi-objective optimization problem resulting in a number
communicating with mobile devices to acknowledge differ- of non-dominated solutions. The multi-objective optimiza-
ent mobility techniques. Low latency required applications tion problem can be studied from different viewpoints with
such as gaming, video streaming, augmented reality, etc. different solution goals while solving it. Thus, it is very hard
is one of the many use cases of fog computing. For a to find a best compromised solution for all the objective
sustainable IoT infrastructure, the cloud will always have functions even with a comprehensive knowledge of the
problem.
• F. Haider, D. Zhang and M. St-Hilaire are with the Department of Systems There are different methods to solve multi-objective op-
and Computer Engineering, Carleton University, Ottawa, Canada timization problems. However, it is a very complex task to
E-mail: marc st [email protected] choose the best method as different methods have various
• C. Makaya is with IBM Research, Yorktown Heights, NY, USA
requirements and can perform differently with different

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2874484, IEEE
Transactions on Cloud Computing
2

The second layer is where the fog nodes reside along with
the core network. The last layer consists of cloud compo-
nents. Below is a description of the three layers.

Layer 1
The first layer is composed of smart objects (e.g., IoT de-
vices) that are collecting and generating data. The devices
can be mobile phones, wireless sensors, vehicles, drones,
etc. with various applications installed transmitting and
receiving data and information to the immediate upper
layer and/or the cloud layer.
Fig. 1: Reference architecture for Fog
Layer 2
problems. Ideally, for the best convergence towards the The fog computing layer is made up of fog nodes and smart
Pareto set, the optimization method that handles the min- devices such as intelligent gateways, routers and Access
imum amount of complexity should be used. But defining Point Name (APN). In this layer, the fog nodes and smart
the amount of complexity is not straightforward. To mit- devices perform tasks like routing, data storing and prepro-
igate that, a number of methods can be used simultane- cessing, caching and packet forwarding towards the cloud.
ously to understand and solve the problem. To address the Additionally, this layer incorporates resource management
multi-objective optimization problem, three different multi- software while maintaining the entire infrastructure. This
objective optimization methods are used to solve the fog layer ensures the QoS to the different fog technology appli-
planning problem. cations as well as the connectivity between the fog nodes
This paper proposes a mathematical model that ad- and the cloud.
dresses those issues for the planning and design of fog
Layer 3
networks. The model simultaneously determines the op-
timal location, the number, and the capacity of the fog The cloud computing layer consists of servers and data
node(s) as well as the interconnection with the cloud while centers with extensive storage and computing capability.
considering the edge device requests. The proposed op- Applications and data which require permanent storage
timization technique uses the fuzzy-based mechanism to and powerful processing are redirected to the cloud. Unlike
find the best compromised result from the non-dominated traditional cloud architecture, the cloud does not serve
solution set. Although network planning has been actively every individual query. The fog computing layer decides the
investigated for IP networks, optical networks, wireless and accessibility and usage of the cloud in an efficient manner.
cellular networks [5], and cloud computing, to the best of
our knowledge, no work has considered the planning and 2.2 Applications and Current Research
design of fog networks. The concept of fog computing has many different important
The rest of the paper is organized as follows. In Section 2, applications. Below we review some of these applications
we review the background with an updated view of the and cite relevant work where fog is playing an important
state-of-the-art in fog computing. In Section 3, we introduce role.
the fog planning problem followed by the mathematical
model. Then, the mathematical model is slightly modified to Vehicular Networks and Smart Grid
fit the three multi-objective optimization methods. Results The authors in [7] propose mobile fog, which is geo-spatially
are shown and analyzed in Section 4 and finally, the paper distributed, large-scale and latency sensitive for future In-
is concluded in Section 5. ternet applications. Two large scale vehicle-to-vehicle (V2V)
application scenarios are simulated. The simulation results
2 BACKGROUND AND R ELATED W ORK show that compare to cloud computing, the fog computing
approach reduces network traffic and latency. A literature
Different research activities from academia and industry
review by the author in [11] discusses the application of the
have explored the concept of fog computing under different
fog computing framework to implement Software-Defined
themes. Towards that end, the OpenFog consortium [6]
Network (SDN) for vehicular networks. Furthermore, the
is working on influencing standards development through
author proposes a multi-layered demand response man-
liaisons with other organizations as well as promoting inno-
agement system connected to fog devices in a smart grid
vation and industry interest in fog computing.
which can minimize the overall power loss as well as reduce
communication costs within a smart grid.
2.1 Architecture
The authors in [7], [8], [9], [10] discuss the programmibility Healthcare
and architecture of fog computing focusing on different The authors in [12] propose a three layer architecture in-
heterogeneous networks. Fig. 1 shows a reference fog archi- tegrating a role model, a layered cloud computing archi-
tecture consisting of different layers. The bottom-most layer tecture, as well as a fog-computing-informed paradigm in
comprises of end devices, gateways and sensors. Additional order to provide a feasible architecture for healthcare and
applications are installed for efficient use of the technology. elderly-care applications. The fog computing layer improves

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2874484, IEEE
Transactions on Cloud Computing
3

the performance of the whole system by providing com- sensitivity. The authors in [20] intent to reduce the service
puting and storage, mobility support, security measures, delay for IoT applications by proposing a delay-minimizing
location awareness and low latency. In [13], FAST, a fog policy. Unlike the model we proposed, the policy considers
computing assisted distributed analytic system to monitor a fixed transmission speed (as the link speed is constant)
fall for stroke mitigation is proposed. Fog computing was and the budget for deployment is not considered. Moreover,
used to design and employ a real-time fall detection sys- the distances between fog nodes and IoT nodes are derived
tem distributing the analytic throughout the network by from the uniformly distributed propagation speed. In our
dividing the detection between the edge devices and servers model, the propagation speed is calculated from the gener-
in cloud. The proposed system attains the high specificity ated distance in conventional way which is more practical.
and sensitivity when investigated with real-world data. In
[14], the authors conduct a case study on Electrocardiogram Concluding Remark
(ECG) feature extraction using fog computing at smart gate- Although we can find several papers on the concept of fog
ways. The experimental results show that more than 90% computing and its planning, the model we put forward in
bandwidth efficiency and low-latency real time response this paper is different from what has been proposed so far.
could be achieved using fog computing. The authors in In fact, most papers dealing with fog computing assume
[15] integrates the “edge device-fog-cloud” architecture to that fog nodes are already deployed and available to use.
propose a system to control and restrain the spread of zika As the deployment of fog computing will play an important
virus. The fog computing is used to reduce the delay and role in many time-sensitive applications such as augmented
communication cost which is normally high for a cloud only reality, tactile Internet, online gaming, etc., proper planning
architecture. and design of fog computing environments is crucial so
that providers can minimize their costs and users enjoy a
Preprocessing and Content Delivery good quality of experience. To that end, the next section
The integration of IoT with cloud computing can be one introduces a new mathematical model for the planning and
of the primary use of fog computing. However, there are design of fog networks.
many challenges in this integration. The authors in [16]
discuss one of the major challenges which is data trimming
or preprocessing. There will be congestion in data center 3 F OG P LANNING M ODEL
and core network when sending huge amount of raw data In this section, a mathematical model is formulated for
generated by IoT environments. A smart gateway-based the fog planning problem. It is important to note that, the
communication is proposed for trimming or preprocessing mathematical model has been derived from the proposed
data before sending it to the cloud. The smart gateway with hierarchical architecture from the OpenFog Consortium [6]
the help of fog computing can provide a better utilization which is a collaborative work between industry, academia
of network and cloud resources. In [17], the authors discuss and end users.
the use of edge servers in fog computing architecture for Given a set of edge user requests, the model simultane-
web optimization. Users will be connected to the Internet ously determines the optimal location, the number, and the
through fog nodes. Each HTTP request made by the user capacity of the fog node(s) as well as the connection of the
will go through the fog nodes where the fog nodes will fog nodes and the cloud. Since the model has two different
reduce the time for web page to load by performing different objectives, the Pareto optimality needs to be established. To
optimization. fulfill the requirement of Pareto optimality, the solution out-
put should guarantee that no better delay can be achieved
Planning and Optmization in Fog Computing without worsening the traffic burden to the cloud.
A recent literature review focusing on state-of-the-art fog
research has noted that the planning and design of fog net- 3.1 Exact Model
works has received very little attention so far [?]. Neverthe-
To formulate the mathematical model, we assume the fol-
less, the authors in [18] introduce an edge capacity planning
lowing information is known:
solution with respect to QoS while minimizing acquisition
costs for the edge clouds. The proposed approach turnouts • There is no existing fog infrastructure meaning that
a satisfied capacity plan for the required QoS while mini- we do the planning from a green field scenario.
mizing the initial deployment cost. However, only the QoS • The location of all edge devices and the possible
takes into account the response delay and different demands locations of each fog node (i.e., x and y coordinates).
for network and system resources. On the other hand, our For each edge device, the generated traffic and the
model takes into account the transmission, propagation and link speed is also known. It is important to note that
processing delay as well as the network traffic while keeping the planning is made from traffic forecast that could
the total deployment cost under the desired budget. In [19], represent the maximum amount of expected traffic
the authors propose a mixed integer linear programming (plus an extra margin for safety).
formulation to minimize the deployment cost for vehicular • The bandwidth availability and the cost of each link
applications in an urban scenario. The model allows only a type for the connection between the fog nodes and
single use case with a strict single-hop communication. In the cloud.
contrast, our fog planning tool is able to consider diverse • The characteristics (i.e., memory, vCPU, cost) of dif-
user requests or use cases with the capability of both single ferent types of fog nodes that may be installed in a
and multi-hop communications depending on the latency network.

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2874484, IEEE
Transactions on Cloud Computing
4

• The location of the cloud. We assume that the cloud • N, set of fog types that can be installed, N =
is located in a remote location and has unlimited {n1 , n2 , ...}
memory and vCPU. Therefore, if an edge device – λn , the total amount of memory available for
cannot be served by the fog, it will be served by the a fog of type n ∈ N
cloud. – αn the number of vCPU available for a fog of
• Depending on the application, the fog nodes send type n ∈ N
a percentage of their total traffic to the cloud. This – θn , the network interface capacity of a fog of
assumption is based on the fact that if an edge device type n ∈ N
is served by a fog node, most of the work will be – β n , the number of nodes available for a fog of
done at the fog node level. However, some data may type n ∈ N
still be sent from the fog to the cloud for various – φn , the cost of a fog of type n ∈ N
reasons such as backup, enhanced services, etc.
• L, set of possible link types that can be used to
The output of the planning problem focuses on the interconnect fog nodes and cloud, L = {l1 , l2 , ...}
following:
– ω l , the bandwidth (in Gbps) of a link of type
• Selecting the optimal geographic location to deploy l∈L
fog nodes in the network (i.e., where to deploy the – ξ l , the price (in $/meter) of a link of type l ∈ L
fog nodes).
• P, set of possible locations to install the fog nodes,
• Defining the optimal number of fog nodes that needs
P = {p1 , p2 , ...}
to be installed in the network.
• Specifying the type (i.e., capacity) of fog nodes to be
Constants
installed in the network.
• Specifying the types of links used to interconnect the • σ , average packet size (in bytes) sent by the edge
various network elements. devices.
• τ , maximum budget (in dollars) allowed for the
The primary goal of the Fog Planning Problem, denoted FPP, deployment of the fog.
is to find the best locations for installing fog nodes so that • t, speed of light (in meters per second).
the network delay experienced by the users and the amount • h, average number of hops packets take from edge
of traffic going towards the cloud are minimized. devices to fog nodes.
• k, average processing delay (in seconds) per hop.
Model Formulation • r, percentage of traffic going to the cloud from fog
nodes.
To model users, we use the notion of clusters where each
cluster represents an agglomeration of user requests [21]. We Functions
are using this approach to reduce complexity as typically,
several users (hundreds or even thousands) are using the • Distance (a, b) is a function that calculates the dis-
cloud at the same time in a given area. tance between points a and b. The value of points a
Before defining the notation, we first illustrate a refer- and b are the x, y coordinates.
ence use case based on the concept of smart homes where • ψ , function that calculates the transmission delay.
each smart home (representing an edge device) is equipped • µ, function that calculates the propagation delay.
with hundreds of different sensors. These sensors measure • γ , function that calculates the processing delay.
different real-time parameters such as energy usage, secu-
rity, alarms and video servers with real-time processing. As Decision Variables
the size and the number of smart homes can vary over a • xnp , a 0-1 variable such that xnp = 1 if and only if a
geographical area, the requirements from the fog nodes will fog node of type n ∈ N is installed at location p ∈ P ;
also vary. Other use cases are available on the OpenFog • yup , a 0-1 variable such that yup = 1 if and only if the
Consortium [6]. edge device cluster u ∈ U is connected to location
p ∈ P;
• vuc , a 0-1 variable such that vuc = 1 if and only if the
Sets
edge device cluster u ∈ U is connected to the cloud;
• U, set of edge device clusters that are present in the • bpl , a 0-1 variable such that bpl = 1 if and only if the
network, U = {u1 , u2 , ...} fog node installed at location p ∈ P is connected to
– λu , the total amount of memory required by a the cloud with a link of type l ∈ L;
cluster of edge devices of type u ∈ U Fig. 2 is a graphical representation of the network
– αu , the total number of vCPU required by a showing the notations described earlier. Fig. 2 illustrates a
cluster of edge devices of type u ∈ U three tier “edge device-fog-cloud” architecture, with links
– θu , the number of packets requested to the fog between each network elements. It can be seen that, de-
nodes or the cloud by a cluster of edge devices pending upon the load, an edge device cluster can either
of type u ∈ U be connected to a fog node or to the cloud. Moreover, when
– κu , the link speed of a cluster of edge devices a fog node is installed at a possible location, it is connected
of type u ∈ U to the cloud.

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2874484, IEEE
Transactions on Cloud Computing
5

Uniqueness Constraints
The uniqueness constraints enforce that at most one fog
node is installed at a given location. In other words, we
cannot install two or more fog nodes at the same location.
X
xnp ≤ 1 (p ∈ P ) (7)
n∈N

Assignment Constraints
The edge device assignment constraints make sure that each
device (i.e. cluster) is assigned to exactly one location (i.e.,
the fog or the cloud).
X
yup + vuc = 1 (u ∈ U ) (8)
p∈P

The node assignment constraints ensure that each in-

stalled fog node is connected to cloud.
X X
xnp = bpl (p ∈ P ) (9)
n∈N l∈L

Capacity Constraints
Fig. 2: Graphical representation of the notation The node capacity constraints for vCPU at the node level
make sure that the number of vCPU required by the edge
devices does not exceed the capacity of the fog node.
Objective Function X X
As shown in Equation (1), the objective function of the FPP, yup au ≤ xnp an (p ∈ P ) (10)
is to minimize the total delay of the network and the amount u∈U n∈N
of traffic sent to the cloud. The node capacity constraints for memory at the node
level make sure that the total amount of memory required
M inimize(Dtotal , T raf f ic) (1) by the edge devices does not exceed the capacity of the fog
node.
As shown in Equations (2) to (5), the total delay (Dtotal )
is the sum of the transmission delay (Dt ), propagation delay X X
(Dn ) and processing delay (Dp ) in the whole network. yup λu ≤ xnp λn (p ∈ P ) (11)
On the other side, Equation (6) represents the total traffic u∈U n∈N

going to the cloud which is the sum of the traffic from all The Network Interface Capacity (NIC) constraints at the
users connected to the cloud plus a percentage of the traffic node interface level ensure that the total inbound band-
coming from the various fogs. width required by the edge devices does not exceed the fog
node capacity.
Dtotal = Dt + Dn + Dp (2)
X X
yup f (θu , σ) ≤ xnp g(θn ) (p ∈ P ) (12)
XX X
u∈U n∈N
Dt = yup ψ + vuc ψ (3)
u∈U p∈P u∈U The Network Interface Capacity (NIC) constraints at the
XX X cloud interface level ensure that the total inbound band-
Dn = yup µ + vuc µ (4) width from the fog nodes to the cloud does not exceed the
u∈U p∈P u∈U link capacity.
XX X
Dp = yup γ + vuc γ (5) X
yup f (θu , σ)r ≤
X
bpl g(ω l ) (p ∈ P ) (13)
u∈U p∈P u∈U
u∈U l∈L
X XX
T raf f ic = vuc θu + bpl θu r (6)
Inventory Constraints
u∈U p∈P l∈L
The inventory capacity constraints make sure that we are
not using more nodes than what is available in the inven-
Formulation of the Constraints
tory.
The model aims to minimize the network delay as well
X
as the traffic entering the cloud. The set of constraints xnp ≤ β n (n ∈ N ) (14)
explained below restricts the minimization function. p∈P

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2874484, IEEE
Transactions on Cloud Computing
6

Budget Constraint 3.2 Models with Pareto Optimality

The budget constraint makes sure that the total cost of Since we model FPP as a multi-objective optimization prob-
the network (fog nodes and links) does not exceed the lem, we need to simultaneously minimize the two objective
maximum allowed budget. functions expressed in Equations (2) and (6). Sometimes,
XX XX the minimum network delay will increase the traffic in the
xnp φn + bpl dist(p, c)ξ l ≤ τ (15) cloud and vice versa. Therefore, to ensure fairness between
n∈N p∈P p∈P l∈L network delay and amount of traffic going to the cloud,
we need to find Pareto optimal solutions. In this paper,
Integrality Constraints we adopt three different methods to solve the fog planning
The integrality constraints state that the domain of the problem.
following decision variables is a set of binary numbers.
3.2.1 Weighted Sum Optimization
xnp ∈ B (n ∈ N, p ∈ P ) (16) The weighted sum strategy converts the multi-objective
problem of optimizing the main objective function into
yup ∈ B (u ∈ U, p ∈ P ) (17) a scalar problem by building a weighted sum of all the
objectives. Therefore, the proposed objective function can
vuc ∈ B (u ∈ U, l ∈ L) (18)
be rewritten as:
bpl ∈ B (p ∈ P, l ∈ L) (19)
norm
M inimize (w1 Dtotal + w2 T raf f icnorm ) (23)
Delay Computation norm
where w1 and w2 are weighted coefficients and Dtotal and
The latencies for the network are computed using the T raf f icnorm are the normalized objective functions as they
transmission, propagation, and processing delays. It would have different scales. Equations (24) and (25) are used to
also be interesting to consider the queuing delay but as normalize the objective functions.
this keeps changing over time, a static model like the one max
norm Dtotal − Dtotal
presented in this paper does not allow to include it. Dtotal = max − D min
(24)
The transmission delay is the time taken for a process Dtotal total
to send all the bits in a packet into the wire. This depends
on the link speed that is used and the packet size that is T raf f icmax − T raf f ic
T raf f icnorm = (25)
to be sent to the user from the fog or cloud. The formula T raf f icmax − T raf f icmin
to calculate the transmission delay is shown below where σ
The weighted coefficients define the contribution of each
is the packet size (bytes) and wl represents the link speed
objective function in the modified model. Even with a
(bytes/sec).
proper knowledge of the problem, it is not easy to specif-
σ ically select the weight of the coefficients in the modified
T ransmission delay (ψ) = (20)
wl objective function. Therefore, an additional constraint (26) is
The propagation delay is the amount of time taken added which states that the sum of the weighted coefficients
to transmit a signal from the sender to the receiver. This w1 and w2 should be 1.
depends on the medium used. Propagation is calculated as w1 + w2 = 1 (26)
distance divided by the speed at which the signal prop-
agates in the medium. Depending on the medium that is Fig. 3 shows the flowchart for solving the FPP using the
used, the propagation speed varies. The propagation speed weighted sum method. In the first step, we generate the
of wireless communications is the speed of light. For copper input for the planning problem. Initially, the value for w1 is
wires, the speed varies from 0.59t to 0.79t [22]. In the set to 1 and w2 is set to 0. Then, the modified model from
proposed model, we use 0.59t for the speed of copper wire. Section 3.2.1 is solved and the returned objective function
values are stored in the Pareto set. Next, the terminating
Distance(a, b) condition is checked (w1 = 0 and w2 = 1) and if it is not
P ropagation delay (µ) = (21)
0.59 ∗ t satisfied, the model is solved again with updated weighted
Each router or switch in the data path adds a finite coefficient values. This step is repeated until the terminating
amount of delay as the packet is received, processed, and condition is satisfied. Once the terminating condition is
then forwarded. This includes the time taken at each layer satisfied, a Pareto optimal set is obtained from the stored
of the TCP/IP down until the bit level layer. Using features objective function values in the Pareto set. The fuzzy-based
that are supported with hardware assistance can greatly mechanism (described in Section 3.3) is used to extract the
reduce latency. Latency with a hardware-assisted switch best compromised result from the Pareto optimal set, which
will be in the 4-to-20 microseconds range [23]. The most is used as the final result for analysis.
reasonable processing delay that can be expected in practice
should be 25 microseconds per hop, which has been used in 3.2.2 Hierarchical Optimization
our calculation. Hierarchical optimization is generally solved by the sequen-
tial optimization method. Here, the objective functions are
P rocessing delay (γ) = k ∗ h (22) ranked in order of importance. From the literature review,

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2874484, IEEE
Transactions on Cloud Computing
7

Fig. 3: Steps to solve the FPP with the weighted sum method
Fig. 4: Steps to solve the FPP with the hierarchical method
it can be assumed that the fog nodes will be used exten-
sively for latency sensitive edge devices [7], [12], [13], [14]. Dtotal as the primary objective function due to the reason
Hence, the model is solved using Traffic as the lone objective explained in Section 3.2.2 and the T raf f ic as an inequality
function with an additional constraint that do not allow the constraint with specified values of traffic in the cloud. We
value of the delay function to exceed a set of prescribed solve the FPP optimization problem in (27) subject to:
fractions of its optimal value. In other words, we try to
improve the less important criteria through minimal loss T raf f ic ≤  (30)
of the most important criterion.
and all other constraints (7)-(19). However, it is not easy to
M inimize (Dtotal ) (27)
find the exact  value for the optimal solution. For this, the
In the first step, we solve Equation (27) with respect to range of T raf f icmax and T raf f icmin is divided into ten
constraints (7)-(19). In the second step, we solve Equation equal intervals and we use the eleven values as the varying
(28) with an additional constraint (29) where δ is the added  value to obtain an optimal Pareto set curve [25].
value for delay. We vary δ with a set of values to obtain a The flowchart to solve the FPP using the trade-off
Pareto optimal front. method is illustrated in Fig. 5. The method uses the same
randomly generated inputs used for the previous two multi-
M inimize (T raf f ic) (28) objective optimization methods. The  value is initially set
Dtotal ≤ min
Dtotal ∗δ (29) to 1 and the modified trade-off model from Section 3.2.3 is
executed. The optimal solutions from the model are stored
The steps for solving the hierarchical method are shown in the Pareto optimal set. In the next step, the terminating
in Fig. 4. Initially, the inputs for the FPP are randomly condition is checked and if it is not satisfied, the problem is
generated using a uniform distribution. Then, the value of solved with a new  value. The model is repeatedly solved
δ is set to 1.0001, the modified hierarchical model is solved until the terminating condition is satisfied with storing all
and the solution is stored in the Pareto optimal set. Next, the the optimal solutions in the Pareto set. When the terminat-
terminating condition is checked and if it is not satisfied, the ing condition is met, an optimal Pareto front is obtained.
problem is solved again with the updated value of δ . When Finally, the fuzzy-based mechanism is used to obtain the
the terminating condition is satisfied, a Pareto optimal set best compromised result for the trade-off method.
is obtained and the best compromised result is found using
the fuzzy decision approach.
3.3 Determining the Best Compromise Solution
3.2.3 Trade-Off Method From each method of the multi-objective model, we obtain a
This method involves optimizing a primary objective and set of non-dominated solutions forming a Pareto front. Since
expressing the other objectives in the form of inequality con- there is an uncertainty in decision makers judgment, it can
straints [24]. In our proposed trade-off model, we consider be assumed there is a fuzziness in the goal of each objective

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2874484, IEEE
Transactions on Cloud Computing
8

The solution with a maximum value of µj is a compro-

mise solution that can be accepted by the decision maker.

3.4 The Hypervolume Indicator

The hypervolume (HV) indicator was first proposed in [27].
In recent years, the hypervolume indicator has been widely
used in multi-objective optimization to evaluate the various
search algorithms.
If an objective space contains solutions that are consid-
ered as points, the n-dimensional space that is contained by
a solution set is called hypervolume. In other words, the
n-dimensional volume of the set relative to some reference
point. The reference point usually refers to the worst pos-
sible point in the solution set. The hypervolume of a set is
the total dominated space of the solutions in the set. The
single unary value of HV gives a measure of the spread of
the solutions along the Pareto front. A set with large HV is
always desirable as it presents a better set of trade-offs.
Given a non-dominated solutions S, for each solution
i  S there is a hypercube vi with a reference point r where
the solution i is the diagonal corners of the hypercube.
With all the union of hypercubes, the hypervolume can be
calculated by:
|S|
HV = ∪i=1 vi (33)
Fig. 5: Steps to solve the FPP with the trade-off method
4 R ESULTS AND A NALYSIS
In this section, we solve the planning problem with the
function. The membership functions define this fuzziness three techniques presented in Section 3 for different problem
by representing the degree of fuzziness in some fuzzy sets sizes. Moreover, we do a complete analysis of the results to
using values in the range [0, 1]. better understand the performance of the planning model.
In the FPP, we adapt a fuzzy-based mechanism [26]
which is used to find out a compromised solution on the
Pareto front of each problem. The fuzzy mechanism looks 4.1 Simulation Setup and Input
at the way the solutions are contributing to each objective To solve the problem, we used CPLEX 12.7 [28]. All the
and assigns a fuzzy variable. When the solutions in a parameters in the optimizer were set to default except that
Pareto front are very close together, the mechanism shows we set a time limit of 1.5 hours as most optimal solutions
a possible way of finding the best compromise solution. In are usually obtained quickly (approx. within 1 hour). This
this mechanism, a membership value for the ith objective means that if the optimal solution is not found after 1.5
of the j th solution in the Pareto font is calculated using the hours, CPLEX will simply return the best solution obtained
membership function as: so far. We run the simulation on a PC with an Intel i5
processor running at 3.00 GHz with a total memory of
 12 GB.
1 if Fi ≤ Fimin
All the planning problems are solved with respect to the

Fimax −Fi
µji = max min if Fimin < Fi < Fimax (31) input specified. Each fog node has a maximum capacity that
 i −Fi
F
0 if Fi ≥ Fimax cannot be crossed. In other words, if one of the requested

parameters like memory is more than the nodes capacity,
µji indicates how well the j th solution is able to satisfy the then the capacity should be increased using a different
ith objective in a Pareto optimal set. The sum of the mem- fog type or send the edge device cluster request to the
bership value of all objectives of the j th solution suggests cloud. The specification of each fog node depends on the
how well it satisfies all the objectives. Given N solutions in service providers and there are many companies like IBM,
a Pareto front and M objective functions for each solution, Rackspace, Amazon AWS, etc. providing cloud services.
the achievement of each solution with respect to all the N Table 1 shows the four types of fog nodes with light to heavy
solutions can be calculated by: configurations that can be used by the solver [29].
The specifications of the different link types used to
M
connect the fog node(s) to the cloud are shown in Table 2.
µji The price of the links depends on their capacity and material
P

µj = i=1
(32) which can be obtained from network sales website [30], [31].
N PM The input parameters for the edge device clusters and
µji
P
j=1 i=1 each edge device within the cluster are tabulated in the

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2874484, IEEE
Transactions on Cloud Computing
9

TABLE 1: Characteristics of the different fog types TABLE 4: Problem size for the small scale FPP

Fog Type # of CPU Memory (GB) NIC (Mbps) Cost ($) Inventory Problem # of edge device clusters # of possible placement locations
1 90 480 360 67200 20 1 10 5
2 180 800 1024 120000 15 2 15 5
3 360 1600 1024 170000 10 3 20 5
4 720 3200 10240 250000 4 4 25 5
5 30 5
6 35 5
7 40 5
TABLE 2: Characteristics of the different link types 8 45 5
9 50 5
10 55 5
Link Type Bandwidth Cost/meter ($) 11 60 5
1 100 Mbps 0.25 12 65 5
2 1 Gbps 2 13 70 5
14 75 5
3 10 Gbps 200
15 80 5
16 85 5
17 90 5
18 95 5
19 100 5
Table 3. The maximum budget of the planning problem 20 150 5
21 200 5
was set to 40% of the maximum cost the fog network can
incur. Instead of a fixed value for the budget, we used a
percentage of the maximum cost because with the increase
of network size (possible placement locations) the maximum 4.2 Small Scale Problems
cost increases and the budget should also be increased
In the first set of problems, the number of edge device clus-
proportionally instead of biasing the small scale problems
ters is varied between 10 and 200 with 5 possible placement
with a large budget. The cost of a fog network includes the
locations. As shown in Table 4, 21 different problem sizes are
fog node cost, the location cost and the link cost between
solved over 4 different instances (i.e. we randomly generate
the connected fog nodes and the cloud. The maximum cost
4 different problems for each size) for a total of 84 problems.
includes the installation of highest capacity of fog node(s)
The results presented in Fig. 6 are the averages over the 4
in all the locations as well as connecting the fog node(s)
different instances and include the best compromised results
and the cloud with the most expensive link type which
for the objective functions.
has the highest speed. The simulation area is chosen to be
The average delay, traffic and solution time comparison
100 km by 100 km. This larger area is chosen because we
of the three methods are graphically represented in figures
envision the fog network to be spanning over a city and
6a, 6b, and 6c respectively. For FPP, we used the 95%
beyond. Two values that make a single point of (x, y) are
confidence interval shown by the vertical bars to examine
randomly generated for the locations of edge device clusters
the reliability of the result obtained. As it can be seen
and possible fog node placements. All other inputs where
from the figures, there is less difference between the three
also randomly generated within a set range.
methods in terms of total network delay and traffic towards
Two different sets of problems (referred to as small scale the cloud for each of the problem. However, the hierarchical
and large scale) are used in the fog planning model. The method attains less delay and more traffic than the other
scale of the problem sizes is determined from the number of two methods for most of the problems. In fact, in only three
possible placement locations (5 vs 10) which also determine problems, the delay attained by the hierarchical method is
the complexity of the problem. The planning problem can greater than the other two methods. This is because the
only be solved by the use of combinatorial methods and as hierarchical optimization is bounded by the delay constraint
the problem size gets larger, the number of possible com- and we do not allow the objective function to go beyond
binations increases exponentially. For example, a network a reasonable delay in the network which compromises
that has 5 possible locations and 20 edge device clusters, the optimal traffic. On the other hand, the weighted sum
there are approximately 520 different possible solutions for method resulted in minimum traffic going to the cloud for
the planning problem and when the possible location is most of the problems compared to the other two methods.
increased the problem becomes exponentially larger with With the increase of problem size, the objective function also
approximately 1020 different possible solutions. increase for all the three methods. This is because more edge
device clusters are introduced in the network resulting in
more traffic requests for the fog nodes for the same number
TABLE 3: The edge device cluster input parameters of possible fog node placement locations. Hence, many of
the edge device clusters must be connected to the cloud.
Simulation area 100x100 km
The weighted sum method and the trade-off method return
Packet size 1500 Bytes similar results for delay and traffic due to the use of the
# of edge device in a cluster U(10,200) maximum and minimum values of the objective function as
# of CPU core U(1,4) formulated in equations 25 and 30.
Memory U(1,40) MB
In general, with the increase of input size, the time taken
# of Packets sent per second U(1,64)
Network access bandwidth U(20,70) Mbps to obtain an optimal solution must increase as the model
must consider a lot of combinations of where to install

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2874484, IEEE
Transactions on Cloud Computing
10

(a) Average delay comparison for small scale (b) Average traffic comparison for small scale (c) Average solution time comparison for
problems problems small scale problems

Fig. 6: Analysis for small scale problems

the fog node, what type of fog node needs to be installed TABLE 5: Average solution comparison among the three
and which edge device cluster should be connected to it. methods
However, from Fig. 6c, it can be seen that the solution time
Weighted Sum Hierarchical Trade-off
increases in a non linear pattern. This is due to the use of Delay Gap Traffic Gap Delay Gap Traffic Gap Delay Gap Traffic Gap
the branch and bound (B&B) algorithm [32] by the solver. Problem (%) (%) (%) (%) (%) (%)
In B&B, at first the search space is recursively split into 1
2
1.5
1
0
0.2
0
0
13
0
1.4
1
26
0.2
smaller search spaces called branching and tries to find the 3 0.65 0 0 3.8 0.65 0
4 0.89 0 0 4.2 0.89 0
minimum objective function in the smaller search space. To 5 1.1 0.81 0 6 1.5 0
6 0.53 0 0 3.3 0.53 0.058
avoid the brute-force search and testing all the candidate 7 0.022 0 0.061 0.82 0 0.57
8 2.1 0 0 5.1 0.68 1.3
solutions, the algorithm uses heuristics to keep track of 9 0.9 0.22 0 4.2 1.1 0
bounds on the minimum that it is trying to find and these 10
11
0.53
1
0
0
0
0
3.4
4.9
0.53
0.84
0.092
0.66
bounds are used to cut back the search space by eliminating 12 1.1 0 0 4.7 0.95 0.36
13 0.033 0 1.1 0.28 0 0.48
the candidate solutions which cannot give optimal solution. 14 0.8 0.12 0 3.8 0.92 0
15 0.54 1.5 0 4.7 1.2 0
To add with that, for the same problem, different multi- 16 0.8 0 0 3 0.78 0.38
17 0 0 16 1.9 16 0.92
objective optimization methods generate different search 18 0.34 0 0 1.3 0.41 0.16
spaces. For instance, from problem seven onward, the hi- 19 0.92 0 0 2.8 0.73 0.52
20 0.25 0.35 0 1.4 0.38 0
erarchical method takes a lot less time compared to at least 21 0.22 0 0 0.62 0.25 0.035
Mean 0.73 0.15 0.82 3.49 1.46 1.51
one of the methods and in some cases both methods. The
reason it takes less time is because the objective function
is traffic towards the cloud which, according to Equation
(6), has less variables and constraint (29) provides a strong
feasible initial solution for the particular method compared
to the other two optimization methods. Moreover, for each
problem, the hierarchical method runs for 8 iterations only
which is 3 less than the other two methods. Although the
confidence interval for the average solution time was also
calculated, it could not be represented in the figure because
of the semi-log scale used in the plot.

4.2.1 Comparison to the Best

Table 5 shows the comparison of the results among the Fig. 7: Average HV indicator for small scale problems
three multi-objective optimization methods for small scale
problems. A 0% implies that the delay or traffic value
obtained by that method for the particular problem is the 4.2.2 HV for Small Scale Problem
minimum among all the three methods, hence, there is no The hypervolume indicator is a widely used performance
percentage difference. The mean percentage gap for the metric for multi-objective optimization. Fig. 7 shows the HV
weighted sum method for average delay and traffic is 0.73% comparison between the three methods for the small scale
and 0.15% respectively. The mean of average delay gap for problems. For different problem sizes we obtain different
the hierarchical method is 0.82% and the traffic is 3.49%. HV ranging from a lowest value of 0.274 to a highest value
For the trade-off method, the mean of average delay gap of 0.783. Each of the problem generates a different search
obtained was 1.46%, whereas, the mean of average traffic space, hence, a randomness can be observed in the figure.
difference was 1.51%. From the results, it can be identified From the figure, it can be observed that the hierarchical
that, for small scale problems, the weighted sum method method results in smaller HV compared to the other two
returns the best optimal results among the three methods. methods for all the problems. This is due to the lack of non-

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2874484, IEEE
Transactions on Cloud Computing
11

TABLE 6: Problem size for the large scale FPP TABLE 7: Average solution comparison among the three
methods over 4 instances for large scale problems
Problem # of edge device clusters # of possible placement locations
22 10 10 Weighted Sum Hierarchical Trade-off
23 15 10 Delay Gap Traffic Gap Delay Gap Traffic Gap Delay Gap Traffic Gap
24 20 10 Problem (%) (%) (%) (%) (%) (%)
25 25 10 22 0.013 0 0.019 0 0 1.2
26 30 10 23 0.74 0 0 4.3 0.67 5.7
27 35 10 24 0.55 0 0 4.1 0.44 5
25 0.56 3.3 0 7.7 0.96 0
28 40 10 26 1 0 0 3.7 0.68 1
29 45 10 27 0 0 0.22 8.7 0.57 6.2
30 50 10 28 1.6 0 0 10 1.1 2.6
29 0.67 0 0 8.1 0.63 0.56
31 55 10 30 0.94 0 0 8.3 0.65 3
32 60 10 31 1.9 0 0 12 1.9 0.65
32 2.3 0 0 4.9 0.94 1.2
Mean 0.93 0.3 0.022 6.53 0.78 2.46

dominated member in the Pareto front, as the hierarchical

method can at most generate only eight points or less proposed with respect to a delay sensitive network. It can
in the Pareto front. Moreover, the delay constraint does be clearly observed from the figure, that the traffic in the
not allow to explore as many combinations as the other cloud increases linearly with the problem size for the first
two methods can do. The weighted sum method and the 5 problems. However, problem 27 interrupts the pattern.
trade-off method have similar HV for its respective Pareto The pattern again continues for the last 5 problems. If
front for each problem and most of the HV is above 0.5 the number of instances is increased, although the average
which is better than the hierarchical method. However, the slightly deviates from the pattern, the confidence intervals
trade-off method outperforms the weighted sum method are still following the pattern resulting a smoother pattern.
by a slight margin proving that it has more diverse non- The average solution times are illustrated in Fig. 8c for the
dominated solutions and better convergence towards the three methods over four instances. As the vertical axis in the
true approximated Pareto front among all the three meth- figure is in logarithmic scale, the confidence interval could
ods. The following section explores the effect of increasing not be illustrated. The figure shows that there is a non linear
the problem size to a larger scale. increasing pattern for the average solution time for all the
problems. The randomness in the solution time relates to
the theory that the combinational space are shortened by
4.3 Large Scale Problems
the solver using a heuristic. As the figure depicts, for the
The second set of problems shows how the three methods same problem size there is a visible difference in solution
behave when the number of possible placement locations time for the three methods. For example, to solve problem
is increased. The number of possible placement locations 25, the weighted sum takes 6,020 seconds, the hierarchical
is increased from 5 to 10 for a range between 10 to 60 solves in 807 seconds and the trade-off method takes 10,445
edge device clusters. The number of edge device cluster seconds. Each method generates a different search space for
is stopped at 60 because after 45 edge device clusters, an the same problem, hence solving the same problem size with
optimal solution was not guaranteed as several iterations different solution time.
where not returning the optimal solution within the time
limit. A total of 44 problems were solved with 11 different 4.3.1 Comparison to the Best
problem sizes as shown in Table 6 and the average of each Table 7 represents the comparison of results among the three
problem size was used for the analysis of the problem set. multi-objective optimization methods for large scale prob-
We used the same CPLEX and hardware environment as lems. The weighted sum method returns a mean percentage
defined in Section 4.1. gap of 0.93% for the average delay and a gap of 0.3% for
Fig. 8a graphically represents the average delay compar- the average traffic. The hierarchical method solves the large
ison between the three methods for large scale problems scale problems with a mean percentage difference of 0.022%
over four different instances along with the 95% confi- for the average delay, which is the lowest among the three
dence interval. As it can be noticed from the figure, as methods and a highest mean percentage difference of 6.53%
the problem size increases, the delay in the network also for the average traffic. Finally, the trade-off method result
increases. We can observe a linearity in the increment of in a mean percentage gap of 0.78% for the average delay
delay for problems 22 to 29, but the pattern is disrupted and a mean of 2.46% gap for the average traffic. Although,
by problem 30 for which the delay in the network is less the hierarchical method returns the best results for the delay
than the previous problem size. The pattern continues for value in the network, the weighted sum method performs
problem 31 and 32. However, we believe that if we run the best with a lowest mean percentage difference of 1.23%.
each problem for more than four instances, a more linear
pattern could be seen without any interruption. The average 4.3.2 HV for large scale problems
traffic comparison between the three methods for the four Fig. 9 portrays the average HV for large scale FPP for the
different instances of the second problem set is illustrated three methods. The highest HV of 0.9396 was achieved
in Fig. 8b. The traffic obtained from the weighted sum and by the hierarchical method, whereas the trade-off methods
trade-off methods are very similar, whereas the hierarchical gives the lowest HV of 0.2993. It can be observed that the
method attains a slightly higher traffic for every problem. hierarchical and the trade-off method show an opposite pat-
This is because, in the FPP, the hierarchical method was tern with the increase of problem size. For the first problem,

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2874484, IEEE
Transactions on Cloud Computing
12

(a) Average delay comparison for large scale (b) Average traffic comparison for large scale (c) Average solution time comparison for
problems problems large scale problems

Fig. 8: Analysis for large scale problems

be used for situations were the decision maker can select

the best result from the set of non-dominated solutions,
instead of using the fuzzy-based approach. The weighted
sum method generated the best Pareto optimal set for the
large scale problems.
The FPP is a NP-hard problem and the solution time
increases exponentially with the increase of input size. A
viable improvement can be the use of multi-objective heuris-
tic algorithms like Strength Pareto Evolutionary Algorithm
(SPEA) or Non-dominated Sorting Genetic Algorithm II
(NSGA-II). Although, optimal solutions cannot be guaran-
teed, a significant improvement can be achieved in terms
of computation time. Another future direction can be the
Fig. 9: Average HV indicator for large scale problems expansion of the fog network. Once the fog network is
planned and implemented, over time, more edge device
clusters may be added to the network. The proposed model
the hierarchical method gives the best Pareto front but keeps can be extended to an expansion model, where new fog
on getting worse as the problem size increases and it is the nodes can be added or replaced without changing the whole
opposite for the trade-off method. Compared to the other network infrastructure. Finally, once fog nodes are in place,
two methods, the weighted sum gives a more consistent set the real time management of fog ressources is also an
of non-dominated solutions for all the problem sizes. Hence, interesting avenue to explore.
it can be deduced that the weighted sum generates the best
dominated space of the solutions in a Pareto front.
R EFERENCES
[1] C. S. M. Babou, D. Fall, S. Kashihara, I. Niang, and Y. Kadobayashi,
5 C ONCLUSION AND F UTURE W ORK “Home edge computing (hec): Design of a new edge computing
In this paper, we proposed a new mathematical model for technology for achieving ultra-low latency,” in Edge Computing –
EDGE 2018, S. Liu, B. Tekinerdogan, M. Aoyama, and L.-J. Zhang,
the planning and design of fog networks. The problem Eds. Cham: Springer International Publishing, 2018, pp. 3–17.
simultaneously finds the optimal location, capacity and the [2] G. Li, J. Wang, J. Wu, and J. Song, “Data processing delay opti-
type of fog nodes, as well as the optimal connection between mization in mobile edge computing,” Wireless Communications and
Mobile Computing, vol. 2018, no. 6897523, p. 9, 2018.
the fog node(s) and the cloud. The optimization model [3] P. Mach and Z. Becvar, “Mobile edge computing: A survey on
minimizes the network delay and the traffic in the cloud architecture and computation offloading,” IEEE Communications
which are two of the most important parameters in fog Surveys Tutorials, vol. 19, no. 3, pp. 1628–1656, thirdquarter 2017.
networks. It can be observed that as the input size increases, [4] J. Lee and J. Lee, “Hierarchical mobile edge computing architec-
ture based on context awareness,” Applied Sciences, vol. 8, no. 7,
the objective functions and the solution time for each multi- 2018.
objective optimization method also increase. The objective [5] M. St-Hilaire, “Topological planning and design of UMTS mobile
functions show a linear increase while the solution time networks: a survey,” Wireless Communications and Mobile Comput-
ing, vol. 9, no. 7, pp. 948–958, 2009.
increases in a non linear form. This illustrates that the FPP
[6] “OpenFog Consortium,” https://www.openfogconsortium.org/,
is a NP-hard problem and further increase in the problem (Accessed on 13/09/2018).
size will increase the solution time exponentially. [7] K. Hong, D. Lillethun, U. Ramachandran, B. Ottenwälder, and
In term of the methods, the weighted sum method was B. Koldehofe, “Mobile fog: A programming model for large-scale
applications on the internet of things,” in Proceedings of the Second
able to return the best trade-off results for the delay and the ACM SIGCOMM Workshop on Mobile Cloud Computing, ser. MCC
traffic. The hierarchical method was able to achieve better ’13. New York, NY, USA: ACM, 2013, pp. 15–20. [Online].
delay but the traffic was worse compared to the other two Available: http://doi.acm.org/10.1145/2491266.2491270
methods. For small scale problems, the trade-off method [8] G. Orsini, D. Bade, and W. Lamersdorf, “Computing at the mo-
bile edge: Designing elastic android applications for computation
generated the best set of non-dominated solutions. It shows offloading,” in 2015 8th IFIP Wireless and Mobile Networking Confer-
that, for small scale problems, the trade-off method can ence (WMNC), Oct 2015, pp. 112–119.

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2874484, IEEE
Transactions on Cloud Computing
13

[9] M. Peng, S. Yan, K. Zhang, and C. Wang, “Fog-computing-based [28] IBM, IBM ILOG CPLEX Optimization Studio CPLEX User’s Manual,
radio access networks: issues and challenges,” IEEE Network, 2011.
vol. 30, no. 4, pp. 46–53, July 2016. [29] “Amazon ec2 instance types amazon web services (aws),”
[10] R. Vilalta, V. Lopez, A. Giorgetti, S. Peng, V. Orsini, L. Velasco, https://aws.amazon.com/ec2/instance-types/, (Accessed on
R. Serral-Gracia, D. Morris, S. D. Fina, F. Cugini, P. Castoldi, 07/01/2018).
A. Mayoral, R. Casellas, R. Martinez, C. Verikoukis, and R. Munoz, [30] I. Kateeb, K. Alotaibi, L. Burton, and M. S Peluso, “The funda-
“Telcofog: A unified flexible fog and cloud computing architecture mental component of telecommunications cabling,” pp. 452–467,
for 5g networks,” IEEE Communications Magazine, vol. 55, no. 8, pp. 01 2013.
36–43, 2017. [31] “Managed dedicated server hosting — rackspace hosting
[11] I. Stojmenovic, “Fog computing: A cloud to the ground support experts,” https://www.rackspace.com/dedicated-servers, (Ac-
for smart things and machine-to-machine networks,” in 2014 cessed on 08/12/2017).
Australasian Telecommunication Networks and Applications Conference [32] J. Clausen, “Branch and bound algorithms – principles and exam-
(ATNAC), Nov 2014, pp. 117–122. ples,” 1999.
[12] V. Stantchev, A. Barnawi, S. Ghulam Muhammad, J. Schubert, and
G. Tamm, “Smart items, fog and cloud computing as enablers of
servitization in healthcare,” vol. 185, pp. 121 – 128, 02 2015.
[13] Y. Cao, S. Chen, P. Hou, and D. Brown, “Fast: A fog computing
assisted distributed analytics system to monitor fall for stroke
mitigation,” in 2015 IEEE International Conference on Networking,
Architecture and Storage (NAS), Aug 2015, pp. 2–11.
[14] T. N. Gia, M. Jiang, A. M. Rahmani, T. Westerlund, P. Liljeberg, and Faisal Haider received his Bachelor of Sci-
H. Tenhunen, “Fog computing in healthcare internet of things: ence degree in Electrical and Electronics Engi-
A case study on ECG feature extraction,” in 2015 IEEE Interna- neering from American International University-
tional Conference on Computer and Information Technology; Ubiquitous Bangladesh in 2013. He received his Master of
Computing and Communications; Dependable, Autonomic and Secure Applied Science degree from the Department of
Computing; Pervasive Intelligence and Computing, Oct 2015, pp. 356– Systems and Computer Engineering from Car-
363. leton University, Canada, in 2018. His research
[15] S. Sareen, S. K. Gupta, and S. K. Sood, “An intelligent interests include fog/edge computing, network
and secure system for predicting and preventing zika virus planning and design, network optimization and
outbreak using fog computing,” Enterprise Information Systems, cloud computing.
vol. 11, no. 9, pp. 1436–1456, 2017. [Online]. Available:
http://dx.doi.org/10.1080/17517575.2016.1277558
[16] M. Aazam and E. N. Huh, “Fog computing and smart gateway
based communication for cloud of things,” in 2014 International
Conference on Future Internet of Things and Cloud, Aug 2014, pp.
464–470.
[17] J. Zhu, D. S. Chan, M. S. Prabhu, P. Natarajan, H. Hu, and
F. Bonomi, “Improving web sites performance using edge servers
in fog computing architecture,” in 2013 IEEE Seventh International Decheng Zhang received his Bachelor of En-
Symposium on Service-Oriented System Engineering, March 2013, pp. gineering degree from the Department of Infor-
320–323. mation and Communication Engineering, Bei-
[18] Y. Xiao, M. Noreikis, and A. Yl-Jaiski, “Qos-oriented capacity plan- jing University of Posts and Telecommunications,
ning for edge computing,” in 2017 IEEE International Conference on China, in 2013. He received his Master of Ap-
Communications (ICC), May 2017, pp. 1–6. plied Science degree from the Department of
[19] G. Premsankar, B. Ghaddar, M. D. Francesco, and R. Verago, Systems and Computer Engineering at the Car-
“Efficient placement of edge computing devices for vehicular leton University, Canada, in 2018. His research
applications in smart cities,” 2017. interests include fog/edge computing, fog net-
[20] A. Yousefpour, G. Ishigaki, and J. P. Jue, “Fog computing: Towards work planning, optimization research and cloud
minimizing delay in the internet of things,” in 2017 IEEE Interna- computing.
tional Conference on Edge Computing (EDGE), June 2017, pp. 17–24.
[21] M. S. Elbamby, M. Bennis, and W. Saad, “Proactive edge com-
puting in latency-constrained fog networks,” in 2017 European
Conference on Networks and Communications (EuCNC), June 2017,
pp. 1–6.
[22] M. Cvijetic and I. Djordjevic, Advanced Optical Communication
Systems and Networks, ser. Artech House applied photonics series.
Artech House, 2013. [Online]. Available: https://books.google. Marc St-Hilaire joined Carleton University in
ca/books?id=cW0m1LtSMj8C 2006 upon completion of his PhD in Com-
[23] C. Systems, “Design best practices for latency optimization,” Cisco puter Engineering, from École Polytechnique of
Systems, Inc., CA, USA, Whitepaper, 2007. Montréal. He is currently an associate profes-
[24] V. Chankong and Y. Haimes, Multiobjective decision making: theory sor with the School of Information Technology
and methodology, ser. North-Holland series in system science with a cross appointment with the Department of
and engineering. North Holland, 1983. [Online]. Available: Systems and Computer Engineering at Carleton
https://books.google.ca/books?id=ZIg-AQAAIAAJ University. Dr. St-Hilaire is conducting research
[25] G. Mavrotas, “Effective implementation of the -constraint method on various aspects of wireline and wireless com-
in multi-objective mathematical programming problems,” Appl. munication systems. More precisely, he is in-
Math. Comput., vol. 213, no. 2, pp. 455–465, Jul. 2009. [Online]. terested in network planning and infrastructure,
Available: http://dx.doi.org/10.1016/j.amc.2009.03.037 network protocols, network interconnection, and performance analysis.
[26] B. Panigrahi, V. Pandi, R. Sharma, S. Das, and S. Das, “Multi- With more than 120 publications, his work has been published in several
objective bacteria foraging algorithm for electrical load dispatch journals and international conferences. Finally, Dr. St-Hilaire is actively
problem,” vol. 52, pp. 1334–1342, 02 2011. involved in the research community. In addition to serving as a member
[27] E. Zitzler and L. Thiele, “Multiobjective optimization using of technical program committees of various conferences, he is equally
evolutionary algorithms - a comparative case study,” in involved in the organization of several national and international confer-
Proceedings of the 5th International Conference on Parallel ences and workshops. He is also a senior member of the IEEE.
Problem Solving from Nature, ser. PPSN V. London, UK,
UK: Springer-Verlag, 1998, pp. 292–304. [Online]. Available:
http://dl.acm.org/citation.cfm?id=645824.668610

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2018.2874484, IEEE
Transactions on Cloud Computing
14

Christian Makaya is currently a researcher at

IBM T.J. Watson Research Center, NY, USA.
Prior to joining IBM, he was a senior research
scientist at Telcordia Technologies, NJ, USA and
a visiting researcher at Ericsson Research, Mon-
treal, Canada. His work has been a catalyst
behind several new initiatives and technologies
resulting in delivery of high-value capabilities to
products and services. For his technical contri-
butions, he has been recognized by several high-
prestige internal awards by IBM. Dr. Makaya
leads several technical research activities in the areas of distributed
systems and analytics with the mission of delivering deep technical
breakthroughs. The focus of his current research interests is on dis-
tributed AI and ML, edge computing, Internet of Things (IoT), network
functions virtualization (NFV), policy-based management systems, and
cyber-security. Dr. Makaya has authored numerous technical papers in
peer-reviewed journals and conferences, and filled several patents. He
received his Ph.D. in Computer Engineering from Polytechnique Mon-
treal (2007). Dr. Makaya is an active member and volunteer of IEEE and
serves on the Industry Outreach Board of IEEE Communication Society
(ComSoc). He served as the co-chair of IEEE Young Professionals for
the IEEE Princeton/Central Jersey Section.

2168-7161 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.