Computer Communications 182 (2022) 88–97

Contents lists available at ScienceDirect

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

A novel proactive controller deployment protocol for 5G-enabled

software-defined Vehicular Networks
Noura Aljeri a,b ,∗, Azzedine Boukerche a
a PARADISE Research Laboratory, EECS, University of Ottawa, Canada
b Gulf University for Science and Technology, Kuwait

ARTICLE INFO ABSTRACT

Keywords: With the vast expansion of data loads across numerous safety and infotainment applications, the software-
5G-enabled Vehicular Networks defined paradigm has become a critical component for 5G-enabled and intelligent vehicular networks. By
Autonomous Driving transforming the network’s rigid hardware devices into programmable units, SDVN offers a broad view of the
Distributed SDVN
entire network status and a standard interface between heterogeneous wireless access technologies. However,
Controller Placement Problem
deploying a centralized control unit carries several problems, such as bottleneck issues and densification
Time-Series Prediction
Sustainability
issues. A distributed control plane is a potential alternative to a centralized control plane. Still, it poses many
questions about the optimal location of local control units and the number of control units required for a given
network structure. In this article, we propose an efficient proactive controller deployment and assignment
strategy for distributed 5G-enabled software-defined Vehicular Networks based on predictive vehicles’ traffic
flow information and communication delays between switch-enabled roadside units. A thorough performance
evaluation is conducted under different vehicular network’s mobility scenarios. We assess the performance of
the proposed method in terms of communication latency and load on the subsequent control units and their
assigned set of switches. The proposed proactive clustering method is compared to different forms of controller
placement methods. The results showed that the proposed approach reduced the network’s communication
delays while maintaining a balanced distributed load among controllers over the simulation time.

1. Introduction changes in network topologies, and traffic flow variations and density
changes [7]. One unique networking paradigm emerged as a solu-
The current worldwide shift toward sustainable and autonomous tion to address those challenges in 5G-enabled vehicular networks,
driving culture is deeming a more intelligent, efficient, and reliable referred to as Software-Defined Vehicular Networks (SDVN) [8–11].
communication standardization of network protocols and services. The SDVN is designed based on the concept of control/data separation
next generation of vehicular networks is envisioned to play an essential in order to allow for broader and centralized control on the network
role in designing heterogeneous wireless access technologies integra- devices management, configuration, and deployment without the need
tion including Cellular LTE/5G/6G, WiFi 6, WiMAX, as well as Satellite of technical personnel intervention. However, the SDVN’s centralized
access stations. Henceforth, enabling vital and time-sensitive services controlling approach introduces some challenges, especially within ve-
and applications to passengers and drivers, as well as traffic control hicular networks environment, such as single point of failure and rapid
management systems [1]. Several proposals [2] investigated possible vehicles migration in dense networks; thus, exhausting and overloading
solutions to mitigate the challenges in intelligent heterogeneous wire- the network control entity. One possible solution is to elevate the
less networks through spectrum expansion and separation of network burden on the centralized controller to be locally handled by several
control and data planes, among others. trusted access points devices. Thus, ensuring scalable network design,
The fifth generation (5G)-based Vehicular networks’ environment, in which tasks are distributed among local control units and reported
in particular, brings along multiple issues that need to be considered to global controllers when needed. This architecture is referred to as
when designing efficient network protocols and applications such as the distributed software-defined vehicular networks [12–14]. However,
network’s infrastructure deployment, resources management and al- distributed SDVNs raise several questions to its design in terms of how
location [3], routing [4], mobility management [5], and more [6]. many local control units are selected and their locations, as well as
Other concerns involves the dense deployment and short-range com- how to partition all switch-enabled roadside units based on the desired
munication of infrastructure units, the rapid mobility of vehicles, number of local control units [15–18].

∗ Corresponding author.
E-mail addresses: naljeri@uottawa.ca (N. Aljeri), boukerch@site.uottawa.ca (A. Boukerche).

https://doi.org/10.1016/j.comcom.2021.09.024
Received 1 April 2021; Received in revised form 18 August 2021; Accepted 20 September 2021
Available online 1 October 2021
0140-3664/© 2021 Published by Elsevier B.V.
N. Aljeri and A. Boukerche Computer Communications 182 (2022) 88–97

In this article, we propose a novel proactive controllers deployment reduce network latency, implementation expense, and energy con-
and assignment protocol based on traffic flow estimation for distributed sumption [22]. Table 1 presents a comparison of the recent related
5G-enabled software-defined vehicular networks. The proposed tech- work on the controller placement problem compared to our proposed
nique utilizes the vehicle’s mobility and traffic flow information to work. Other studies focused on optimizing reliability [23]. Nonetheless,
dynamically adapt the network’s topology and control-to-switch assign- given the constraints imposed by vehicles’ continuous mobility and
ment in order to provide efficient network performance and balanced connectivity limitations within vehicular networks, an effective alloca-
distribution of load on controllers. tion strategy is needed to minimize the difficulties inherent in such a
The remainder of this paper is organized as follows. Section 2 scenario. The optimal location of controllers in an urban/highway road
reviews several strategies introduced in the literature that address the structure is determined by a number of variables, including contact
centralized controller’s deployment problem and points out possible latency, demand on infrastructure units, hotspot areas managed by the
solutions toward finding the optimal deployment scheme. Section 3 units, vehicle movement and traffic between network entities, as well
presents the proposed adaptive controller deployment and clustering as the potential capacity of the selected units.
technique and preliminaries, followed by a detailed description of the
incremental k-mean clustering method and the traffic flow prediction 3. The proposed proactive adaptive controller selection and clus-
design. In Section 4, we evaluate the proposed scheme under various tering scheme
scenarios and variants. Lastly, Section 5 concludes the presented work
and points out future work direction. This section discusses the preliminaries and main components of
the proposed adaptive controller deployment strategy based on a mod-
2. Background ified incremental k-mean clustering technique to select the controller’s
location and their corresponding set of switches.
In this section, we present a general overview of the design and
characteristics of 5G-enabled software-defined vehicular networks, in 3.1. Preliminaries
addition to the recent research studies that tackle the controller-to-
Over the past years, we have exhibited a tremendous increase in
switch deployment and clustering issue addressing in particular vehic-
the number of network devices and wireless technologies, all deployed
ular network’s conditions and scenarios.
within what we refer to now as heterogeneous wireless networks.
The architecture of a centralized SDN controller poses a range of
Henceforth, a standard framework that can ease the control and man-
challenges in ultra-dense networks, especially in the rapidly evolving
agement of such devices is required more than ever. Software-defined
vehicular ecosystem and rapid mobility of vehicles [19]. Additionally,
networks became an essential factor in the design of efficient hetero-
using a single controller produces a variety of problems, like a single
geneous networks, specifically vehicular networks. Software-defined
point of failure. Due to the controller’s failure, data transmission is
vehicular networks consist of three principal planes, application, con-
delayed, and the data drop rate increases. That can be problematic in
trol, and data plane. Typically, the control plane acts as a proxy
time-critical safety applications and services such as disaster warnings
between the application and data plane and resides at the server-side.
and environmental surveillance. Additionally, using a single unified
In contrast, the data plane consists of roadside units and vehicles,
controller can be insufficient to accommodate the large volume of data
referred to as switch-enabled devices. However, when we look closely
packets arriving from base stations, cars, and several IoT gadgets. Each
at the characteristics of vehicular networks environment, from the rapid
switch-enabled infrastructure unit anticipates that a particular individ-
mobility of vehicles to its varied traffic densities over time, having a
ual, such as the SDN controller, can tolerate a minimal delay period.
centralized control unit is inclined to become a single point of failure
Finally, since we are dealing with a vehicular setting, the continuous
and overloaded. Therefore, it impacts the scalability and efficiency
movement of vehicles between communication points would often
of networks communication. In this article, we focus on designing a
entail controller awareness to maintain track of the vehicles’ mobility. distributed control structure based on multiple local controllers that
As a result, a significant amount of overhead would be added to the is able to manage the dense heterogeneity environment in 5G-enabled
network. Henceforth, many research studies have addressed the issue vehicular networks.
by proposing various SDN-enabled vehicular network architectures that The deployment of SDN controllers can be represented by an undi-
require either a hierarchical controller system [15] or the assistance of rected connected graph 𝐺 with 𝑆 switch-enabled units and 𝐿 links,
the Edge- and Cloud-Computing paradigms [20]. such that 𝐺(𝑆, 𝐿) is constructed. Each link 𝑙𝑖,𝑗 defines a direct link
By using additional controllers in the network architecture, the between switch 𝑖 and switch 𝑗. The set of 𝑆 switches represent all
issue of a single point of failure is eliminated, and the controllers’ available switches and also the selected controllers subset 𝐶. A 𝐺-graph
load is reduced [21]. However, determining the number of local con- includes 𝑁 switches of which 𝐾 are acting as controllers. Vehicles
trollers available, their position, and how to group several switches can communicate with switch-enabled roadside units (S-RSU) through
to each controller poses another obstacle, resulting in unbalanced Vehicle-to-Vehicle (V2V) and Vehicle-to-Infrastructure V2I communi-
load management and an excessive expense for local control unit cation protocols. In software-defined structure, we assume only V2I
deployment. communication through OpenFlow (OF) protocol [8]. As for switch-
In the literature, two essential methods are studied: static and enabled roadside units, they are connected through backhaul network
dynamic migration strategies [22]. Fixed deployment strategies are connections.
usually based on the intended outcome in terms of network latency So, let us start by defining the set of controllers 𝐶, which is a subset
or load. Current solutions include hard/soft clustering, partitioning, of 𝑆 and consist of 𝑐1 , 𝑐2 , ..𝑐𝑘 controllers. Each controller is associated
hierarchical controllers, and integrated solutions [10]. Most clustering with another subset of switches 𝐸 ∈ 𝑆, such that 𝐸 = {𝑒̂1 , ..𝑒̂𝑘 } and each
techniques are unsupervised clustering method [22], which transforms subset 𝑒̂𝑘 are associated with the controller 𝑐𝑘 . We also assume that a
the SDN network architecture into a graph theory problem. It operates switch 𝑠 ∈ 𝑆 belongs to one master controller subset 𝑒̂ ∈ 𝐸.
by clustering a series of switches to satisfy the given objective function.
The switch-migration approach is more flexible, which traces the net- 𝐶 = {𝑐1 , 𝑐2 , ..𝑐𝑘 }, 𝑘 = 1, … , 𝐾 (1)
work’s stability and migrates switches between controllers based on a 𝐸 = {𝑒̂1 , 𝑒̂2 , ..𝑒̂𝑘 }
series of metrics such as latency, cost, and quality of service, among (2)
others [24]. 𝑒̂𝑖 = {𝑠1 , 𝑠2 , … , 𝑠𝑗 }, 𝑠𝑗 ∈ 𝑆
Many research experiments on the SDN controller deployment prob- To define the most appropriate topology setup in a distributed
lem have been suggested for wired and wireless area networks to software-defined vehicular network, we need to acquire specific metrics

89
N. Aljeri and A. Boukerche Computer Communications 182 (2022) 88–97

Table 1
Controller placement strategies.
Related work Deployment Approach Metric Objective
Heller et al. [22] Static -minimum k-center propagation Reduce network
latencies latency
Liao et al. [23] Static Density-based Path hop count Optimizing
clustering reliability
Yang et al. [24] Dynamic Heuristic Load and migration Load balancing
optimization cost
problem
Wang et al. [25] Dynamic stable matching Response time Reduce controller
problem Traffic overhead response time
Wang et al. [26] Static Clustering-network E2E and controller Reduce E2E and
partitioning queue latency controller queue
latency
Liyanage et al. [27] Static p-median facility communication Reduce E2E delay,
location problem delay and RSU packet overhead
location priority
Our work Dynamic Adaptive Traffic-flow Reduce network
incremental k-means estimation latency and load
Communication balancing
latency, and load

that define the links’ stability, reliability, and availability. Based on The propagation delay is measured by the distance of channel’s
such information, we can construct the set of controllers and associated link between two points (𝑖, 𝑗) and the channel’s speed. In wireless
links as described in (1), (2). The link between switches can be defined communications the speed of the channel reflects the speed of the light
based on multiple factors, including the distance between two switches, (i.e., 3𝑥108 m∕s, while in wired links the Ethernet speed is around
the communication delay, or the stability factor between the two 2𝑥108 m∕s.
switches.
𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑖,𝑗
The set of controllers and associated switches can be constructed 𝑝𝑟𝑜𝑝𝑑 (𝑖, 𝑗) = (5)
using clustering techniques, splitting, or hierarchical approach [28]. 𝑠𝑝𝑒𝑒𝑑
That set of controllers can then be re-evaluated and assigned based The transmission delay 𝑡𝑟𝑎𝑛𝑠𝑑 is the total time it takes for all bits
on specific observations from the network, such as vehicles’ density, in one packet (size of the packet is 𝑝𝑎𝑐𝑘𝑒𝑡𝑠 ) to be transmitted by one
velocity, power consumption, traffic flow, and so on. In terms of point in bit rate 𝑏∕𝑟
clustering techniques, several methods can be adapted to address the
𝑝𝑎𝑐𝑘𝑒𝑡𝑠
controller deployment problem. The hard clustering methods, which 𝑡𝑟𝑎𝑛𝑠𝑑 (𝑖, 𝑗) = (6)
𝑏𝑖𝑡𝑟𝑎𝑡𝑒
gives a clear partitioning of the graph nodes, and soft clustering allows
the nodes (i.e., switches) to be part of two clusters (i.e., control unit Using the link delay as the bases of graph link between switches, we
subset). can now describe the relationship between switch-enabled RSUs and
their neighborhood in an 𝑁 × 𝑁 matrix as follows
𝑐 ∗= 𝑎𝑟𝑔𝑚𝑖𝑛𝑃 (𝐶 = 𝑐|𝑆 = 𝑠), 𝑐 = 1, … , 𝑘 (3)
⎛ 1 𝑑1,2 ⋯ 𝑑1,𝑛 ⎞
Following the definition of the graph network described above, ⎜ ⎟
𝑎 1 ⋯ 𝑑2,𝑛 ⎟
we can define an objective function that optimizes the assignment 𝑅(𝐸)𝑑 = ⎜ 2,1 (7)
⎜ ⋮ ⋮ ⋱ ⋮ ⎟
problem between controllers and switches based on several factors. ⎜𝑑 𝑑𝑚,2 ⋯ 1 ⎟⎠
⎝ 𝑚,1
The number of controllers, latency within a cluster, density of vehi-
cles, location of S-RSUs, and the connectivity degree between S-RSUs, Since we are dealing with a connected graph representation of the
among others. For example, The location of S-RSUs can be prioritized network, each switch can reach another switch in the graph through
based on intersection or hotspots areas. Undoubtedly, as we increase multiple directions. Thus, we assign the shortest path value between
the number of deployed controllers, we also impact each controller’s any two nodes to be the relationship delay link between them. That is,
load. Thus, making the relationship between the number of controllers the link between two switches 𝑖, 𝑗 ∈ 𝑆 are described as 𝑙𝑖𝑗 ∈ 𝐿.
and objective factors interchanging. If we include more controllers in When we consider vehicles as part of the graph design, they will be
the network, we risk having excessive migrations of vehicles between represented by their links to the nearest roadside unit it is connected
control units much more often, which leads to higher disruption in with at time interval ▵ 𝑇 .
services. { }
𝑑(𝑣, 𝑠), attached 𝑣 ∈ 𝑉 𝑒ℎ 𝑠 ∈ 𝑆
𝑉 𝑒ℎ𝑅𝑣,𝑠 (𝑡) = (8)
3.1.1. Link connectivity measurement 0, otherwise
Two main types of links are available between switches 𝑠 ∈ 𝑆,
namely wireless links, and wired links. The links between vehicles and 3.1.2. Load measurement
infrastructure units are wireless 𝑤𝑖 and communication between units The second measurement that need to considered when we consider
can be either wired or wireless. In order to evaluate the link delay the objective function to find the best number of controllers, their
between them, we can describe it as the total path delay between two location, and associated set of switches, is the load on each node. The
nodes. That delay reflects the time it takes one packet from switch 𝑠𝑖 load on controllers can be defined as the number of switches assigned
to reach switch 𝑠𝑗 as follows to each control unit, and the used bandwidth by connected vehicles to
each of these switches. At any time 𝑡, a group of vehicles are connected
𝑑(𝑖, 𝑗) = 𝑝𝑟𝑜𝑝𝑑 + 𝑡𝑟𝑎𝑛𝑠𝑑 + 𝑝𝑟𝑜𝑠𝑑 + 𝛽 (4)
to the nearest switch-enabled RSU in which they are either using certain
which includes the propagation delay 𝑝𝑟𝑜𝑝, transmission delay 𝑡𝑟𝑎𝑛𝑠, communication bandwidth with the connected switch. Hence, we can
processing delay 𝑝𝑟𝑜𝑠, and a delay error factor 𝛽. measure the amount of load each switch is carrying based on the

90
N. Aljeri and A. Boukerche Computer Communications 182 (2022) 88–97

number of connected vehicles 𝑀 and the used bandwidth at time-step Algorithm 1 Incremental K-means clustering method
▵ 𝑡.
Require: 𝑘,𝐾𝑚𝑎𝑥, 𝐺(𝑆, 𝐿),𝑝𝑎𝑡ℎ, 𝜆, 𝑐𝐿𝑖𝑠𝑡
1 ∑ 1: 𝑘 = 1
𝑙𝑜𝑎𝑑(𝑡) = 𝑏𝑎𝑛𝑑𝑤𝑖𝑑𝑡ℎ𝑗 ∀𝑗 ∈ 𝑉 𝑒ℎ𝑠 (𝑖) (9)
𝑁 2: Initialize 𝑠𝐿𝑖𝑠𝑡, 𝑐𝐿𝑖𝑠𝑡, 𝑜𝑀𝑒𝑎𝑛
where 𝑉 𝑒ℎ𝑠 (𝑖) represent the set of vehicles connected to switch 𝑠(𝑖) and 3: Get average delays from each switch to all other switches
𝑏𝑎𝑛𝑑𝑤𝑖𝑑𝑡ℎ𝑗 represents the total used bandwidth by each vehicle 𝑗 at 4: 𝑐𝐿𝑖𝑠[0] ← 𝑚𝑒𝑑𝑖𝑎𝑛(𝑝𝑎𝑡ℎ)
time-step ▵ 𝑡. 5: while True do
6: while 𝐼𝑡𝑒𝑟 < 𝑀𝐴𝑋 or 𝑐𝑜𝑛𝑣𝑒𝑟𝑔𝑒 do
3.2. Incremental k-mean method 7: 𝑠𝐿𝑖𝑠𝑡 ← ∅ ⊳ Re-assign switches to nearest controller
8: for each 𝑠 ∈ 𝑆 do
In this step, the partition of switch-enabled RSUs into 𝐾 number 9: 𝑛𝑝𝑎𝑡ℎ𝑠 ← 𝑝𝑎𝑡ℎ(𝑠, 𝑐), 𝑐 ∈ 𝑐𝐿𝑖𝑠𝑡
of clusters representing the controllers regions while minimizing the 10: 𝑛𝑒𝑎𝑟 = 𝑚𝑖𝑛(𝑛𝑝𝑎𝑡ℎ𝑠).𝑖𝑛𝑑𝑒𝑥 ⊳ Minimum delay path
average latency among different controllers’ regions and maintaining a 11: 𝑠𝐿𝑖𝑠𝑡[𝑛𝑒𝑎𝑟].𝑎𝑑𝑑(𝑠)
relatively balanced load among them. The objective function that we 12: for 𝑐𝑙𝑎 ∈ 𝑠𝐿𝑖𝑠𝑡 do ⊳ update controllers locations 𝑐𝐿𝑖𝑠𝑡
want to reach is as follows 13: for 𝑏𝑠 ∈ 𝑐𝑙𝑎 do
∑∑ ∑
14: 𝑎𝑣𝑔𝑆𝑏𝑠 = 𝑚𝑒𝑎𝑛 𝑝𝑎𝑡ℎ(𝑠𝑏𝑠 , 𝑠𝑗 ), 𝑗 ∈ 𝑐𝑙𝑎, 𝑏𝑠 ≠ 𝑗
𝐽 = 𝑎𝑟𝑔 min(1∕𝑁) 𝛼𝑘𝑖 𝑑(𝑐𝑘 , 𝑠𝑖 )𝜎𝑘𝑖 (10)
15: 𝑐𝐿𝑖𝑠𝑡(𝑐𝑙𝑎).𝑢𝑝𝑑𝑎𝑡𝑒𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝑙𝑒𝑟(𝑚𝑖𝑛(𝑎𝑣𝑔𝑆).𝐼𝑑)
where 𝑑 defines the shortest path delay between switch 𝑠𝑖 and con- 16: 𝑐𝐿𝑖𝑠𝑡(𝑐𝑙𝑎).𝑢𝑝𝑑𝑎𝑡𝑒𝐷𝑒𝑙𝑎𝑦(𝑚𝑖𝑛(𝑎𝑣𝑔𝑆).𝑣𝑎𝑙𝑢𝑒)
troller 𝑐𝑘 , and 𝜎 is the relationship coefficient 17: 𝑓 𝑢𝑟𝑡ℎ𝑒𝑠𝑡 ← 𝑚𝑎𝑥(𝑎𝑣𝑔𝑆)
{ } 18: end for
1, for i switch connected to controller j
𝜎𝑖𝑗 = (11) 19: end for
0, otherwise
20: if old centroids list equals new centroids list then
Given that only switch-enabled roadside units can be controllers, 21: break
we devise an updated k-mean clustering process. The standard 𝑘-mean 22: end if
clustering process consists of four stages. Begin with initializing the 23: end while
centroids points 𝑐1 , 𝑐2 , … , 𝑐𝑘 to be one of the provided switches set 24: Find the Sum of path delays ∀𝑐 ∈ 𝑐𝐿𝑖𝑠𝑡
points 𝑆, which can be chosen randomly or in a greedy approach. 25: 𝑛𝑀𝑒𝑎𝑛 = 𝑔𝑒𝑡𝑀𝑒𝑎𝑛(𝑐𝐿𝑖𝑠𝑡, 𝑠𝐿𝑖𝑠𝑡)
Then measure the objective function between the selected centroids and 26: if 𝑘 >= 𝑚𝐾 or 𝑑𝑖𝑓 𝑓 (𝑛𝑀𝑒𝑎𝑛, 𝑜𝑀𝑒𝑎𝑛) < 𝜆 then
every other switch 𝑠. After this, for each 𝑠𝑖𝑛𝑆, attach it to the centroid 27: return
cluster that is nearest to it. 28: end if
29: 𝑜𝑀𝑒𝑎𝑛 = 𝑛𝑀𝑒𝑎𝑛
1 ∑∑
𝑆 𝐾
𝛼 𝑡 = 𝑎𝑟𝑔 min 𝛼𝑘𝑖𝑑(𝑐𝑘𝑡−1 , 𝑠𝑖 ) (12) 30: 𝐶𝑛𝑒𝑤 ← 𝑚𝑎𝑥(𝑓 𝑢𝑟𝑡ℎ𝑒𝑠𝑡).𝐼𝑑
𝑁 𝑖=1 𝑘=1
31: add 𝐶𝑛𝑒𝑤 to list of 𝑐𝐿𝑖𝑠𝑡
Then, using the collection of RSUs in each cluster 𝐶𝑗 , we recalculate 32: 𝑘=𝑘+1
each cluster’s centroid 𝑐𝑗 such that the total objective function’s to 33: end while
the new centroid 𝑐𝑗 is minimize. Given that RSUs have fixed locations
and links among each other, the new centroid unit must belong to the
switches set. As a result, the average amount of path latency from each Algorithm 2 Adaptive incremental clustering method
connection point 𝑠 in 𝐶𝑗 to its neighboring switches in the same cluster Require: 𝑘,𝐾𝑚𝑎𝑥, 𝐺(𝑆, 𝐸),𝑝𝑎𝑡ℎ, 𝜆, 𝑐𝐿𝑖𝑠𝑡
is computed. 1: if 𝑐𝐿𝑖𝑠𝑡 == ∅ then
1 ∑ 2: call 𝐴𝑙𝑔𝑜𝑟𝑖𝑡ℎ𝑚(1)
𝐴𝑠 = 𝑑(𝑠, 𝑥) (13) 3: return
𝑁𝑗 𝑥∈𝐶
𝑗
4: else
𝑐𝑗 (𝑛𝑒𝑤) = 𝑎𝑟𝑔 min(𝐴𝑠 ∈ 𝐶𝑗 ) (14) 5: keep previous centroids list
6: goto: 𝐴𝑙𝑔𝑜𝑟𝑖𝑡ℎ𝑚(1) line: 5
where 𝑁𝑗 represents the total number of switches in a control group 𝐶𝑗 7: end if
with 𝑗 = 1, … , 𝐾. After certain number of iterations are made or the gap
between the current chosen centroid and the previous one has crossed
a certain threshold, a final set of clusters and their respective centroids
is derived. Also, it is noteworthy that the value of the updated objective overall total delays in each cluster. The method of assigning switches
function will never increase and it will stay the same in a worst case to clusters is replicated before convergence is achieved. Henceforth, a
scenario. new cluster centroid is located from the pool of switches to be the one
The aforementioned method generates 3-tuple sets, including 𝐶 with the most prolonged (worst) number of delays to its neighboring
clusters, 𝐾 control units, and 𝑆 switches within each cluster 𝑐 ∈ 𝐶. points within the same cluster. However, this process is terminated if
To find the optimum K value, the classical 𝑘-mean approach in- the gap between the current and previous average delay of centroids is
volves a repeated number of executions and randomly chosen initial less than a certain value. Otherwise, we recalculate the set of points in
centroids. To address this problem, we adjust the approach so that each cluster and the centroids associated with them.
clusters are defined incrementally before they achieve the optimum 𝐾
value. The mechanism of the proposed incremental k-mean clustering 3.3. Traffic flow prediction
approach is presented in Algorithm 1. First, we choose a median switch-
enabled roadside unit 𝑠 as the centroid of the first 𝑘 cluster. In this case, Using the aforementioned clustering techniques, the 3-tuple sets
the median switch is calculated based on the prior knowledge of the are generated to describe the most appropriate topology for a given
network topology structure to evaluate the closeness of each switch to network at time 𝑡. However, in a vehicular network environment, which
all others. Then, using the basic k-mean rule, add switches to the closest is highly dynamic, we need to re-evaluate the generated topology based
cluster centroid and recalculate the centroid point for each cluster. on real-time measurement of the network’s conditions. That includes
Following that, the cluster centroid is re-evaluated in proportion to the the traffic flow measurement on each S-RSU region that impacts the

91
N. Aljeri and A. Boukerche Computer Communications 182 (2022) 88–97

Fig. 1. Vehicles’ arrival to RSU zone.

overall delay and load on the controllers. To do so, using the current
network’s state and measurement will not help us avoid future services
interruptions and heavy loads; thus, a prediction mechanism is needed Fig. 2. Cologne city environment and mobility.
to estimate future traffic conditions and measurement and adjust the
topology network ahead of time.
The established initial clustering algorithm can be used as a base
topology setup that can be further expanded or reduced depending on
the estimated measurement of the network conditions. The estimation
of traffic flow can be done through various prediction methods based
on historical data collections such as neural network models, Markov
models, or even a simple moving average method.

3.3.1. Dataset processing

In order to measure the traffic flow density on each S-RSU region,
we need to first understand the vehicles mobility model and charac-
teristics over the network. Assuming vehicles arrive at the edge of an
RSU’s communication range over time 𝑡. Plotting each vehicle arrival as
ℎ1 , ℎ2 , .. over equal time intervals 𝑡1 , 𝑡2 , .., then the number of vehicles
arriving in each time slot is integer such that 2,1,2, and 2 vehicles
arrived in time slot 𝑡1 , 𝑡2 , .., as illustrated in Fig. 1. Considering a Poisson
distribution, the probability density function is as follows:
𝜇 𝑥 𝑒−𝜇
𝑝(𝑥) = (15)
𝑥!
where 𝑝(𝑥) is the probability of event 𝑥 occurring in the time interval,
and 𝜇 is the expected rate of occurrence of event 𝑥 in that time interval.
Fig. 3. Ottawa city environment and mobility.
When we define the arrival rate of vehicles in a roadside unit’s
(RSUs) communication range as the number of vehicles’ arriving to the
boarder of the RSUs’ communication per unit of time. Then, we can
measure the inter-arrival time (i.e., headway) of an RSU region by the 3.4. Topology change trigger method
time between each arrival into the region and the next.
Since the vehicles’ arrival is usually a random process, in which Each S-RSU can track and estimate the traffic flow within its com-
vehicles either arrive together in a multi-lane scenario, or they arrive munication range and use predictive traffic flow measurements to
sparsely. Then, vehicles’ arrival can be modeled statistically either notify the controllers ahead of time and trigger a new topology cluster-
by the time interval between two successive arrivals or the number ing method. The topology change is triggered by the super controller,
of vehicles arriving over a time interval. Because of the stochastic which evaluates the collected measurements from each local control
nature of vehicles’ arrival rate, we can model the inter-arrival time unit, which in return collects measurements from its group of switches.
through mathematical distribution. Assuming vehicles are arriving at At the time of trigger, the super controller calls a new clustering
time 𝑡1 , 𝑡2 , .., then the difference between two consecutive time is process, as described in Algorithm 1, and generates a list 𝑐ℎ𝑎𝑛𝑔𝑒𝐿𝑖𝑠𝑡
referred to as headway and is represented by two main elements, the that contains the new assigned controllers/switches. The method of
occupancy time (i.e., duration of passing) and the time gap between processing the mechanism of switch-migration is currently not the
two vehicles (i.e., ℎ1 = 𝑡2 − 𝑡1 , ℎ2 = 𝑡3 − 𝑡2 , …). Although the headways scope of this article.
are not consistent, yet they follow some random distribution and over
different regions, different distributions may exhibit certain arrival
4. Performance evaluation
pattern.
In similar context, the vehicles’ mobility characteristics is an impor-
tant factor in the selection of which S-RSU the vehicle is connected at The performance evaluation of the proposed scheme is divided into
time 𝑡. The vehicles’ velocity, direction, and location are key attributes three main parts besides the simulation parameters settings. We first
to characterize the vehicles best suited connection. Using the vehicles’ evaluate the performance of variants of clustering techniques that will
direction and velocity can determine the most appropriate S-RSU for lay out the first deployment structure. The subsequent evaluation is
connection with reasonable residual time. However, this is not the focus based on the prediction techniques for traffic flow estimation, in which
of this work, as we are only interested in the average number of vehicles we utilize the time-series Long Short-Term Memory (LSTM) [29] pre-
that are connected to an RSU at time 𝑡 to represent the traffic flow diction model to derive an estimate. Finally, the evaluation of static and
density on each unit. dynamic deployment strategies is compared based on several metrics.

92
N. Aljeri and A. Boukerche Computer Communications 182 (2022) 88–97

Table 2
Simulation parameters.
Parameter Value
Mobility model Cologne dataset - Ottawa city
Number of vehicles 700 K–4 K
Vehicles’ speed 0–50 m/s–0–30 m/s
Simulation time 3 h
RSU coverage 1 km–500 m
Time window 5-10-30 min
Channel speed Wireless- 3 × 108 s and Wired- 2 × 108 s
Data rate Wireless DSRC (6Mbps), Wired Ethernet (100 Mbps)

matrix is derived containing the shortest path delay from each switch-
enabled RSUs to another (see Figs. 4 and 5). The chosen parameters are
Fig. 4. S-RSU connectivity graph (Cologne). presented in Table 2.
The original area of Cologne city covered is around 400 km2 with
24 h mobility dataset including 700 𝐾 vehicles trips. In our set of
experiments, we sampled the dataset between 6:00 AM to 9:00 AM and
down-town area of 10x10 km center. As for Ottawa city, we generate
three hours of mobility for a range between 100 to 7000 vehicles
based on the Krauss car-following model. The speed of vehicles ranges
between 0-30m∕s. The mobility of vehicles is simulated in both cities
in order to gather total traffic flow of vehicles over time-steps. The
connection or association of vehicles and switch-enabled RSUs is based
on the Always Best Connection (ABC) at each time-step. Therefore,
we can generate at a time-window, a global snapshot of the network’s
status and traffic flow in each area.
The simulation lasts three hours, in which we analyze the total link
duration between vehicles and a switch-enabled RSU area, as presented
in Figs. 6 and 7. Based on that, the results indicated an average link
Fig. 5. S-RSU connectivity graph (Ottawa). duration between 5 to 15min, which helps to determine the most ap-
propriate time window to re-evaluate the controller’s status and gather
new vehicles’ traffic flow information. The vehicles’ flow over each
4.1. Scenarios RSU, displayed in Fig. 8, shows that in the case of Cologne mobility,
the vehicles’ traffic density starts very low during the early morning
period and gradually increases in peak time when people are going to
The chosen evaluation scenarios are based on two different cities, work and school. The generated mobility of Ottawa city shows different
Ottawa downtown and Cologne city. The deployment of roadside units traffic flow distribution. It sharply starts with high traffic flow densities
is based on actual data deployment of Ottawa and Cologne city, as seen at the beginning of the simulation and fluctuates a bit in the middle,
in Figs. 2 and 3. As for the mobility of vehicles, we use SUMO [30] to then gradually reduces until all vehicles reach a stop. That indicates
generate Ottawa’s mobility and Cologne realistic mobility traces [31]. that the generated mobility of SUMO does not necessarily reflect actual
mobility behavior, in which vehicles can exhibit certain traffic flows
The knowledge and existence of a real snapshot of the network’s
at a specific time of the day (i.e., peak hours). However, there is
connectivity graph is important and since it is not provided, we assume always chaarees of unexpected anomalies, especially during citywide
that two static infrastructural units (i.e., switch-enabled roadside units) events, disasters, and traffic congestions. With this context, we can
are connected (at one-hop manner) if they are within each other use this knowledge to measure adaptive controllers’ deployment and
coverage range 𝑅. This helps us to generate the connectivity graph 𝐺. assignment strategies; thus, avoid overloaded controllers and service
Therefore, all switch-enabled RSUs are connected and an adjacency interruptions.

Fig. 6. Cologne’s Average vehicles’ residual time with each RSU region.

Fig. 7. Ottawa’s Average vehicles’ residual time with each RSU region.

93
N. Aljeri and A. Boukerche Computer Communications 182 (2022) 88–97

Fig. 8. Average traffic flow based on S-RSU deployment over the simulation time.

Fig. 9. Cologne’s convergence factor.

Fig. 10. Ottawa’s convergence factor.

4.2. Clustering evaluation the maximum delay is simplified to be the maximum hop limit (𝛾-hop)
that a switch-based RSU can have to reach the control unit.
The evaluation of the clustering techniques is two-fold. We op- In Figs. 9 and 10, we study the impact of the threshold and hop
timize the parameter selection for the proposed incremental k-mean selection as stopping criteria. The incremental clustering technique
clustering method and compare two stopping criterion clustering tech- will no longer add more control units to the set of clusters when the
niques regarding the number of controllers, average delay, and load on requirements are met.
controllers. In Fig. 9(a), we investigate the effect of selecting a Mean Square
Error (MSE) threshold on the clustering approach in an offline setting
4.2.1. Parameter optimization and without knowledge of the vehicles’ density. As expected, increasing
To assess the new adaptive controller deployment and clustering the threshold value in the TD-based clustering results in an increase
technique, we must first optimize the parameters of the clustering in the number of controllers. Additionally, when there are fewer con-
method to find the optimal number of controllers and their co-located trollers in the graph, a significant delay is observed. This is because
switches. Based on the stopping criterion, two ways are considered: each controller is handling a greater number of switch-enabled RSUs.
threshold-based (T) and maximum hop-based (M). As previously men- Choosing the appropriate threshold value is simply a function of the
tioned, calculating the optimum number of controllers is an NP problem intended outcome. In our example, we believe that a threshold value
that may be solved by defining a stopping criterion. The threshold- of 0.08 to 0.15 produced a rather excellent result to the number of
based criterion is identified as the Mean Squared Error (MSE) similarity control unit and RSU load on each controller. In the case of MD-based
rate between two consecutive sets of clusters. The clustering process is clustering, as illustrated in Fig. 9(b), the higher the controller’s latency
terminated when the error gap value approaches 𝜆. The incremental limit, the fewer control units are required. However, this comes at the
clustering is terminated when no controller has an average delay more cost of adding delay and load to the control units. As for Ottawa city’s
significant than 𝛾 in terms of the maximal hop criterion. In this article, scenario, as presented in Fig. 10, similar behavior is observed, in which

94
N. Aljeri and A. Boukerche Computer Communications 182 (2022) 88–97

the higher the MSE threshold is set, the higher the delay is with lower
numbers of control units.

4.3. Prediction evaluation

In this set of experiments, the estimation of traffic flow measure-

ments is presented using the LSTM time-series prediction technique.
The first step towards time-series prediction is preprocessing
datasets based on a selected window and gap to produce supervised
time-series sequences. Choosing Ottawa city’s dataset as a test scenario,
we split the data into training and testing based on (80%–20%) divi-
sion. The collected data is based on each S-RSU traffic flow information
gathered using the generated SUMO mobility traces and the known
location and communication range of each S-RSU.
The gap between each data instance is 5 min, and the time-series
sequence window is selected to be 3, 4, and 5. The reason behind
choosing a 5-minute gap is that we have seen that vehicles reside within
the S-RSU’s communication range between 5-15 min. Thus, the traffic
flow of each S-RSU will remain the same during this period. As for the
sequence window, which indicates the number of data input required
by the predictor to estimate future lookups, it is selected based on the
elapsed time while collecting the data sequence. Hence, for window
3, the predictor will have three values, each reflects 5 min period.
Accordingly, the predictor will be using between 15 to 25 min of traffic
flow data inputs before producing an estimation.
Fig. 11 demonstrates the root mean squared error (RMSE) of the
LSTM predictor for sequences 3, 4, and 5, respectively. The LSTM
configuration used in this experiment is adam optimizer and Relu
activation function, with 20 hidden layers and 500 epochs. We used
Keras neural network library to test and evaluate the prediction of
traffic flow. During the training phase, the LSTM model is given the
traffic flow sequences from each selected S-RSU for training and then
tests the model on unseen dataset sequences. The result shown in
Figs. 11(a) to 11(c) illustrates the change in RMSE over each addition
of sequence up to the blue box border. The blue box indicates the
unseen dataset evaluation. The resulted average RMSE in each window
scenario is shown in Table 3.

4.4. Deployment evaluation

In this set of experiments, we compare the static and adapt cluster-

ing techniques in terms of communication delay, load on controllers,
and migration costs.
The static and adaptive clustering evaluation is based on two main
comparisons depending on the stopping factor, MSE threshold (static-
T) and Maximum 2-hop (static-M). Also, another variant is added
to consider the load balancing metric in the clustering technique:
delay/load-based clustering (DL) and maximum hop/load-based clus- Fig. 11. LSTM time-series window selection evaluation.
tering (ML).
Fig. 12 shows the traffic flow and end-to-end delay metric for all
clustering variants under Ottawa city mobility. The results show that
result shows that the number of controllers is adapting to the estimated
the static threshold-based clustering produced the worse flow and delay
traffic flow information, in which it increases during peak times and
over simulation time. The reason is that the threshold-based technique
decreases in others. Moreover, in Fig. 13(a), we notice that the number
only considers one aspect to cluster controllers and switches: the link
of controllers starts with nine control units and drops after minute 50,
delay between switch-enabled RSUs. Also, because of the static nature
which shows that the adaptive clustering technique manages to reduce
of the clustering method, it does not consider the vehicles’ traffic flow
the needed number of controllers at that time based on the low traffic
information in different time-steps and thus highly impacted to any
flow estimation.
change in the number of vehicles density in terms of load and delay.
As for the adaptive clustering methods, all of them showed a reduction
in average delay and traffic load on control units. Moreover, when 5. Conclusions
including the load balancing metric to the objective function of the
clustering technique, the results showed more stable load on controllers In this article, we proposed a proactive controllers deployment and
during all time-steps. assignment protocol based on traffic flow estimation for 5G-enabled
Fig. 13 plots the change in the number of deployed controllers over software-defined vehicular networks. The proposed technique utilizes
time-steps for DL and ML-based clustering, which adaptively re-cluster vehicles’ mobility and traffic flow information to dynamically adapt the
switches using traffic flow information and load-balancing metrics. The network topology and control-to-switch assignment to provide efficient

95
N. Aljeri and A. Boukerche Computer Communications 182 (2022) 88–97

Table 3
Comparison of RMSE in both training and testing phases.
Input 3 4 5
Output 3 2 1 3 2 1 3 2 1
Train 28.131 23.129 18.004 31.845 25.933 17.519 34.216 28.059 21.587
Test 28.473 21.742 18.705 34.146 26.311 27.570 35.043 19.754 26.844

Fig. 12. Comparison between static and adaptive deployment strategies.

Fig. 13. Change in number of controllers over Ottawa mobility.

network performance and balanced load distribution on controllers. Acknowledgments

The performance evaluation was done using two different mobility
scenarios of Ottawa city and Cologne datasets. We first investigated the This work was partially supported by NSERC, Canada Research
mobility characteristics in each scenario and reported on their traffic
Chairs Program and NSERC-CREATE TRANSIT.
flow pattern. Moreover, the optimization of the modified incremental
k-mean algorithm showed the best selection of a threshold value using
empirical analysis of a various range of possibilities. Additionally, References
we implemented a centralized LSTM time-series prediction model for
training and testing the traffic flow prediction. The evaluation of the [1] N. Aljeri, A. Boukerche, Mobility management in 5G-enabled vehicular networks:
prediction model is based on selecting the data instants’ gap and Models, protocols, and classification, ACM Comput. Surv. 53 (5) (2020).
window. Finally, a comparison between the static deployment of con- [2] X. Cui, X. Gao, Y. Ma, W. Wang, A controller deployment scheme in 5G-IoT
network based on SDN, EURASIP J. Wireless Commun. Networking 2020 (1)
trollers and the adaptive method is evaluated. The results indicate that
(2020) 248, http://dx.doi.org/10.1186/s13638-020-01853-8.
the proposed technique is successful in minimizing the communication
[3] R. Yu, J. Ding, X. Huang, M.-T. Zhou, S. Gjessing, Y. Zhang, Optimal resource
delay between vehicles and control units while maintaining an average sharing in 5G-enabled vehicular networks: A matrix game approach, IEEE Trans.
balanced load among controllers. Veh. Technol. 65 (10) (2016) 7844–7856.
[4] Q. Su, K. Jaffres-Runser, G. Jakllari, C. Poulliat, An efficient content delivery
CRediT authorship contribution statement infrastructure leveraging the public transportation network, in: MSWiM ’16,
2016, 338–347.
Noura Aljeri: Conceptualization, Methodology, Writing and edit- [5] N. AlJeri, A. Boukerche, An efficient movement-based handover prediction
ing. Azzedine Boukerche: Conceptualization, Methodology, Writing scheme for hierarchical mobile IPv6 in VANETs, in: Proceedings of the 15th
ACM International Symposium on Performance Evaluation of Wireless Ad Hoc,
and editing.
Sensor, & Ubiquitous Networks, in: PE-WASUN1́8, ACM, 2018, pp. 47–54.
[6] I.F. Akyildiz, S. Nie, S.-C. Lin, M. Chandrasekaran, 5G roadmap: 10 key enabling
Declaration of competing interest
technologies, Comput. Netw. 106 (2016) 17–48.
[7] N. Aljeri, A. Boukerche, An adaptive traffic-flow based controller deployment
The authors declare that they have no known competing finan- scheme for software-defined vehicular networks, in: Proceedings of the 23rd
cial interests or personal relationships that could have appeared to International ACM Conference on Modeling, Analysis and Simulation of Wireless
influence the work reported in this paper. and Mobile Systems, MSWiM ’20, 2020, pp. 191–198.

96
N. Aljeri and A. Boukerche Computer Communications 182 (2022) 88–97

[8] ONF, Software-defined networking: The new norm for networks, 2012, [30] M. Behrisch, L. Bieker, J. Erdmann, D. Krajzewicz, SUMO - Simulation of urban
URL https://www.opennetworking.org/images/stories/downloads/sdn- mobility: an overview, in: The 3rd Intl. Conf. on Advances in System Simulation,
resources/white-papers/wp-sdn-newnorm.pdf. 2011, pp. 63–68.
[9] X. Ge, Z. Li, S. Li, 5G software defined vehicular networks, IEEE Commun. Mag. [31] S. Uppoor, O. Trullols-Cruces, M. Fiore, J.M. Barcelo-Ordinas, Generation and
55 (7) (2017) 87–93. analysis of a large-scale urban vehicular mobility dataset, IEEE Trans. Mob.
[10] A. Boukerche, N. Aljeri, Design guidelines for topology management in software- Comput. 13 (5) (2013) 1061–1075.
defined vehicular networks, IEEE Netw. 35 (2) (2021) 120–126, http://dx.doi.
org/10.1109/MNET.011.2000369.
[11] X. Ge, Z. Li, S. Li, 5G software defined vehicular networks, IEEE Commun. Mag. Noura Aljeri [Member, IEEE] received her Ph.D. in Com-
55 (7) (2017) 87–93. puter Science from the University of Ottawa, Canada. She
[12] J. Qiao, et al., Improving video streaming quality in 5G enabled vehicular is currently a Research Associate at PARADISE Research
networks, IEEE Wirel. Commun. 25 (2) (2018) 133–139. Laboratory, University of Ottawa, Canada, and a Faculty
[13] S. Correia, A. Boukerche, R.I. Meneguette, An architecture for hierarchical Member in the computer science department at Gulf Uni-
software-defined vehicular networks, IEEE Commun. Mag. 55 (7) (2017) 80–86. versity for Science and Technology (GUST), Kuwait. She is
[14] S. Correia, A. Boukerche, Toward a scalable software-defined vehicular network, the recipient of the most outstanding Ph.D. thesis Award
in: GLOBECOM 2017, IEEE, 2017, pp. 1–6. at the University of Ottawa (2020) and the Pierre Laberge
[15] K.S. Atwal, A. Guleria, M. Bassiouni, SDN-based Mobility management and Award (2021) for her research contribution on mobility
QoS support for vehicular ad-hoc networks, in: 2018 (ICNC), IEEE, 2018, pp. management for autonomous and connected vehicular net-
659–664. works. Her current research interest includes smart mobility,
[16] B.P.R. Killi, S.V. Rao, Towards improving resilience of controller placement with topology management, and prediction models for connected
minimum backup capacity in software defined networks, Comput. Netw. 149 and autonomous vehicular networks, intelligent transporta-
(2019) 102–114. tion systems, and mobility networks. She has authored and
[17] Y. Zhang, L. Cui, W. Wang, Y. Zhang, A survey on software defined networking coauthored extensively in those areas and received the best
with multiple controllers, J. Netw. Comput. Appl. 103 (2018) 101–118. paper award in the 16th IEEE AICCSA Conference (2019).
[18] G. Wang, Y. Zhao, J. Huang, W. Wang, The controller placement problem in She served as a technical program committee member for
software defined networking: A survey, IEEE Netw. 31 (5) (2017) 21–27. several ACM and IEEE conferences, including MSWIM, PE-
[19] A. Mahmood, W.E. Zhang, Q.Z. Sheng, Software-defined heterogeneous vehicular WASUN, GLOBECOM, and ICC. She is also on the Editorial
networking: The architectural design and open challenges, Future Internet 11 (3) Board of ACM ICPS.
(2019) 70.
[20] J. Liu, et al., A scalable and quick-response software defined vehicular network
assisted by mobile edge computing, IEEE Commun. Mag. 55 (7) (2017) 94–100. Azzedine Boukerche [FIEEE, FEiC, FCAE, FAAAS, FAAIA]
[21] A. Jalili, M. Keshtgari, R. Akbari, R. Javidan, Multi criteria analysis of controller is a Distinguished University Professor and holds a Senior
placement problem in software defined networks, Comput. Commun. 133 (2019) Canada Research Chair Tier-1 position with the University
115–128. of Ottawa. He is the founding director of the PARADISE
[22] B. Heller, R. Sherwood, N. McKeown, The controller placement problem, in: Research Laboratory and the DIVA Strategic Research Cen-
Proceedings of the First Workshop on Hot Topics in Software Defined Networks, ter, and NSERC-CREATE TRANSIT at the University of
ACM, 2012, pp. 7–12. Ottawa. He has received the C. Gotlieb Computer Medal
[23] J. Liao, et al., Density cluster based approach for controller placement problem Award, Ontario Distinguished Researcher Award, Premier of
in large-scale software defined networkings, Comput. Netw. 112 (2017) 24–35. Ontario Research Excellence Award, G. S. Glinski Award for
[24] Y. Xu, et al., Dynamic switch migration in distributed software-defined networks Excellence in Research, IEEE Computer Society Golden Core
to achieve controller load balance, IEEE J. Sel. Areas Commun. 37 (3) (2019) Award, IEEE CS-Meritorious Award, IEEE TCPP Leadership
515–529. Award, IEEE ComSoc ComSoft and IEEE ComSoc ASHN
[25] T. Wang, F. Liu, J. Guo, H. Xu, Dynamic SDN controller assignment in data Leadership and Contribution Award, and the University of
center networks: Stable matching with transfers, in: IEEE INFOCOM 2016-the Ottawa Award for Excellence in Research. He serves as
35th Annual IEEE International Conference on Computer Communications, IEEE, an Editor-in-Chief for ACM ICPS and Associate Editor for
2016, pp. 1–9. several IEEE transactions and ACM journals and is also a
[26] G. Wang, Y. Zhao, J. Huang, Y. Wu, An effective approach to controller Steering Committee Chair for several IEEE and ACM inter-
placement in software defined wide area networks, IEEE Trans. Netw. Serv. national conferences. His current research interests include
Manag. 15 (1) (2017) 344–355. sustainable sensor networks, autonomous and connected ve-
[27] K.S.K. Liyanage, M. Ma, P.H.J. Chong, Controller placement optimization in hicles, wireless networking and mobile computing, wireless
hierarchical distributed software defined vehicular networks, Comput. Netw. 135 multimedia, Qo Service provisioning, performance evalu-
(2018) 226–239. ation and modeling of large-scale distributed and mobile
[28] C. Reddy, B. Vinzamuri, A survey of partitional and hierarchical clustering systems, and large scale distributed and parallel discrete
algorithms, in: Data Clustering, Chapman and Hall/CRC, 2018, pp. 87–110. event simulation. He has published extensively in these
[29] N. Aljeri, A. Boukerche, A performance evaluation of time-series mobility areas and received several best research paper awards for
prediction for connected vehicular networks, in: Proceedings of the 16th ACM his work. He is a Fellow of IEEE, a Fellow of the Engineering
Symposium on QoS and Security for Wireless and Mobile Networks, Q2SWinet Institute of Canada, the Canadian Academy of Engineering,
’20, 2020, pp. 127–131. and the American Association for the Advancement of
Science.

97