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/TVT.2014.2323693, IEEE Transactions on Vehicular Technology
1

A Fuzzy Multi-metric QoS-balancing Gateway

Selection Algorithm in a Clustered VANET to LTE
Advanced Hybrid Cellular Network
Ghayet el mouna ZHIOUA†∗ , Nabil TABBANE∗ , Houda LABIOD† and Sami TABBANE∗
† InfRes - Telecom ParisTech - France
∗ MEDIATRON - Higher School of Communication of Tunis (Sup’Com) - Tunisia

Email: {zhioua, houda.labiod}@telecom-paristech.fr, {nabil.tabbane, sami.tabbane}@supcom.rnu.tn

Abstract—Intelligent Transportation Systems (ITS) attract estimate that by 2027 almost 100% of the vehicles will be
nowadays research community and the automotive industry, equipped with On Board Units (OBUs). OBUs are devices
aiming to provide not only more safety in the transportation that provide communications among neighboring vehicles, i.e
systems but also other high QoS based services and applications
for their customers. In this work, we propose a cooperative traffic Vehicle to Vehicle (V2V) or between vehicles and nearby
transmission algorithm in a joint VANET - LTE Advanced hybrid fixed equipments (called also Road Side Units (RSU)), i.e.
network architecture that elects a gateway to connect the source Vehicle to Infrastructure (V2I) communications [1]. Thus,
vehicle to the LTE Advanced infrastructure under the scope Intelligent Transportation Systems (ITS) attract not only re-
of Vehicle to Infrastructure (V2I) communications. The origi- search community but also the automotive industry. Recently,
nality of the proposed Fuzzy QoS-balancing Gateway Selection
(FQGwS) algorithm is the consideration of QoS traffic classes they focus their efforts to grow vehicular communication and
constraints for electing the gateway. Our algorithm is a multi- networking into maturity by moving it from research field into
criteria and QoS based scheme optimized by performing the real implementation, aiming to provide not only more safety in
fuzzy logic for making the decision over the appropriate gateway. the transportation systems but also other high quality of service
Criteria are related to the Cluster Head and gateway candidates (QoS) based services and applications for their customers.
Received Signal Strength, load and the vehicle to vehicle Link
Connectivity Duration. Simulation results demonstrate that our Obviously, VANET networks have to overcome some issues
algorithm presents better results than the deterministic scheme and challenges related to their specific characteristics, such as
for gateway selection. Moreover, results show the efficiency of the very dynamic network topology related to vehicles high
FQGwS algorithm as it adapts its gateway selection decision to velocity, to ensure acceptable Vehicle-to-Vehicle (V2V) and
cluster density and to relative velocity of source node. Vehicle-to-Infrastructure (V2I) communications. Most of the
solutions proposed to handle these issues are based on the
I. I NTRODUCTION creation of dynamic clusters to self-organize the vehicular
IEEE 802.11-based Vehicular Ad hoc (VANET) networks network where the dynamic clustering formation can be either
have been widespread due to their relevant attractive features in a decentralized or centralized way. Clustering is to group the
such as self-organization and the decentralized administration. nodes into homogeneous sets named clusters. Each cluster has
Although VANET networks are considered to be a subset of one Cluster Head (CH) elected from the cluster members’ that
MANETs networks, they have some advantages over these controls flows and signaling inside the cluster specially for V2I
later. Typically VANET nodes do not have battery limitations communications. Typically, the members of one cluster have
and benefit from more processing power and storage space. some common characteristics, e.g. near coordinates, velocities,
The great potential of this technology has been acknowledged same direction, etc.
with the establishment of ambitious research programs on ve- An initial set of services for use by V2I systems on
hicular communication systems worldwide, such as European European highways has been defined within the European
eSafety framework, numerous United States V2V and V2I Commission project on COOPerative systEms for intelligent
projects, and the Japanese Smartway and Advanced Safety Road Safety (COOPERS)[2]. There are safety-critical services,
Vehicle programs. Moreover, vehicular communication and such as accident warning, and roadwork information, and
networking present an active field of standardization activities convenience services, such as journey time and road charging
worldwide, such as ISO TC204, IEEE (802.11p and 1609.x) services. The focus of this paper is on defining an algorithm
and SAE DSRC in the USA, ETSI TC ITS and CEN WG278 aiming to select the appropriate gateway for a given source
in Europe and ARIB T-75 in Japan. In addition to these stan- vehicle under the scope of V2I communications and based
dardization efforts, considerable evolution inside the vehicles on the QoS class of the traffic to be transmitted to the
themselves is observed. In fact, future vehicles are expected infrastructure. In this paper, we focus not only on safety-
to be equipped with high efficiency computing systems and critical and convenience services delivered on highways via
multiple wireless communication interfaces. According the V2I architecture but also on other QoS classes delivering
ETSI 102 638 technical report, in 2017, 20% of the run- real time, streaming or best effort services. An efficient al-
ning vehicles will have communication capabilities and they gorithm based on a QoS-balancing scheme will be proposed.

0018-9545 (c) 2013 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/TVT.2014.2323693, IEEE Transactions on Vehicular Technology
2

In addition, the infrastructure in V2I communications can terminal problems may occur. The most common solution
be supported by various wireless technology interface, i.e. adopted for this problem is the clustering. Clustering in
IEEE 802.11p, 802.16m, LTE Advanced, etc. In this work, VANET aims at organizing vehicles into groups based on some
we consider an LTE Advanced interface for many reasons. In specific common characteristics. Using that technique can lead
fact, according to [3], it is more advantageous to use a cellular to more nodes’ coordination and fewer inter-nodes interfer-
infrastructure rather than WiFi based infrastructure, for V2I ence. The main challenge for a VANET clustering algorithm is
communications. On the other side, LTE Advanced standard to maintain cluster stability for the longest period otherwise the
is characterized by very relevant and advanced system features performance will be degraded due to the frequent re-clustering
as compared to other systems that meets the requirements of operations. Besides, clustering in VANETs requires selecting
IMT Advanced project for the 4th G of cellular systems [4]. a Cluster Head (CH) to be responsible for coordinating the
Moreover, LTE Advanced cellular network is a well designed members of the cluster. This process is carried out by each
system that offers high uplink (UL) and downlink (DL) data node belonging to the cluster by broadcasting its information
rates for high and low mobility environments [5]. In this paper, to all other neighboring nodes. Thus, after network information
we propose a cooperative traffic transmission algorithm in collection and based on a specific CH selection algorithm,
a joint LTE Advanced-VANET hybrid network architecture see section III, these nodes select a CH to coordinate the
where an elected gateway will connect a source vehicle to communication among them. After decision over the CH is
the LTE Advanced infrastructure. Some references that do completed and cluster nodes informed, the CH will be able to
not consider ad hoc nodes clustering, elect the gateway after communicate directly to all other cluster members and may
information broadcast and gathering from all nodes leading to act also as the relay node of communications to other cluster
have a high amount of overhead in the network. In a clustered members and other nodes in different clusters.
architecture, in spite of its role of initiating the communication Several clustering algorithms have been proposed in lit-
and controlling the flow of signaling messages within the erature. In [6], the cluster formation is based on direction
cluster, the cluster head could be also set as the gateway to of vehicles, where vehicles moving in the same direction
the infrastructure for source vehicles of its cluster. Obviously, belongs to the same cluster. Authors in reference [7] propose
centralizing the connectivity to the infrastructure for nodes a clustering algorithm for a heterogeneous network based
belonging to one cluster is beneficial. The key advantage on vehicular and UMTS cellular networks. The clustering is
is to decrease the cellular network resources consumption, based on three criteria: direction of movement, UMTS received
by multiplexing distinct source nodes’ flows by one gateway signal strength (RSS) and 802.11p wireless transmission range.
that handles to send them to the infrastructure. However, this Neighboring vehicles having the same direction of movement
scheme generates serious issues such as inducing CH overload, and an RSS higher than a specific threshold becomes be-
increasing end-to-end delay as compared to the direct link (i.e. longing to the same cluster. The drawback of this approach
send directly to the eNodeB) which is intolerable for delay is that one cluster could be composed of a high number
sensitive services. Moreover, the cluster head might not be the of nodes which generates a huge amount of overhead for
optimal gateway to the infrastructure as almost all algorithms cluster maintenance. Algorithm proposed in [8] is a clustering
consider only VANET layer features for CH election and protocol that does not use special control packets dedicated
forget reflection about the infrastructure layer features such to perform clustering but it builds the cluster and maintains
as base station load and received signal quality. Therefore, it it based on data traffic forwarding. This algorithm is suitable
is essential to propose a method that considers traffic priority for a dense network with high mobility because the cluster
and its QoS constraints and on the other side take aware of maintenance is dependent of the traffic. Moreover, it is not
both VANET and infrastructure features. affected by the increase of control overhead caused by frequent
The remainder of this paper is structured as follows. We changes of cluster members. However, this algorithm does not
first present, in Section II, an overview of VANET clustering consider relative velocity metric which causes a decrease of
algorithms. Then, in Section III, we present the most relevant the lifetime of the cluster. Clustering method proposed in [9]
related works of gateway selection algorithms in a clustered is build upon protocol of reference [8] and extends the lifetime
and non-clustered architecture. Sections V and VI describe of the network by balancing energy consumption among the
our system model and the fuzzy logic modeling method. We network nodes. Reference [10] proposes MOBIC algorithm
address also gateway selection challenge by presenting our that uses a special mobility metric for cluster formation phase
gateway selection algorithm based on fuzzy logic modeling. where each mobile node sends two consecutive messages
Finally, simulation results are presented and future works are to each of its neighbors to compute their relative speeds.
highlighted in Sections VII and VIII. Then, each mobile node broadcasts this information to its
neighbors. The drawback of this method is the need for
extra explicit message exchanges among mobile nodes for
II. VANET C LUSTERING
maintaining the cluster structure. Thus, with frequent network
As in a classical MANET network, VANET nodes have topology changes resulting to frequent clustering update, clus-
dynamic connectivity and self organizing features. However, ter maintenance overhead would increase drastically, consum-
with the increasing of the number of nodes, where each ing high portion of the bandwidth. In [11], authors propose
node handles its own decentralized routing and neighborhood a hierarchical clustering technique where cluster members
connectivity maintenance tasks, serious scalability and hidden are grouped into subsets of slaves nodes and cluster relay

0018-9545 (c) 2013 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/TVT.2014.2323693, IEEE Transactions on Vehicular Technology
3

nodes and a Cluster head that is in the top of the hierarchy. In clustered networks, clustering algorithms are used in
This algorithm generates a huge amount of overhead due the VANETs to ensure stability and increase link lifetime be-
clustering hierarchy maintenance. In [12], authors propose a tween vehicles belonging to the same cluster. There are basic
density based clustering algorithm that takes into account the clustering and CH election techniques that were proposed
effects of multi path fading. The cluster formation is based in literature such as highest degree [19] and lowest Id [20]
on the weight metric which takes into consideration the link algorithms which are not effectively efficient as they may
quality and the traffic conditions. A position based clustering generate frequent re-clustering. In [6], as cluster formation
technique is proposed in [13] where the cluster structure is is based on direction of vehicles, the first vehicle moving
determined by the geographic position of the nodes. The in that direction will be elected as CH. Using this method,
stability of the system is improved by electing the vehicles vehicles with high relative speed will generate frequent CH
having a longer trip as cluster heads. Despite it seems that reselection which causes additional overhead in the cluster.
this solution gives stable clusters, performances simulations In [21] authors propose to elect one CH in each segment of
and evaluation do not consider sparse and jammed traffic a road based on geographical information collection which
conditions which are very frequent in VANET environment. are provided by the infrastructure. This algorithm fails to
A similar approach is defined in [14]; where clusters are address cluster stability and cluster maintenance. In reference
formed based on vehicles position in the road. However, [22], the cluster head is elected based an additive metric
proposed algorithm is very limited as it does not address of three criteria: network connectivity level (based on the
the cluster maintenance and CH election challenge. Another maximum number of vehicles that are directly connected to
position based clustering algorithm that performs hierarchical considered vehicle and on the vehicles on the same traffic flow)
and geographical data collection and dissemination is proposed and average distance and velocity levels. Authors considers
in [15]. The cluster formation in this algorithm is based on urban scenarios characterized by several lanes and intersection
the division of the road into segments. Its performances are architecture, e.g. going straight through or turning left or right.
affected by the mandatory availability of an infrastructure. Gathering information over the roads’ topology and flows
Moreover, it generates high overheads for V2V and V2I intersections requires accurate positioning systems which is
communications. not always achieved. Moreover, the high dynamicity of the
vehicular nodes and the random drivers’ reactivity causes
III. R ELATED WORKS limits to this approach. In reference [10], authors elects the
In this paper, we address the gateway selection challenge cluster head based on relative velocity. The relative velocity
in a clustered VANET architecture. The gateway will relay is computed using a ratio of received signal strength of two
traffics of a source vehicle to the infrastructure under the scope successive HELLO messages. The CH is then the node with
of V2I communications. In a clustered VANET architecture, the lowest relative velocity variance. This is an interesting
the CH is the default gateway to the infrastructure for all approach for electing the CH, however, there is more accurate
source nodes belonging to its cluster. However, in a non- schemes for computing relative velocity than the HELLO
clustered architecture, the gateway is elected after information messages based approach. In [11], after handling hierarchical
broadcast and gathering from all mobile nodes in the network. clustering technique, authors propose to elect the CH as the
In this section, we present some algorithms that have been slave node that received three synchronous messages. This
proposed for gateway, respectively CH selection in a non- method does not consider vehicles’ movement dynamicity and
clustered, respectively clustered mobile ad hoc architecture. causes huge amount of overhead to the messages exchange for
For a non-clustered ad hoc networks, in reference [16], the hierarchical cluster maintenance. In [7], the cluster head
authors proposed an algorithm for gateway selection based on is designed as the vehicular node that is in the middle of the
choosing the mobile node with shortest hops from source node. cluster, at equal distance from the border nodes. A source
One single metric that combines physical hops and, virtual vehicle will then select one of the elected CH as gateway to
hops relative to congestion and contention levels, is used. This the infrastructure.
algorithm has got one limitation which is the use of the NAV To the best of our knowledge, there is no proposed algorithm
timer to compute virtual hops for contention measurements for gateway selection to infrastructure in a clustered VANET
which is not an easy parameter to evaluate concretely. In [17], architecture that considers traffic class priority. In this paper,
authors discuss the issues associated with the selection of mo- we propose a cooperative traffic transmission algorithm in a
bile gateways in an integrated MANET-UMTS heterogeneous joint VANET - LTE Advanced hybrid network architecture
network. They use simple additive weighting techniques to that elects a gateway to connect the VANET source vehicle
select an adequate gateway based on residual energy, UMTS to the LTE Advanced infrastructure under the scope of V2I
signal strength and mobility speed of the gateway candidates. communications. LTE Advanced standard might be widely
In [18], metrics used to select the gateway for interconnecting adapted by numerous operators as the next generation of
the MANET with the infrastructured network are remaining their cellular networks. We propose then a multi-criteria and
energy, mobility, and number of hops based on a simple QoS related attributes approach used to make a decision
additive weighting method. Node with the highest weight will of the appropriate gateway for source vehicle to the LTE
be selected as a gateway. These three metrics are not enough Advanced infrastructure. The originality of our algorithm is the
to select the optimal gateway as the link from the gateway to consideration of QoS traffic classes constraints for electing the
the infrastructure is not considered. gateway. The next section presents the adopted system model.

0018-9545 (c) 2013 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/TVT.2014.2323693, IEEE Transactions on Vehicular Technology
4

IV. S YSTEM M ODELING V2V : IEEE 802.11p LTE

V2I : LTE Advanced
Advanced
The system model is based on a hybrid network architecture CH : Cluster Head EPC
VN : Vehicular Node
that consists of two systems: an LTE Advanced infrastructure S1 LTE Advanced
and a VANET network. The architecture comprises IEEE Network
802.11p based VANET vehicles and LTE Advanced eNodeBs
interconnected through X2 interface and connected via S1
X2 X2
interface to core network, i.e. Evolved Packet Core (EPC),
see Fig.1. We assume that all vehicles are equipped with
OBUs that contain an IEEE 802.11p interface. Each vehicle
Source VN
uses its VANET interface to communicate with its neighboring
VANET
vehicles based on IEEE 802.11p standard. IEEE 802.11p
Network
inter-vehicle communications are possible even if vehicles are
inside eNodeB coverage. The OBU could also contain an CH
LTE interface to communicate directly with the infrastructure.
Simultaneous two interfaces communications do not cause
harmful interferences as they lie on distinct non-overlapping Figure 1. System Model
channels in two different spectrum regions.
We made the assumption that the VANET network is already
clustered and the cluster head (CH) is elected. Clustering of QGwS algorithm by performing the fuzzy logic because
and cluster head selection techniques are out of the scope of the inherent strength of fuzzy logic in solving problems
of this paper. The goal of this work is to make the decision exhibiting imprecision and the fact that many of the terms
of a source vehicle upon the gateway to select to connect used for describing radio signals are fuzzy in nature [24].
to the LTE Advanced infrastructure. The gateway could be The proposed algorithm is named Fuzzy QoS based Gateway
the cluster head of its cluster, another Gateway candidate Selection (FQGwS) algorithm. Fuzzy logic can be viewed
(GWC) vehicle or there may be also no gateway between as a theory for dealing with uncertainty about a complex
the source and the infrastructure where the source vehicle system, and as an approximation theory. The fuzzy logic
will be directly attached to the eNodeB, depending on traffic has two objectives: first, it develops computational methods
classes. Obviously, sending the traffic of each source vehicle that can perform reasoning and problem solving tasks that
directly to the infrastructure without defining a gateway is in require human intelligence, and second it explores an effective
theory the ideal scheme. However, nowadays cellular networks trade-off between precision and the cost in developing an
still suffer from the lack of resources in spite of numerous approximate model of a complex system [24]. Fuzzy inference
proposed solutions for the optimization of radio resources system is a computing framework based on the concepts of
allocation. Thus, decreasing cellular network resources con- fuzzy set theory, fuzzy if-then rules, and fuzzy reasoning. It
sumption, by aggregating distinct source nodes’ flows to one is known as the process that draws conclusions from a set of
default gateway, i.e. the CH, that handles to send them to facts using a collection of rules. In this work, we use Mamdani
the infrastructure seems to be an advantageous solution for a fuzzy inference system [25]. The Mamdani fuzzy system
clustered VANET architecture. However, despite it overcomes is composed of following modules: the fuzzifier, the fuzzy
radio resources lack dilemma, this scheme generates serious rules base, the fuzzy inference engine and the defuzzifier. We
issues like inducing CH overload and increasing end-to-end propose to implement FQGwS algorithm in the vehicles having
delay, which is intolerable for delay sensitive, real time and OBUs. The next section lists the input parameters of FQGwS
alerts services, etc. Therefore, the key challenge we address in algorithm.
this work is how to select for a source vehicle the appropriate
V2I gateway while considering at the same level the QoS A. Input parameters
traffic class, the CH features and the cellular network load.
The next section presents our proposed algorithm which is FQGwS algorithm input parameters are divided into two
optimized by the integration of the fuzzy logic. categories: attributes and criteria.
1) Attributes: In this paper, we consider only one attribute
which is the type of traffic to be transmitted by a source vehicle
V. F UZZY Q O S- BALANCING G ATEWAY S ELECTION to the infrastructure. we define three classes of traffic: Class 1,
D ECISION Class 2 and Class 3. As the source vehicle is a VANET node,
An efficient gateway selection algorithm have to take aware generated traffic types are governed by IEEE 802.11p standard.
of the QoS features of traffic to be transmitted to the in- On the other side, the infrastructure is an LTE Advanced
frastructure. We designed a QoS-balancing gateway selection cellular network. Thus, to achieve systems’ compatibility, a
algorithm (QGwS) in a VANET - LTE Advanced hybrid mapping between IEEE 802.11p access categories and LTE
network where the decision over the gateway depends on Advanced priority classes, using the QoS Class Indentifier
the traffic to be transmitted to the infrastructure [23]. It is (QCI) [26] is done. The mapping is performed based of QoS
an adaptive multi-criteria multi-attribute gateway selection constraints (packets average delay budget and packets loss) of
decision algorithm. In this work, we optimize gateway decision VANET and LTE Advanced standards. Traffic mapping is pre-

0018-9545 (c) 2013 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/TVT.2014.2323693, IEEE Transactions on Vehicular Technology
5

Class 1 Class 2 Class 3 Signal Strength (RSS) and the Load (L) in order to select
QoS Classes
(Voice) (Streaming) (Data) the best base station from this diversity set. It is based on
IEEE 802.11p a SAM to minimize the computational efforts that causes
BE and
Standard Voice AC Video AC delays and resources consumption. In fact, using only
BK ACs
Designation the RSS to select the best BS is not sufficient. The load
1-3-5-7 2 is also an important criterion for base station selection
4-6-8-9
LTE-Adv QCI Conversational, Streaming, in nowadays cellular networks, as important as RSS. It
Background
Signaling Interactive describes the resources occupation ratio of the eNodeB.
<100ms LCS X2I between X and the Infrastructure is computed using
Average < 50 ms (1), where X ∈ {Source, CH, GwC}.
< 150 ms < 300 ms
Delay Budget real-time  
gaming RSSi T hrLoad
LCSX2I = max + (1)
Table I i∈[1,n] T hrRSS Li
T RAFFIC MAPPING
where n is the number of eNodeBs in the diversity set, RSSi
is the Received Signal Strength of eN odeB i , Li is the load
of eN odeB i and T hrRSS (= 41.76 dBm) and T hrLoad are
sented in Table I. We define only three QoS category classes thresholds of respectively RSS and Load criterion.
because for first deployments, a majority of 4G operators will Therefore, source vehicle selects the best eNodeB from the
likely start with three basic service classes containing voice, diversity set, with the highest LCSS2I computed using (1), to
and best effort data classes. In the future, dedicated bearers ensure the best compromise between the RSS and the Load.
offering premium services such as high-quality conversational Without the loss of generality, for source vehicles that are
video can be introduced into the network. Moreover, for sake not equipped with an LTE interface, LCSS2I criteria will be
of simplicity, in this paper, we will call Class 1 traffic as Voice considered as very low.
traffic, Class 2 as Streaming and Class 3 as Data traffic. • CH to LTE Advanced Infrastructure Link Connectivity
Strength (CH2I LCS): is the LCS between the cluster
In addition to the delivery of information and entertainment, head and the selected eNodeB. In fact, the CH selects
i.e. infotainment, services, the role of typical V2I communi- the eNodeB with the highest LCSCH2I computed using
cations includes the provisioning of safety related, real-time (1), where X = CH.
and alerts services, such as speed limit information, traffic jam, • Load of the CH: it is the occupation ratio of the buffering
safe distance and accident warnings, etc [2]. All these services queue of the cluster head.
aim to prevent accidents by directly providing information to • Source Vehicle to CH Link Connectivity Duration
vehicles’ drivers. Moreover, ITS protocols are designed for the (Source2CH LCD): The LCDS2CH between the source
5.850- to 5.925-GHz band, divided into one central control vehicle and the CH is computed using (2).
channel (CCH) and six service channels (SCHs) where CCH
LCD criterion reflects well the stability of the link and its life-
is dedicated to the transmission of traffic safety messages,
time between two vehicles. In fact, considering only relative
whereas SCHs are dedicated to the transfer of various ap-
velocity is not sufficient to characterize link stability between
plication data. IEEE 802.11p standard allows the use of the
two mobile nodes as vehicles positions, moving directions and
four traffic classes for SCH and CCH channels with a MAC
VANET transmission range impact also on the connectivity
layer based on IEEE 802.11e enhanced distributed channel
duration of the V2V link. LCDij between vehicle i and
access (EDCA) [27]. We propose to classify traffic safety
vehicle j is computed using (2), inspired from [28]:
applications as real-time systems. However, communicating
real-time messages require a predictable system that is able p
to deliver the message before the deadline. Therefore, inside (α2 + γ 2 )R2 − (αδ − βγ)2 − (αβ + γδ)
LCDij = (2)
the class 1 traffics, safety and alerts services have priority α2 + γ 2
over other conversational voice or live streaming services. Our where, α = vi cosθi − vj cosθj , β = xi − xj , γ = vi sinθi −
work considers traffic priority class for choosing the adequate vj sinθj , δ = yi − yj . (xi , yi ), respectively (xj , yj ), is the
gateway to the infrastructure under V2I context. Cartesian coordinates of two neighboring vehicles i and j and
2) Criteria: Criteria are input parameters that are either that have an inclination of θi , respectively θj (0<θi , θj <2Π)
measured or received by the source vehicle from neighboring with respect to the x-axis and moving at vi , respectively vj ,
vehicles: speed. R is IEEE 802.11p wireless transmission range. For
• Source Vehicle to LTE Advanced Infrastructure Link LCDS2CH , we suppose that the CH and the source vehicle
Connectivity Strength (Source2I LCS): As source mo- are adjacent, otherwise LCDS2CH = min {LCDij } , 1 ≤
bile vehicle could detect one or several eNodeBs with i, j ≤ n, where n is the number of hops between the source
various loads, source vehicle build a diversity set con- and the CH.
taining detected BSs. We define the LCSS2I criterion, • Gateway candidate to LTE Advanced Infrastructure
computed using formula (1), which is a simple additive Link Connectivity Strength (GwC2I LCS): There could
metric (SAM) of two important metrics: the Received be another gateway candidate (GwC) for the source

0018-9545 (c) 2013 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/TVT.2014.2323693, IEEE Transactions on Vehicular Technology
6

vehicle with better LCS, load and LCD as compared 2) Decision phase : The second phase is the decision phase
to the CH. Therefore, each Ordinary Vehicle (OV) that where the gateway to the infrastructure have to be elected. The
experience high RSS (>RSSth ), low buffering queue load figure 2 illustrates the diagram of FQGwS proposed algorithm
(<Lth ) and high LCD (>LCDth ) becomes a potential that will make the decision upon the adequate gateway to
GwC (pGwC), see paragraphs 2 and 3 of section V-B2 select to attach the source vehicle to the infrastructure. In
for more details. The latter will elect and then keep a this phase, if Source2I LCS criterion is acceptable and
track of the best GwC, and uses it as an input for FQGwS the traffic type is Voice, the source vehicle will be attached
algorithm for making final gateway decision. directly to the best BS without defining a gateway to the
For GwC2I LCS criterion, the gateway candidate selects infrastructure. This decision is taken because Class 1 traffics
the eNodeB with the highest LCSGwC2I computed using are very delay sensitive traffics and not resources’ greedy.
(1). Source vehicle will check CH criteria for Class 1 traffic only if
• Load of the Gateway candidate: It is the occupation ratio Source2I LCS criterion is not acceptable. However, with the
of the buffering queue of the gateway candidate. remaining traffic classes, in spite of Source2I LCS criterion,
• Source Vehicle to gateway candidate Link Connectivity the source vehicle will always check the CH features.
Duration (Source2GwC LCD): It is the LCD between the
source vehicle and the gateway candidate. LCDS2GwC is 1) Cluster Head criteria check:
computed based on (2). With Class 2 and Class 3 traffics, source vehicle verifies
the metrics of the default gateway, i.e. the CH. In fact,
if the CH metrics values, i.e. CH2I LCS, CH load
B. Proposed algorithm
and Source2CH LCD, are acceptable, the decision
In this subsection, we present our proposed algorithm for over the gateway is immediately taken, without the
gateway selection, under the scope of V2I communications. solicitation of other gateway candidates. The gateway to
FQGwS algorithm is a decentralized scheme performed by a the infrastructure is then the CH. In fact, the CH has got
source vehicle that have to connect to the LTE Advanced in- acceptable features when the membership function of
frastructure. It is a QoS-balancing scheme as gateway selection each of the three criteria is equal to one, i.e. LCSCH2I >
is based on the traffic class to transmit to the infrastructure. LCSth , CH load <Lth and LCDS2CH >LCDth relative
FQGwS is performed in three phases: Data collection phase, to the CH threshold features of Table II.
Decision phase and Maintenance phase. 2) Gateway Candidates solicitation :
1) Data Collection phase : Criteria are provided using If the levels of the three criteria of the CH are not accept-
network measurements handled either by the source vehicle able (as compared to the specified threshold of TableII),
and/or participant vehicles (e.g. CH, OVs). Afterward, gateway source vehicle broadcasts a GATEWAY SOLICITATION
decision is performed. The first step is the data collection REQuest message through the cluster to take track of
phase, where dual interface source vehicle triggers measure- potential OVs that may become gateway candidates.
ments on its LTE Advanced and IEEE 802.11p network cards. In fact, each OV with acceptable criteria respond with
On the LTE Advanced interface, source vehicle measures the a GATEWAY SOLICITATION RESponse, containing
RSS of neighboring base stations and checks their loads. It its metrics features (GwC2I LCS, load, velocity and
builds then, the diversity set which contains all BSs it senses position (used to compute LCD by source vehicle)).
and sorts them based on the LCS metric computed using They become potential GwCs (pGwCs). In other terms,
(1). Afterward, source vehicle selects the eNodeB with the pGwCs are OVs with high LCSGwC2I (>LCSth ) and
highest Source2I LCS to ensure a compromise between RSS low load (<Lth ) relative to pGwC threshold features of
and Load. On the IEEE 802.11p interface, as we focus on a TableII. Afterward, source vehicle collects information
VANET clustered architecture, source vehicle is already aware over all pGwCs and to use it to take the decision
of the CH features, position and velocity. Thus, it deduces over the gateway candidate to select as a gateway to
Source2CH LCD. For remaining criteria, source vehicle the infrastructure based on an additive metric weights
requests the CH about CH2I LCS and load and sends it computing.
to the source vehicle. Note here that GATEWAY SOLICITATION REQuest
and GATEWAY SOLICITATION RESponse messages
QoS Classes Class 1 Class 2 Class 3 are a 6-bytes packets containing the node LCS (1 byte),
RSS th -90dBm -90dBm -90dBm load (1 byte), velocity (1 byte) and position (3 bytes). If
CH Lth 70% 65% 75% it is a request message, LCS, load, velocity and position
LCDth 50 s 80 s 80 s fields contains 0. If it is a response message, these fields
RSS th -120dBm -120dBm -120dBm contains the computed value of each criterion.
pGwC Lth 90% 85% 95% 3) Gateway Candidate Selection :
LCDth 30 s 60 s 60 s Based on criteria collection of all pGwCs, source vehicle
Table II
computes and affects a weight to each pGwC. Weights
FQG W S ALGORITHM PARAMETERS computing are performed using (3), which is a SAM of
the three criteria. Criteria are divided into two categories:
positive and negative. A positive criterion is a criterion

0018-9545 (c) 2013 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/TVT.2014.2323693, IEEE Transactions on Vehicular Technology
7

Start

Measurement of BSs’ RSS Repeat BSs Measurements

Built Diversity Set / sort Diversity Set

Select the best BS (SAM=RSS & Load)

LTE Advanced
Interface
Source2I LCS

Acceptable Not Acceptable

Traffic Class Traffic Class

Streaming Data Voice

Voice Streaming Data

CH2I LCS (**) CH2I LCS

CH2I LCS (*)

Acceptable Not Acceptable Acceptable Not Acceptable

Acceptable Not Acceptable

CH Load CH Load
CH Load
Acceptable Not Acceptable VANET
Acceptable Not Acceptable Acceptable Not Acceptable
Interface

Source2CH LCD 0
Source2CH LCD
Source2CH LCD
Acceptable Not Acceptable
Acceptable Not Acceptable
Acceptable Not Acceptable
GwC2I LCS / GwC Load /
GwC2I LCS / GwC Load /
Source2GwC LCD
Source2GwC LCD

Acceptable Not Acceptable Acceptable Not Acceptable

Gateway Decision 1 : Gateway Decision 3: Gateway Decision 2 :

Gateway Decision 2 : Gateway Decision 3: Gateway Decision 2 :
No Gateway, Send traffic to Send traffic to the CH
Send traffic to the CH Send traffic to Send traffic to the CH
Send traffic to the the selected GWC The gateway is the CH
The gateway is the CH the selected GWC The gateway is the CH
Base Station directly

CH is the proxy Decision

CH is the proxy GWC is the proxy from source vehicle Phase
Send UL Grant
message – from source vehicle from source vehicle to the LTE-Adv
Source Vehicle to the LTE-Adv to the LTE-Adv infrastructure
attached to the BS infrastructure infrastructure
Timer initiation
Timer initiation Timer initiation

Transmission

LTE-Adv interface Timer =LCDS2CH - ξ Timer =LCDS2CH - ξ

measurement Reports Timer =LCDS2GwC - ξ
with Periodic CQI feed
backs If more
If more If more No Maintenance
No No End packets
packets End packets Phase
to transmit
No to transmit to transmit
CQI > CQITh Yes
Yes Yes
Re-iterate in step Re-iterate in step
Yes “CH2I LCS (**)” “CH2I LCS (**)”

Figure 2. FQGwS algorithm

0018-9545 (c) 2013 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/TVT.2014.2323693, IEEE Transactions on Vehicular Technology
8

that is considered better while increasing, such as LCS Low Medium High
Low Medium High
and LCD. However, a negative criterion is a criterion 1
1
that goes worse as it increases, like the load. Therefore, 0,8
0,8
the weight Wi of the potential gateway candidate i 0,6
0,6
(pGW Ci ) can be formulated like in (3). 0,4
X Mij X Mjth 0,4
0,2
Wi = + (3)
j
Mjth Mik 0,2
0
k 60 65 70 75 80 85 90 95 100
0
-150 -140 -130 -120 -110 -100 -90 -80 -70 -60 Load membership functions
where, M ij is the positive criterion and M ik is the negative RSS membership functions (in dBm) Class 1 Traffic (in %)
criterion. LCDth and Lth are defined in TableII. LCSth is -
(a) (b)
2.17 for the three traffic classes.
Low Medium High Low Medium High
The source vehicle keeps track of the best pGwC with
1 1
the highest weight, named the Gateway candidate (GwC).
0,8 0,8
GwC2I LCS, load and Source2GwC LCD of the GwC
0,6 0,6
are then used to compute the gateway decision as input of
0,4 0,4
FQGwS algorithm. 0,2 0,2
This algorithm is the baseline for making the decision 0 0
over the gateway. As it can be deduced, a source vehicle 55 60 65 70 75 80 85 90 95 65 70 75 80 85 90 95 100
Load membership functions Load membership functions
checks in priority the CH state (CH2I LCS, load and Class2 Traffic (in %) Class 3 Traffic (in %)
Source2CH LCD). If the latter ensures efficient proxy
(c) (d)
capabilities, it will be selected as the gateway and source low high
low high
vehicle won’t send the GATEWAY SOLICITATION REQuest
1 1
message. Otherwise, source vehicle will solicitate other pG- 0,8
0,8
wCs. Thanks to this step, gateway efficient decision is taken 0,6 0,6
without leading to useless extra overhead due to GATEWAY 0,4 0,4

SOLICITATION REQuest and GATEWAY SOLICITATION 0,2 0,2

RESponse messages. 0 0
40 50 60 70 80 90 100
0 10 20 30 40 50 60 70 80
3) Maintenance phase: After making the decision over LCD membership functions (in s) LCD membership functions (in seconds)
the gateway, source vehicle initiates its Timer and starts Class 1 traffic Class 2 & Class 3 Traffics

transmitting. If there is no gateway between source vehicle (e) (f)

and the infrastructure, the link maintenance phase is handled
by periodic Channel Quality Indicator (CQI) reports. However, Figure 3. Membership functions
if a gateway is elected, the transmitting period will last
LCDS2CH − ε or LCDS2GwC − ε depending on the selected
gateway, afterward, if there is more packets to be transmitted, membership-degree. These results will be used then to decide
source vehicle will re-iterate the process in the CH check phase of the vehicle gateway to the infrastructure based on rules base.
to re-elect the appropriate gateway for the remaining packets. Note here that the membership function of X2I LCS, where
X ∈ {Source, CH, GwC}, is not presented as the RSS
C. Fuzzification and the Load membership functions are more significant and
illustrative. However, while performing FQGwS algorithm, the
In the fuzzy logic, the fuzzifier transforms the inputs values LCS membership function is used. It is concluded from both
into degrees of match with linguistic values. The fuzzifier RSS and loads membership functions and by using (1).
>
maps a crisp point, x = [x1 , x2 , x3 , ..., xn ] ∈ U , into a The fuzzification step is performed directly after measuring
fuzzy set A = (x1 , µA (x1 )), (x2 , µA (x2 )), ..., (xn , µA (xn )) the values of each proposed criteria. It will generate one or
in U where µA : U −→ [0, 1] is the membership function of several membership degrees from each value for one or several
the fuzzy set A and µA (xi ) is the membership degree of xi membership functions which will be applied into the next step
in A. of rules evaluation. Values range of each curve is fixed based
In our system model, input parameters (Source2I LCS, on the standard scenario when the degree is equal to 1, and
CH2I LCS, CH load, Source2CH LCD, GwC2I LCS, based on extrapolation between the best and worst scenarios
GwC load and Source2GwC LCD) collected by the source in the fuzzy part when the degree is different from 1.
vehicle are fuzzified using the predefined input membership
functions shown in the figures 3 (a, b, c, d, e, f). The fuzzi-
fication transforms the input value to names and degrees of D. Rules evaluation
membership in the functions. Each curve presents the possible After the fuzzification step of the input values, fuzzified
memberships degrees for each values set of that criterion. For generated values are used to evaluate the rules to obtain the
example for one input parameter computed or received by Fuzzy Gateway Decision. The fuzzy rule base contains a
source vehicle, using the membership function figure, we can number of fuzzy IF-THEN rules. If a fuzzy system has n
determinate the coordinate of this value into the curve and inputs and a single output, its fuzzy rule Rj is of the general
therefore identifying each couple of membership-function and form:

0018-9545 (c) 2013 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/TVT.2014.2323693, IEEE Transactions on Vehicular Technology
9

FQGwD Output CH GW Send to the BS Repeat Measurements

Membership F GwD <0.125 0.125<F GwD <0.625 F GwD >0.625
function
Table III
D EFUZZIFICATION TABLE OF C LASS 1 TRAFFIC

FGwD Output Send to the BS CH GW GwC Repeat Measurements

Membership function F GwD <0.125 0.125<F GwD<0.55 0.55<F GwD<0.625 F GwD>0.625
Table IV
D EFUZZIFICATION TABLE OF C LASS 2 / 3 TRAFFICS

IF x1 is A1j , x2 is A2j , ... xn is Anj THEN y is Bj , We note here that the gateway that could be elected for a
where A1j and Bj are fuzzy sets in Ui ⊂ R and V ⊂ R, Class 2 traffic might not be the same if the source vehicle has
>
respectively. The variables x = [x1 , x2 , x3 , ..., xn ] ∈ U1 × to transmit a Class 1 or Class 3 traffic (c.f. TableII) as QoS
U2 × U3 ... × Un and y ∈ V are called the input and the output requirements are not the same for the three traffic classes.
linguistic variables, respectively. Therefore, our algorithm takes advantages of the elasticity of
The fuzzy inference engine performs the inference opera- some traffic classes to make an efficient gateway decision and
tions on the fuzzy rules. Fuzzy logic principles are used to to balance the QoS classes over the cluster.
combine fuzzy IF-THEN rules in the rule base, and fuzzy sets
in U = U1 × U2 × U3 ... × Un are mapped into fuzzy sets in E. Defuzzification
V . Let A ∈ U be the input to the fuzzy inference engine, and
The last step is defuzzification which is used to determine
let the fuzzy rule be represented as the fuzzy implication
R = R1j × R2j × R3j ... × Rnj −→ Gj = C in U × V . the value of the Fuzzy Gateway Decision (F GwD). In our
Then, each fuzzy IF-THEN rule determines a fuzzy set system model, we consider the Centroid defuzzification tech-
D = Bj ∈ V using the composition: nique. This method is also known as center of gravity or
µD (u, w) = sup ∈ U [µR−→C (x, y)4µA (x)] center of defuzzification area. This technique was developed
x by Sugeno in 1985. This is the most commonly used technique
In our system model, for Class 1 traffic, there are four and is very accurate. The centroid defuzzification technique is
input parameters. The three input parameters (Source2I LCS, computed using this formula:
CH2I LCS and CH load) are each composed of three fuzzy ´
sets (High, Low and Medium) and the Source2CH LCD µi (x) x dx
F GwD = ´ (4)
input parameter is composed of only two fuzzy sets (High µi (x) dx
and Low). Therefore the maximum possible number of rules
where F GwD is the defuzzified output, it is the member-
used to build the rule base is
ship degree of output, µi (x) is the aggregated membership
Nr = 33 x 21 = 54 rules
function and x is the output variable. The only disadvantage of
For Class 2 and Class 3 traffics, there are seven in-
this method is that it is computationally difficult for complex
put parameters. The five input parameters (Source2I LCS,
membership functions. However, in our system membership
CH2I LCS, CH load, GwC2I LCS, GwC load) are each
functions have a simple trapezoid shape. Table III and table IV
composed of three fuzzy sets (High, Low and Medium)
illustrate membership degrees of output F GwD used to make
and the Source2CH LCD and Source2GwC LCD input
the final decision over the gateway. Membership functions of
parameters are composed of only two fuzzy sets (High and
the F GwD output are built after collection of experimental
Low). Therefore the maximum possible number of rules used
simulation result of several scenarios where the decision over
to build the rule base is
the gateway is trivial. Depending on the value of F GwD
Nr = 35 x 22 = 972 rules
Rules table is build on IF-THEN basis. It contains at most and according to the table III and table IV, the gateway is
Nr rules defining the FQGwS decision of input parameters selected. In the next section, the FQGwS decision is illustrated
belonging to diverse membership-functions after fuzzification by a realistic example and compared to deterministic gateway
step. Nr seems to be a high number of rules. However, while selection algorithm.
building rules data base, several rules can be combined in only
one rule. For example, if the Source2I LCS, CH2I LCS, VI. P ERFORMANCES E VALUATION
GwC2I LCS are all Low (in the fuzzy set), it becomes In this section, we present simulation results to confirm the
useless to check CH load, GwC load, Source2CH LCD and validity of our proposed approach. Simulations are performed
Source2GwC LCD as neither the CH nor the GwC could in our testbed. Through these simulations, we compare our
be a gateway for the source vehicle. Thus analytically there is FQGwS algorithm with the deterministic gateway selection
32 x 22 =36 rules having the same output but concretely these scheme in a clustered VANET architecture. The testbed con-
rules could be combined in only one rule which is «Repeat tains a Matlab module where FQGwS algorithm attributes,
measurements». criteria and rules are implemented. Membership functions

0018-9545 (c) 2013 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/TVT.2014.2323693, IEEE Transactions on Vehicular Technology
10

VANET Interface LTE Advanced Interface

Parameters Values Parameters Values
Physical/Mac protocols IEEE 802.11p Physical Layer protocols LTE Advanced
Transmission range R = 250m Carrier frequency 2 GHz
Preamble Length 144 bits eNobeB Trans Power 15 W
PLCP Header Length 48 bits Queue/LTEQueue qos True
Beacon Interval 100ms Source / GwC / CH - Data Rate : 10Mb
CW Min / Max 15 / 1023 eNodeB link LTEQueue/ULAirQueue
802.11p Data Rate 11 Mb/s LTEQueue/ULAirQueue
Interface Queue Type DropTail/PriQueue eNodeB - Serving GW Data Rate : 1000Mb
Interface Queue Length 50 packets link LTEQueue/ULS1Queue
Radio Propagation LTEQueue/ULS1Queue
P G Gr h2t h2t
Number of vehicles N v 15 - 40 Pr (d) = t td4 L
Model
Traffic Simulator VanetMobiSim N BS 3
Mobility Scenarios IDM_IM / IDM_LC System Loss (L) 1 dB
Data Rate (Mb/s): Tran/Rec Antenna High
0.064/ 0.2/ 0.3 1.5 m / 20 m
Class 1 / 2 / 3 (ht / hr )
Packet Size (bytes): Tran/Rec Antenna Gain
160 / 1300 / 1000 1 dB / 1 dB
Class 1 / 2 / 3 (Gt / Gr )
NS 20 α (Confidence Interval) 0.5
ε 1s
Table V
S IMULATION PARAMETERS S ETTINGS

of LCS, Load, LCD and the output F GwD for the CH, vehicles, where vehicles moving in the same direction belongs
GwC and Source V ehicle for each traffic class are defined to the same cluster. The first vehicle moving in each direction
in this module. Rules base based on membership functions will be elected as CH. This algorithm considers intersection
fuzzy sets is also built and integrated into the module for scenarios, while in our work we focus on the generalization of
the 3 traffic classes. Network performances simulation are the C-DRIVE to use it as a comparison base for CH election
performed using the Network Simulator NS2.33 [29]. We algorithms. In reference [7], the CH is the vehicle in the middle
implement in our testbed an IEEE 802.11p package in order of the cluster at equal distance, in terms of number of hops,
to enable VANET communications among high-speed vehicles from vehicles in the border of the cluster, called here Middle
and an LTE module to enable 4G communications [30]. Vehi- algorithm. On the other hand, the infrastructure is composed
cles mobility is generated using VanetMobiSim [31]. Table V of N BS eNodeBs positioned in the border of the highway
lists parameters of VANET and LTE modules. Specifically, where vehicles are attached to the eNodeB offering highest
a source vehicle first collects the input parameters of the LCS S2I , see Table V.
FQGwS algorithm using NS2.33 module. Then, these input In the traffic simulation scenario, a source node that belongs
parameters are integrated into the Matlab fuzzy module to to the cluster uses the FQGwS or the deterministic scheme for
perform decision over the gateway by considering the class of gateway selection. In the deterministic scheme, i.e. C-DRIVE
the traffic to be transmitted to the infrastructure. The fuzzy or Middle without FQGwS, the selected gateway is always the
module generates then the output of the FQGwS. FQGwS CH. With FQGwS algorithm, source vehicle takes the decision
gateway decision is afterward plugged into the simulator and over the gateway using our fuzzy module and after data
compared to the default deterministic scheme. collection. C-DRIVE and Middle without FQGwS algorithms
are compared to C-DRIVE and Middle with FQGwS schemes
in order to evaluate the performances of our FQGwS scheme
A. Simulation Scenario
in terms of delay, throughput, Packet Loss (PL), overhead and
In the simulation scenario, we consider a section of a LCD and RSS improvement. Simulation scenarios have been
straight multiple-lane highway. Vehicles are equipped with a repeated N S times by varying source vehicle velocity. We
Global Positioning System (GPS) receiver. Thus, we assume plot simulation results using confidence interval (CI) with a
that each vehicle knows its own position and velocity. The certitude of 95% (see α (confidence interval) in Table V).
vehicles are organized into clusters where a CH is elected at
the top of each cluster. A cluster is composed of N v vehicles.
Clustering and CH election schemes are out of the scope of B. Impact of Cluster Density
this paper. Hence, we consider two clustering and CH election In this section, we evaluate the performances of our algo-
algorithms [6], [7] recently proposed in literature for VANET rithm while increasing the cluster density. We compare two
networks. In reference [6], authors propose the C-DRIVE deterministic schemes, noted in the legends C-DRIVE w/o
algorithm for cluster formation that is based on direction of FQGwS and Middle w/o FQGwS, where the gateway to the

0018-9545 (c) 2013 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/TVT.2014.2323693, IEEE Transactions on Vehicular Technology
11

200 300 C-DRIVE w

Average Delay Class 2 Traffic (s)

Average Delay Class 1 Traffic (s)
0,35 0,4

Class 3 Traffic (kbps)

C-DRIVE w FQGwS

Class 2 Traffic (kbps)

0,3 0,35 170 260
FQGwS C-DRIVE w/o
0,25 0,3
C-DRIVE w/o 140 220 FQGWS
0,25
0,2 FQGWS
0,2 110 180 Middle w
0,15 Middle w FQGWS
0,15
0,1 FQGWS 80 140
0,1 Middle w/o
0,05 Middle w/o FQGWS
0,05 50 100
FQGWS
0 0 10 20 30 40 10 20 30 40
10 20 30 40 10 20 30 40 Cluster Density Cluster Density
Cluster Density Cluster Density

Figure 6. Average Throughput Vs Cluster Density for Class 2 & Class 3

Figure 4. Average Delay Vs Cluster Density for Class 1 & Class 2 traffics traffics

45 80
Throughput Average (kbps)

C-DRIVE w 50 50 C-DRIVE w
Packet Loss Average (%)

40
70 FQGwS FQGwS

Class 2 Traffic (%)

Class 3 Traffic (%)

35 40 40
30 60 C-DRIVE w/o C-DRIVE w/o
25 FQGWS 30 30
50 FQGWS
20
Middle w 20 20 Middle w
15 40
10 FQGWS FQGWS
30 10 10
5 Middle w/o Middle w/o
0 20 FQGWS 0 0 FQGWS
10 20 30 40 10 20 30 40 10 20 30 40 10 20 30 40
Cluster Density Cluster Density Cluster Density Cluster Density

Figure 5. Packet Loss and Throughput Averages Vs Cluster Density for Class Figure 7. Packet Loss Average Vs Cluster Density for Class 2 & Class 3
1 traffic traffics

infrastructure is always the CH. that C-DRIVE algorithm without FQGwS scheme presents
In figure 4, we plot the average delay of Class 1 and Class high amount of packet loss as the CH is positioned in front of
2 traffics while increasing the cluster density. The computed the cluster members causing an increase in the number of hops
delay is the average of the delay of the all received packets. to attend the gateway to the infrastructure. In figure 5, we plot
The figure shows clearly that with FQGwS scheme lower de- also the throughput average of Class 1 traffic while increasing
lays (<0.5ms) are generated as compared to deterministic ap- the number of vehicles in the cluster. The throughput remain
proaches, i.e. C-DRIVE w/o FQGwS and Middle w/o FQGwS constant for C-DRIVE and Middle w FQGwS near to 64kb/s.
for both Classes which exceeds the delay constraints, see Table However, w/o FQGwS algorithm, the data rate decreases
I. This is due to the fact that the selected gateway has better while increasing the cluster density. The throughput decreases
features of VANET and LTE Advanced interfaces than the reaches the half of initial data rate for Middle algorithm in a
CH. In fact, FQGwS algorithm is decision efficient as it takes cluster density of 40 vehicles.
into consideration at the same level the LTE Advanced and In figure 6, we plot the throughput average of Class 2 and
VANET features. Thus, if the LTE Advanced link properties, Class 3 traffics while increasing the cluster density. Class 2
i.e. BS load and source vehicle RSS, are acceptable, it is more throughput constraint is 0.2Mb/s and Class 3 constraint is
suitable to have no intermediate gateways and relays for Class 0.3Mb/s. According to the figure, C-DRIVE and Middle w
1 traffic type as, for delay sensitive traffics such as alerts, voice FQGwS algorithms reach the throughput constraints as LCS,
or real time gaming, the lower the delay is the better the QoS LCD and load are baseline criteria of our algorithm. However,
perceived by the users is. Without FQGwS algorithm, for Class w/o FQGwS algorithm the throughput decreases considerably
1 and Class 2 traffics, as the number of vehicle increases, the while cluster density increases.
delay increases because the CH will handle and manage more In figure 7, we plot the packet loss average of Class 2
vehicles. However, with FQGwS, the delay is almost invariant and Class 3 traffics while increasing the number of vehicles
as the algorithm takes aware of the load of the either the CH in the cluster. Without FQGwS algorithm, as the number of
and the gateway candidates for gateway selection. vehicle increases, the PL increases because the CH will handle
In figure 5, we first plot the average packet loss for Class and manage more vehicles. However, with FQGwS, the PL is
1 traffic while increasing the cluster density. With FQGwS almost invariant as the algorithm takes aware of the load of
packets drop is lower than 0.05%, in opposite to algorithms the either the CH and the gateway candidates for gateway
without FQGwS which increases the PL average as the number selection.
of vehicles belonging to the cluster increases. Without FQGwS We plot in figure 8, the percentage of control packets
scheme, buffering queues of the CH becomes rapidly over- overhead in the network from the total number of packets
loaded while the number of handled nodes increases causing send by the source vehicle against the number vehicles in the
considerable packets losses. However, with FQGwS, one of the cluster. Though, this is generally the trend, Middle and C-
main criteria for gateway selection is the buffering queues load DRIVE for Class 1 traffics show less control packets overhead
of the either the CH and/or the gateway candidate leading to as compared to the other traffic Classes (Class 2 and 3)
an adequate gateway selection. Moreover, curves show clearly due to the fact that only the CH is solicited for to be a

0018-9545 (c) 2013 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/TVT.2014.2323693, IEEE Transactions on Vehicular Technology
12

Control Packets Overhead (%)

70 250 C-DRIVE w
C-DRIVE Class 1

Class 2 Traffic (kbps)

Class 1 Traffic (kbps)
2,5 60 FQGWS
200
2 50 C-DRIVE w/o
Middle Class 1 150
40 FQGWS
1,5
30 100
1 C-DRIVE Class 2 & 3 Middle w
20 FQGwS
50
0,5 10
Middle Class 2 & 3 0 0 Middle w/o
0 FQGWS
10 20 30 40 0 5 10 15 20 25 0 5 10 15 20 25
Cluster Density Source Node to CH relative Velocity (m/s) Source Node to CH relative Velocity (m/s)

Figure 8. Overhead Vs Cluster Density Figure 10. Throughput Average Vs Relative Velocity for Class 1 & Class 2
traffics

90 6
Metrics Improvement (%)

80 300 C-DRIVE w

Class 3 Traffic (kbps)

5
LCS Improvement (%)
70 RSS C-DRIVE 15 250 FQGwS
60 4 C-DRIVE w/o
LCD C-DRIVE 20 200
50 FQGWS
40 3 150
RSS Middle
30 30 Middle w
2 100
20 LCD Middle FQGWS
10 1 40 50
Middle w/o
0 0
0 FQGWS
10 20 30 40 0 50 100 150 200
Cluster Density Cluster Density
LCD (s)
(a) (b)

Figure 9. (a) Metrics Improvement Vs Cluster Density - (b) FQGwS Vs Flat Figure 11. Throughput Average Vs LCD for Class 3 Traffic
Architecture

or presents a bad LCS.

potential gateway for Class 1 traffics, no other pGwCs are
considered. Indeed, Class 1 traffics exhibit around 1.5% of C. Impact of Nodes’ Velocity
overhead packets as compared to the 2.5% for Class 2 and
In this section, we evaluate the impact of CH to source
Class 3 traffics.
vehicle relative velocity and LCD upon the performances of
In figure 9-(a), we plot the metrics improvement while com-
the algorithm.
paring our algorithm to the C-DRIVE and Middle schemes.
Figure 10 illustrates the throughput average of Class 1 and
Two metrics are considered in this paper: RSS and LCD.
Class 2 traffics while the relative velocity between source
LCD metric is the link connectivity duration between the
vehicle and the CH increases. As it can be noticed from
source node and the selected gateway (either the CH or the
the figure, throughput average is much higher with FQGwS
GwC). According to the figure, there is an improvement of
algorithm than the one without FQGwS algorithm as LCD
more than 10% in terms of RSS for Middle algorithm and
criterion, which considers relative velocity is one of the input
about 20% for C-DRIVE. It is important to choose a gateway
criteria of our algorithm. Moreover, the throughput remains
that has high RSS as it will interface the VANET layer with
invariant while increasing the relative velocity. It is decreased
the LTE Advanced layer. An improvement of about 80% for
by almost the half while not using FQGwS algorithm for both
Middle scheme and 90% for C-DRIVE scheme of the LCD
clustering schemes.
metric is performed with FQGwS thanks to the fact that our
We plot, in figure 11, the throughput average versus LCD
algorithm selects the gateway with a low relative velocity and
for Class 3 traffics. The figure shows better results of FQGwS
high LCD.
algorithm as compared to deterministic approach. This is due
We compare our proposed FQGwS algorithm with the flat to the fact that the selected gateway has better features than
architecture where the VANET network is not clustered. In the CH leading to better simulation results. In fact, as the
a flat architecture, the source vehicle sends directly to the LCD between the source node and the gateway decreases, the
eNodeB. In figure 9-(b), we plot the improvement of the LCS relative velocity increases, therefore, the V2V link become
criterion when using the FQGwS algorithm as compared to the
flat architecture. An improvement of more than 5% is noticed,
it increases while the cluster density increases because the set 100 120 C-DRIVE w/o
of pGwCs becomes bigger, increasing the chance to find a 100 FQGWS
Class 3 Traffic (%)

80
Class 1 Traffic (%)

80 C-DRIVE w
better GwC. Note that with our FQGwS algorithm the direct 60 FQGwS
60
link is considered as the first choice to check for sending the 40 Middle w/o
40 FQGWS
traffic to the infrastructure. If the latter presents acceptable
20 20 Middle w
criteria, traffic will be sent directly to the infrastructure for FQGwS
0 0
class 1 traffic without CH solicitation. In fact, the advantage 0 50 100 150 0 50 100 150
LCD (s) LCD (s)
of using a clustered architecture is to propose an alternative
way for V2I communications in the situation where the direct
link from the source vehicle to the infrastructure is not possible Figure 12. Packet Loss Average Vs LCD for Class 1 & Class 3 Traffics

0018-9545 (c) 2013 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/TVT.2014.2323693, IEEE Transactions on Vehicular Technology
13

Scenario Specifications C-Drive Middle

Parameters Values w w/o w w/o
FQGwS FQGwS FQGwS FQGwS
RSS / Load Source Node -98.57 dBm / 81 %
CH 30 100 40 100
Source Node Velocity 25.2 ms
Class 1 GwC 0 0 0 0
RSS / Load eNodeB CH -100.56 dBm / 80 %
eNB 70 0 60 0
CH Velocity 19.6 ms
CH 25 100 25 100
CH Load 85 % Class 2 GwC 45 0 60 0
LCD CH 85.57 s eNB 30 0 15 0
GwC LCS -1.33 CH 20 100 40 100
Load GwC 77.4 % Class 3 GwC 50 0 60 0
LCD GwC 77.08 s eNB 30 0 0 0
Table VI Table VIII
S CENARIO S PECIFICATIONS Q O S-BALANCING G ATEWAY S ELECTION PERCENTAGE

Class 1 Class 2 Class 3

w Direct (F GwD= 0.6) (F GwD = 0.48) gathering and statistics over previous figures.
FQGwS link <-> GWC <-> CH According to Table VIII, for Class 1 traffics, more than
w/o half of traffics are send directly to the infrastructure without
CH CH CH
FQGwS defining a gateway for the source vehicles. However for Class
Table VII 2 traffics, only 25% of the traffic is handled by the CH. The
FQG W S G ATEWAY D ECISION GwC has got the highest percentage to be selected as gateway
to the infrastructure. For Class 3 traffics, more than 50% of
the traffics are handled by the GwC.
unstable. However, with our algorithm, thanks to the LCD
criterion, source vehicle to gateway link stability is achieved. VII. C ONCLUSIONS
Figure 12 illustrates the PL average versus LCD for Class This paper presented an original algorithm based on a
1 and Class 3 traffics. The figure shows better results of fuzzy logic for gateway selection from a VANET network
FQGwS algorithm as compared to deterministic approach. to the LTE Advanced infrastructure. This new approach has
This is due to the fact that the selected gateway has better been compared to the standard deterministic approache that
features than the CH leading to better simulation results. In uses the CH as a default gateway. Two clustering and CH
fact, as the LCD between the source node and the gateway selection algorithms have been studied and simulation results
decreases, the relative velocity increases, therefore, the V2V show that our protocol performs better results in terms of
link become unstable. However, with our algorithm, source delay and packet loss than the deterministic approach for both
vehicle to gateway link stability is achieved. algorithms. Moreover, simulations show that an efficient CH
election algorithm is important to ensure good performances as
D. QoS-balancing for Gateway selection with C-DRIVE higher packet loss averages have been observed
In this section, we point out the QoS-balancing feature of than with middle algorithm. In our future works, we will
FQGwS algorithm while deciding over the gateway to select focus on studying the performances of FQGwS algorithm
for a source vehicle depending on the Class of the traffic and investigating its adaptability while performing cluster
to be transmitted to the infrastructure, under the scope of head handovers in a clustered architecture and optimizing the
V2I communications. To this end, we consider an example gateway solicitation phase where selected gateway candidate
of scenario in which, we present and compare the gateway might be identical for several source nodes belonging to the
decision of both the FQGwS and the deterministic algorithms. same cluster specially in case of small clusters.
Table VI presents the input criteria of the scenario. Ac-
cording to table VII, depending on the class of the traffic to R EFERENCES
be transmitted to the infrastructure, the gateway decision is [1] D. Camara, C. Bonnet, and F. Filali, “Propagation of Public Safety
different. If it is a Class 2 traffic, it is better to use the GwC Warning Messages A Delay Tolerant Network Approach,” IEEE WCNC,
than the CH as a gateway. If it is a Class 3 traffic, it is better to pp. 1–6, 2010.
[2] P. Belanovic, D. Valerio, A. Paier, T. Zemen, F. Ricciato, and C. Meck-
use the CH as a gateway to the infrastructure. This is due to the lenbrauker, “On Wireless Links for Vehicle-to-Infrastructure Communi-
fact that LCS and load membership functions for Class 2 and cations,” IEEE Transactions on Vehicular Technology, vol. 59, Issue 1,
Class 3 do not present same limits of the fuzzy set. This feature pp. 269 – 282, 2010.
[3] C.-M. Chou, C.-Y. Li, W.-M. Chien, and K. chan Lan, “A Feasibility
leads to have, in this scenario, the CH is a good gateway for a Study on Vehicle to Infrastructure Communication: WiFi vs. WiMAX,”
Class 3 traffic and not for a Class 2 Traffic. Therefore, QoS- International Conference on Mobile Data Management: Systems, Ser-
balancing feature of FQGwS algorithm is achieved. vices and Middleware (MDM), pp. 397–398, 2009.
[4] G. e. m. Zhioua, H. Labiod, N. Tabbane, and S. Tabbane, “LTE
Table VIII presents percentage of each gateway selection Advanced Relaying Standard: A survey ,” Springer Wireless Personal
decision. These percentage are computed using information Communications Journal, vol. 72, Issue 4, pp. 2445–2463, April 2013.

0018-9545 (c) 2013 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/TVT.2014.2323693, IEEE Transactions on Vehicular Technology
14

[5] “ITU-R M.2134,” http://www.itu.int.

[6] N. Maslekar, M. Boussedjra, J. Mouzna, and H. Labiod, “Direction based
clustering algorithm for data dissemination in vehicular networks,” IEEE
Vehicular Networking Conference (VNC), pp. 1–6, 2009.
[7] A. Benslimane, T. Taleb, and R. Sivaraj, “Dynamic clustering-based Ghayet el mouna ZHIOUA obtained her
adaptive mobile gateway management in integrated vanet-3g heteroge- PhD degree jointly from Sup’Com (the Higher School of
neous wireless networks,” IEEE JSAC, vol. 29, Issue 3, pp. 559–570,
2011. Communication of Tunis- TUNISIA) and Telecom Paris-
[8] T. Kwon, M. Gerla, V. Varma, M. Barton, and T. Hsing, “Efficient Tech (Paris-FRANCE). She received her Master’s degree in
Flooding with Passive Clustering - An Overhead-Free Selective Forward Telecommunication from Sup’Com. She had publications in
Mechanism for Ad Hoc/Sensor Networks,” Proceedings of the IEEE, vol.
91, Issue 8, pp. 1210 – 1220, Aug 2003.
refereed International Conferences, such as IEEE WCNC,
[9] D. Elghnami, T. Kwon, and A. Hafid., “GRIDS: Geographically Re- IEEE WiMob, ACM MSWiM. Her research interests are:
pulsive Insomnious Distributed Sensors - An Efficient node selection 3GPP LTE, 4G Heterogeneous Wireless Networks, Vehicular
mechanism using Passive Clustering,” IEEE WIMOB, pp. 241–246, Communications, IEEE 802.11p, IEEE 802.11s standard and
2008.
wireless mesh communications.
[10] P. Basu, N. Khan, and T. Little, “A Mobility Based Metric for clustering
in Mobile ad hoc Networks,” International Conference on Distributed
Computing Systems, pp. 413–418, 2001.
[11] E. Dror, A. C., and Z. Lotker, “Fast Randomized Algorithm for Hierar-
chical Clustering in Vehicular Ad-Hoc Networks,” IFIP Med-Hoc-Net,
pp. 1–8, 2011.
[12] S.Kuklinski and G.Wolny, “Density Based Clustering Algorithm for
Nabil TABBANE is currently an Associate
VANETs,” International Journal of Internet Protocol Technology, vol. Professor in the Computer Networks Department at Higher
4, Issue 3, pp. 149–157, Sep 2009. School of Communications, Tunis: SUP’COM. He received his
[13] Z.Wang, L.Liu, M.Zhou, and N. Ansar, “A Position Based Clustering PhD degree in SUP’COM and in the University of Versailles
Technique for Ad Hoc Intervehicle Communication,” IEEE Transactions
on Systems Man and Cybernetics - Applications and reviews, vol. 38, Saint-Quentin-en-Yvelines (France) in joint guardianship. His
pp. 201–208, 2008. PhD research was on stochastic models for forecasting the
[14] M.Jerbi, S. Senouci, T. Rasheed, and Y. Doudane, “An Infrastructure- quality of service in multi-media ad hoc networks. His research
Free Traffic Information System for Vehicular Networks,” IEEE VTC interests are in ad hoc and sensor networks.
Fall, pp. 2086–2090, 2007.
[15] I.Salhi, M.Cherif, and S.Senouci, “A new Architecture for Data Collec-
tion in Vehicular Networks,” IEEE ICC, pp. 1–6, 2009.
[16] Y. Fu, K.-M. Chan, K.-S. Tan, and B.-S. Yeo, “Multi-Metric Gateway
Discovery for MANET,” IEEE VTC Spring, pp. 777 – 781, 2006.
[17] R. Manoharan, S. Rajarajan, S. Sashtinathan, and K. Sriram, “A Novel Houda LABIOD received a PhD degree in
Multi-hop B3G Architecture for Adaptive Gateway Management in Computer Science from the University of Versailles Saint-
Heterogeneous Wireless Networks,” IEEE WiMob, pp. 447–452, 2009.
Quentin-en-Yvelines (France) in 1998. From 1999 to 2000,
[18] F. Setiawan, S. Bouk, and I. Sasase, “An optimum multiple metrics
gateway selection mechanism in MANET and infrastructured networks she worked as an assistant researcher at Eurecom Institute
integration,” IEEE WCNC, 2008. in Sophia-Antipolis (France) in the Mobile Communications
[19] A. Ramalingam, S. Subramani, and K. Perumalsamy, “Associativity Unit. Since September 2000, she is an associate professor
based cluster formation and cluster management in ad hoc networks,” at Telecom ParisTech in the INFRES (Computer Science
International Conference On High Performance Computing (HiPC),
2002. and Network) department. She is an author of a book titled
[20] S.Basagni, “Distributed clustering for ad hoc networks,” International “De Wi-Fi à Bluetooth” (Hermès Editions, 2004). Her main
symposium on Parallel Architectures, Algorithms and Networks, (I- research interests include optimization of mobile and wireless
SPAN), pp. 310–315, 1999.
networking and mobile ad hoc networks (unicast routing,
[21] I. Salhi, M. Cherif, and S. Senouci, “A New Architecture for Data
Collection in Vehicular Networks,” IEEE ICC, pp. 1–6, 2009. multicast routing, security, QoS routing).
[22] M. S. Almalag and M. C. Weigle, “Using Traffic Flow for Cluster For-
mation in Vehicular Ad-hoc Networks,” IEEE Local Computer Networks
(LCN), pp. 631 – 636, Oct 2010.
[23] G. e. m. Zhioua, H. Labiod, N. Tabbane, and S. Tabbane, “An Efficient
QoS based Gateway Selection Algorithm for VANET to LTE Advanced Sami TABBANE is currently a Professor in
Hybrid Cellular Network,” ACM MSWiM, pp. 353–356, 2012.
the Computer Networks Department at Sup’Com. He is a
[24] T. Bchini, N. Tabbane, S. Tabbane, E. Chaput, and A. Beylot, “Fuzzy
logic based layers 2 and 3 handovers in IEEE 802.16e network,” Journal specialist of mobile communication systems. He has graduated
of Computer Communications, vol. 33, Issue 18, pp. 2224–2245. from Ecole des Mines de Paris in 1998 and got a PhD of
[25] J.-S. Jang and C.-T. Sun, “Neuro-fuzzy modeling and control,” Proceed- Telecom ParisTech in 1991. He started his carreer at France
ings of the IEEE, vol. 83, pp. 378 – 406, March 1995.
Télécom in 1992-94 and was recruted by Sup’Com in 1994.
[26] 3GPP TS 23.203 V12.0.0, March 2013.
He has achieved many missions for ITU in the domain of
[27] J. Misic, G. Badawy, and V. Misic, “Performance Characterization for
IEEE 802.11p Network With Single Channel Devices,” IEEE Transac- mobile networks planning, management and training as well
tions on Vehicular Technology, vol. 60, Issue 4, pp. 1775 – 1787, 2011. as spectrum management in regulatory agencies. He is the co-
[28] C. Andre, S. Susana, and S. Peter, “On the Connection Availability author of “Réseaux GSM” (Hermes, 1995) and “Ingénierie des
between Relay Nodes in a VANET,” IEEE Globecom, pp. 181–185,
2010. Services de Télécommunications” (Hermes-Lavoisier, 2005).
[29] “The Network Simulator NS2.33,” http://www.isi.edu/nsnam. He is the author of “Réseaux Mobiles” (Hermes, 1997),
[30] http://code.google.com/p/lte-model/updates/list. “Handbook of Mobile Networks” (Artech House, 2000), “In-
[31] “VanetMobiSim Project,” Website: http://vanet.eurecom.fr/. génierie des Réseaux Cellulaires” (Hermes-Lavoisier, 2002).

0018-9545 (c) 2013 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.