254 JOURNAL OF COMMUNICATIONS SOFTWARE AND SYSTEMS, VOL. 4, NO.

4, DECEMBER 2008

A Quality of Service Driven Approach for

Clustering in Mobile Ad hoc Networks Based on
Metrics Adaptation: Looking Beyond Clustering
Zouhair El-Bazzal, Michel Kadoch, Basile L. Agba, Mohamad Haidar, and François Gagnon

Abstract: Recently, research topics are focusing on clustering information which facilitates the spatial reuse of resource and
approaches for Ad hoc networks due to their effectiveness in increase the system capacity. Each CH will act as a temporary
building a virtual backbone formed by a set of suitable base station within its zone or cluster and communicates with
clusterheads (CH) to guarantee the communications across other CHs. Thus, packets for route finding may only spread
clusters. In this paper, we propose a clustering approach to elect
suitable nodes’ representatives and to store minimum topology
among CHs instead of flooding among all nodes. On the other
information by reducing the propagation of routing information hand, the topology change information caused by movement
which facilitates the spatial reuse of resource and increase the of some nodes is limited in adjacent clusters, not in the whole
system capacity. The clusters must adapt dynamically to the network. Therefore, any clustering scheme should be adaptive
environment changes, we also propose a distributed maintenance to such changes with minimum clustering management
procedure that allows managing nodes’ adhesion, nodes’ handoff overhead incurred by changes in the network topology.
and CHs’ re-election. Based on our analytical model used to
estimate the quality of service (QoS) parameters, we implement To establish a cluster, traditional clustering algorithms
an admission control algorithm to determine the number of suggest CH election exclusively based on nodes’ IDs or
members inside a cluster that can be accommodated while location information and involve frequent broadcasting of
satisfying the constraints imposed by the current applications. control packets, even when network topology remains
This might effectively drive congestion avoidance on the CH and unchanged. Most recent work takes into account additional
interclusters load-balancing to achieve better network resource metrics (such as energy and mobility) and optimizes initial
utilization. The obtained results will help us to readjust the
clustering without taking into consideration the QoS
clustering algorithm metrics in order to provide better
maintenance and QoS guarantees depending on the used
parameters like “cluster achievable throughput, cluster delay
applications. Through numerical analysis and simulations, we and transmissions’ number per packet”. In many situations, re-
have studied the performance of our model and compared it with clustering procedure is hardly ever invoked; hence initially
that of other existing algorithms. The results demonstrate better elected CHs soon waste their batteries due to serving the other
performance in terms of number of clusters, number of handoffs, members for longer periods of time. In addition, a topology
number of transitions (state change) on CHs, QoS parameters, control mechanism is required to mitigate the vulnerability of
load balancing and scalability. We also observed how the such clusters due to node joining/leaving and link failures. It
connectivity and the stability are maximized when the number of aims to reduce interference and energy consumption, to
nodes increases in presence of the mobility.
increase the effective network capacity, and to reduce the end
Index Terms: Ad hoc networks, clusters, maintenance, to end delay.
quality of service, scalability. As election of optimal clusterheads is an NP-hard problem
[1], many heuristic mechanisms have been proposed.
I. INTRODUCTION Centralized algorithms rely on the assumption that the elected
Clustering is a promising approach for mimicking the CH is responsible of the cluster’s maintenance. However,
operation of the fixed infrastructure and managing the these algorithms suffer from single point (CH) of bottleneck
resources in mobile Ad hoc networks. An Ad hoc network is especially in highly mobile environments; hence initially
characterized by a collection of wireless nodes which elected CHs have to collect excessive amounts of information
communicate with each other using high-frequency radio and soon reach battery exhaustion. On the other hand,
waves. These nodes arbitrarily and randomly change their distributed algorithms are more adaptive to mobility due to the
locations and capabilities without the aid of any fixed fact that the maintenance is done in collaboration between all
infrastructure. The main objective of clustering is to elect the nodes where each node relies on the local information
suitable nodes’ representatives, i.e. CHs and to store minimum collected from the nearby nodes. Although the distributed
topology information by reducing the propagation of routing manner is preferred for MANET, it lacks a major drawback in
achieving and guarantying a strong connectivity between the
Manuscript received March 03, 2008, and revised December 14, 2008. nodes. The performance changes greatly for small and large
Authors are with the Department of Electrical Engineering, Ecole de clusters and depends strongly on the formation and
technologie superieure, University of Quebec, Canada (email: zouhair.el- maintenance procedures of clusters which should operate with
bazzal@etsmtl.ca). minimum overhead, allowing mobile nodes to join and leave

1845-6421/08/8009
EL-BAZZAL et al.: A QUALITY OF SERVICE DRIVEN APPROACH 255

without perturbing the membership of the cluster and (LID) [4, 5, 6] chooses the node with the lowest ID as a
preserving current cluster structure as much as possible. On clusterhead, the system performance is better than Highest-
the other hand, under high traffic load conditions, admission Degree in terms of throughput. However, those CHs with
control becomes necessary in order to provide and maintain smaller IDs suffer from the battery drainage, resulting short
the QoS of existing members. Our approach differs from lifetime of the system. Based on an analytical performance
others in that it is based on the clusters’ capacity and it uses model, Lian et al. [7] have concluded that Lowest-id performs
the link lifetime instead of the node mobility for the better than MCC in terms of bandwidth consumption and
maintenance procedure. We refer this to the fact that the node induced control overhead. Least Clusterhead Change
mobility metric does not affect the election of a CH as much Algorithm (LCC) [8] allows minimizing clusterhead changes
as the link stability metric does. It also provide an efficient that occur when two CHs come into direct contact. In such a
technique to estimate and derivate the number of members case, one of them will give up its role and some of the nodes
inside a cluster with respect to the allowed saturation in one cluster may not be members of the other CH’s cluster.
throughput, the delay and the packet error rate when the Therefore, some nodes must become CH while causing a lot of
protocol DCF (Distribution Coordination Function) is used to re-elections because of the propagation of such changes across
access the channel. The estimation methodology builds on the the entire network.
existence of a mathematical relationship between the number
3hBAC (3-hop Between Adjacent Clusterheads) [9] have
of competing members, the packet collision probability
introduced the concept of clusterguests which are some nodes
encountered on the shared medium and the packet arrival rate
that may not be connected to any CH, but can access a few
at each member.
clusters with the help of member nodes. The elected CHs are
In this paper, we propose a weight based Efficient 3-hops away from each other. First, the node with the highest
Clustering Algorithm (ECA) which underlay on node stability degree is elected as CH and forms the first cluster. Then, the
level, QoS parameters requested within a cluster as well as on formation of other clusters is done in parallel in the whole
sophisticated distributed maintenance procedures. Our network. However, the stationary assumption is required for
objectives are yielding low number of clusters, maintaining cluster formation in order to decide the first cluster. DDCA
stable clusters, achieving load-balancing and scalability. In (Distributed Dynamic Clustering Algorithm) [10] uses α, t
this manner, we propose an analytical model to estimate the criteria that indicate that every node in a cluster has a path to
QoS parameters in order to determine the cluster size that can every other node during some time period with a probability
be accommodated while satisfying the constraints imposed by greater than α regardless of the hop distance between them. A
the applications. This will help us to readjust the used node can join a cluster if it has a mutual path to satisfy the
parameters of the clustering algorithm in order to provide above criteria between itself and the CH of that cluster.
better quality of service guarantees depending on the used Otherwise, the node creates a new cluster just to cover itself.
applications. The results show that the proposed clustering DDCA can adaptively adjust its cluster size, considering the
model provides better performance in terms of number of same stability level α. In low mobility, it forms large-sized
formed clusters, number of handoffs, average number of clusters, but in highly mobile network, it diminishes the
transition (state change) on CHs, QoS parameter, load cluster size.
balancing, and scalability (connectivity) when compared to
that of other weight based algorithms. The paper is organized MOBIC [11] uses a new mobility metric; Aggregate Local
as follows. In Section II, we review several approaches Mobility (ALM) for CH election. ALM is computed as the
proposed previously. Section III presents the cluster formation ratio of received power levels of successive transmissions
model. Section IV presents the proposed analytical model. (periodic hello messages) between a pair of nodes, which
Section V presents the distributed maintenance model. Section means the relative mobility between neighboring nodes. It is
VI presents the performance analysis of the model. Finally, easy to see that MOBIC is effective for group mobility
Section VII concludes the paper. behavior, in which a group of nodes moves with similar speed
and direction. Thus, an elected CH can normally promise the
II. OVERVIEW OF THE EXISTING APPROACHES low mobility with respect to its member nodes. However, if
nodes move randomly and change their speeds from time to
Many approaches have been proposed for the election of
time, the performance of MOBIC will be greatly degraded.
clusterheads ad hoc networks. Maximum Connectivity
DLBC (Degree Load Balanced Clustering) [12] periodically
Clustering (MCC) [2] is based on the degree of connectivity.
runs the clustering to keep the number of nodes in each cluster
A node is elected as CH if it is the highest connected node.
around a system parameter named Elected Degree (ED),
This is not suitable in dynamic network topologies where the
which indicates the optimum number of nodes that a CH can
degree of connectivity changes rapidly. The Highest-Degree
handle. This mechanism tries to balance the load between the
[3] uses the degree of a node as a metric for the selection of
CHs. However, DLBC will cause frequent clustering
clusterheads. The degree of a node is the number of neighbors
invocations because the movement of mobile nodes and
each node has. The node with maximum degree is chosen as a
consequent link setup/break results in dynamic variation of
clusterhead; since the degree of a node changes very
node degree. In addition, how to decide the used system
frequently, the CHs are not likely to play their role as
parameters is not discussed in DLBC.
clusterheads for very long. In addition, as the number of
ordinary nodes in a cluster is increased, the throughput drops The Distributed Clustering Algorithm (DCA) [13] and
and system performance degrades. The Lowest-Identifier Distributed Mobility Adaptive clustering algorithm (DMAC)
256 JOURNAL OF COMMUNICATIONS SOFTWARE AND SYSTEMS, VOL. 4, NO. 4, DECEMBER 2008

[14] are enhanced versions of LID; each node has a unique is not possible to design orthogonal spreading codes due to
weight instead of just the node’s ID, these weights are used for multipath effect which make these transmissions suffer
the selection of CHs. A node is chosen to be a clusterhead if different time delays. In this case, the cross-correlation
its weight is higher than any of its neighbor’s weight; between codes cannot be neglected. We also take into
otherwise, it joins a neighboring clusterhead. The DCA makes consideration these assumptions:
an assumption that the network topology does not change a) There is a common signaling CDMA code named
during the execution of the algorithm. Thus, it is proven to be “sig_code” used by all nodes (new arrival nodes, existing
useful for static networks when the nodes either do not move nodes and CHs) for signaling messages (i.e. node wish to
or move very slowly. The DMAC algorithm, on the other join/leave a cluster, CH admits a node, CH rejects a node,
hand, adapts itself to the network topology changes and etc.). This might help in separating the payload traffic from
therefore can be used for any mobile networks. However, the the signaling traffic.
assignment of weights has not been discussed in the both b) There is a common CDMA code named
algorithms and there are no optimizations on the system “intercluster_code” for interclusters’ communications.
parameters such as throughput and power control. CHs will communicate between them using this unique
code.
The Weighted Clustering Algorithm (WCA) [15] is based
c) CH assigns a CDMA code exclusive to the cluster named
on the use of a combined weight metric that takes into account
“intracluster_code”. All nodes that are accepted as
several parameters like the node-degree, distances with all its
belonging to the cluster must use this code to communicate
neighbors, node speed and the time spent as a clusterhead.
with the CH. No other nodes can interfere with the
Although WCA has proved better performance than all the
throughput of these nodes. Other nodes that will be
previous algorithms, it lacks a drawback in knowing the
rejected by the CH must align themselves with other
weights of all the nodes before starting the clustering process
clusters or create a new cluster.
and in draining the CHs rapidly. Based on assigned weights
similar to those used in WCA, the authors in [16] proposed the We assume that the total number of “intracluster_codes” is
formation of k-clusters (member is k-hops away from its CH) large, so that neighboring clusters are not assigned the same
by limiting the size of each cluster to S nodes. The “intracluster_code”. However, a spatial reuse of these codes
maintenance is distributed among all the nodes for load may be used. With a large number of codes, it is unlikely for a
balancing purposes. The simulations show better performance node to have an interfering node which is on the same channel
when comparing to LID and LCC. However, the assignment but associated to another CH and using a different quasi-
of k and S values has not been discussed, the maintenance orthogonal “intracluster_code” (Gold, Kasami). CHs also
seems very complicated and the induced overhead is very maintain a list of the intracluster_codes used in the
high. neighboring clusters in order to avoid interferences and to
reduce the collisions (bit error rate).
III. CLUSTER FORMATION PROCEDURE
A. Preliminaries and Network Design C. State of a Node Inside the Network

An ad hoc network is formed by a set of nodes and links For cluster formation, we allocate IDs for the nodes and
that can be represented by a graph G V, E . V represents the set for the CHs. The 2-tuplet id , id H forms a unique identifier
of mobile nodes v . E represents the set of links between the for every node in the ad hoc network. Every node in the
connected nodes. The network cardinality is represented by: cluster will have information about its CH so that it can
n |V| and , | , , communicate across the cluster. More formally, every node
,
where , represents the received power from v at v . Assume v V will be identified by the following state:
that N v is the set of k-hops neighboring nodes of v and v : id , id H, w , intracluster_code, intercluster_code, sig_code, counter
∆ v is the cardinality of N v .The clustering is a graph
The weight parameter w is periodically calculated by each
partitioning problem with some added constraints and metrics.
node in order to indicate the suitability of a node for playing
More formally, we look for the set of vertices of CH’s role. The ‘counter’ is maintained by each node in order
cardinality N |C| such that: and to calculate the QoS level inside the cluster.
, | (1) D. Weight Calculation and Clusterhead Election
Vuong et al. [18] concluded that the construction of k-clusters Each node should be able to compute its weight based on
in an ad hoc network is a NP-Complet problem. For that several clustering metrics. The CH’ election is based on the
reason, we are focusing on the case of 1-clusters, where 1. weight values of the nodes. In comparison with other
More formally: , | , . approaches, we do not use neither the sum of distance for the
largest range coverage, nor the cumulative time during which
B. Code Assignment and Node States the node acts as a CH. The metrics used are the following:
Enabling CDMA-based solutions for ad hoc is fraught with a) The node degree ∆ : defined as the actual number
challenges, due to the absence of centralized control. An ad of neighbors for . It allows estimating the connectivity
hoc network is time-asynchronous system because it is degree and the position of that node regarding the other
generally not feasible to have a common time reference for all nodes.
the transmissions that arrive at the receiver side. In addition, it b) The transmission power : which allows electing the node
EL-BAZZAL et al.: A QUALITY OF SERVICE DRIVEN APPROACH 257

the can cover the largest range. A. Modeling the DCF Channel Parameters
c) The average speed : which is calculed as:
Note that the collisions and transmission errors can only
1 occur between the members of the same cluster due to the use
of the same CDMA intracluster_code inside that cluster. The
communications in neighboring clusters that use a distinct
Where , represents the node’s coordinate of at the quasi-orthogonal intracluster_code do not affect at all those of
moment . the current cluster. Therefore, to model the exponential
d) The remaining battery power : which allows extending backoff schema implemented in DCF 802.11 MAC layer, we
the lifetime of nodes by relinquish the role as a CH in case know that for each packet transmission, a node initializes a
of insufficient battery power. backoff time which is a random integer uniformly distributed
over the interval 0, W 1 . The value of W is called
Once these metrics are calculed, every node must calculate
contention window, and depends on the number of failed
its combined weight which is a good indication for its
transmissions of the packet. At the first transmission attempt,
suitability to be a CH. The node which has the best weight
the value of W is equal to CW called the minimum
value within the 1-hop neighboring is elected as CH. The
contention window.
weight calculation is done periodically and locally as
Let p be the probability that the transmitted packet faces a
following:
collision in the channel due to two or more nodes transmitting
; 1 (2) simultaneously in the same slot. In this case, after each
unsuccessful transmission, the value of W is doubled, up to a
We suppose that the nodes inside each cluster have the same
maximum value CW 2 CW where m represents the
needs. The idea is to combine each of the above system
number of unsuccessful attempts for this packet, i.e., the
parameters with certain weighing factors depending on the
maximum backoff stage. Once W reaches CW , it keeps this
system needs. The flexibility of changing the weighting
value until it returns to its initial value CW . Now, we can
factors helps us apply our algorithm to various scenarios. For
derive the general probability that the contention window W
example, in low mobility environment (conference room), we
that a node chooses and is given by:
can privilege the remaining battery parameter, thus the factor
‘b’ can be adjusted on the nodes. In a military environment p 1 p for W 2 CW
where the battery energy is always available, we can adjust the P Window W (3)
factor ‘d’ in order to elect the less mobile node as CH. On the p for W CW
other hand, we believe that the node average speed does not
reflect the relative mobility of that node. However, we still use The authors in [19] derived the collision probability p for
that parameter for initial CH election, because it is impossible the case of saturated network where a transmitting node has
to estimate the relative mobility when the node is alone in the persistently a queue of packets to send, so each incoming
zone without any other reference point (neighbors). During packet is immediately backlogged, i.e. it is preceded by a
maintenance procedures, the re-election is more sophisticated; backoff. At saturation, each packet is backlogged immediately.
we use the historical of the received power signals from the The average backoff window in the saturated case is given by:
one-hop neighboring to determine the relative mobility of a W 2W 2 W 2 W
node regarding the others. This mechanism is detailed in the 1 p p 1 p p 1 p p
2 2 2 2
section V. W
(4)

IV. ANALYTICAL MODEL FOR ADMISSION CONTROL BASED B. Modeling the Intraclusters Performance Parameters
ON CLUSTER QOS LEVEL
In this paper, we extend the model proposed in [19] to
Under overloading conditions in the network, the load obtain an approximate expression for collision probabilities in
balancing and the admission control are required in order to case of non-saturated arrival rates. Therefore, we use Poisson
provide a certain level of QoS for the participant nodes. The data source instead of saturated data source. We consider a
estimated knowledge of the number of members sharing an cluster with "N" members operating in discrete time where
802.11 cluster might effectively drive congestion avoidance on each member could be represented as an M/M/1 queuing
the CHs and inter-clusters load-balancing to achieve better system with an infinite storage, the packet arrival process is a
network resource utilization. First, we start by modeling the Poisson memoryless processes with a rate given by λ packets
intraclusters and interclusters performance parameters based while the packet service process of the network has an
on the code assignment procedure mentioned previously. exponential distribution with first moment μ which will be
More specifically, we investigate the impact of active nodes explained in subsequent sections. The probability that the
and wireless channel collisions on the performance of the packets’ interface queue is empty could be approximated by:
DCF. Based on this model, the cluster will be able to admit or π node 1
λ
.
reject the nodes by restricting the input traffic being admitted μ

to the system in order to maintain the QoS of the existing A packet is backlogged if the system is non-empty at the
members. instant of arrival. When an arbitrary arrival occurs, we
258 JOURNAL OF COMMUNICATIONS SOFTWARE AND SYSTEMS, VOL. 4, NO. 4, DECEMBER 2008

N
approximate the probability that the cluster (N members in P successful transmission α Nq 1 q , we also can have
N N
λ N P unsuccessful transmission β 1 1 q Nq 1 q ,
steady state) is empty by: ́ cluster 1 . and P idle slot γ 1 q N
. By resolving (6), we can obtain
μ
Then, the backoff window is 0 for any arbitrary packet the relationship between N and T as:
with probability ́ cluster , and it is backlogged with
probability 1 ́ cluster . Therefore, the average backoff (9)
window size for general (non-saturated) arrival rates is given
by: The intracluster delay D is defined by the average time from
the point when a packet becomes head of the node’s queue to
1 1 (5) the point when it is successfully received by the destination,
i.e., when a positive acknowledgment is received. We model
Let T be the saturation throughput inside the cluster, this delay without considering the waiting time in the packets’
defined as the expected time needed to transmit the data
payload with respect to idle, collision and header transmission interface queue before transmitting: D γ α . The
μ
time, during a cycle of frame exchange. A cycle of frame service rate μ is expressed via the average time required for
exchange consists of several collision cycles and one successful packet transmission, thus:
successful data frame transmission plus header transmission μ (10)
and idle times. γ β ∆ N

(6) ∆ represents the time that a node has to wait after an

unsuccessful transmission, before starting to re-sense the
α P successful transmission channel. Therefore ∆ SIFS ACK_Timeout . Owing to the
Where β P unsuccessful transmission assumption of independence at each retransmission, we can
γ P idle slot calculate N (the average number of unsuccessful
transmission of a packet) and approximate the average rate of
α represents the time that the channel is captured with a
successful node transmission, β represents the duration of a successful transmission per packet N which follows a
unsuccessful transmission, i.e., the time that the channel is geometric distribution having an average of . Knowing
captured by the node with errors (collisions on the same
that N defines the average number of node attempts to
intracluster_code) and γ represents the duration of a time slot.
The value of the parameters α and β differs depending on the successfully transmit its packet, i.e., the node has been
access model. Assuming that the packets are data fragment exposed to ( N 1 unsuccessful transmission before to
only, that means that there is no fragmentation. Thus, for basic successfully transmit its packet; therefore: N 1.
access mechanism: Finally, we have:
α 2 ∆
(11)
β (7)
C. Modeling the Interclusters Performance Parameters
represents the channel bit rate and ε represents the As for the interclusters communications, the CH must
average propagation delay of any frame on the channel. All route all intraclusters traffic; we assume that the CH has a
these parameters are defined in the IEEE 802.11 standard [20]. single antenna that serves either for transmission, or for
Assuming that at an instant t, an arrival to an idle node will reception (half duplex channel). Consequently, the CH cannot
not be backlogged even if there are some other backlogged communicate simultaneously with its members and with other
nodes having non empty queues. Now, considering the fact neighboring CH. When listening for intraclusters and
that the cluster contains N members and only those with a interclusters communications, the CH throughput will be
nonempty queue can actually collide with packets from other affected not only by its members but also by its neighboring
nodes. CHs. Note that the average number of clusters in the network
On the other hand, assuming q the probability that a node
(if the nodes are uniformly distributed) can be calculated as .
transmits in a randomly chosen slot in , the probability p
that a transmitted packet faces an unsuccessful transmission on On the other hand, when the CH is in transmission mode
the channel in a given slot will be equivalent to the probability and all the neighboring CHs are not transmitting; the CH does
that at least one of the N 1 remaining nodes transmits in not have to wait to immediately transmit a packet on the
the same time slot. In other words, if we assume that each intercluster_code channel. That means that the packet will be
remaining node transmits a packet with probability q in the directly transmitted if the CH is not in reception mode of
same time slot, thus the probability p that a collision occurs is incoming traffic on the intracluster_code channel and if the
given by: neighboring CHs do not interfere with its own transmissions
on the intercluster_code channel. We consider our CHs as
1 1 (8) traffic aggregation points, thus they will be saturated by
transmitting packets and serving the other nodes. Therefore,
the average backoff window size of a CH in saturated case is
N
Knowing that p 1 1 q , therefore: given by:
EL-BAZZAL et al.: A QUALITY OF SERVICE DRIVEN APPROACH 259

́ ́
(12)
́

Under the same conditions, the packet collision probability for

interclusters communications can be obtained by the following
equation:
́ N
́ 1 1 ́ ́ ́
(13)

On the other hand, assuming q́ the probability that a CH

transmits in a randomly slot in W, thus ṕ 1 1 q́ N .
Following the same definitions given in (6) while taking into
account the average number of clusters in the network, the
achievable interclusters throughput can be obtained by the
following equation:
́ ́ N
N
(14)
́ ́ N ́ N ́ ́ N ́ N
N N

By following the same definition given in (11), the

interclusters delay is defined as the average waiting time for a
successful transmission on the interclusters_code channel:
́ γW β ∆
D γW α ́
(15)

The average end to end delay " " depends on the average TABLE I
number of hops separating the source node from the MESSAGES INVOLVED IN THE CLUSTERING ALGORITHM
destination node:
Type of the message Message description
2 1 (16) hello ( id , id
N H, w , counter) To update the tables of the nodes
V. DISTRIBUTED MAINTENANCE PROCEDURE Join ( id , id H) To affiliate a cluster
A. Neighboring Discovery Mechanism Accept ( id , id H, w ) The CH accepts a Join
Every node (member or CH) has to maintain an Reject ( id , id H, w ) The CH rejects a Join
‘intracluster_table’ wherein the information on the local
members including the weights of each node is stored. CH_request ( id ) The node declares itself as CH
However, the CHs maintain another clusterhead information CH_response(
table ‘intercluster_table’ wherein the information about the id H , intracluster_code The CH accepts a CH_Request
other CHs is stored. In complex networks, the nodes must Join_accept ( id , id H, w , counter) The node accepts the Accept
coordinate between each other to update their tables. The hello
messages (TTL set to 1 hop) are used to complete this role. A CH_ACK The CH adds the node as a
(id , id H, w , intracluster_code, counter)
hello contains the full state of the node including its weight; it member
is periodically exchanged either between CHs or between each Database_info( id , id H , w , counter) The current CH sends the
CH and its members in order to update the database to a new elected CH
Database_ACK
‘intercluster_tables’ and the ‘intracluster_tables’ respectively. ( id , id H , w , counter)
The new elected CH accepts the
received database
Once a wireless node is activated, it has ( NULL , CH_change ( id ) The CH notifies a CH change
since it does not belong to any cluster. The node continuously
monitors the channel until it figures out that there are some CH_info ( id , id H , w , counter) The CH accepts the presence of a
new CH in the network
activities in its 1-hop neighborhood. This is due to the ability
leave ( id , id H) The node leaves the cluster
to receive the hello messages from other nodes in the network.
The node still has no stable state until it is fully identified by
the reception of the CH identifier, the intracluster_code and C. QoS Driven Procedure for Nodes Admission Control
the counter values. Table I illustrates the messages used for
the maintenance. The reason that a CH accepts or rejects a Join depends on
the ability to provide and guarantee the QoS level. We
B. New Arrival Nodes Maintenance Mechanism proposed an admission control mechanism based on the
Every node ( NULL broadcasts a Join message analytical model presented previously in order to estimate the
(TTL set to 1 hop) on the sig_code channel in order to join the achievable throughput and delay. This mechanism will be only
most powerful 1-hop clusterhead. Thus, it waits either for an implemented on the CHs and applied during the maintenance
Accept or for a Reject from a CH. More formally, this so that they can take decisions to accept and/or reject the
mechanism can be written as the following: requests while guarantying an acceptable QoS level. More
260 JOURNAL OF COMMUNICATIONS SOFTWARE AND SYSTEMS, VOL. 4, NO. 4, DECEMBER 2008

formally, this procedure can be written as the following:

D. Clusterheads Maintenance Mechanism

A CH has to calculate periodically its
F. Member Nodes Maintenance Mechanism
weight, and sends periodically hello messages to its members
and to the neighboring CHs in order to update the After joining a cluster, the node declares itself as a
intraclusters_tables and interclusters_tables respectively. member of this cluster; it calculates and sends periodically its
When the CH receives a Leave message, it updates the weight in hello messages towards its CH. More formally, this
intraclsuter_table and broadcasts a hello message to its mechanism can be written as the following:
members and to its neighboring CHs to inform them that its
previous ‘counter’ was decremented. When the CH receives a
hello from a neighboring CH, it updates the interlcuster_table.
If the hello’s source is a node member, the CH updates the
intracluster_table and verifies the weights. In case of better
weight is found, the CH must invoke the re-election
procedure. We restrict this procedure to the CHs in order to
simplify the maintenance and the complexity of the cluster
management. The re-election does not necessarily mean that a
new CH must be elected even if there is a member node
having a better weight, we will explain in details this
procedure in subsequent section.
When the CH receives a Join ( ) from a new
arrival node or a Join ( ) from a node which belongs
to another cluster , the CH must invoke the handoff
procedure explained in subsequent section in order to admit or
to reject the request based on the admission control
mechanism presented previously. On the other hand, in order
to maximize the scalability of the architecture, if the CH
receives a CH_Request from a node desiring to be CH, it must
accept the request without performing any admission control;
the CH sends a CH_Response to the node, updates the
intercluster_table and broadcasts a hello message to the
neighboring CHs. Thus, an interclusters link is established on
the intercluster_code channel.

E. Handoff Procedure
We have tried to balance the load between all the clusters
in the network in order to give more flexibility to the members
by allowing them to leave a weak CH and join another one G. Re-election Procedure
which seems stronger than the current one. However, in order The re-election is not periodically invoked; it is performed
to maintain a certain QoS level, the solicited CH must perform by the CH just in case of a best received weight, it allows
an admission control so that the handoff is not accepted in the minimizing the generated overhead and the utilization of
detriment of the QoS level of the existing members. More nodes’ resources. As we explained above, the re-election may
formally, this procedure can be written as the following: not result a new CH, it depends on the stability of the new
EL-BAZZAL et al.: A QUALITY OF SERVICE DRIVEN APPROACH 261

node for playing the CH’s role. In the case where a new CH lower is the throughput. This condition remains valid until the
must be elected, the procedure should be soft and flexible value of the window “W” is approximately equal to 500.
while copying the databases from the old to the new CH. We
We see that an higher value of ‘W” tends to give better
limit the execution of the algorithm where there is a CH’s
throughput performance in the case of large members’ number
change to minimize the effect on the whole topology. Thus the
in the cluster, while it drastically penalizes the throughput in
furthest nodes are not affected by any problem which occurs in
the case of small members’ number. On the other hand, we
other clusters. More formally, this procedure can be written as
neglect the probability of having great values of W in case of
the following:
small number of members, since the probability of collisions
in this case is very small in comparison to the case of large
number of members.

1,00

intracluster throuphput (Mbps)

0,95
0,90
0,85
0,80
0,75
0,70 N = 5, m = 6, data = 8184
N = 15, m = 6, data = 8184
0,65
N = 25, m = 6, data = 8184
0,60
32 232 432 632 832 1032
window size (slots)
VI. IMPLEMENTATION AND PERFORMANCE ANALYSIS
Fig. 1. Cluster throughput versus backoff window size
A. Simulation Environment and Parameters
This behavior is also seen in figure 2, the amount of
We have used the parameters specified in IEEE 802.11b channel time wasted in collisions is extremely large for a small
specifications [20] which define the frame sizes of the MAC value W and a large number of members. Large value of W
layer and the Direct Sequence Spread Spectrum (DSSS) may, in fact, increase the delay; because for small W, the
physical layer parameters used the previous analytical model. amount of idle slot times per packet transmission is very low.
In the rest of the paper, the channel bit rate has been This value becomes significant only when W gets greater and
assumed equal to 1 Mbps (for a range of 90 meters) without the number of members is small. When the number of
RTS/CTS. On the other hand, the generation of appropriate members is large, the throughputs of large windows are very
scenarios with realistic mobility pattern is very challenging close, but the cluster delay deteriorates much more severely
when designing ad hoc networks. Agba and al. [21] have for smaller contention windows.
developed a simulation tool including a scenarios generator
21
and a propagation modeler. It allows a complete description of N = 5, m = 6, data = 8184
environmental parameters, mobility model parameters and 20 N = 15, m = 6, data = 8184
intracluster delay (msec)

other simulation settings. 18 N = 25, m = 6, data = 8184

The simulations scenarios used in this paper were 17
randomly generated based on the Random Waypoint mobility
model (RWM). For the physical layer, a semi-deterministic 15
channel model that takes into account the path-loss calculation 14
with respect of 3D environment parameters is used. The model
also allows defining the maximum radio range as the radius of 12
a mobile when operating at full transmission power and 11
having an effective communication range. The network size is
9
500m x 500m. The number of nodes used in the whole
network varies between 20 and 100. All the nodes follow the 32 232 432 632 832 1032
window size (slots)
RWM used in the scenario generator with speed ranging from
Fig. 2. Cluster delay versus backoff window size
3 to 10 Km/h. The hello interval was set to 5 seconds and the
simulations were run for 300 seconds. On the other hand, figure 3 shows that the number of
transmissions per packet significantly increases as the window
B. Cluster Performance Analysis and Discussion
size decreases. This effect is much more significant for large
Figure 1 shows that the throughput strongly depends on the number of members. Therefore, we can settle the fact that the
number of members inside the cluster. We realize that, in most clustering of an ad hoc network cannot be concluded by the
cases, the greater is the number of the active members, the randomly execution of an algorithm regardless the
262 JOURNAL OF COMMUNICATIONS SOFTWARE AND SYSTEMS, VOL. 4, NO. 4, DECEMBER 2008

specifications of the applications used inside the cluster. By

using a sophisticated algorithm like ours (ECA), we can adjust 0,92

achievable throughput (Mbps)

the parameters of this algorithm in order to generate a certain 0,90
number of clusters which fulfill the requirement of each 0,88
member in terms of the required throughput, the tolerated 0,86
delay and the allowed retransmission number per packet. 0,84
0,82
0,80
0,80 0,78 Theoretical: n=20, data=8184
0,76 Simulation: n=20, data=8184
no of retransmission per packet

0,70 N= 5, m = 6, data = 8184

N = 15, m = 6, data = 8184 0,74 Theoretical: n=100, data=8184
0,60 N = 25, m = 6, data = 8184 0,72 Simulation: n=100, data=8184
0,50 0,70

0,40 30 60 90 120 150 180 210 240 270 300

transmission range (meters)
0,30
Fig. 5. Achievable throughput (theoretical vs. simulation)
0,20
C. Performance Analysis of the Maintenance Procedures
0,10
Figure 6 shows that for small ranges, most of nodes remain
0,00 out of each other’s transmission range, thus the number of
32 232 432 632 832 clusters is relatively high and the network may become
window size (slots) disconnected because there are no other choices. When
Fig. 3. Average no. of retransmission per packet vs. window size transmission range increases, more nodes can hear each other.
The average number of clusters formed decreases and the
clusters become larger in size. When the transmission range is
Figure 4 plots the achievable throughput versus the packet very small, most of nodes form one node cluster which only
size for three different cluster sizes. We see that when the consists of itself.
maximum channel bit rate is equal to 1Mbps, the On the other hand, when the transmission range begins to
throughput efficiency increases (approaches to ) as the be larger, mobile nodes tend to remain in the range of their
packet size increases. The situation is explained by neighbors, the clusters are less dynamic and the number of
considering that the time spent for frame transmission is CHs’ transitions decreases as depicted in figure 7. In
decreased as the data rate increases but the time overhead disconnected networks, we can conclude a high rate of
spent on DIFS, SIFS and the backoff delay remains the same. transitions on CHs. However, we argue that this will not affect
We can conclude that the choice of a smaller number of network performance as this will only occur when the network
members performs better in the whole cluster. We fixed that is disconnected (A disconnected network is unable to function
number to 15 nodes to define the cluster size “N” with respect too).
to the QoS parameters requested by the current applications. 60
n = 20 nodes
50 n = 40 nodes
n = 60 nodes
avg. no. of clusters

1
40 n = 80 nodes
intracluster throughput (Mbps)

0,9 n = 100 nodes

30
0,8
20
0,7
0,6 10

0,5 0
N = 5, m = 6, w = 32
N = 15, m = 6, w = 32 30 60 90 120 150 180 210 240 270 300
0,4 transmission range (meters)
N = 25, m = 6, w = 32
0,3 Fig. 6. Avg. number of clusters vs. transmission range

500 2000 3500 5000 6500 8000 9500 11000 We have also tried to balance the load between all the
packet size (bits) clusters; we have based on the numerical results obtained from
Fig. 4. Cluster throughput versus packet size the proposed analytical model to choose the optimal cluster
size in respect to the QoS level. The results shown in figure 8
Figure 5 shows a strong correlation between the theoretical show that none of the clusters has reached a 100 % of load
and simulation throughput results for different transmission rate. The load is balanced between the different clusters even
range and network density. The average deviation of the when we increase the number of nodes in the network. This
simulation from theory is very small. This comes to validate allows avoiding congestion on CHs and guarantying the
the analytical model proposed in this paper. scalability issue.
EL-BAZZAL et al.: A QUALITY OF SERVICE DRIVEN APPROACH 263

7 high transmission range (more than 250 m), WCA generates

n = 20 nodes less CH than ECA but to the detriment of a large number of
6 n = 40 nodes
transition on each CH (figure 11), where the stability is one of
CH transition state

n = 60 nodes
n = 80 nodes the important criteria in clustering because the frequent
5 changes of CH adversely affect the performance of the
n = 100 nodes
4 clustering algorithm.
35 ECA [60 m]
3 WCA [60 m]
30 ECA [180 m]
2 WCA [180 m]

avg. no. of clusters

25 ECA [300 m]
WCA [300 m]
1 20
30 60 90
120 150 180 210 240 270 300
transmission range (meters) 15
Fig. 7. Avg. no. of CH transition state vs. transmission range 10
Figure 9 shows that the probability that every node can 5
reach the other nodes is very high. The situation is improved
when the topology is denser. By increasing the transmission 0
range, the CH will be able to establish more intercluster links; 20 40 60 80 100
which improves the global connectivity in the network. We network density (nodes)
can also remark fewer fluctuations when the number of nodes Fig. 10. Avg. no.of clusters vs. network density
increases in the network; this is also a good indication on the As shown in figure 11, with 100 nodes in the ad hoc
stability level and the scalability of the approach. network and for a transmission range equal to 180m, the
90 proposed algorithm produced about 50.0% to 83.3% less
n = 20 nodes
80 transitions on each CH than WCA. In the WCA algorithm,
n = 40 nodes
70 n = 60 nodes
WCA will keep changing with changes in topology. The CH
of WCA algorithm relinquishes its position when another node
cluster load rate

60 n = 80 nodes
50 n = 100 nodes having better weight joins the cluster. In our algorithm, the
CH has to verify the suitability of a new election even if a new
40
node having better weight has joined the cluster. As a result,
30
our algorithm gives better performance in terms of stability
20 when the node density in the network is high.
10
0 25
30 60 90 120 150 180 210 240 270 300
transmission range (meters) 20
CH transition state

Fig. 8. Cluster load rate vs. transmission range ECA [60 m]

15 WCA [60 m]
100 ECA [180 m]
90 WCA [180 m]
10 ECA [300 m]
connectivity factor in %

80 WCA [300 m]
70
5
60
50
0
40 n = 20 nodes
30 n = 40 nodes 20 40 60 80 100
20 n = 60 nodes network density (nodes)
n = 80 nodes Fig. 11. Avg. no. of CH transition state vs. network density
10 n = 100 nodes
0
As shown in figure 12, for a transmission range of 120
30 60 90 120 150 180 210 240 270 300
transmission range (meters)
meters, the number of handoffs increased when varying the
number of nodes in the network for both our algorithm and
Fig. 9. Connectivity factor vs. transmission range
WCA. As the number of nodes increased, the increasing rate
We also compare the performance of our approach with the slowed down in ECA, which was not the case in WCA. For a
corresponding performance of the WCA algorithm while the node speed varying between 3 and 10 km/h, when there were
nodes are moving under the same conditions. In figure 10, we 20 nodes in the network and for the same transmission range
note that the performance difference is small between WCA (120 m), the proposed algorithm produced 61.5% less
and ECA with respect to the average number of clusters. This handoffs than WCA. When the number of nodes was increased
is because both algorithms are variations of a local weight to 100, our algorithm gave 66.5% less handoffs than WCA for
based clustering technique that forms one-hop clusters. For the same node speed.
264 JOURNAL OF COMMUNICATIONS SOFTWARE AND SYSTEMS, VOL. 4, NO. 4, DECEMBER 2008

35 Simulation results indicated that the model agrees well with

ECA [range=120 m] the behavior of the algorithm. For future work, we are
30 WCA [range=120 m] planning to investigate the model in presence of clusterheads
avg. no. of handoffs

that are MIMO capable; more research must be done in case of

25
heterogeneous nodes inside the network, especially where the
20
nodes follow distinct mobility models.

REFERENCES
15
[1] Das B., Bharghavan V., "Routing in ad-hoc networks using
10 minimum connected dominating sets," IEEE International
Conference on Communications (ICC Montreal 97), vol. 1, Jun.
5 1997, pp. 376-380.
20 40 60 80 100 [2] Parekh A.K., "Selecting routers in ad-hoc wireless networks,"
network density (nodes) Proceedings of the SBT/IEEE International Telecommunications
Fig. 12. Avg. no. of handoffs vs. network density Symposium, Aug. 1994.
[3] Gerla M., Tsai J. T. C., "Multicluster, Mobile, Multimedia Radio
Network," ACM/Baltzer Wireless Networks Journal 95, vol. 1,
Figure 13 shows the connectivity improvement of ECA in Oct. 1995, pp. 255-265.
comparison with WCA. We also remark more fluctuations in [4] Baker D.J., Ephremides A., "A distributed algorithm for
WCA; ECA tends to stabilize more quickly. This is explained organizing mobile radio telecommunication networks,"
by the higher number of transitions on the WCA’s CHs, Proceedings of the 2nd International Conference on Distributed
causing more frequent breaks of links and reducing the overall Computer Systems, Apr. 1981, pp. 476-483.
stability level. [5] Baker D.J., Ephremides A., "The architectural organization of a
100 mobile radio network via a distributed algorithm," IEEE
Transactions on Communications, Nov. 1981, pp. 1694-1701.
90
[6] Ephremides A., Wieselthier J. E., Baker D. J., "A Design Concept
80 for Reliable Mobile Radio Networks with Frequency Hopping
connectivity factor in %

70 Signaling" Proceedings of the IEEE, vol. 75, no. 1, Jan. 1987, pp.
60 56-73.
[7] Lian, Jie, Kshirasagar Naik et Gordon B. Agnew. 2007. « A
50
Framework for Evaluating the Performance of Cluster
40 WCA [n = 20 nodes] Algorithms for Hierarchical Networks ». In IEEE/ACM
30 ECA [n = 20 nodes] Transactions on networking. vol. 15, no 6.
WCA [n = 60 nodes] [8] Chiang C. C., Wu H. K., Liu W., Gerla M., "Routing in Clustered
20 ECA [n = 60 nodes] Multihop, Mobile Wireless Networks with Fading Channel,"
10 WCA [n = 100 nodes]
ECA [n = 100 nodes] IEEE SICON, Singapore, Apr. 1997.
0 [9] Yu, J. Y. et P. H. J. Chong. 2003. « 3hBAC (3-hop between
30 60 90 120 150 180 210 240 270 300 Adjacent Clusterheads): a Novel Non-overlapping Clustering
transmission range (meters) Algorithm for Mobile Ad hoc Networks ». In Proceedings IEEE
Fig. 13. Connectivity factor vs. transmission range Pacrim’03. (Victoria, 2003), vol. 1, p. 318-321.
[10] McDonald, Bruce et Taib F. Znati. 2001. « Design and
VII. CONCLUSION AND FUTURE WORK Performance of a Distributed Dynamic Clustering Algorithm for
This paper has presented a Quality of Service Driven Ad-Hoc Networks ». In Proceedings 34th Annual Simulation
Symposium. p. 27-35.
Approach for Clustering in Mobile Ad hoc Networks Based on
[11] Basu P., Kham N., Little T. D. C., "A mobility based metric for
Metrics Adaptation. It has the flexibility of assigning different clustering in mobile ad hoc networks," International Conference
weights and takes into account a combined metrics to form on Distributed Computing Systems Workshop, Apr. 2001.
clusters automatically. An analytical model has been proposed [12] Amis, Alan D. et Ravi Prakash. 2000. « Load-Balancing
in order to estimate the QoS parameters when we follow a Clusters in Wireless Ad hoc Networks ». In Proceedings 3rd
specific CDMA code assignment mechanism in the whole IEEE Application-Specific Systems and Software Engineering
network. We have concluded that the cluster performance Technology. p. 25–32.
strongly depends on the number of members. Limiting the [13] Basagni S., "Distributed clustering for ad hoc networks,"
number of nodes inside a cluster allows restricting the number Proceedings of International Symposium on Parallel
Architectures, Algorithms and Networks, Jun. 1999, pp. 310-315.
of nodes catered by a clusterhead so that it does not degrade
[14] Basagni S., "Distributed and mobility-adaptive clustering for
the MAC functioning. multimedia support in multi-hop wireless networks," Proceedings
A clusterhead with constrained energy may drain its of Vehicular Technology Conference, VTC, vol. 2, fall 1999, pp.
battery quickly due to heavy utilization; in order to spread the 889-893.
[15] Chatterjee M., Das S.K., Turgut D., "WCA: A Weighted
energy usage over the network and achieve a better load Clustering Algorithm for Mobile Ad Hoc Networks," Cluster
balancing among clusterheads, re-election of the clusterheads Computing Journal, vol. 5, no. 2, Apr. 2002, pp. 193-204.
may be a useful strategy; the algorithm is executed only when [16] Venkataraman, Gayathri, Sabu Emmanuel et Srikanthan
there is a demand. Also, if a node is moving away from the Thambipillai. 2007. « Size-restricted cluster formation and cluster
clusterhead, then the algorithm is flexible and cheap enough to maintenance technique for mobile ad hoc networks ». In
be applied iteratively as the network configuration changes. International Journal of Network Management. p. 171-194.
EL-BAZZAL et al.: A QUALITY OF SERVICE DRIVEN APPROACH 265

[17]] G. Bianchi, "Performancee analysis of the IEEE 802.11 8 Basile L. Aggba (M’ 2005) was born in Kara, K
distributed cooordination funcction", IEEE Journal on Seelected Togo. He recceived his B.S in physics in 1998
Areas in Commmunications, Vool. 18, No. 3, 20000, pp. 535-5447. (university of Lomé, Togo o). He receivedd the
[18]] Vuong, Thai H. P. et Dung T. Huynh. 20000. « Adapting d-hop M.S degrree in high frequuency
dominating setts to topology changes in Add hoc networkss ». In telecommuniccations in 2001 1 and he comppleted
Ninth Internatiional Conferennce on Compuuter Communiccations the Ph.D. deggree in high frequencies electrronics
and Networks. p. 348-353. and optoelecctronics in 20 004 (Universitty of
[19]] Y. Tay and K.
K Chua, "A cappacity analysis for the IEEE 802.11
8 Limoges, Fraance). From 20 005 to 2007, he is
MAC protocoll", Wireless Networks, Vol. 7, No. 2, 20001, pp. Postdoctoral Fellow at Elecctrical Engineeering
159-171. Deparrtment (École de Technologiie Supérieure, Montreal
M - Cannada)
[20]] IEEE Standarrd for Wireless LAN Mediium Access Control C wheree he mainly worked
w on tactical ad hoc networks
n projeect in
(MAC) and Phyysical Layer (PPHY) Specificattions, 1999. partneership with Ultra
U Electroniccs (TCS). Sincce 2007, Dr. Agba
[21]] Agba L., Gaggnon F., Kouki A., "Scenarios generator for ad a hoc workeed as Senior Researcher
R in teelecommunicatiions at the Research
networks," Intternational Symmposium on Inndustrial Electrronics, Institu
ute of Hydro-QQuebec, IREQ. His main ressearch interestss are:
Montreal, Canaada, Jul. 2006. channnel modeling paarticularly in hiigh voltage env
vironments, wirreless
netwo orks design, deployment
d sofftware for wirreless systemss and
antennna design.

Mohamad Haaidar received his B.S. degree in

Zouhair El Bazzal
B received the B.S. deggree in Electrical Enngineering fro om the Lebaanese
Computer Sciences froom the Lebbanese American Unniversity, Byblo os, Lebanon in 2000.
2
University, Beirut, Lebanoon in 1999, thee M.S The M.S. deggree was receiv ved from Roossevelt
degree in tellecommunicatioon networks inn 2001 University in Telecommuniccations, Chicagoo, IL,
from the Ecole
E Superieurre des Ingenieeurs à in 2002 and P Ph.D. in Applieed Science fromm the
Beirut, Sainnt-Joseph Univeersity, Lebanonn, and University off Arkansas at Little
L Rock (UAALR),
the Ph.D. deegree in engineering in 20088 from Little Rock, AR, in 2007. His major fieeld of
the University of Quebec, Canada. Since 2006, research is in wireless netwoorking and
he is working as a lecturer in
i the Departmeent of commmunications. Previously, he diid a one year post-doc
p at Ecoole de
Softwware and Elecctrical Engineeering at the Éccole de Technnologie Technnol,ogie Superrieure, Montreeal, Quebec, in i 2008 wherre he
Supéérieure, Montreeal. Also, Dr. El-Bazzal
E is cuurrently workinng as a developed a power control algoritthm for Cognitive Wireless Mesh M
Reseearcher at Ericssson Canada Innc. His main reesearch interests are: Networks. Prior to that, he workeed as a Develo opment Engineeer at
wireeless networkinng and communnications, Longg Term Evoluttion in IBM, Durham, NC, teacher assistaant at UALR, Little L Rock, ARR, and
3GPPP, IP mobillity, load ballancing, admiission control,, and WiMeesh/WiMax RF F Engineer Annalyst on a UA ALR-Alltel-Mottorola
deveelopment of Pollicy and qualityy of service toolls for LTE. projecct. His researchh interests are looad balancing, channel assignmment,
MAC C and PHY layyers design in W WLAN, Cogniitive Wireless MeshM
Michel Kadocch (S’67, M’777, SM’04) recceived Networks, Ad Hooc Networks, sensor netwo orks, and WiMax
W
the B. Eng froom Sir George Williams Univversity enviro
onments.
(Canada) in 1971, the M. Eng from Caarleton
(Canada) in 1974,
1 MBA froom McGill (Caanada) François Gagnon (S’87-M’8 87-SM’99) was born
in 1983 and thet Ph.D from Concordia (Caanada) in Québec Ciity, QC, Canad da. He received the
in 1991. He is a full proofessor at Ecoole de B.Eng. and Ph.D. degrrees in elecctrical
technologie supérieure ETS S (Canada) annd the engineering from École Polytechniquee de
director of thee Master Progrram in engineeering. Montreal, Moontreal, QC, Canada.
C Since 1991,
He is also ann adjunct profeessor at Concoordia he has been a Professor in the Departmeent of
Univversity (Canadaa). He is presenntly working onn Cross layer, and
a on Electrical E Engineering at a the Écolee de
Reliiable multicastt in wireless Ad hoc and WiMax netw works. Technologie Supérieure, Montreal.
M He has
Proffessor Kadoch has
h published manym articles and
a is the authoor of a chaired the deepartment from 1999 to 2001 and
a is
bookk « Protocoles et réseaux loccaux ». He hass been involveed for now the holder of the Ultra Elecctronics (TCS)) Chair in Wirreless
manny years at ITU
U-T as a speciaal rapporteur annd with the inddustry Comm t same univeersity. His reseearch interest covers
munication at the
nam
mely Teleglobe Canada, CAE, and Communnication Canadda. He wireleess high-speed communicationns, modulation,, coding, high-sspeed
has been a consultant with Harriss, Bell South, BCB Tell, Concerrt and DSP implementationns, and militarry point-to-point communicattions.
Brittish Telecom UK, as well as thhe Commonw wealth He haas been very innvolved in the creation of thee new generatioon of
Teleecommunicationn Organization on issues relateed to installatioon and high-ccapacity line-oof-sight militaryy radios offereed by the Canadian
optimization of SO
ONET, telephonne and high speeed data networkks and Marco oni Corporatioon, which is now Ultra Electronics
E Tactical
servvices. Comm munication Systems.