IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO.

6, NOVEMBER 2007 3309

Clustering-Based Multichannel MAC Protocols for

QoS Provisionings Over Vehicular Ad Hoc Networks
Hang Su and Xi Zhang, Senior Member, IEEE

Abstract—Making the best use of the Dedicated Short Range (V2V) communication capability will allow large-scale sens-
Communications multichannel architecture, we propose a cluster- ing, decision, and control actions in support of these objec-
based multichannel communications scheme that can support not tives. The allocation of 75 MHz in the 5.9-GHz band that is
only public-safety message delivery but also a wide range of future
multimedia (e.g., video/audio) and data (e.g., e-maps, road/vehicle licensed for Dedicated Short-Range Communications (DSRC)
traffic/weather information) applications. Our proposed scheme [1], which supports seven separate channels, may also enable
integrates clustering with contention-free and/or -based medium the future delivery of rich multimedia contents to vehicles at
access control (MAC) protocols. In our scheme, the elected cluster- short-to-medium range via either V2V or vehicle-to-roadside
head vehicle functions as the coordinator to collect/deliver real- (V2R) links in Vehicular Ad Hoc NETworks (VANETs).
time safety messages within its own cluster and forward the
consolidated safety messages to the neighboring cluster heads. In While there has been a large body in the literature studying
addition, the cluster-head vehicle controls channel assignments both V2V [2]–[6] and V2R [7], [8] networks, there are several
for cluster-member vehicles transmitting/receiving nonreal-time advantages of using V2V-based VANETs as compared with the
traffics, which makes the wireless channels more efficiently utilized V2R-based VANETs. First, the V2V-based VANET is more
for vehicle-to-vehicle (V2V) nonreal-time data transmissions. Our flexible and independent of the roadside conditions, which is
scheme uses the contention-free MAC within a cluster and the
contention-based IEEE 802.11 MAC among cluster-head vehicles particularly attractive for most developing countries or remote
such that the real-time delivery of safety messages can be guar- rural areas where roadside infrastructures are not necessarily
anteed. Under our proposed scheme, we develop an analytical available/furnished. Second, V2V-based VANET is less expen-
model to study the delay for the consolidated safety messages sive than the V2R-based one since it does not need expensive
transmitted by the cluster-head vehicles. Based on this analytical roadside infrastructures. Third, V2V-based VANET can avoid
model, we derive the desirable contention-window size, which can
best balance the tradeoff between the delay of safety messages and the fast fading, short connectivity time, high frequent handoffs,
the successful rate of delivering safety messages. The extensive etc., that are caused by the high relative-speed difference be-
simulation results show that, under various highway traffic scenar- tween the fast-moving vehicles and the stationary base stations.
ios, our proposed scheme can efficiently support the nonreal-time Finally, the V2V-based VANET much better fits vehicle-related
traffics while guaranteeing the real-time delivery of the safety applications, which only needs to exchange data/information
messages.
among neighboring vehicles within their nearby areas. Moti-
Index Terms—Clustering algorithms, Dedicated Short-Range vated by the aforementioned observations, in this paper, we will
Communications (DSRC), multichannel medium-access control focus on V2V-based VANETs.
(MMAC), protocol designs, time–division multiple access
(TDMA), vehicle-to-vehicle (V2V) communications, vehicular The data that are transmitted over the VANETs can be
ad hoc networks (VANETs). classified into real-time (such as safety messages and video/
audio signals) and nonreal-time traffics (such as e-maps and
road/vehicle–traffic/weather information), which impose the
I. I NTRODUCTION
diverse quality of service (QoS) requirements for VANET

I NTELLIGENT transport system architecture provides a

framework for the much needed overhaul of the highway
transportation infrastructure. The immediate impacts include
designs. Supporting the delay-bound QoS is challenging when
the VANET is under contention-based (e.g., IEEE 802.11 proto-
col) environments, where the packet delay and data-congestion
alleviating vehicle ndash;traffic congestion and improving op- level increase dramatically as the total number of vehicles
erations management in support of public safety goals, such contending for the common wireless media (and, thus, the
as collision avoidance. Equipping vehicles with various kinds collision rate) gets large. On the other hand, clustering [9], [10]
of on-board sensors and instrumenting the vehicle-to-vehicle is an efficient technique to reduce data congestion and sup-
port QoS over wireless networks. To provision QoS over our
V2V networks and reduce data congestion, under the DSRC
Manuscript received February 28, 2007; revised July 6, 2007 and July 14,
2007. This work was supported in part by the National Science Foundation multichannel architecture, we propose a distributed cluster-
CAREER Award under Grant ECS-0348694. The review of this paper was based multichannel communications scheme that integrates
coordinated by Prof. X. Shen. the clustering with contention-free/contention-based medium
The authors are with the Networking and Information Systems Labo-
ratory, Department of Electrical and Computer Engineering, Texas A&M access control (MAC) protocols.
University, College Station, TX 77843 USA (e-mail: [email protected]; Our proposed scheme mainly consists of the following three
[email protected]). core protocols: First, the Cluster Configuration Protocol groups
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. all vehicles in the same direction into clusters, each containing
Digital Object Identifier 10.1109/TVT.2007.907233 a cluster-head vehicle that is elected. This is viable since the

0018-9545/$25.00 © 2007 IEEE

3310 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 6, NOVEMBER 2007

vehicles flowing in the same direction share similar speeds TABLE I

DEFINITIONS OF THE SEVEN DSRC CHANNELS FOR OUR SCHEME
and the moving patterns, which are regulated by the traffic
laws and road structures/constructions, resulting in the rela-
tively stable topology in each cluster. Second, the Intercluster
Communication Protocol dictates the transmissions of real-time
safety messages and nonreal-time traffics among clusters over
two separate IEEE 802.11 MAC-based channels, respectively.
Third, the Intracluster Coordination and Communication Pro- is the dedicated control channel for delivering safety messages
tocol employs multichannel MAC (MMAC) algorithms for and announcements, Ch172 is the high-availability-and-low-
each cluster-head vehicle to conduct the following two major latency channel for vehicle safety and high-priority applica-
tasks within its own cluster: 1) collecting/delivering safety tions, and the remaining five are unreserved service channels
messages from/to cluster-member vehicles using the upstream [11]. Complying with the DSRC’s seven-channel bandplan, we
time-division multiple-access (TDMA)/downstream-broadcast define the particular functions for these seven channels in our
method and 2) allocating available data channels to cluster- scheme, which are summarized by Table I. Specifically, the
member vehicles for nonreal-time traffics. definitions of these seven channels in our scheme are given
Based on our proposed scheme, we apply the regenerative as follows: Ch178 is the intercluster control (ICC) channel,
process technique to develop an analytical model for investi- Ch174 is the intercluster data (ICD) channel, Ch172 is the clus-
gating the safety messages’ transmission delay and successful ter range control (CRC) channel, and the remaining channels
delivery rate, which identifies the tradeoff between them in (Ch176, 180, 182, and 184) are the cluster range data (CRD)
terms of the contention-window size. Then, using the developed channels.
analytical model, we further derive the desirable contention- In our proposed scheme, each vehicle is equipped with two
window size, which can best balance the tradeoff between sets of transceivers,1 which are denoted by Transceiver I and
the delay of safety messages and the successful rate of deliv- Transceiver II, respectively, which can operate simultaneously
ering safety messages. The extensive simulation results that on different channels. As shown in Fig. 1, our proposed scheme
were obtained show that our proposed scheme can efficiently works as follows: The vehicles within a nearby proximity
support vehicular nonreal-time data communications while form a cluster, where one of them is elected to act as cluster
guaranteeing real-time delivery of safety messages over the head based on the given election rules (refer to Section III-A
VANETs. for details). In each cluster-head vehicle, one transceiver uses
The rest of this paper is organized as follows: Section II contention-free TDMA-based MAC over the CRC channel to
presents the system architecture. Section III proposes our dis- collect and deliver safety messages as well as control packets
tributed cluster-based multichannel communications scheme. within this cluster, while the other transceiver exchanges con-
Section IV develops an analytical model to derive the average solidated safety messages among cluster-head vehicles through
delivery delay of safety messages. Section V derives the suc- contention-based MAC over the ICC channel. In each cluster-
cessful delivery rate of safety messages and also determines member vehicle, one transceiver is dedicated for communicat-
the desirable contention-window size for best balancing the ing with the cluster-head vehicle over the CRC channel within
tradeoff between the delay of safety messages and the suc- its cluster, and the other transceiver can be used to transmit
cessful safety-message delivery rate. Section VI defines the the nonreal-time traffic over one of the ICD/CRD channels that
performance evaluation metrics for our proposed scheme and were assigned by the cluster-head vehicle.
then evaluates our proposed scheme through the extensive As shown in Fig. 2, our proposed scheme handles the
simulations. This paper concludes with Section VII. following three tasks: 1) cluster-membership management;
2) real-time traffic (such as safety messages) delivery; and
3) nonreal-time data communications, such as e-maps down-
II. S YSTEM A RCHITECTURE load and movies download. Then, to accomplish the system
Our proposed scheme aims at supporting QoS for timely functions of our proposed scheme, we develop three differ-
delivery of real-time data (e.g., safety messages and platoon ent protocols: 1) the Cluster Configuration Protocol; 2) the
commands) and increasing the throughput for nonreal-time Intercluster Communication Protocol; and 3) the Intracluster
traffics (e.g., e-maps download and movie downloads) over Coordination and Communication Protocol, as shown in Fig. 2.
the V2V-based VANETs. To achieve these goals, we devel- First, the Cluster Configuration Protocol employs contention-
op the cluster-based multichannel communications scheme un- based MAC over the ICC channel to perform cluster man-
der the infrastructure-free VANET environments. The key of agement tasks, such as joining and leaving a cluster and
our proposed scheme is to integrate the clustering algorithm cluster-head election. Second, the Intercluster Communication
with both the contention-free and contention-based MAC proto- Protocol is responsible for the exchange of safety messages
cols under the DSRC architecture. The Federal Communication and nonreal-time traffics among clusters over the ICC and
Committee mandates the seven-channel bandplan to the DSRC ICD channels, respectively. Third, the Intracluster Coordination
standard for vehicle communications. In particular, DSRC’s and Communication Protocol utilizes the MMAC protocol to
75-MHz bandwidth at 5.9-GHz band is divided into seven chan-
nels, including Ch172, Ch174, Ch176, Ch178, Ch180, Ch182, 1 Note that the cost of two sets of transceivers is practically trivial, as
and Ch184, each spanning 10-MHz bandwidth, where Ch178 compared to the cost of the vehicle itself.
SU AND ZHANG: CLUSTERING-BASED MMAC PROTOCOLS FOR QoS PROVISIONINGS OVER VANETs 3311

Fig. 1. Our proposed cluster-based multichannel communications architecture. The vehicles in white color represent the elected cluster-head vehicles, and the
radio-wave symbol indicates that two vehicles are performing point-to-point communication.

Fig. 2. Cluster-based multichannel communications scheme structure diagram.

arbitrate the communications between the cluster-head and head; 2) quasi-cluster head; 3) cluster member; and 4) quasi-
cluster-member vehicles within a given cluster. Each cluster- cluster member.
head vehicle collects/delivers safety messages and assigns ICD/ The functions of the four states are described as follows:
CRD channels to cluster members by using contention-free First, in the state of CH, the vehicle’s Transceiver I works on
MAC protocol over the CRC channel. Each cluster-member the ICC channel to forward consolidated safety messages to the
vehicle uses one transceiver to exchange the safety messages neighboring clusters, and the Transceiver II is tuned to the CRC
with its cluster-head vehicle. Meanwhile, the cluster-member channel to collect/broadcast safety messages from/to cluster
vehicle uses another transceiver to communicate with its peer members. Second, the quasi-cluster-head state represents that
vehicle within the same cluster over the CRD channel that is this vehicle is neither a cluster head nor a cluster member. In
assigned by its cluster-head vehicle. the quasi-cluster-head state, while Transceiver II is turned off,
the Transceiver I of the vehicle works on the ICC channel, so
that it can also receive and send the safety messages. In fact,
III. F UNCTIONS AND D ESIGNS OF P ROTOCOLS
the quasi-cluster-head vehicles function as real cluster heads,
A finite-state machine (FSM), as shown in Fig. 3, is em- except for the ability in forming clusters.
ployed to precisely describe the principle and operating process Third, when entering the cluster-member state, the vehi-
of our proposed scheme. Each vehicle operates under one and cles tune their Transceiver II’s to the CRC channel where
only one of the following four states at any given time: 1) cluster the cluster-member vehicles receive the consolidated safety
3312 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 6, NOVEMBER 2007

also adds the requesting vehicle into the cluster-member

list. Note that threshold Prth determines the cluster size or
the cluster radius that is denoted by LC (see Fig. 1).
3) Electing the cluster head: when the duration of being a
quasi-cluster head is longer than tj time units. Because
the cluster-head vehicles repeat the ITJ advertisement
messages every tj time unit, a quasi-cluster-head vehicle
can receive the ITJ advertisement messages within tj time
units if it is within the cluster head’s transmission range.
Thus, a quasi-cluster-head vehicle will elect itself as a
cluster head if it cannot receive a valid ITJ advertisement
message within tj time units.
4) Losing contact with the cluster head temporally: when
Fig. 3. FSM of our proposed scheme. T1 and T2 represent Transceiver I and the cluster-member vehicle cannot receive the sched-
Transceiver II, respectively. ule assignment broadcast every T time unit (refer to
Section III-B for details) from the cluster head. If this
messages and send their own safety messages as well as data
condition holds, the state of the vehicle changes from
channel reservation requests. Each cluster-head vehicle uses
cluster member to quasi-cluster member to guarantee the
the centralized multichannel control algorithm (e.g., the Ran-
timely delivery of the safety messages. The vehicles in
dom Scheduling algorithm [12]) to assign appropriate CRD
quasi-cluster-member state can send and receive safety
or ICD channels to cluster members after receiving the data
messages by tuning Transceiver I to the ICC channel.
channel reservation requests. According to the decision on the
On the other hand, the quasi-cluster-member vehicle tries
assignment of CRD/ICD channels by the cluster-head vehicle,
to resume the communications with the previous cluster
the cluster-member vehicles set their Transceiver I’s to either
head by keeping Transceiver 2 operating on the CRC
the corresponding CRD channels for communications with
channel, because the disconnection may be temporally
other cluster members within the cluster or the ICD channel
due to the unreliability of the wireless channel.
for nonreal-time traffic data packet exchange among clusters.
5) Losing contact with the cluster head completely: when the
Finally, the function of the quasi-cluster-member state is to
quasi-cluster-member vehicle cannot receive the sched-
guarantee that the cluster-member vehicles can receive and
ule assignment consecutively for two times. The quasi-
transmit the safety messages by switching Transceiver I’s to
cluster-member vehicle considers that it loses contact
the ICC channel, even if it temporally loses contact with the
with the previous cluster head completely and thus
cluster heads. In the quasi-cluster-member state, the vehicle’s
changes to its state to quasi-cluster head. In the cluster
Transceiver II still monitors the previous CRC channel and tries
head, it will delete this vehicle from the member list if
to resume the communications with the previous cluster-head
it cannot receive packets from this vehicle consecutively
vehicle.
for three times.
6) Discovering the cluster head: when the quasi-cluster-
A. Cluster Configuration Protocol member vehicle receives the schedule assignment from
the cluster head again. The quasi-cluster-member vehicle
The state transitions of the FSM that is depicted in Fig. 3 changes to cluster-member state and resumes the commu-
are controlled by the Cluster Configuration Protocol. There are nications with the cluster head.
seven state-transition conditions for the protocol FSM on each 7) Merged by other cluster: when the cluster-head vehi-
vehicle. cle receives the valid ITJ advertisement message from
1) Entering a highway (Initiate): when vehicles just enter the neighboring cluster head, which has more cluster-
the highway. member vehicles. If the distance between two cluster-
2) Joining a cluster: when a quasi-cluster-head vehicle head vehicles is less than LC , this merging condition
receives the valid advertisement message from nearby holds. Then, the cluster-head vehicle changes to cluster-
cluster-head vehicles. The cluster-head vehicle broad- member state and joins that neighboring cluster head,
casts the invite-to-join (ITJ) advertisement message every and its previous cluster-member vehicles either join that
tj time unit. Once the quasi-cluster-head vehicle, which neighboring cluster head or form another new cluster
does not belong to any cluster, receives the ITJ message, according to the Cluster Configuration Protocol. LC is
it checks the received signal strength, which is denoted by the minimal distance between two cluster heads, which
Pr . The received ITJ message is considered to be valid if can minimize the effect of oscillatory distances between
its signal strength is greater than the predefined threshold nodes.
that is denoted by Prth . When receiving a valid ITJ mes-
sage, the vehicle sends a request-to-join (RTJ) message
B. Intracluster Coordination and Communication Protocol
including the vehicle’s ID and the network address to the
advertising cluster head. After the cluster head receives The Intracluster Coordination and Communication Protocol
the RTJ message, it sends an acknowledgment (ACK) and is based on a MMAC protocol, where each cluster head
SU AND ZHANG: CLUSTERING-BASED MMAC PROTOCOLS FOR QoS PROVISIONINGS OVER VANETs 3313

Fig. 4. Time division in the CRC channel.

employs a scheduling scheme over the CRC channel to cording to N number of vehicles in the cluster. This TDMA
collect/broadcast safety messages and coordinate the cluster- schedule is broadcast back to the cluster-member vehicles in
member vehicles to transfer nonreal-time data within/between the cluster. Second, the cluster members send safety mes-
cluster(s). In the CRC channel, time is partitioned into regular sages and data channel reservation requests to the cluster head
time intervals with equal lengths of T , which is called the during their own assigned time slots. Third, the cluster-head
“repetition period.” Fig. 4 shows the time division in the CRC vehicle consolidates the safety messages that were collected
channel. The repetition period consists of the TDMA upstream from both its cluster members and neighboring cluster heads,
period that is denoted by Tt and broadcast downstream period and makes a decision on the assignment of ICD and CRD
that is denoted by Tb . The length of the time slot that is assigned channels according to the data reservation requests. Then, the
to each member within a cluster, which is denoted by tslot , can cluster-head vehicle broadcasts consolidated safety messages
be determined by and data channel assignments back to the cluster members.
Finally, the cluster-member vehicles switch their Transceiver I
Tt to the assigned channel transmitting/receiving the nonreal-time
tslot = (1)
N traffics. Because only two cluster members are assigned to op-
erate over the same CRD channels, they can use point-to-point
where N is the number of cluster members within the cluster.
communication without any contention. Note that the nonreal-
The TDMA scheme can guarantee that each vehicle within a
time traffics and safety messages can be served concurrently in
cluster has a chance to transmit data every T time unit. Hence, if
our scheme due to two sets of transceivers that are used.
we denote the update interval of safety messages by Tupdate , the
required delivery delay of safety messages by Tsafety , the size
of the safety message by Hsafety , and the channel rate by R, C. Intercluster Communication Protocol
then the timely delivery of the safety messages can be achieved
when the following conditions hold: In the Intercluster Communication Protocol, two types of
 traffics are served on two separate channels between clusters:
 
 T < min T 1 1) the real-time safety messages over the ICC channel and
update , 2Tsafety
. (2) 2) the nonreal-time traffic over the ICD channel. On one hand,
 R ≥ Hsafety cluster heads, quasi-cluster heads, and quasi-cluster members
tslot
use contention-based protocols (e.g., IEEE 802.11) to share the
Research states that the driver reaction time to traffic warning ICC channel. After the cluster-head vehicles collect the safety
signals, such as brake lights, can be on the order of 700 ms and messages from their own clusters, they use the data fusion
longer [13]. Consequently, the update interval of safety mes- technique to consolidate the safety information and then con-
sages should be less than 500 ms. Otherwise, the safety system tend for the ICC channel to forward the processed information
is useless in helping the driver deal with emergency situations. to the neighboring cluster heads. The transmission range for
Hence, a safety message of 200 bytes is typically updated the Intercluster Communication Protocol, which is denoted by
every 500 ms and should be delivered before the generation of LI (see Fig. 1), depends on the intracluster communication
the next new message. Therefore, we set Hsafety = 200 bytes, range. To let two nearby neighboring cluster-head vehicles to
Tupdate = 200 ms, Tsafety = 200 ms, and T = 80 ms in this communicate in one hop, LI ≥ 2LC should hold.
paper. On the other hand, applying the Intracluster Coordination
The operating procedure of the Intracluster Coordination and and Communication Protocol, one vehicle is assigned to the
Communication Protocol can be divided into the following ICD channel in each cluster. By employing the contention-
four phases: First, each cluster-head vehicle creates a TDMA based MAC, those vehicles from different clusters contend for
schedule specifying each vehicle when it can transmit ac- the common ICD channel to transmit/receive the nonreal-time
3314 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 6, NOVEMBER 2007

traffic packets between clusters. They work as gateways to for the delivery of safety message. Without the help of the
forward the packets for the other cluster-member vehicles. GPS, our proposed scheme utilizes the cluster head as the
local centralized controller to achieve synchronization among
all cluster-member vehicles within the cluster.
D. Remarks on the Protocol Designs
1) Interference Between Multiple Clusters: The preceding
IV. D ELAY M ODELING FOR D ELIVERY
discussion describes how the individual clusters communicate
OF S AFETY M ESSAGES
among cluster-member vehicles in that cluster. However, the
transmitted signal in one cluster will affect communications in The most important performance metrics for our proposed
a neighboring cluster since the radio medium is inherently a scheme is the transmission delay of the safety messages. Under
broadcast medium. To reduce this type of interference, different our proposed scheme, the safety messages that are generated
clusters communicate with the different code-division multiple- by the cluster-member vehicles are mainly delivered via three
access (CDMA) codes. Thus, when a vehicle decides to become hops (corresponding to their longest path/delay), which consist
a cluster head, it chooses randomly from a list of CDMA of the transmission over the CRC channel from the cluster-
codes. It informs all the cluster-member vehicles within its member vehicle to its cluster-head vehicle, the transmission
cluster to transmit with this CDMA code during the exchange over the ICC channel from this cluster-head vehicle to its
of ITJ/RTJ packets. The cluster head then filters all received neighboring cluster-head vehicle, and the transmission over the
energy using the given CDMA code. Thus, the neighboring CRC channel from this neighboring cluster-head vehicle to its
cluster’s radio signals will be filtered out without corrupting the targeted cluster-member vehicle in the neighboring cluster. In
data transmissions between vehicles in the cluster. terms of delay bound guarantees, the delays of the first and
2) Why Does the Intracluster Communication Use the third transmissions are upper bounded by using the contention-
TDMA Scheme, Instead of the IEEE 802.11 MAC?: The stan- free TDMA-based MAC protocol, while the second-hop trans-
dard of the IEEE 802.11 protocol provides detailed MAC and mission is achieved through the contention-based IEEE 802.11
Physical layer (PHY) specifications. Being a contention-based MAC protocol and thus is not delay-bound guaranteed. Since
MAC protocol, the IEEE 802.11 degrades its performance due the delays that are caused by the first and third transmissions
to the limitations caused in V2V communications, such as are deterministically upper bounded and thus are much easier to
blocking problem, lack of stability because of high mobility, derive, we will only concentrate on delay modeling and analy-
and delay insensitivity. Among them, delay insensitivity is ses of the second-hop transmission by applying the stochastic
the major concern for V2V communications, where the safety analyses. To formally formulate the delay modeling problem,
messages must be delivered within a bounded delay. we define V (t) as the set of vehicles in the VANET at time t.
On the other hand, although the IEEE 802.11e is a QoS- We then further partition V (t) into two sets: 1) the set of
oriented MAC protocol, it is still not suitable for safety message cluster heads C(t) and 2) the set of noncluster-head nodes
delivery in V2V communications due to the following reasons: V (t) \ C(t).2 For a cluster head i ∈ C(t), we let CHi (t) be set
First, the randomness of the binary backoff mechanism in the of the cluster-member vehicles that belong to the cluster that is
enhanced distributed coordination function (EDCF) mode of associated with cluster-head vehicle i at time t. We define Bi
the IEEE 802.11e leads to random variation of delay [14], as the set of cluster-head vehicles that are operating on the ICC
which is not tolerable for the delay-sensitive applications of channel and within the broadcast area that is covered by cluster-
V2V communications. Second, in heavily loaded networks, head vehicle i, and Hi as the set of cluster-head vehicles that are
the IEEE 802.11e chooses larger initial backoff counter or the hidden terminals from cluster-head vehicle i.
interframe spacing intervals, resulting in the increase of av- Let l and k be two cluster-member vehicles, which belong
erage access delay. Finally, the collision probability shoots to two different clusters under cluster-head vehicle i and
up significantly as the number of packets with the same pri- cluster-head vehicle j, respectively, where cluster-head vehicles
ority contending for the wireless shared media increases. In i and j are located within each other’s radio transmission range.
V2V communications, this kind of scenario may often happen, This implies that l ∈ CHi (t) and k ∈ CHj (t) with j ∈ Bi and
because the safety messages of each vehicle have the same i ∈ Bj . Then, based on our proposed scheme, we can derive
priority. the average delay that is spent to send a safety message from
Considering that the contention-based MAC protocols cluster-member vehicle l to cluster-member vehicle k, which is
(l,k)
cannot guarantee the bounded delay, contention-free MAC denoted by T d , as follows:
protocols, such as TDMA, are more suitable for delay-sensitive
(l,k) (l,i) (i,j) (j,k)
V2V communications. A set of TDMA-based MAC proto- Td = T m2h + T h2h + T h2m (3)
cols have been proposed to avoid the inherent randomness
(l,i)
and delay unpredictability of the IEEE 802.11. The Universal where T m2h is the average delay for a safety message that
mobile telecommunications system Terrestrial Radio Access is sent from the cluster-member vehicle l to its cluster-head
(j,k)
with Time Division Duplexing (UTRA-TDD) was proposed vehicle i, T h2m is the average delay for a safety message
for the Fleenet project [6]. The UTRA-TDD is a distributed that is sent from cluster-head vehicle j to its cluster-member
TDMA technology where Global Positioning System (GPS)
helps achieve synchronization between vehicles. In this paper, 2 For convenience, we use vehicle, node, and terminal interchangeably in the
we also use the TDMA-based scheme as the MAC protocol rest of this paper.
SU AND ZHANG: CLUSTERING-BASED MMAC PROTOCOLS FOR QoS PROVISIONINGS OVER VANETs 3315

(i,j)
vehicle k, and T h2h is the average time that is spent by cluster-
head vehicle i to send the consolidated safety message from
cluster-head vehicle i to its neighboring cluster-head vehicle j.
Note that cluster-head vehicle i must use the ICC channel to
broadcast the consolidated safety messages to its neighboring
cluster-head vehicles. If we define si as the average time that
is spent by cluster-head vehicle i to broadcast the consolidated
safety messages to all neighboring cluster-head vehicles over
the ICC channel, then, provided that there are no collisions,
we have Fig. 5. Worst case of the cluster topology when LC = 250 m and Li =
100 m. The white, dark, and shaded dots represent vehicle i, the vehicles within
∆ (i,j1 ) (i,j2 ) broadcast area, and the hidden terminals, respectively.
si = T h2h = T h2h , ∀j1 , j2 ∈ Bi and i = j. (4)
(l,i) (k,j) the received safety messages, by η(n), where n is the number
From Fig. 4, we know that T m2h ≤ T and T h2m ≤ T .
of received safety messages plus the cluster-head vehicle’s
Hence, the maximum total delivery delay of safety messages,
(l,k) own safety message. We consider the worst case where the
which is denoted by tm , can be expressed as cluster-head vehicles do not compress the safety messages, i.e.,
η(n) = 1. Hence, the average size of the consolidated message
t(l,k)
m = si + 2T. (5)
that was generated by cluster-head vehicle i, which is denoted
To assure the timely delivery of the safety messages, the max- by H i , can be expressed as
imum total delay of safety messages should be less than the
(l,k) H i = (|CHi | + 1) Hsafety (7)
required delivery delay of safety message, i.e., tm ≤ Tsafety ,
which implies, by using (5), that where |CHi | is the average number of cluster-member vehicles
within that cluster under cluster-head vehicle i, and Hsafety is
si = t(l,k)
m − 2T ≤ Tsafety − 2T = tbd (6)
the packet size of the safety message.
∆ Let |Bi | be the number of cluster-head vehicles within the
where tbd = Tsafety − 2T is the delay upper bound for the broadcast area of cluster-head vehicle i and |Hi | be the number
safety message delivery between two neighboring cluster-head of hidden terminals of cluster-head vehicle i. Clearly, the larger
vehicles. the |Bi | and |Hi |, the more intensive the channel contention
Now, we analyze the transmission delay of the consolidated over the ICC channel. As shown in Fig. 5, max{|Bi |} =
safety message through the ICC channel. As discussed in 2LI /LC , and max{|Hi |} = 2LI /LC , which yield the
Section III-C, the main participants that are associated with most intensive channel contention over the ICC channel as the
the intercluster safety message delivery include cluster-head ve- worst case. Now, we develop an analytical model to derive
hicles, quasi-cluster-head vehicles, and quasi-cluster-member the average delay for delivering the safety messages in this
vehicles. There are three types of packets that are transmitted worst case.
over the ICC channel. The first type is the consolidated safety As described in Section III-C, the cluster-head vehicles em-
messages that were generated and transmitted by the cluster- ploy the IEEE 802.11 MAC protocol to broadcast consolidated
head vehicles. The second type is the safety messages of safety messages. See [15] for the detailed and general descrip-
the quasi-cluster-head vehicles and the quasi-cluster-member tions of the IEEE 802.11 MAC protocol. In particular, when
vehicles. The third type is the cluster management packets a cluster-head vehicle has a packet to transmit, it monitors the
that was sent by the quasi-cluster-head vehicles. Because the ICC channel. If the channel is busy, the cluster-head vehicle
arrival rate of safety messages is much larger than that of waits for a random number of time slots before it can transmit.
cluster management packets and the size (200 bytes) of the Since, the value of the backoff counter, which is denoted
safety message is also ten times larger than that (20 bytes) by BO, is chosen uniformly from the interval of [1, CW ],
of the cluster management packet, the safety message traf- the probability that the backoff counter is equal to n can be
fics always dominate the ICC channel usage, and thus, it is written as
reasonable to omit the cluster management packets (which is
also validated by our simulation results that were obtained in 1
Pr{BO = n} = , 1 ≤ n ≤ CW (8)
Section VI-C). Moreover, because the number of quasi-cluster- CW
head and quasi-cluster-member vehicles are much fewer than
where CW is the maximum contention-window size. Thus, the
that of the cluster-head vehicles, we can also neglect the
average backoff window size, which is denoted by W , can be
messages that were transmitted by quasi-cluster-head and
given by
quasi-cluster-member vehicles. That is, we mainly consider
the consolidated safety messages that were transmitted by 1 + CW
the cluster-head vehicles since they always dominate the ICC W = . (9)
2
channel usage.
We denote the safety message compression/consolidation In the saturation IEEE 802.11 based network, the proba-
rate function, with which the cluster-head vehicles consolidate bility that cluster-head vehicle i attempts to transmit a safety
3316 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 6, NOVEMBER 2007

message, which is denoted by αi , is determined by [16] or failing (collided) transmission of other vehicles, which is
denoted by γi and can be determined by
1
αi = . (10) αi
W γi =  (15)
1− j∈Bi (1 − αj ) + αi
For nonsaturated networks, following the work of [17], the
probability that cluster-head vehicle i attempts to transmit a where Bi is the set of cluster-head vehicles that are located
safety message, which is denoted by βi , is determined by within the broadcast area of cluster-head vehicle i over the ICC
channel. Letting Qi be the random number of busy time slots
λsi that are occupied by other neighboring vehicles between the
βi = λsi αi = (11) consecutively successful transmissions of vehicle i, we have
W
that Qi is following the geometric distribution, i.e.,
where λ = 1/Tupdate is the arrival rate of the safety messages,
and si is the average delay that is spent by cluster-head vehicles Pr{Qi = n} = γi (1 − γi )n−1 . (16)
i to transmit a consolidated safety message, as defined by (4).
The authors in [18] observed that the successful transmission Then, we can derive E[ei ] as follows:
process for a particular node can be considered as a regenerative
(1 − γi )m
process. On average, each node successfully transmits one E[ei ] = E[Qi ]m =
packet during the cycle time of this regenerative process, which γi
includes the following: 1) the backoff time; 2) the successful  
transmission time; and 3) the channel busy time. The average m
= 1− (1 − αj ) (17)
delay, which is si defined by (4), that is spent by vehicle i to αi
j∈Bi
transmit a safety message is equivalent to the average cycle time
of regenerative process for vehicle i (i.e., the period between
where m is given in (13). Thus, combining and applying
two consecutively successful transmissions by vehicle i). Thus,
(10)–(14) and (17), we can obtain the numerical results for αi ,
we have
βi , and, most importantly, si .
si = E[mi + bi + ei ] = E[mi ] + E[bi ] + E[ei ] (12)
V. S UCCESSFUL D ELIVERY R ATE OF S AFETY M ESSAGES
where mi is the fixed time that is spent by cluster-head vehicle If we define pi as the probability that cluster-head vehicle
i to transmit a safety message, bi is the random backoff time i successfully broadcasts its buffered safety message to all its
interval that is observed by cluster-head vehicle i between its neighboring vehicles whose Transceiver I’s are listening on ICC
two consecutive successful transmissions, and ei is the random channel, then we have
time interval when cluster-head vehicle i observes that the
E[bj ]m
channel is busy due to other nodes’ successful or collided trans-
missions between its two consecutive successful transmissions. pi = (1 − P ERij ) (1 − βj ) (1 − βj ) sj
(18)
j∈Bi j∈Bi j∈Hi
First, let us derive E[mi ] and E[bi ]. Since the size of the
safety message is fixed, the time that is taken by cluster-head where Bi is the set of cluster-head vehicles that are operating on
vehicle i to transmit a safety message is a constant, which is the ICC channel and within the broadcast area that is covered
denoted by m. Thus, E[mi ] can be expressed as by vehicle i, Hi is the set of cluster-head vehicles that are the
hidden terminal from vehicle i, and P ERij is the packet error
∆ Hi rate that vehicle j cannot decode the packet from vehicle i
E[mi ] = m = TH + + DIF S (13)
R due to the wireless-channel quality dropping (e.g., fading and
noise).
where R is the channel rate, H i that is given by (7) is the Then, we further study the asymptotical properties of pi
packet length of the consolidated safety messages, TH is the when the contention-window size CW becomes very large. If
time that is spent to transmit the packet head of the consolidated we let CW go to infinity, then, from (17), we obtain
message, and DIF S is the distributed coordination function 
(DCF) interframe space. It is also easy to obtain E[bi ] by using 1− j∈Bi (1 − αj ) m
lim E[ei ] = lim
CW →∞ CW →∞ αi
 |Bi | 

E[bi ] = W σ (14) 1
= lim W 1 − 1 −
CW →∞ W
where σ is the length of time slot over the ICC channel. |Bi |    n
 |Bi | 1
To obtain E[ei ], we introduce the probability that the next = lim W (−1)n−1
CW →∞ n W
transmission is a successful transmission by cluster-head ve- n=1
hicle i, given that the last time slot is busy due to successful = |Bi | (19)
SU AND ZHANG: CLUSTERING-BASED MMAC PROTOCOLS FOR QoS PROVISIONINGS OVER VANETs 3317

where |Bi | is the number of vehicles within the broadcast area

that is covered by vehicle i. Then, using (11), we have
 
λsi
lim (1 − βi ) = lim 1−
CW →∞ CW →∞ W
  
E(bi ) + E(ei ) + m
= lim 1−λ
CW →∞ W
  
W σ + E(ei ) + m
= 1 − λ lim
CW →∞ W
(a)
= 1 − λσ (20) Fig. 6. Algorithm to determine the desirable CW , where  is a small positive
real number (0 <  < 1).
where equation of (a) in (20) holds due to (19). In addition,
when CW goes to infinity, applying (19), we have livery rate of safety messages. Therefore, we do not need to
  distinguish among the different cluster-head vehicles for the
E(bi )m m following discussions, and thus, we can denote the average
lim = lim = m. (21)
CW →∞ si CW →∞ 1 + E[ei ] + m σW σW
delay of safety messages for all cluster-head vehicles by s,
the average successful delivery rate of safety messages by p,
Then, applying (19) and (20), the limit of (1 − βi )E(bi )m/si can and the upper bound of the successful delivery rate for safety
be rewritten as messages by p .
E(bi )m E(bi )m
Our analyses indicate that, when CW increases, the suc-
lim log(1−β )
lim (1−βi ) si
=e CW →∞ si i
cessful delivery rate of safety messages p becomes higher,
CW →∞
  but the safety message delivery delay s also becomes larger.
(a) limCW →∞ E(bi )m [limCW →∞ log(1−βi )]
=e si This introduces a tradeoff between successful delivery rate p
=em log(1−λσ) and delivery delay s for delivering the safety messages as
functions of the size of the contention window CW . However,
= (1−λσ)m (22)
to achieve better QoS provisioning for delivering the safety
where equation of (a) in (22) holds due to the propo- messages, both the smaller s and larger p are desired. In addi-
sition that, if limx→∞ f (x) and limx→∞ g(x) exist, then tion, our analyses further indicate that, although both s(CW )
limx→∞ f (x)g(x) = [limx→∞ f (x)][limx→∞ g(x)]. Again, us- and p(CW ) are monotonically increasing functions of CW ,
ing the same proposition and (20) and (22), we can obtain the the increasing rate of s(CW ) gets larger, but the increasing
limit of p if the contention-window size approaches infinity, rate of p(CW ) becomes smaller, as CW increases. In addition,
which is denoted by p , as follows: based on our analyses, p(CW ) converges to its upper bound
(p ) quickly, when CW increases. This implies that, for any
pi = lim pi given delay-bound QoS requirement for the safety message
CW →∞
delivery, there is typically a wide range, which is called the
= (1 − P ERij ) lim (1 − βj ) allowable CW range, to choose the values for CW , which
CW →∞
j∈Bi j∈Bi can meet the given reliability QoS requirement for the safety
 E[bj ]m

message delivery. The lower and upper bounds of the allowable
· lim (1 − βj ) sj
CW range are determined by the given reliability QoS require-
CW →∞
j∈Hi
ment and the delay-bound QoS requirement, respectively, if
= (1 − P ERij ) (1 − λσ) (1 − λσ)m the reliability QoS and delay QoS requirements are given with
j∈Bi j∈Bi j∈Hi the reasonable values such that the allowable CW range is
|Bi |+|Hi |m not an empty set. Thus, if we choose the lower bound of this
= (1 − λσ) (1 − P ERij ) (23)
j∈Bi allowable CW range as the desirable CW value, we can further
decrease s significantly without violating the reliability QoS
where |Bi | and |Hi | are the number of vehicles within the requirement.
broadcast area that is covered by cluster-head vehicle i and Based on the preceding observations, we develop the simple
the number of hidden terminals for cluster-head vehicle i, but practically useful Algorithm 1 listed in Fig. 6 to search
respectively. Note that pi is a monotonically increasing func- for the desirable contention-window size, which is denoted
tion of CW , and thus, (23) yields the upper bound of the by CW ∗ and can best balance the tradeoff between s and
successful delivery rate for safety messages, which is pi when p. Specifically, the desirable CW ∗ is equal to the minimum
CW approaches infinity. As shown in (23), the value of the CW that satisfies both s < tbd and (p/p ) > 1 − +, where + is
upper bound pi is determined by the wireless networks’ channel a small real-valued number (0 < + < 1) called the reliability
qualities, network topologies, safety message arrival rate, etc. QoS requirement. By tuning up the value of +, we can get the
Because we only consider the cluster-head vehicles, on desirable tradeoff between s and p.
average, all cluster-head vehicles share the same properties, To further demonstrate the tradeoff between s(CW ) and
including both the safety message delay and successful de- p(CW ), we study an example of the average cluster size of
3318 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 6, NOVEMBER 2007

TABLE II TABLE IV
PARAMETERS USED IN THE ANALYSIS AND SIMULATIONS PARAMETERS USED IN OUR PROTOCOL EVALUATION METRICS

VI. S IMULATION E VALUATIONS

A. Performance Evaluations Metrics
To quantitatively analyze the performance of our proposed
protocols, we use both topology-based metrics and data-transfer
metrics. Topology-based metrics, such as average cluster-head
lifetime, average relative speed compared to the cluster-head
within a cluster, average relative speed among cluster heads,
and average cluster size, can provide information about our
Cluster Configuration Protocol. On the other hand, the data-
transfer metrics include the following: 1) cluster manage-
ment overhead; 2) the average delay of safety messages; and
3) the probability of safety-message delivery failure. We further
explain these metrics as follows: For convenience of presenta-
tion, Table IV summarizes all the parameters that were used
to calculate our performance evaluation metrics in the discus-
sions here.
Fig. 7. Tradeoff between the safety messages delay (dash-dotted line) and the 1) Average cluster-head lifetime: We use the average
successful delivery rate of safety messages (solid line). The solid line marked by cluster-head lifetime to measure the performance of the Cluster
p (corresponding to the left y-axis) represents the successful delivery rate of the
safety message. The dash-dotted line marked by s (corresponding to the right
Configuration Protocol. Intuitively, larger cluster-head lifetime
y-axis) represents the delay of the safety message among cluster-head vehicles. implies more stable cluster topologies, leading to less overheads
The dashed horizontal line marked by p
represents the upper bound for the that are caused by the cluster management. Let BCi (t) be a
successful delivery rate of safety messages. The dotted horizontal line labeled
with tbd represents the delay threshold for the safety messages. The maximum
cluster-head transition indicator such that
acceptable CW is calculated by using tbd . The desirable CW ∗ is derived 
by using Algorithm 1 in Section V. |Hi | = |Bi | = 4. The average cluster  BCi (t) = 1, when vehicle i changes from a
size is 8. noncluster head to a cluster head . (24)

BCi (t) = 0, otherwise
TABLE III
DERIVED CW ∗ ’S WITH DIFFERENT CLUSTER SIZES WITH  = 10−4 For node i, {i ∈ C(t)} ∧ {BCi (t) = 1}, the cluster-head life-
time, which is denoted by CHLi (t), is the time interval from
the time when node i becomes a cluster head at t to the time
when it is merged by the other cluster head. The average cluster-
head lifetime, which is denoted by CHL, can be expressed as

S |V (t)|
∆ i=1 CHLi (t)
eight cluster-member vehicles and + = 10−4 . By using the
CHL = |V (t)| . (25)
t=0 i=1 BCi (t)
parameters that are listed in Table II, we obtain the numerical
results for analyzing the tradeoff, as shown in Fig. 7. As pre- 2) Average relative speed compared to the cluster head
viously discussed, when CW increases, the slope of the solid within a cluster: It represents the topology stability of clusters.
line, denoting the successful delivery rate of safety messages, The less the average relative speed of the cluster member
becomes flatter, but the dash-dotted line, which represents the compared to that of the cluster head, the more stable the cluster.
delivery delay of safety messages, increases faster. By running If we denote the average relative speed compared to that of the
Algorithm 1, Fig. 7 shows that the desirable CW ∗ is equal to cluster head within a cluster by RSW C, then we get
880 for + = 10−4 . By adopting this desirable CW ∗ , we can
1
achieve the guaranteed reliability QoS (+ = 10−4 ) while reduc- RSW C =
∆
|C(t)| S
ing s as much as possible (see the delay QoS gain in Fig. 7). For
  
+ = 10−4 , we also obtain the desirable CW ∗ ’s corresponding to −→ −→ 

S  j∈CHi (t) SP i (t) − SP j (t)
the different cluster sizes by executing Algorithm 1, which are · . (26)
|CHi (t)|
listed in Table III. t=0 i∈C(t)
SU AND ZHANG: CLUSTERING-BASED MMAC PROTOCOLS FOR QoS PROVISIONINGS OVER VANETs 3319

3) Average relative speed among cluster heads: It measures C. The Simulations on the Impact of Highway Traffics
the global topology of the vehicular networks. The average rel-
We simulate our proposed protocols by using the Matlab-
ative speed among cluster heads, which is denoted by RSCH,
based event-driven simulator in this section. Our simulation
can be calculated by
experiments take various highway traffic parameters into con-
   sideration, which include highway density, vehicular speed, and
−→ −→ 
∆ 1
S
i,j∈C(t)∧i=j  SP i (t) − SP j (t) data traffic parameters. We assume that each channel has the
RSCH = . (27) same data transmission rate of R = 6 Mb/s. Table II summa-
S t=0 |C(t)|2
rizes the parameters that were used in our simulations.
We start with studying the impact of the highway traffic
4) Average cluster size, which is denoted by CS: It mainly parameters on the performance of our proposed protocols.
depends on the traffic density and can be obtained by As indicated by the authors in [20]–[22], the mobility has a
large impact on the performance of mobile ad hoc networks
due to the variations of vehicular network topologies. We are
1 
S
∆
CS = |CHi (t)| . (28) also interested in how the different highway traffic conditions
S t=0 affect the average relative speed compared to the cluster-head
i∈C(t)
vehicle, the average relative speed among the cluster-head
vehicles, the average cluster size, and the average cluster-
5) Cluster management overhead: Because the cluster man- head lifetime.
agement messages and the safety messages among cluster- We consider the results for three different cases in terms of
head vehicles are transmitted over the same ICC channel, to densities of vehicles, including low density, medium density,
better characterize the proposed protocols, we need to study the and high density. The vehicle density cases that were con-
cluster management overhead, which can be determined by the sidered include 12, 24, and 40 vehicle/km/lane, on average,
ratio of the total messages of RTJ and ITJ to the total messages corresponding to the low-density, medium-density, and high-
over the ICC channel. density cases, respectively. For each traffic density, the average
6) The delay of the safety message: The time that is used preferred speed of vehicles varies from 20 to 50 m/s, and the
for a safety message to be delivered to all the vehicles in the preferred speed variance is set to 6 m2 /s2 . In each scenario
circular area that is centered at the transmitter vehicle with the (with distinct preferred speed and traffic density), we run the
radius equal to LI (see Fig. 1). simulations ten times to obtain the mean value as the final
7) The probability of safety-message delivery failure: It is performance metric.
defined as the probability that a given safety message will not be Fig. 8 shows the impact of the vehicular preferred speed on
received by all the vehicles in the circular area that is centered the cluster-topology variation with different traffic densities.
at the transmitter vehicle with the radius equal to LI . From Fig. 8(a), we observe that the average relative speed
among cluster heads increases as the average preferred speed
gets larger for the cases of low- and medium-density traffics.
B. The Highway Traffic Model This is because the larger speed leads to more rapid topol-
The highway traffic model that is used in this paper is built ogy variations. On the other hand, the average relative speed
up based on the car-following model that was proposed in almost remains the same for the high-density traffic. This is
Simone 2000 [19], which is mainly composed of the desired expected because the vehicular speed is significantly limited by
gap function and longitudinal control function. In the highway the high-density traffic according to the car-following model
traffic model, every vehicle has its own preferred speed, which even when the preferred speed is high. Fig. 8(b) shows the
the vehicle tries to reach if the conditions are satisfied (e.g., relative speed compared to the cluster head within a cluster
having enough safe distance). The preferred speed is based on against the average preferred speed. The larger the average
the normal distribution with an average preferred speed and a preferred speed, the bigger the relative speed compared to the
variance of the preferred speed. cluster head within a cluster, resulting in the more unstable
In our highway traffic model, we assume that the vehicles run cluster topology, which is also reflected in Fig. 8(c). From
along a three-lane-per-direction circular loop with a perimeter Fig. 8(d), we observe that the average preferred speed has
of 2000 m. The period for each vehicle to stay on the highway little impact on the cluster size. In fact, the cluster size is
is based on the normal distribution with its mean equal to dominated by the traffic density. Then, we can determine the
2100. This implies that, on average, the vehicles will run on the desirable contention-window size CW ∗ for different highway
highway for 2100 s or 35 min, on average, before they “exit” traffic scenarios by using the relations between the average
the highway. A new vehicle is assumed to enter the highway cluster size that was obtained in simulations and CW ∗ , which
every 2100 s and run at a random position on the highway loop is listed in Table III.
initially. The lane width is set to be 5 m, and the width of the Fig. 9 shows the cluster management overhead against the
highway buffer zone between lanes of the opposite directions is average preferred speed. For the low and medium traffic den-
20 m. When the vehicles arrive at the end of the highway, they sities, the cluster management overhead becomes larger, as the
wrap around from the beginning position of the same lane of average preferred vehicular speed increases. This can also be
the highway. observed by carefully examining Fig. 8(a)–(c), which indicates
3320 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 6, NOVEMBER 2007

Fig. 8. Impact of vehicular speed on the cluster-topology variations with different traffic densities. (a) Relative speed among cluster-head vehicles. (b) Relative
speed compared to cluster-head vehicles within a cluster. (c) Cluster-head lifetime. (d) Cluster size.

that the higher average preferred vehicular speed leads to the

more frequent cluster topology variations. However, for the
low- and medium-density traffics, in spite of the increasing
trend of cluster management overhead, the cluster management
overhead is no more than 13%, even when the preferred vehicu-
lar speed reaches 50 m/s. On the other hand, for the high-density
traffic, the average preferred speed only has slight effects on the
cluster management overhead because the cluster topology is
almost independent of the average preferred speed as previously
discussed.
From Fig. 10(a), we observe that the safety message delay
increases linearly with the average preferred speed for the low
and medium traffic densities, while the safety message delay
under high traffic density changes little as the average preferred
speed increases. The safety message delay of lower traffic
density is higher than that of medium traffic density, given the
same average preferred speed. We also notice that, even in the
Fig. 9. Cluster management overhead against the preferred speed with differ- worst case (i.e., the average preferred speed is 50 m/s under low
ent traffic densities. density traffic), the average message delay is 151 ms, which
SU AND ZHANG: CLUSTERING-BASED MMAC PROTOCOLS FOR QoS PROVISIONINGS OVER VANETs 3321

Fig. 10. Performance evaluations of our proposed scheme under different traffic densities. (a) Safety message delivery delay against the preferred speed.
(b) Safety message delivery failure rate against the preferred speed.

is still lower than the required safety message delivery time protocols [24], [25] are not suitable in the V2V communication
Tsafety . environments for the following reasons. The accurate synchro-
The probability of safety-message delivery failure increases nization that is required by MMAC is very difficult to achieve,
slightly with the growth of the average preferred speed and is no although GPS can be used to realize coarse synchronization.
more than 9 × 10−4 in the worst case, as shown in Fig. 10(b). Moreover, the synchronization is more difficult to realize in
From Fig. 10(a) and (b), we observe that our proposed scheme a chain topology, which is often the most typical situation in
can achieve the low and stable safety-message delivery failure highway vehicle traffic flows.
probability as well as the low average safety message delay un- In the original DCA, one transceiver always operates over
der different traffic scenarios. There are two reasons for our per- a dedicated control channel, and the other transceiver can be
formance superiority: First, the number of cluster-head vehicles switched to any data channels in an on-demand manner. The
contending for the ICC channel to transmit consolidated safety Request-To-Send (RTS) and Clear-To-Send (CTS) packets are
messages is greatly reduced due to the clustering techniques transmitted over the control channel to reserve the data channel
that were used in our scheme, which can significantly decrease for data transmission. Specifically, upon receiving the RTS, the
the packet collisions and increase the successful broadcast rate. receiver decides on a channel and adds this channel reservation
Second, the cluster-member vehicles within a cluster exchange information to the CTS. Then, data and ACK packets are
the safety messages without contention, leading to the timely transmitted over the agreed data channel.
safety-message delivery within the cluster. To support the safety message delivery, we make the neces-
sary modifications on the original DCA. The safety messages,
which need to be broadcast, have to be transmitted on the
D. The Simulations on the Impact of Nonreal-Time Traffics
dedicated control channel, because only the control channel is
We also conduct simulations to study the impact of nonreal- continually monitored by each vehicle terminal. Therefore, in
time traffic on the performance of our proposed scheme by the modified DCA, which is called V2V-oriented DCA (V2V-
using the same simulation parameters as used in the previous DCA), the control channel is used not only for data-channel
section. In the following simulations, we compare our proposed reservation but also for the delivery of safety messages. Ch178
scheme with the other existing schemes. The highway traffic serves as the control channel in V2V-DCA. We do not differ-
parameters are set as follows: The average preferred vehicular entiate the priorities between RTS/CTS and the safety message
speed is 35 m/s, the preferred vehicular speed variance is in the V2V-DCA. In other words, the contention of the control
6 m2 /s2 , and the traffic density is medium. channel is performed in the IEEE 802.11 DCF mode, instead of
Because the original IEEE 802.11 MAC only works on the the EDCF mode. The size of each nonreal-time traffic packet is
single channel, the comparison results will not be fair if we set to be 512 bytes.
only compare the IEEE 802.11 with our proposed scheme. Fig. 11 shows the performance comparison of the probability
Hence, we adopt another MMAC protocol, which is based on of safety-message delivery failure among the IEEE 802.11
Dynamic Channel Assignment (DCA) [23]. Like our proposed MAC (operating on Ch178), V2V-DCA, and our proposed
scheme, the DCA also utilizes two sets of transceivers, which scheme, with the nonreal-time traffic loads varying. From
make it fair to compare the DCA-based scheme with our pro- Fig. 11, we can observe that the probability of safety-message
posed scheme. Note that the single-transceiver-based MMAC delivery failure in our proposed scheme is much lower than
3322 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 6, NOVEMBER 2007

nication scheme, which consists of three major protocols, to re-

duce data congestion and support QoS for real-time delivery of
safety messages while efficiently utilizing wireless bandwidth
over V2V networks. Under our proposed scheme, most safety
message traffic is exchanged within a cluster in the TDMA
and broadcast manner. Only the cluster-head vehicles need
to send the consolidated safety messages over a contention-
based channel. Hence, our scheme can significantly improve
the timely and reliable delivery of safety messages as compared
to the case using the IEEE 802.11 and V2V-DCA. In addition,
we developed the analytical model to study the delay and suc-
cessful rate for transmitting the consolidated safety messages
by the cluster-head vehicles. Based on the analytical model, we
also derived the desirable contention-window sizes for different
highway traffic scenarios. The simulation results obtained show
that our proposed scheme can not only achieve timely and
reliable delivery of safety messages but also efficiently sup-
Fig. 11. Probability of safety-message delivery failure against nonreal-time port other nonreal-time traffics, under variant highway traffic
traffic arrival rate. The size of each nonreal-time traffic packet is 512 bytes. scenarios.
that in the IEEE 802.11 MAC or V2V-DCA scheme, regardless
of the nonreal-time traffic load. In addition, the probabili- R EFERENCES
ties of safety-message delivery failure in the IEEE 802.11 [1] Standard Specification for Telecommunications and Information Ex-
and V2V-DCA get much higher than those for our proposed change Between Roadside and Vehicle Systems—5 GHz Band Dedicated
scheme when the nonreal-time traffic load becomes heavier, Short Range Communications (DSRC) Medium Access Control (MAC)
and Physical Layer (PHY) Specifications, ASTM Int., ASTM E2213-03,
as shown in Fig. 11. When the nonreal-time arrival rate is Jul. 2003.
equal to or larger than 50 packet/s, the probability of safety- [2] X. Zhang, H. Su, and H. H. Chen, “Cluster-based multichannel communi-
message delivery failure already approaches 1, which implies cations protocols in vehicle ad-hoc networks,” IEEE Wireless Commun.,
vol. 13, no. 5, pp. 44–51, Oct. 2006.
that the IEEE 802.11 MAC virtually cannot provision any real- [3] X. Yang, J. Liu, F. Zhao, and N. H. Vaidya, “A vehicle-to-vehicle com-
time delivery of safety message. These preceding observations munication protocol for cooperative collision warning,” in Proc. Mobiq-
are expected for the following reasons: First, the number of uitous, 2004, pp. 114–123.
[4] Q. Xu, T. Mak, J. Ko, and R. Sengupta, “Vehicle-to-vehicle safety mes-
vehicles contending for the control channel in our proposed saging in DSRC,” in Proc. VANET, Oct. 2004, pp. 19–28.
scheme is much fewer than that in the IEEE 802.11 or V2V- [5] T. Elbatt, S. Goel, G. Holland, H. Krishnan, and J. Parikh, “Cooperative
DCA, because only the elected cluster-head vehicles, instead collision warning using dedicated short range wireless communications,”
in Proc. VANET, Los Angeles, CA, Sep. 2006, pp. 1–9.
of all vehicles, contend for the control channel in our proposed [6] H. Hartenstein, B. Bochow, A. Ebner, M. Lott, M. Radimirsch, and
scheme. Thus, the successful transmission probability of safety D. Vollmer, “Position-aware ad hoc wireless networks for inter-vehicle
messages in our proposed scheme is much higher than that in communications: The fleenet project,” in Proc. MobiHoc, pp. 259–262.
[7] T. K. Mak, K. P. Laberteaux, and R. Sengupta, “A multi-channel vanet
the IEEE 802.11 or V2V-DCA. Second, our proposed scheme providing concurrent safety and commercial services,” in Proc. VANET,
uses two dedicated channels (ICC and CRC) and contention- Cologne, Germany, Sep. 2005, pp. 1–9.
free TDMA/Broadcast scheme over the CRC channel to deliver [8] H. Reumerman, M. Roggero, and M. Ruffini, “The application-based
clustering concept and requirements for intervehicle networks,” IEEE
safety messages, so that the load of the nonreal-time traffic Commun. Mag., vol. 43, no. 4, pp. 108–113, Apr. 2005.
only has a slight impact on the delivery of safety messages, [9] C. R. Lin and M. Gerla, “Adaptive clustering for mobile wireless net-
resulting in an almost constant probability of safety-message works,” IEEE J. Sel. Areas Commun., vol. 15, no. 7, pp. 1265–1275,
Sep. 1997.
delivery failure, as shown in Fig. 11, regardless of the nonreal- [10] W. Chen and S. Cai, “Ad hoc peer-to-peer network architecture for vehicle
time traffic load. In contrast, unlike our proposed scheme, the safety communications,” IEEE Commun. Mag., vol. 43, no. 4, pp. 100–
IEEE 802.11 MAC has only one channel, and thus, it has 107, Apr. 2005.
[11] FCC Report and Order: FCC-03-324, Feb. 2004. [Online].
to transmit both the safety messages and nonreal-time traffic Available: http://hraunfoss.fcc.gov/edocs_public/attachmatch/FCC-
in the same single channel. On the other hand, for the V2V- 03-324A1.pdf
DCA scheme, although the transmission of safety messages [12] R. Chipalkatti, Z. Zhang, and A. S. Acampora, “High speed communi-
cation protocols for optical star coupler using WDM,” in Proc. Infocom,
and nonreal-time traffic data packets are transmitted through 1992, pp. 2124–2133.
different channels, the safety messages still have to contend [13] M. Grenn, “‘How long does it take to stop?’ Methodological analysis
with the control packets (i.e., RTS/CTS) of nonreal-time traffic of driver perception-brake times,” Transp. Hum. Factors, vol. 2, no. 3,
pp. 191–216, 2000.
for the control channel, leading to the high probability of safety- [14] H. Zhu and I. Chlamtac, “Performance analysis for IEEE 802.11e EDCF
message delivery failure. service differentiation,” IEEE Trans. Wireless Commun., vol. 4, no. 4,
pp. 1779–1788, Jul. 2005.
[15] Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY)
VII. C ONCLUSION Specifications, Nov. 1999. IEEE Standard 802.11—1999.
[16] G. Bianchi, “Performance analysis of the IEEE 802.11 distributed coordi-
Complying with the DSRC’s seven-channel bandplan, we nation function,” IEEE J. Sel. Areas Commun., vol. 18, no. 3, pp. 535–547,
proposed and analyzed a cluster-based multichannel commu- Mar. 2000.
SU AND ZHANG: CLUSTERING-BASED MMAC PROTOCOLS FOR QoS PROVISIONINGS OVER VANETs 3323

[17] K. Medepalli and F. Tobagi, “Towards performance modeling of IEEE Xi Zhang (S’89–SM’98) received the B.S. and M.S.
802.11 based wireless networks: A unified framework and its applica- degrees in electrical engineering and computer sci-
tions,” in Proc. IEEE INFOCOM, 2006, pp. 1–12. ence from Xidian University, Xi’an, China, the M.S.
[18] K. Medepalli and F. Tobagi, “Throughput analysis of IEEE 802.11 wire- degree in electrical engineering and computer sci-
less LANs using an average cycle time approach,” in Proc. IEEE GLBE- ence from Lehigh University, Bethlehem, PA, and the
COM, 2005, pp. 3007–3011. Ph.D. degree in electrical engineering and computer
[19] M. M. Minderhoud, Simone 2000, Simulation Model of Motorways With science (electrical engineering systems) from the
Next Generation Vehicles, Mar. 2002. Technical specification. University of Michigan, Ann Arbor.
[20] T. Camp, J. Boleng, and V. Davies, “A survey of mobility models for He is currently an Assistant Professor and the
ad hoc networks research,” in Proc. ACM/IEEE MOBICOM, Aug. 2002, Founding Director of the Networking and Informa-
pp. 483–502. tion Systems Laboratory, Department of Electrical
[21] F. Bai, N. Sadagopan, B. Krishnamachari, and A. Helmy, “Modeling path and Computer Engineering, Texas A&M University, College Station. From
duration distributions in MANETs and their impact on reactive routing 1984 to 1989, he was an Assistant Professor and the Founding Director of
protocols,” IEEE J. Sel. Areas Commun., vol. 22, no. 7, pp. 1357–1373, the Division of Computer Systems Engineering, Department of Electrical
Sep. 2004. Engineering and Computer Science, Beijing Information Technology Engi-
[22] V. Naumov, R. Baumann, and T. Gross, “An evaluation of inter-vehicle ad neering Institute, Beijing, China. He was a Research Fellow with the School
hoc networks based on realistic vehicular traces,” in Proc. ACM MobiHoc, of Electrical Engineering, University of Technology, Sydney, Australia, and
May 2006, pp. 108–119. the Department of Electrical and Computer Engineering, James Cook Uni-
[23] S. Wu, C. Lin, Y. Tseng, and J. Sheu, “A new multi-channel MAC pro- versity, Queensland, Australia, under a Fellowship from the Chinese National
tocol with on-demand channel assignment for multi-hop mobile ad hoc Commission of Education. He was a Summer Intern with the Networks and
networks,” in Proc. ISPAN, 2000, pp. 232–237. Distributed Systems Research Department, AT&T Bell Laboratories, Murray
[24] J. So and N. Vaidya, “Multi-channel MAC for ad hoc networks: Han- Hills, NJ, and with AT&T Laboratories Research, Florham Park, NJ, in 1997.
dling multi-channel hidden terminals using a single transceiver,” in Proc. He has published more than 100 research papers. He is an Editor for Wiley’s
MobiHoc ’04, May 2004, pp. 222–233. Journal on Wireless Communications and Mobile Computing, an Editor for
[25] H. Su and X. Zhang, “An efficient single-transceiver CDMA-based MAC the Journal of Computer Systems, Networking, and Communications, and an
protocol for wireless networks,” in Proc. IEEE INFOCOM, May 2007, Associate Editor for John Wiley’s Journal on Security and Communications
pp. 1487–1495. Networks. His research interests include wireless networks and communica-
tions, mobile computing, cross-layer optimizations for QoS guarantees over
mobile wireless networks, effective capacity and effective bandwidth theories
for wireless networks, DS-CDMA, MIMO-OFDM and space-time coding,
adaptive modulations and coding, wireless diversity techniques and resource
allocations, wireless sensor and ad hoc networks, cognitive radio and cooper-
ative communications/relay networks, VANETs, multichannel MAC protocols,
wireless and wired network security, wireless and wired multicast networks,
channel coding for mobile wireless multimedia multicast, network protocols
design and modeling, statistical communications theory, information theory,
random signal processing, and control theory and systems.
Prof. Zhang is a Member of the Association for Computing Machinery
(ACM). He received the U.S. National Science Foundation CAREER Award in
2004 for his research on mobile wireless and multicast networking and systems
and the Texas Engineering Experiment Station Select Young Faculty Award for
Excellence in Research Performance from the Dwight Look College of Engi-
neering, Texas A&M University, College Station, in 2006. He is as an Editor for
the IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, an Associate
Editor for the IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, an Asso-
ciate Editor for the IEEE COMMUNICATIONS LETTERS, and a Guest Editor for
the IEEE WIRELESS COMMUNICATIONS MAGAZINE for the special issue on
the “next generation of CDMA versus OFDMA for 4G wireless applications”
He has frequently served as a Panelist on the U.S. National Science Foundation
Research Proposal Review Panels, a Panelist on the Cross-Layer Optimized
Wireless Networks and Multimedia Communications at the IEEE International
Conference on Computer Communications and Networks (ICCCN) 2007 and
WiFi-Hotspots/WLAN, and a QoS Panel at the IEEE International Conference
on Heterogeneous Networking for Quality, Reliability, Security, and Robust-
ness (QShine) 2004. He is serving or has served as the Cochair of IEEE
Global Communications Conference (Globecom) 2008—Wireless Communi-
cations Symposium and the Cochair for the IEEE International Conference on
Communications (ICC) 2008—Information and Network Security Symposium,
respectively, the Symposium Chair of the IEEE/ACM International Cross-
Layer Optimized Wireless Networks Symposium (IWCMC) 2006, 2007, and
2008, respectively, the Technical Program Committee (TPC) Chair of the
IEEE/ACM International Wireless Communication and Mobile Computing
Hang Su received B.S. and M.S. degrees in electrical Conference (IWCMC) 2006, 2007, and 2008, respectively, the Poster Chair
engineering from Zhejiang University, Hangzhou, of the IEEE Conference on Computer Communications (INFOCOM) 2008, the
China, in 2002 and 2005, respectively. He is cur- Student Travel Grants Cochair of the IEEE INFOCOM 2007, the Panel Cochair
rently working toward the Ph.D. degree at Texas of the IEEE ICCCN 2007, the Poster Chair of the IEEE/ACM International
A&M University, College Station. Symposium on Modeling, Analysis and Simulation of Wireless and Mobile
He is currently a Research Assistant with the Systems (MSWiM) 2007 and the IEEE QShine 2006, and the Publicity Chair
Networking and Information Systems Laboratory, of the IEEE/ACM QShine 2007 and the IEEE WirelessCom 2005. He has
Department of Electrical and Computer Engineering, served as a TPC Member for more than 50 IEEE/ACM Conferences, including
Texas A&M University. He was a Software Engineer the IEEE INFOCOM, IEEE Globecom, IEEE ICC, IEEE Wireless Commu-
with Nokia Research Center, Hangzhou, in 2005. His nications and Networking Conference (WCNC), IEEE Vehicular Technology
research interests include wireless sensor networks Conference (VTC), IEEE/ACM QShine, IEEE International Symposium on
and vehicular ad hoc networks with emphasis on design and analysis of MAC a World of Wireless, Mobile and Multimedia Networks (WoWMoM), IEEE
and routing protocols. ICCCN, etc.