7708 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 65, NO.

9, SEPTEMBER 2016

Design of Analytical Model and Algorithm for

Optimal Roadside AP Placement in VANETs
Bo Zhang, Xiaohua Jia, Fellow, IEEE, Kan Yang, and Ruitao Xie

Abstract—The emerging vehicular ad hoc network (VANET) a competitive technology for Internet access in VANETs. To
enables vehicles to access the Internet through roadside access provide reliable Internet access, more roadside access points
points (APs). An important issue in system deployment is de- (APs) are desired. On the other hand, deploying an AP is very
termining how many roadside APs shall be installed on a road.
However, the existing works fail to provide rigorous and accurate costly, and the Internet service provider (ISP) wants to deploy
analysis for VANETs. In this paper, we propose a general structure as few APs as possible to save money. Therefore, it is important
for Internet access in VANETs. It allows both real-time traffic to determine the minimum number of roadside APs required on
and delay-tolerant traffic to be delivered to users in the most a road for a given user quality of service (QoS) requirement.
efficient ways. An analytical model is also proposed to analyze In WLAN systems, an extensive amount of work was pro-
the system performance with random arrival of the vehicles. We
finally develop an AP placement algorithm based on theoretical posed to study the system performance with multihop relays
results derived from the model to deploy the minimal number [3], [4] and AP placement [5]. In [5], an AP placement method
of roadside APs with quality-of-service (QoS) guarantees. The is proposed to cost-effectively deploy APs with sustainable
simulation results have demonstrated the accuracy of the proposed energy supply and allocate network resources to meet the QoS
analytical model and the efficiency of the proposed algorithm. requirements of users. However, these solutions cannot be used
Index Terms—Analytical model, delay-tolerant traffic, real-time directly in this paper, because the mobility of the vehicles
traffic, roadside access point (AP) placement, vehicular ad hoc makes the roadside AP deployment problem very different
network (VANET). from the WLAN systems. For example, as a vehicle moves
on the road, its distance to an AP is constantly changing but
I. I NTRODUCTION predictable. The transmissions can be scheduled at a certain
time point with a maximum data rate or at a certain location

W ITH the increasing popularity of various types of In-

ternet applications, there is a growing demand for In-
ternet access by services for vehicles, such as car navigation,
with minimum interference to the system. Moreover, to the
best of the authors’ knowledge, there is no analytical model
that can accurately model the system performance in VANETs,
multimedia streaming, voice/video chatting, and web brows- which considers both real-time traffic data and delay-tolerant
ing. The vehicular ad hoc network (VANET) has emerged to traffic data.
enable vehicles to access the Internet via vehicle-to-vehicle In this paper, we develop an analytical model for system
(V2V) and vehicle-to-infrastructure communications. Although performance evaluation in VANETs. Based on the results of the
the cellular-based access techniques (e.g., third generation and analytical model, we also propose an efficient AP placement
long-term evolution) can provide reliable and ubiquitous In- protocol for VANETs that enables the ISPs to deploy the
ternet access to vehicles, it is costly and cannot satisfy the minimal number of roadside APs for Internet services with
ever-increasing growth of mobile data traffic [1]. Wireless local given users’ QoS requirements. The original contributions of
area network (WLAN) systems (e.g., WiFi) can be deployed this paper are listed as follows.
with low cost and high performance, and recent research has
demonstrated that the WiFi-enabled mobile devices on board
can access the Internet when vehicles are moving in the cov- 1) We propose a general structure for Internet access in
erage of WiFi hotspots, which is often referred to as drive- VANETs, which caters for two classes of traffic: real-
thru Internet access [2]. Therefore, the WLAN has become time traffic and delay-tolerant traffic. This structure uti-
lizes different strategies to deliver these two classes of
traffic with optimal system throughput, while preserving
Manuscript received February 15, 2015; revised July 6, 2015; accepted fairness among the vehicles.
September 2, 2015. Date of publication October 15, 2015; date of current 2) We develop an analytical model for VANETs with In-
version September 15, 2016. The review of this paper was coordinated by
Dr. C. Chen. (Corresponding author: Kan Yang.) ternet access, which estimates the system performance
B. Zhang, X. Jia, and R. Xie are with the Department of Computer Science, with randomized distances between the vehicles. This is
City University of Hong Kong, Kowloon, Hong Kong (e-mail: Bo.Zhang@ the first work that analyzes the system throughput with
my.cityu.edu.hk; [email protected]; [email protected]).
K. Yang is with the Department of Electrical and Computer Engineering, combined real-time traffic and delay-tolerant traffic.
University of Waterloo, Waterloo, ON N2L 3G1, Canada (e-mail: kan.yang@ 3) We present a binary search algorithm to find the maxi-
uwaterloo.ca). mum length of a road segment that an AP can serve for
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. given users’ QoS requirements, and thus, the entire road
Digital Object Identifier 10.1109/TVT.2015.2491358 can be served by using the minimum number of APs.
0018-9545 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
ZHANG et al.: ANALYTICAL MODEL AND ALGORITHM FOR OPTIMAL ROADSIDE AP PLACEMENT IN VANETs 7709

Fig. 1. Segment of the road in VANETs.

The remainder of this paper is organized as follows. We efficient direct transmissions at short distances. Thus, they are
discuss the related work in Section II and present the system not economically efficient for the ISPs.
model and the problem statement in Section III. A general In this paper, we propose a system structure for two classes of
system structure that caters for both real-time traffic and delay- traffic: real-time traffic and delay-tolerant traffic. By separating
tolerant traffic is proposed in Section IV. We develop an analyt- the network traffic into two classes, there is no connectivity
ical model of the given traffic delivery strategies in Section V requirement for delay-tolerant traffic. Thus, we are able to take
and a planning algorithm for network operators in Section VI. full advantage of the mobility of the vehicles and propose the
We present the simulation results in Section VII and, finally, delivery strategies with optimal theoretical throughput for both
conclude this paper in Section VIII. classes of traffic. The corresponding analytical model and AP
placement algorithm can help the ISPs deploy the minimal
number of roadside APs with QoS guarantees.
II. R ELATED W ORK
III. S YSTEM M ODEL AND P ROBLEM S TATEMENT
The idea of utilizing WiFi technology to provide low-cost
Internet services for moving vehicles was introduced by Ott and In VANETs, APs are deployed along the road, as the RSU,
Kutscher in [2], which is called the drive-thru Internet system. to provide Internet services to the moving vehicles. Since a
In the proposed system, a limited number of APs are placed on long road is mostly straight, we assume that the road in our
city roads to provide occasional islands of connectivity to the consideration is a straight line and can be divided into segments
moving vehicles. where one roadside AP (also called a roadside unit) is installed
There are many existing works that studied the architecture in each segment. Our task is to determine the minimum number
or feasibility [6]–[8] of drive-thru Internet systems. Some of of roadside APs required to serve the whole road for the given
them tried to improve the system performance by utilizing users’ QoS requirements. This objective is equivalent to finding
different medium access control or network layer technologies the maximum length of a road segment such that one AP is
[6]–[17]. However, most of them consider the connectivity of installed in the middle of the segment and that the QoS require-
the vehicles to be more important than the system throughput ments can be met. Therefore, we will focus on a subsystem
and make all vehicles in the transmission range of an AP share that contains a segment of road, as shown in Fig. 1. In this
the wireless medium, which degrades the average performance. subsystem, one roadside AP serves all vehicles in this road
Some works have been also proposed for wireless coverage segment. This road segment is called the service segment and
issues in WLANs [5] and wireless sensor networks [18]–[24]. In denoted by S. The length of the service segment is denoted by
[5], an AP placement method is proposed to cost-effectively de- dS , which could be much longer than the transmission range of
ploy APs with sustainable energy supply and allocate network an AP. The vehicles in a service segment may connect to the AP
resources to meet the QoS requirements of users. In VANETs, directly or through some intermediate vehicles. We assume that
the AP placement problem becomes a challenging issue due to all antennas installed on the vehicles and the AP are of the same
the mobility of vehicles. height, because the vehicle antennas are likely to be installed at
The traffic data in VANETs may consist of both real-time the similar height, and compared with the horizontal distance
data [25]–[27] and delay-tolerant data [28]. In [29], Ota et al. of the vehicles, the antenna height difference can be usually
proposed a correlation-based approach to predict urban traffic negligible. Thus, the distance of any two nodes (i.e., vehicle or
information. They also proposed a cooperative downloading AP) can be represented by the difference in their x-coordinate
algorithm to minimize an average delivery delay of each user values, and the system can be represented in a 1-D line.
request while maximizing the amount of data packets down- To construct an analytical model, without loss of generality,
loaded from the roadside unit (RSU) [27]. In [30], continuous- we make the following assumptions: We assume that the vehi-
time Markov networks are applied to analytically capture the cles arrive at Poisson distribution with the arrival rate λ, and all
operation of carrier-sense multiple access with collision avoid- vehicles move toward the same direction with the same speed v,
ance (CSMA/CA) networks. Some other works introduced for example, platooning on a highway. The vehicles going the
V2V relay schemes [31]–[33] to extend the service range opposite direction will be studied separately in the same way.
of roadside APs and allow the vehicles to maintain real- We assume that all vehicles and the AP use the maximum
time network connections rather than intermittent connections. allowed transmit power in WiFi systems and, thus, have the
However, the major problem of V2V relay is the system per- same transmission range dtx . For any two nodes vi and vj that
formance. These schemes do not consider the delay-tolerant are within the transmission range, there is a communication link
traffic and allocate much of the wireless spectrum time frame li,j . The transmission data rate ri,j of the link is determined by
for long-distance multihop transmissions rather than the more the distance di,j of the two nodes.
7710 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 65, NO. 9, SEPTEMBER 2016

TABLE I where Rrt and Rdt are the theoretical per-vehicle throughput
N OTATIONS
for real-time and delay-tolerant traffic, respectively, and tdelay
is the worst case of delay of delivering a message from a vehicle
to a destination.
If dS is short, it can surely satisfy the QoS requirements, but
requires more APs to cover the entire road; if dS is too long,
QoS requirements cannot be met due to too many users to share
the bandwidth of the AP and too many hops of transmissions
for the vehicles at either end of the segment. There exists an
optimal value of dS . We will present the method to find this
optimal dS in the following sections.

IV. G ENERAL S YSTEM S TRUCTURE FOR V EHICULAR A D

H OC N ETWORKS W ITH I NTERNET ACCESS
We use the protocol interference model in this paper. Let dinf In this work, we consider two classes of traffic: real-time
denote the interference range of wireless communications. We traffic and delay-tolerant traffic. Since these two classes of
define the following interference. traffic have different delivery requirements, we will introduce
different delivery strategies to maximize the system throughput.
• There is interference between a node vi and a node vj
iff di,j ≤ dinf .
• There is interference between a node vi and a link lj,k A. Real-Time Traffic Delivery
iff vi interferes with either vj or vk .
In the delivery of real-time traffic, every vehicle needs to
• There is interference between a link li,j and a link lk,l
maintain real-time network connection to the AP. This real-time
iff either vi or vj interferes with link lk,l .
connection can be either via a direct link or via some intermedi-
ate relays. It is a complicated issue to select the relay nodes for a
We consider two classes of network traffic in the system: vehicle, because it depends on the following two factors: 1) the
1) real-time traffic, which is expected to be delivered in real total wireless spectrum time being occupied by transmitting one
time, such as the traffic of instant message, online game, and packet by a direct link or several relay links (this is also referred
voice or video chat; and 2) delay-tolerant traffic, which allows to as packet transmission time in the following paragraphs) and
the messages to be delivered within a tolerable delay (i.e., 2) the interference level of each related link. In the case of a
from a few seconds to 1–2 min), such as the traffic for e-mail, line topology, the traffic is always transmitted along the same
large-file downloading, and multimedia streaming (with suffi- straight path. Since every node has the same transmission range,
cient caching). Some of the basic notations are summarized selecting any relay node would have the similar interference to
in Table I. the system. Thus, we will focus on minimizing the packet trans-
Bandwidth fairness among the vehicles is another objective mission time. We do not consider the packet-level overhead
of the system design. The vehicles that are far away from the (e.g., wireless interframe protection spaces and frame headers)
AP are often disadvantaged in receiving the end bandwidth. In in the calculation of packet transmission time, because these
this paper, we guarantee that all vehicles in the same service overheads heavily depend on the wireless protocol used and the
segment get a similar amount of traffic. We assume that no transmission scheduling. Neglecting these overheads enables us
vehicle is idle in the system. That is, all vehicles have demand to calculate the system raw data rate, which reflects the actual
for both real-time and delay-tolerant traffic. This is an extreme performance well. The existing packet-level analytical models
case of the system traffic, which can be used to evaluate the can be applied to this system when we need higher accuracy.
system performance in the worst case. Let p denote the standard data size of a packet. We introduce
The problem of our consideration is defined as follows. the packet transmission time ti,j over a direct link li,j as
Given the vehicle arrival rate λ and the QoS requirements of follows:
the vehicles (including the throughput requirement for real-time
p
traffic Ruser_rt and delay-tolerant traffic Ruser_dt and the delay ti,j = . (1)
bound tuser ), we are going to determine the maximum length ri,j
dS of the service segment that an AP can serve. Thus Equation (1) is general for any node pair of vi and vj . For the
case that vi has no direct link with vj , we will have ri,j = 0,
Given λ, Ruser_rt , Ruser_dt , tuser and ti,j = ∞.
max dS Consider a vehicle vi that is far away from the AP, and
there are i − 1 vehicles between vi and the AP, namely,
s.t. Rrt ≥ Ruser_rt vi−1 , vi−2 , . . . , v2 , v1 according to their distance to the AP,
Rdt ≥ Ruser_dt where v1 is the closest to the AP. When vi communicates with
an AP, the packet may be relayed through multiple hops. Thus,
tdelay ≤ tuser the packet transmission time between the AP and vi is the
ZHANG et al.: ANALYTICAL MODEL AND ALGORITHM FOR OPTIMAL ROADSIDE AP PLACEMENT IN VANETs 7711

summation of the packet transmission time of all relay links.

Let Ti denote the optimal packet transmission time between the
AP and vi and v0 denote the AP. Note that T0 is 0. The Ti value
for other vehicles can be calculated by the recursive method
shown as follows:

Ti = min {ti,j + Tj }. (2)

0≤j<i

If the minimum Ti is achieved with the help of node vj , we call

vj the optimal next-hop node of vi .
The optimal relay path with minimal packet transmission
time can be computed by using (2). However, in real systems,
the vehicles move at high speed, and their optimal relay path Fig. 2. Loading zone for vehicle B.
would change from time to time. It is not efficient to let the
vehicles recalculate their optimal next hops before every trans- rate delivery strategy, which makes the AP always transmit
mission. To solve this problem, we define the link efficiency ei,j or receive data on its highest available data rate. This strategy
for link li,j as the ratio of the packet transmission time to the achieves the theoretical optimal system throughput. However,
distance of the link, i.e., as the distance between the vehicles is not the same, some
di,j vehicles may have a longer period of time during which they
ei,j = . (3) are closest to the AP, whereas others have a shorter period of
ti,j
time being closest to the AP. To preserve the throughput fairness
A link with better link efficiency indicates that the link among the vehicles, we will present a carry-and-forward deliv-
uses shorter transmission time to transmit the packet for a ery strategy that balances the bandwidth among the vehicles
unit distance. Thus, selecting the links with the highest link without affecting the efficiency of the optimal highest data rate
efficiency in transmissions would help achieve minimal packet delivery strategy. The combined use of these two strategies
transmission time. We can select the next-hop node for a vehicle would achieve the optimal system throughput while preserving
vi as follows: 1) Find the node vj , 0 ≤ j < i that achieves the fairness among the vehicles.
highest link efficiency ei,j , and 2) let vj be the next-hop node of 1) Highest Data Rate Delivery: The performance of a net-
vi . This is called the max-efficiency relay strategy. This strategy work system is measured by the total amount of data that
would achieve a similar result as the recursive method defined are finally delivered to the destinations. As the VANET with
by (2) when the distance between the AP and vi is long. Internet access works in infrastructure mode, all traffic in the
The next-hop node of any vehicle selected by the max- system is destined to or originated from the AP. When the AP
efficiency relay strategy is stable and does not change for a transmits or receives signals, if it is not interfered by the other
relatively long period of time. This is because this strategy links, and always selects the link with the highest data rate,
selects the next-hop node mainly based on the intervehicle link the performance of the entire network would be theoretically
distance and the data rate, which does not change that often, at maximum.
compared with the AP–vehicle distance. Only when a vehicle Consider the transmission policy that the AP always com-
enters in or exits from the direct transmission range of an municates with the closest vehicle in the region. We define
AP, or switches between different service segments, shall the the loading zone of a vehicle vi as the part of the road where
next-hop relay node be updated. Thus, the max-efficiency relay vi is the closest vehicle to the AP. As shown in Fig. 2, there
strategy can achieve the similar relay throughput with much less is a loading zone sB for vehicle B. Different vehicles have
computation and operation cost. different loading zones. When a vehicle enters its loading zone,
it becomes the closest vehicle to the AP, and the AP only
communicates with it. A loading zone can be further divided
B. Delay-Tolerant Traffic Delivery
into two subzones by the location of the AP, namely, the left
Many types of Internet traffic are delay tolerant. For example, loading zone and the right loading zone. The length of the left
in most cases, a 1–2-min delay in e-mail delivery is acceptable. loading zone of vehicle vi is equal to half the distance between
For the scenario of large-file downloading, as the transmission vi and the vehicle ahead of it. For example, the length of sBL
process of the file is expected to be relatively long, a short delay is equal to half the distance between vehicles A and B, because
in delivery can be tolerated. Even in the case of online media when vehicle B enters sBL , it becomes closer to the AP than
streaming, if a sufficient length of media content is cached, the vehicle A. Similarly, the length of the right loading zone of vi
delivery of the rest of the content is also delay tolerant, because is equal to half the distance between vi and the vehicle after it.
a tolerable delay would not affect the playback of the media. For different vehicles, the length of their loading zones can
The delay-tolerant nature of the traffic enables us to make be different. For example, if vehicle B is close to both vehicles
use of the mobility of the vehicles. The transmissions between a A and C, sB would become very short. Since the amount of data
vehicle and the AP can be postponed until this vehicle becomes received by a vehicle from the AP is determined by the length
the closest to the AP, where the data rate between this vehicle of its loading zone, the highest data rate delivery policy is not
and the AP is the highest. This is called the highest data fair to the vehicles with a short loading zone.
7712 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 65, NO. 9, SEPTEMBER 2016

2) Carry-and-Forward Delivery: We introduce a carry-and- A. Throughput of Real-Time Traffic

forward delivery strategy for downlink traffic of the vehicles,
The real-time traffic is connection oriented. Thus, the
such that the bandwidths of the vehicles could be balanced,
throughput of real-time traffic mainly depends on the packet
whereas the highest theoretical throughput achieved by the
transmission time between every vehicle and AP, as well as
highest data rate delivery strategy is not affected.
the interference situation. We will first estimate the packet
Initially, the AP estimates the average size of data Cavg
transmission time.
that can be transmitted in a loading zone. Then, the AP uses
According to the max-efficiency relay strategy, a vehicle
the carry-and-forward delivery strategy between the vehicles to
selects a link with the highest link efficiency to transmit the
ensure that the size of data finally delivered to every vehicle
traffic between it and the AP. In practice, since less number
is at least ρCavg (i.e., 90% Cavg ). The vehicles with a longer
of relay hops is usually more efficient in traffic relay, the best
loading zone may be selected to carry some downlink traffic
length of a relay hop is the transmission range dtx . We define
for some other vehicles that have a shorter loading zone and
the relay distance (rd) of a vehicle to be the distance to its
forward the traffic to the destination when they move out of the
best available vehicle for relay. The relay distance of vi is
interference range of the AP. The carry-and-forward delivery
determined by the locations of the vehicles between vi and
strategy does not affect the efficiency of the highest data rate
the AP. As the vehicles arrive in a random process, the relay
delivery strategy, because the transmissions of the forwarding
distance is also a random variable between 0 and dtx . Let
traffic are always scheduled outside the AP’s interference range. pdf
frd (d) denote the probability density function (pdf) of the
In particular, the parameter ρ is configurable. A larger ρ
relay distance. It can be used to calculate the probability density
would result in better balanced traffic but with longer inter- pdf
of any given relay distance d. frd (d) can be formulated as
vehicle forwarding time and distance, whereas a smaller ρ
would result in less balanced traffic but with shorter intervehicle pdf
frd (d) = λ × e−λ(dtx −d) , 0 ≤ d ≤ dtx . (4)
forwarding time and distance. We will analyze the performance
of different values of ρ in the simulation section. The average packet transmission time of a direct link with
The detailed steps of the carry-and-forward delivery strategy distance d is
are as follows.
p
1) The AP transmits downlink traffic to vehicles sequentially t(d) = , 0 ≤ d ≤ dtx . (5)
r(d)
using the highest data rate delivery policy.
2) If the length of a loading zone of a vehicle is not sufficient r(d) is the data rate of a link where the distance of two end
to transmit ρCavg data, the entire loading zone will be nodes is equal to d.
used to transmit data for itself, and the rest of the data In multihop relay systems, the traffic may be relayed in
will be queued at the AP. multiple links. Let T (d) denote the average packet transmission
3) If the length of a loading zone of a vehicle vi is long time between vehicle vi and the AP with distance d using
enough to transmit ρCavg data, the ρCavg data of this a multihop relay. Suppose vk is the relay node for vi , the
vehicle will be transmitted first, and the queued data on packet transmission time between vi and the AP is equal to the
the AP will be then sent to vi . If there is no queued data, packet transmission time between vi and vk , plus the packet
vi can make use of this period of time to transmit some transmission time between vk and the AP. This could be a
extra data for itself. recursive calculation, because vk may be also far away from
4) When a vehicle moves out of the interference range of the AP and has to use some relays for communication. As the
the AP, if it carries some data for the other vehicles, it relay distance between vi and vk is a random variable, we can
will forward the data to them. calculate T (d) using the following recursive method:
5) If the requested data are not delivered after a predefined  d
tx pdf
deadline, the data will be redelivered by the AP through 0 (T (d−x)+t(x)) frd (x)dx, if d > dtx
T (d) =
the multihop real-time relay connection instead. t(d), if 0 ≤ d ≤ dtx .
(6)
In multihop relay systems, any two interfering links cannot
transmit simultaneously. Suppose we have the amount of traffic
V. A NALYTICAL M ODEL FOR D ELIVERY OF R EAL -T IME
scheduled to be transmitted on each link and a timeframe T for
T RAFFIC AND D ELAY-T OLERANT T RAFFIC
transmitting all traffic on the link. Let LI (vk ) denote the set of
As we divide the road into segments, the optimal AP place- links that interfere with node vk . The interference constraint is
ment problem is equivalent to the problem of finding the maxi- defined as follows.
mum length dS of a service segment S. Here, we will present an For any link set LI (vk ), the summation of the transmission
analytical model to calculate the expected per-vehicle through- time required for transmitting the traffic on these links shall
put of any given dS . This analytical model is essential for the never exceed T .
planning algorithm of optimal AP placement. Since we utilize To deliver real-time traffic, every vehicle in the system
different delivery strategies for the two classes of traffic, we will maintains a real-time connection to the AP either via a direct
calculate the theoretical per-vehicle throughput Rrt for real- communication link or via some intermediate relays. These
time traffic and Rdt for delay-tolerant traffic separately. multihop communication routes form a tree, where the AP is the
ZHANG et al.: ANALYTICAL MODEL AND ALGORITHM FOR OPTIMAL ROADSIDE AP PLACEMENT IN VANETs 7713

root. The throughput of such a multihop system highly depends Since the length of a subzone (i.e., a left loading zone or a
on the interference of the links. It is obvious that the bottleneck right loading zone) is equal to half of the vvd, it is also a random
cdf
of the network is at the AP, because all traffic is originated from variable. Let fszd (d) denote the cdf of a subzone’s distance
or destined to the AP. Therefore, when calculating the packet (szd). It can be formulated as follows:
transmission time between any two nodes, we can consider only
the total transmission time required by the links in set LI (AP). cdf
fszd cdf
(d) = fvvd (2d) = 1 − e−λ2d . (11)
By doing so, the interference constraint can be easily verified.
Let TAP (d) denote the transmission time required by LI (AP) 2) Size of Data Transmitted in a Loading Zone: When a
of transmitting a packet between a vehicle and the AP with vehicle moves inside any subzone with distance d, its distance
distance d. TAP (d) can be expressed as to the AP changes either from 0 to d or from d to 0. Let
C(d) denote the total size of traffic that can be transmitted in
TAP (d) a subzone with distance d. Since we assume that the vehicle
moves at an even speed v, C(d) can be calculated as follows:
⎧ d
⎪
⎪
tx
TAP (d−x)frd pdf
(x)dx, if d ≥ dinf + dtx
⎪
⎪ 0 d
⎪
⎪ d−d inf pdf
⎨0  T AP (d−x)f rd (x)dx C(d) = r(x)dt
dtx pdf
= + d−d (T (d−x) + t(x))frd (x)dx, if dinf < d
⎪
⎪ inf 0
⎪
⎪ < dinf + dtx
⎪
⎪
⎩T (d), if 0 ≤ d < dinf . d
x
(7) = r(x)d
v
0
For a vehicle inside the service segment of an AP, its distance
d
to the AP is constantly changing. The average transmission 1
time required by LI (AP) of transmitting a packet between this = r(x)dx. (12)
v
vehicle and the AP is 0

 dS  dS
Equation (12) can be used to calculate the size of data that can
0
2
TAP (x)dx 2 0
2
TAP (x)dx
Tavg = = . (8) be transmitted within a subzone with distance d. Given a size of
dS dS
2 data c, we can use the reverse function of (12) to calculate the
length of the subzone required, i.e.,
According to the Poisson distribution, the average number
of vehicles in a service segment is λdS . That is, the expected d = C −1 (c). (13)
transmission time used for every vehicle to transmit a packet
with the AP is Tavg λdS . Thus, the per-vehicle throughput of As the length of a subzone is a random variable, the size
real-time traffic is of data that can be transmitted in a subzone is also a random
p p variable. Let z denote a random loading zone and (1/2)z a
Rrt = =  dS . (9) cdf
subzone (i.e., half of a loading zone). Let f(1/2)z (c) denote the
Tavg λdS 2λ 0 2 TAP (x)dx cdf of the size of data c that can be transmitted in a subzone. It
can be expressed by substituting (13) into (11) as follows:
B. Throughput of Delay-Tolerant Traffic −1

f cdf
1
cdf
z (c) = fszd C (c)
2
The delay-tolerant traffic is delivered intermittently. It is
transmitted from the AP to a vehicle only when this vehicle = 1 − λe−λ2C
−1
(c)
. (14)
is inside its loading zone. To calculate the expected per-vehicle
throughput Rdt for delay-tolerant traffic, we will first calculate
We could derive the corresponding pdf as follows:
the average size of data transmitted in a loading zone. Since a
vehicle is only served once by an AP in a service segment and  
d f cdf
1 (c)
the total time a moving vehicle stays in this service segment is f pdf (c) = 2z
. (15)
1
known, we will be able to calculate the per-vehicle throughput. 2z dc
1) Length of a Subzone: Since we assume that the arrival of
the vehicles follows Poisson distribution with the arrival rate λ, Let fzpdf (c) denote the pdf of the size of data c that can be
the distance between any two consecutive vehicles follows the transmitted in a full loading zone. It represents the probability
exponential distribution with parameter λ. Let fvvdcdf
(d) denote density of any given size of data c to be transmitted within a
the cumulative distribution function (cdf) of the vehicle-to- full loading zone. Since a loading zone contains two individ-
vehicle distance (vvd) between any two consecutive vehicles. ual subzones, it can be formulated as the self-convolution of
pdf
This function can be defined as follows: f(1/2)z (c), i.e.,
 
cdf
fvvd (d) = 1 − e−λd . (10) fzpdf (c) = f pdf
1
z
∗ f pdf
1
z
(c). (16)
2 2
7714 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 65, NO. 9, SEPTEMBER 2016

Let Cavg denote the average size of data that can be transmit- this strategy enables the vehicles to postpone the intervehicle
ted in a random loading zone. Cavg can be then calculated as relay traffic until they get out of the interference range of an AP.
follows: The worst case delay in delivering a message to a destination
∞ vehicle is the summation of the two types of delay, i.e.,
Cavg = fzpdf (x)xdx. (17) tdelay = tdelay_hd + tdelay_cf . (21)
0
As the users’ QoS requirement for delivery delay is defined
3) Average Per-Vehicle Throughput: Consider the case that to be tuser , we must ensure that the worst-case delay tdelay is
all vehicles move at an even speed v in the system, and there is smaller than tuser .
a service segment of an AP with distance dS . For any moving For the highest data rate delivery strategy, given the distance
vehicle, the time of this vehicle to stay in this service segment dS of a service segment and the moving speed v of the vehicles,
is dS /v. Since this vehicle is only served once by the AP during we can calculate tdelay_hd as follows, which represents the
this period of time and the average size of data received is Cavg , worst case interval that a vehicle cannot access any AP:
we are able to calculate the average per-vehicle throughput of
delay-tolerant traffic as follows: dS
tdelay_hd = . (22)
Cavg v v
Rdt = = Cavg . (18)
dS dS Suppose the vehicles are numbered as v1 , v2 , . . . , vn from the
v
right side to the left side of the road. According to the process
of delay-tolerant traffic delivery introduced in Section IV, once
C. Total Throughput
a vehicle vk cannot receive the desired amount of data ρCavg ,
We now have the theoretical per-vehicle throughput for both the vehicle after it (i.e., vk+1 ) would offer help. If these two
real-time and delay-tolerant traffic delivery strategies. We will vehicles cannot receive the desired amount of data 2ρCavg , the
jointly use these two strategies to increase the flexibility of vehicle after them (i.e., vk+2 ) would offer help. This process
the system. Let α denote the ratio of system transmission time continues until a vehicle vk+x arrives, and the total amount
allocated for real-time traffic and β for delay-tolerant traffic. of data received by the vehicles vk , vk+1 , . . . , vk+x is greater
We can update calculations of Rrt and Rdt as follows: than (x + 1)ρCavg . We define a cooperative group to be a
p sequence of consecutive vehicles, such that the total amount of
Rrt = α  dS data transmitted from the AP to these vehicles is larger than
2λ 0 2 TAP (x)dx the total size of data that shall be guaranteed for them. We
v
Rdt = βCavg . (19) define the size of a cooperative group to be the number of vehi-
dS cles in it.
They can be used to verify whether the theoretical per-vehicle For the carry-and-forward delivery strategy, there are three
throughput for both classes of traffic satisfies the users’ QoS types of delay as follows.
requirements Ruser_rt and Ruser_dt . 1) In case that a vehicle does not receive sufficient data
We can calculate the total per-vehicle throughput Rtotal as in its loading zone, the transmission is postponed, and
follows: the traffic will be delivered to another vehicle in the
cooperative group. Let Navg denote the average size of
Rtotal = Rrt + Rdt a cooperative group, the worst case of this delay can be
p v expressed as Navg /λv.
= α  dS + βCavg . (20)
λ 0 TAP (x)dx dS 2) The intervehicle traffic is carried by some vehicles until
they move out of the interference range of the AP. This
Note that the total transmission time allocated Γ = α + β delay can be expressed as dinf /v.
can be less than 100%, because we only consider the vehicles 3) It takes some time for a vehicle to forward the traffic
moving toward the same direction. The vehicles in the opposite to a destination vehicle. Let Crelay denote the average
direction are considered as under a different system, and they size of relayed data transmitted in a cooperative group,
share the same wireless spectrum. this delay can be expressed as (Crelay /p) × T ((Navg ) −
1/λ), where (Navg − 1)/λ is the worst-case distance of
D. Delay Bound Analysis of Delay-Tolerant Traffic intervehicle traffic relay in a cooperative group.
Apart from the throughput, the delay bound of traffic delivery The average delivery delay of the carry-and-forward delivery
would also affect the service quality of the system. There are strategy can be then calculated as follows:

two types of delay in delivering a message to a destination Navg dinf Crelay Navg − 1
vehicle. The first type of delay is caused by the highest data rate tdelay_cf = + + ×T . (23)
λv v p λ
delivery strategy, which is denoted by tdelay_hd , because this
strategy always selects the vehicle with the highest data rate to We will calculate Navg and Crelay in the following para-
receive data from an AP and postpones the transmissions of the graphs. Suppose a vehicle is the first vehicle of a cooperative
other vehicles. The second type of delay is caused by the carry- group (i.e., the rightmost vehicle), and it does not receive
and-forward strategy, which is denoted by tdelay_cf , because sufficient data within its loading zone (i.e., there must be at least
ZHANG et al.: ANALYTICAL MODEL AND ALGORITHM FOR OPTIMAL ROADSIDE AP PLACEMENT IN VANETs 7715

pdf Since we are able to calculate the per-vehicle throughput for

two vehicles in this cooperative group). Let fz− (c) denote the
pdf of the size of traffic this vehicle receives. It can be expressed any given dS , we can use a binary search method to calculate
as follows: the optimal dS .

pdf fzpdf (c), if 0 ≤ c ≤ ρCavg
fz− (c) = (24) A. Bounds of dS
0, if c < 0 or c > ρCavg .
For the given β and Ruser_dt , we can use (19) to calculate
For the kth vehicle in a cooperative group, if it does not the maximum dS that an AP can support. Suppose we allocate
receive sufficient traffic within its loading zone (i.e., there are all transmission time to delay-tolerant traffic, i.e., β = Γ, and
at least k + 1 vehicles in this cooperative group), the pdf of the only consider the QoS requirement of delay-tolerant traffic.
size of traffic that all these k vehicles receive is The calculated maximum dS can be regarded as the upper-
  bound distance, because by adding the consideration of the QoS
pdf
pdf f (k−1)z− ∗ f pdf
z (c), if 0 ≤ c ≤ kρCavg requirement of real-time traffic, an AP can no longer serve so
fkz− (c) = large an area with QoS guarantees, and the length of the service
0, if c < 0 or c > kρCavg .
segment would definitely reduce. Let dS+ denote the upper-
(25)
bound distance of dS , it can be calculated as follows:
The probability that the size of a cooperative group N is v
greater than k (i.e., at least k + 1) is dS+ = ΓCavg . (30)
Ruser_dt
∞
pdf The lower bound dS− can be assumed to be 0 in the initial
P (N > k) = fkz− (x)dx, k ≥ 1. (26)
state.
0

In particular, P (N > 0) = 1. The probability that the size of a B. Binary Search of dS

cooperative group N is equal to k is
After we find the upper and lower bounds of dS , we can
P (N = k) = P (N > k) − P (N > k − 1), k ≥ 1. (27) use the binary search method to find the optimal length dS , as
shown in Algorithm 1. We can then optimally place the APs on
The average size of a cooperative group Navg for traffic relay the road with distance dS .
can be then calculated as follows:
∞
 Algorithm 1 BiSearch dS
Navg = P (N = k) × k. (28)
0 Input dS+  the upper-bound distance of dS
Input dS−  the lower-bound distance of dS
In a cooperative group, the total size of relayed data trans- Output dS  the optimal length of dS
mitted by the vehicles is equal to the total size of relayed data 1: while Rrt is not close enough to Ruser_rt do
received by them. Thus, to calculate Crelay for a vehicle, we 2: dS_mean ← (dS+ + dS− )/2;  the current segment
can just calculate the total size of data that is required by the length
vehicles with insufficient length of loading zones, i.e., 3: Calculate the minimum β required to make Rdt ≥
Ruser_dt ;
 avg
ρC
4: Calculate transmission time ratio α = Γ − β;
Crelay = Navg fzpdf (x)(ρ ∗ Cavg − x)dx. (29) 5: Calculate the expected throughput for real-time traffic
0 Rrt with α;
6: if Rrt > Ruser_rt then
VI. P LANNING A LGORITHM FOR O PTIMAL 7: dS− = dS_mean ;  as the AP has the potential to pro-
ACCESS P OINT P LACEMENT vide service to a larger area
8: end if
The objective of the optimal AP placement problem is equiv- 9: if Rrt < Ruser_rt then
alent to finding the maximum length of a service segment with 10: dS+ = dS_mean ;  as the AP cannot support this length
one AP that the users’ QoS requirements can be met. Thus, of service segment
the planning algorithm presented in this section will focus on 11: end if
the maximum length of a service segment. The problem is 12: end while
defined as follows. The network settings (i.e., speed v, arrival 13: return dS = dS_mean ;
rate λ, and allocatable transmission time ratio Γ) and the QoS
requirements Ruser_rt and Ruser_dt are given. We are going to
determine the maximum distance dS of the service segment, as
VII. S IMULATION R ESULTS
well as the optimal ratio α and β used for either real-time traffic
or delay tolerant traffic, such that the QoS requirements can be In the simulation, we created a one-way highway model of
satisfied. 10 000 m in SUMO 0.23, where the highway has four lanes
7716 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 65, NO. 9, SEPTEMBER 2016

TABLE II
PARAMETERS U SED IN S IMULATION

and three entrances. We simulate the traffic on the highway for

half an hour. In the simulated traffic data set, the vehicles enter
the highway randomly from different entrances. The arrival rate
and the speed of the vehicles change from time to time. The
average vvd is 44.29 m, and average speed is 24.78 m/s (i.e.,
89.2 km/h). To evaluate the accuracy of the analytical model,
we developed a simulator by C#. The wireless transmission Fig. 3. Comparison of average per-vehicle throughput.
parameters used in the simulation are listed in Table II. The
data rate between any two nodes with distance d is assumed
to follow the Shannon–Hartley theorem and can be calculated
as follows:

r(d) = 0.332B(Prx (d) − PN ) (31)

where Prx (d) denotes the received signal strength (in dBm) at
distance d and can be calculated as follows:

Prx (d) = Ptx + Abase − a × 10 × log10 (d). (32)

The vehicles in the highway use the delivery strategies

presented in Section IV for both real-time traffic and delay-
tolerant traffic. As we do not consider the packet-level traffic
scheduling in this work, the simulator evaluates the system
performance by packet transmission time between the nodes Fig. 4. Distribution of per-vehicle throughput for real-time traffic.
and does not simulate/schedule the transmissions at the packet
level. That is, the system performance is measured in RAW data B. Throughput Distribution
rate (the theatrical data rate without considering the packet-
As the vehicles arrive in a random process, the number of
level overhead).
vehicles in a service segment can be different from time to time.
Fig. 4 shows the distribution of per-vehicle throughput for real-
time traffic when dS = 3000 m and when α = 100%. There
A. Average Throughput
is no throughput balance algorithm used in real-time traffic
We first simulate the cases of the vehicle system with dif- delivery. That is, when the density of the vehicles in a service
ferent service segment distance dS values. We allocate 50% segment becomes higher, the throughput of all vehicles in this
of the transmission time for real-time traffic and another 50% service segment would decrease. However, for most cases, the
of the transmission time for delay-tolerant traffic. As shown in vehicles can still achieve reasonable throughput. For example,
Fig. 3, the analytical results are very close to the simulation for more than 90% of the cases, the vehicles could achieve at
results, particularly for the real-time traffic (with only 0.11% least 80% of the expected throughput.
difference). The simulation result is 4.79% better than the In a delay-tolerant traffic delivery system, the situation be-
analytical result on delay-tolerant traffic. This is because the comes different, because we utilize the carry-and-forward strate-
vvd in the same lane tends to be smoothed out in a long run gy to balance the traffic between vehicles to preserve throughput
(i.e., the vehicles will not become too close or too far away). fairness. Fig. 5 shows the distribution of per-vehicle throughput
Note that the scaling of the y-axis is in logarithm. As dS grows, for delay-tolerant traffic when dS = 3000 m, β = 100%, and
the throughput of delay-tolerant traffic could be more than ten ρ = 90%. That is, we balance the traffic among the vehicles
times higher than real-time traffic. This is because the delay- to make every vehicle achieve at least 90% of the expected
tolerant traffic of a vehicle is always delivered through a direct throughput. As shown in the figure, more than 67% of the
link at the highest data rate, whereas the real-time traffic of a vehicles achieve the throughput around the lower bound, while
vehicle can be relayed for many hops. some vehicles achieve higher throughput.
ZHANG et al.: ANALYTICAL MODEL AND ALGORITHM FOR OPTIMAL ROADSIDE AP PLACEMENT IN VANETs 7717

VIII. C ONCLUSION
In this paper, we have developed a general system structure
to maximize the system performance of vehicle networks. For
real-time traffic, the links with maximum link efficiency were
used to a construct multihop relay tree rooted from the AP to
deliver the traffic. For delay-tolerant traffic, the highest data
rate delivery and the carry-and-forward strategy were utilized to
deliver the traffic with the optimal throughput while preserving
fairness among the vehicles. An analytical model was proposed
to evaluate the system performance, and a binary search algo-
rithm was presented to enable the operators to deploy a minimal
number of roadside APs while satisfying the QoS requirements
of the vehicles.

R EFERENCES
Fig. 5. Distribution of per-vehicle throughput for delay-tolerant traffic. [1] N. Cheng, N. Lu, N. Zhang, X. S. Shen, and J. W. Mark, “Vehicular
WiFi offloading: Challenges and solutions,” Veh. Commun., vol. 1, no. 1,
pp. 13–21, Jan. 2014.
[2] J. Ott and D. Kutscher, “Drive-thru Internet: IEEE 802.11b for automobile
users,” in Proc. IEEE INFOCOM, 2004, vol. 1, pp. 362–373.
[3] G. Bianchi, “Performance analysis of the IEEE 802.11 distributed coordi-
nation function,” IEEE J. Sel. Areas Commun., vol. 18, no. 3, pp. 535–547,
Mar. 2000.
[4] B. Zhang and X. Jia, “Multi-hop collaborative relay networks with con-
sideration of contention overhead of relay nodes in IEEE 802.11 DCF,”
IEEE Trans. Commun., vol. 61, no. 2, pp. 532–540, Feb. 2013.
[5] Z. Zheng, L. X. Cai, M. Dong, X. Shen, and H. V. Poor, “Constrained
energy-aware AP placement with rate adaptation in WLAN mesh net-
works,” in Proc. IEEE GLOBECOM, 2011, pp. 1–5.
[6] V. Bychkovsky, B. Hull, A. Miu, H. Balakrishnan, and S. Madden,
“A measurement study of vehicular internet access using in situ Wi-Fi net-
works,” in Proc. ACM MobiCom, New York, NY, USA, 2006, pp. 50–61.
[7] D. Hadaller, S. Keshav, T. Brecht, and S. Agarwal, “Vehicular oppor-
tunistic communication under the microscope,” in Proc. ACM MobiSys,
New York, NY, USA, 2007, pp. 206–219.
[8] A. Balasubramanian, R. Mahajan, A. Venkataramani, B. N. Levine, and
J. Zahorjan, “Interactive WiFi connectivity for moving vehicles,” Proc.
SIGCOMM, vol. 38, no. 4, pp. 427–438, Aug. 2008.
[9] M. Wang et al., “Asymptotic throughput capacity analysis of VANETs
Fig. 6. Average size of the cooperative group with different ρ values. exploiting mobility diversity,” IEEE Trans. Veh. Technol., vol. 64, no. 9,
pp. 4187–4202, Sep. 2015.
[10] M. Wang et al., “Real-time path planning based on hybrid-VANET-
C. Selection of ρ enhanced transportation system,” IEEE Trans. Veh. Technol., vol. 64,
no. 5, pp. 1664–1678, May 2015.
In the delivery of delay-tolerant traffic, the carry-and-forward [11] N. Cheng et al., “Opportunistic spectrum access for CR-VANETs: A
game-theoretic approach,” IEEE Trans. Veh. Technol., vol. 63, no. 1,
strategy is used to relay the traffic between the vehicles to pp. 237–251, Jan. 2014.
ensure that every vehicle would have at least a ratio ρ of average [12] T. Yamada, R. Shinkuma, and T. Takahashi, “Performance evaluation of
per-vehicle throughput. This ratio ρ represents the extent of the inter-vehicle packet relay for fast mobile road-vehicle communication,”
in Proc. IEEE CCNC, 2008, pp. 212–216.
desired balance of traffic among the vehicles. A larger ρ would [13] W. L. Tan, W. C. Lau, O. Yue, and T. H. Hui, “Analytical models and
result in better balanced traffic but with a longer intervehicle performance evaluation of drive-thru Internet systems,” IEEE J. Sel. Areas
relay distance, whereas a smaller ρ would result in less balanced Commun., vol. 29, no. 1, pp. 207–222, Jan. 2011.
[14] J. Eriksson, H. Balakrishnan, and S. Madden, “Cabernet: Vehicular con-
traffic but with a shorter intervehicle distance. tent delivery using WiFi,” in Proc. ACM MobiCom, New York, NY, USA,
Fig. 6 shows the average size Navg of cooperative groups and 2008, pp. 199–210.
the worst case carry-and-forward relay distance with different [15] J. Ott and D. Kutscher, “A disconnection-tolerant transport for drive-
thru Internet environments,” in Proc. IEEE INFOCOM, 2005, vol. 3,
ρ values. As ρ grows, there would be more vehicles to receive pp. 1849–1862.
an insufficient amount of data and more traffic to be relayed. [16] S. Yang, C. K. Yeo, and B. S. Lee, “Maxcd: Efficient multi-flow schedul-
ing and cooperative downloading for improved highway drive-thru Inter-
Thus, a vehicle would take longer hops to find another vehicle net systems,” Comput. Netw., vol. 57, no. 8, pp. 1805–1820, 2013.
to relay data, which results in larger size of cooperative groups. [17] S. Yang, C.-K. Yeo, and B.-S. Lee, “Maxcd: Max-rate based coopera-
As shown in the figure, when ρ is selected to be higher than tive downloading for drive-thru networks,” in Proc. IEEE ICCCN, 2012,
pp. 1–7.
95%, the average relay distance dramatically increases, and the [18] Q. Yang, S. He, J. Li, J. Chen, and Y. Sun, “Energy-efficient probabilistic
traffic may need to be relayed outside the service segment. area coverage in wireless sensor networks,” IEEE Trans. Veh. Technol.,
Selecting ρ around 90% would be a good choice, because it vol. 64, no. 1, pp. 367–377, Jan. 2015.
[19] S. He, J. Chen, X. Li, X. S. Shen, and Y. Sun, “Mobility and intruder prior
helps preserve reasonable fairness of the vehicles, whereas the information improving the barrier coverage of sparse sensor networks,”
size of the cooperative group does not significantly increase. IEEE Trans. Mobile Comput., vol. 13, no. 6, pp. 1268–1282, Jun. 2014.
7718 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 65, NO. 9, SEPTEMBER 2016

[20] S. He et al., “Maintaining quality of sensing with actors in wireless Xiaohua Jia (F’13) received the B.Sc. and M.Eng.
sensor networks,” IEEE Trans. Parallel Distrib. Syst., vol. 23, no. 9, degrees from the University of Science and Tech-
pp. 1657–1667, Sep. 2012. nology of China, Hefei, China, in 1984 and 1987,
[21] S. He, J. Chen, D. K. Yau, H. Shao, and Y. Sun, “Energy-efficient capture respectively, and the D.Sc. degree in information
of stochastic events under periodic network coverage and coordinated science from The University of Tokyo, Tokyo, Japan,
sleep,” IEEE Trans. Parallel Distrib. Syst., vol. 23, no. 6, pp. 1090–1102, in 1991.
Jun. 2012. He is currently a Chair Professor with the De-
[22] J. Chen et al., “Dynamic channel assignment for wireless sensor networks: partment of Computer Science, City University of
A regret matching based approach,” IEEE Trans. Parallel Distrib. Syst., Hong Kong, Kowloon, Hong Kong. His research in-
vol. 26, no. 1, pp. 95–106, Jan. 2015. terests include cloud computing and distributed sys-
[23] J. Chen et al., “On energy-efficient trap coverage in wireless sensor tems, computer networks, wireless sensor networks,
networks,” ACM Trans. Sens. Netw., vol. 10, no. 1, article 2, pp. 1–12, and mobile wireless networks.
Nov. 2013. Dr. Jia was an Editor of the IEEE T RANSACTIONS ON PARALLEL AND
[24] S. He et al., “Energy provisioning in wireless rechargeable sensor net- D ISTRIBUTED S YSTEMS during, Wireless Networks, the Journal of World Wide
works,” IEEE Trans. Mobile Comput., vol. 12, no. 10, pp. 1931–1942, Web, and the Journal of Combinatorial Optimization during 2006–2009. He
Oct. 2013. was the General Chair of the 2008 ACM International Symposium on Mobile
[25] B. Bellalta, E. Belyaev, M. Jonsson, and A. Vinel, “Performance eval- Ad Hoc Networking and Computing, the Technical Program Committee (TPC)
uation of IEEE 802.11 p-enabled vehicular video surveillance system,” Cochair of the 2009 IEEE International Conference on Mobile Ad hoc and
IEEE Commun. Lett., vol. 18, no. 4, pp. 708–711, Apr. 2014. Sensor Systems, the Area Chair of the 2010 IEEE Conference on Computer
[26] N. Lyamin, A. V. Vinel, M. Jonsson, and J. Loo, “Real-time detection Communications (INFOCOM), the TPC Cochair of the 2010 IEEE Global
of denial-of-service attacks in IEEE 802.11p vehicular networks,” IEEE Telecommunications Conference Ad Hoc and Sensor Networking Symposium,
Commun. Lett., vol. 18, no. 1, pp. 110–113, Jan. 2014. and the Panel Cochair of IEEE INFOCOM 2011.
[27] K. Ota, M. Dong, S. Chang, and H. Zhu, “MMCD: Cooperative down-
loading for highway VANETs,” IEEE Trans. Emerging Topics Comput.,
vol. 3, no. 1, pp. 34–43, Mar. 2015.
[28] J. N. Isento et al., “Vehicular delay-tolerant networks: A novel solution
for vehicular communications,” IEEE Intell. Transp. Syst. Mag., vol. 5,
no. 4, pp. 10–19, Winter 2013.
Kan Yang received the B.Eng. degree from the
[29] K. Ota, M. Dong, H. Zhu, S. Chang, and X. Shen, “Traffic information
University of Science and Technology of China,
prediction in urban vehicular networks: A correlation based approach,” in
Hefei, China, in 2008 and the Ph.D. degree from City
Proc. IEEE WCNC, 2011, pp. 1021–1025.
University of Hong Kong, Kowloon, Hong Kong,
[30] B. Bellalta et al., “Throughput analysis in CSMA/CA networks us-
ing continuous time Markov networks: A tutorial,” inWireless Net- in 2013.
working for Moving Objects. Berlin, Germany: Springer-Verlag, 2014, He is currently a Postdoctoral Fellow with the
pp. 115–133. Broadband Communications Research group, De-
[31] J. Zhao, T. Arnold, Y. Zhang, and G. Cao, “Extending drive-thru data partment of Electrical and Computer Engineering,
access by vehicle-to-vehicle relay,” in Proc. ACM VANET, New York, University of Waterloo, Waterloo, ON, Canada. In
NY, USA, 2008, pp. 66–75. 2012, he was a Visiting Scholar with the State Uni-
[32] J. Yoo, B. Choi, and M. Gerla, “An opportunistic relay protocol for versity of New York at Buffalo, Buffalo, NY, USA.
vehicular road-side access with fading channels,” in Proc. IEEE ICNP, His research interests include cloud security and privacy, big data security,
2010, pp. 233–242. cloud data mining, cryptography, social networks, vehicular ad hoc networks,
[33] J. Zhao, Y. Zhang, and G. Cao, “Improving drive-thru access to roadside smart grids, wireless communication and networks, and distributed systems.
infrastructure by lightweight relay,” in Proc. IEEE Mobile Netw. Veh.
Environ., 2007, pp. 144–149.

Bo Zhang received the B.Sc. degree in electronic in-

formation science and technology from Sun Yat-sen
Ruitao Xie received the B.Eng. degree from the
University, Guangzhou, China, in 2009 and the
Beijing University of Posts and Telecommunica-
Ph.D. degree from City University of Hong Kong,
tions, Beijing, China, in 2008 and the Ph.D. de-
Kowloon, Hong Kong, in 2013.
gree in computer science from City University of
He is currently a Software Engineer Google.
Hong Kong, Kowloon, Hong Kong, in 2014.
He was a Postdoctoral Fellow with the Department
She is currently a Senior Research Associate with
of Computer Science, City University of Hong Kong.
the Department of Computer Science, City Univer-
His research interests include vehicular ad hoc
sity of Hong Kong. Her research interests include
networks, wireless ad hoc networks, collaborative
cloud computing, distributed systems, and wireless
relay networks, and multiple-input–multiple-output
sensor networks.
systems.