Content uploaded by Song Guo
Author content
All content in this area was uploaded by Song Guo on Mar 16, 2015
Content may be subject to copyright.
1
Opportunistic Offloading of
Deadline-Constrained Bulk Cellular Traffic in
Vehicular DTNs
Hong Yao, Member, IEEE, Deze Zeng, Member, IEEE, Huawei Huang, Student Member, IEEE,
Song Guo, Senior Member, IEEE, Ahmed Barnawi, Member, IEEE, and Ivan Stojmenovic, Fellow, IEEE
Abstract—The ever-growing cellular traffic demand has laid a heavy burden on cellular networks. The recent rapid development
in vehicle-to-vehicle communication techniques makes vehicular delay-tolerant network (VDTN) an attractive candidate for traffic
offloading from cellular networks. In this paper, we study a bulk traffic offloading problem with the goal of minimizing the cellular
communication cost under the constraint that all the subscribers receive their desired whole content before it expires. It needs to
determine the initial offloading points and the dissemination scheme for offloaded traffic in a VDTN. By novelly describing the content
delivery process via a contact-based flow model, we formulate the problem in a linear programming (LP) form, based on which an
online offloading scheme is proposed to deal with the network dynamics (e.g., vehicle arrival/departure). Furthermore, an offline LP-
based analysis is derived to obtain the optimal solution. The high efficiency of our online algorithm is extensively validated by simulation
results.
Index Terms—Cellular Traffic Offloading, Opportunistic Networks, Minimum Offloading Problem, Delay Tolerant Networks
✦
1 INTRODUCTION
Cellular communication is facing a critical challenge
of explosively increasing traffic demands due to the
massive growth of mobile devices, e.g., mobile phones,
tablets, cars, etc. According to a recent report from Cisco,
the overall mobile data traffic is expected to grow to 11.2
exabytes per month by 2017, a 13-fold increase over 2012
[1]. The immense mobile traffic demand has imposed a
heavy burden on current cellular networks with limited
spectrum. Considerable effort has been made to tackle
this issue from various aspects, most of which rely
on sophisticated new physical-layer techniques, e.g. [2],
with a high capital expenditure. Recently, an efficient al-
ternative, generally known as cellular traffic offloading,
which migrates the traffic from the cellular networks
to other networks, has attracted much attention in the
literature [3]–[8].
Meanwhile, the development in intelligent transporta-
tion makes vehicles biggest cellular resource consumers.
The intelligence on vehicles highly relies on the available
information, such as realtime traffic condition, augment-
ed map, etc. Besides, contents like podcasts and commer-
cials are also required to disseminate to certain vehicles
•H. Yao and D. Zeng are with China University of Geosciences, Wuhan,
China. E-mail: yaohong@cug.edu.cn, dazzae@gmail.com
•H. Huang and S. Guo are with the School of Computer Science and En-
gineering, The University of Aizu, Japan. E-mail: {d8152101, sguo}@u-
aizu.ac.jp
•A. Barnawi is with King Abdulaziz University, Jeddah, Saudi Arabia.
E-mail: ambarnawi@kau.edu.sa
•I. Stojmenovic is with SIT, Deakin University, Melbourne, Australia; King
Abdulaziz University, Jeddah, Saudi Arabia; and SEECS, University of
Ottawa, Canada. E-mail: Stojmenovic@gmail.com
in some emerging vehicular applications. In a representa-
tive networking scheme, delay-tolerant contents are dis-
seminated through a Vehicular Delay-Tolerant Network
(VDTN) [9]–[12], where vehicles “carry-and-forward”
data packets cooperatively upon transmission opportu-
nities via Vehicle-to-Vehicle (V2V) communication (e.g.,
Dedicated Short Range Communications (DSRC) [13],
802.11p [14]) when two vehicles come into the reciprocal
communication range of each other [15]–[17]. VDTNs
are infrastructureless since they do not rely on any in-
frastructure but simply explore the V2V communication
capabilities inherent in modern vehicles. VDTNs are also
regarded as opportunistic networks because the contact-
s, or the transmission opportunities, between vehicles
happen in a dynamical and unexpected manner. The
recent development in VDTN has made it an appealing
candidate for cellular traffic offloading as the offloaded
content can be disseminated over a VDTN in a cost-
efficient way.
The offloading model for disseminating a content to
certain interested subscribers (vehicles) can be described
by a representative scenario, but not limited to it, as
shown in Fig. 1. The content to be distributed is ini-
tially offloaded to only a few subscribers capable of
cellular communication (e.g., d0,d1and d4), which then
cooperatively disseminate it to other vehicles (e.g., d2,d2)
in a “carry-and-forward” manner. Because the direct
contacts to those vehicles by subscribers are rare, some
vehicles serve as helpers (e.g., h0,h2,h3and h4) to relay
the content such that the cellular network cost can be
reduced and the delivery efficiency will be improved
significantly.
Although offloading from cellular networks to oppor-
2
h2
h0
h4
d0d1
h1
Helper Subscriber
d2
h3
h5
d4
d3
Fig. 1. Offloading cellular traffic to a cooperative oppor-
tunistic VANET
tunistic networks has been investigated, to the best of
our knowledge, all existing studies only focus on the
scenario that the content can be transmitted by one
contact (e.g., [3]–[6], [8], [12], [18]). In practice, bulk
data, e.g., a podcast, can hardly be completely forwarded
from one vehicle to another during their limited contact
duration.
In this paper, we are motivated to investigate the min-
imum offloading problem for a large deadline-constrained
content. That is, the number of data transmitted from
cellular networks to all content subscribers in a VDTN
is minimized while the content must be fully received
before it expires at each subscriber. Our main contribu-
tions are summarized as follows:
•We study the minimum offloading problem for
bulk data dissemination in VDTNs. To represent the
deadline-constrained bulk data dissemination pro-
cess in a VDTN, we invent the concept of Deadline
Trimming Contact Graph (DTCG), in which each
transmission opportunity (i.e., contact) is modeled
by a vertex. The content dissemination process is
then described using a flow-based model on the
constructed DTCG.
•We propose an online offloading algorithm based on
a linear programming (LP) formulation. A specula-
tive DTCG is first constructed with the predictions
of future contact events before the deadline. To deal
with the dynamics (e.g., vehicle departure, arrival,
overtaking, etc.), the LP formulation will be rebuilt
accordingly when an event makes the correspond-
ing DTCG invalid.
•We investigate the optimal offloading using an of-
fline flow-based analysis according to the contact
traces, by which a deterministic DTCG is constructed.
Specially, we carefully design a scheme to deal with
the cases that two vehicles encounter multiple times.
Experimental results validate the high efficiency of
our online algorithm by the fact that it performs
much close to the offline optimal results.
This paper proceeds as follows. Section 2 reviews the
related work. Section 3 presents the traffic offloading
model and problem statement. Section 4 proposes a
flow-based online algorithm for the minimum offloading
problem. Section 5 gives the optimum offloading scheme
TABLE 1
Notations
Notation Meaning
Nmobile nodes (i.e., vehicles) in the network
D D ⊆N, content subscribers
dii= 1,2,··· ,|D|,di∈D, a subscriber
H H ⊂N, helper nodes
Kcontent size
Tcontent deadline
G G = (V,E), deadline trimming contact graph
Vvertexes in G
Eedges in G
eαβ eαβ ∈E, an edge from αto β
cαβ weight of edge eαβ
hu, v, tia contact event between uand vat time t,u, v ∈N
fd
αβ a flow destined to dfrom vertex αto β,α, β ∈V
φ(α, β)the number of coded packets received
by the common node of αand β
ϕ(di)the number of coded packets received by di
dA(β)out-degree of vertex βto repetitive-contact
vertex group A
based on an offline analysis. Intensive simulation results
confirming the proposed algorithm are presented in Sec-
tion 6. Finally, concluding remarks are given in Section
7.
For the conveniences of the readers, the major nota-
tions used in this paper are listed in Table 1.
2 RELATED WORK
With the growing popularity of mobile accesses to the
cellular networks, it is an urgent need to find sub-
stitution networks to migrate such traffic burden. Op-
portunistic networks have attracted increasing interests.
Han et al. [3], [6] investigate a target-set selection prob-
lem to minimize the data traffic over cellular networks
for information delivery in mobile social networks. Li
et al. [4] propose a greedy algorithm to maximize the
total traffic that could be offloaded under the constraints
of buffer size and delivery deadline. Then, Barbera et
al. [19] analyze encounter-based social graphs accord-
ing to various metrics to elect socially important “VIP
delegates”, through whom 3G traffic could be offloaded
to the other users. Whitbeck et al. [18], [20] propose a
scheme called Push-and-Track to determine how many
copies of a message should be injected, when, and to
whom and examine the achievable offload ratio depend-
ing on the freshness constraints. More recently, Chuang
et al. [8] apply social-analysis and propose a community-
based method to select the initial sources that shall be
distributed with message via cellular communication.
Besides the offloading issue in conjunction with social
analysis, some work also starts to consider offloading
the cellular traffic to VDTNs. Yan et al. [12] propose a
data dissemination framework called DOVE aiming at
minimizing the data delivery delay to a fixed number
of receivers in VDTNs. Kashihara et al. [7] preliminar-
ily validate the efficiency of data offloading to public
transport vehicles for alleviating heavy traffic load over
3
mobile networks but they do not specify how to dissem-
inate the data packets to VDTNs.
Recently, other studies argue that the cellular traffic
can be offloaded to WiFi network. For example, Di-
matteo et al. [21] propose and evaluate their integrated
architecture, in which the data traffic is migrated from
cellular networks to metropolitan WiFi access points in
the opportunistic networking paradigm. Lu et al. [22]
mention that facing the challenges of mobile data ex-
plosion, current cellular networks do not have sufficient
accommodating capacity corresponding to the exponen-
tial growth of mobile data requirements. They consider
that data can be transmitted between end users via
opportunistic peer-to-peer WiFi communications. They
further propose a Subscribe-and-Send architecture and
an opportunistic forwarding protocol aiming to offload
the cellular data.
By a comprehensive literature survey, we conclude
that existing work shares a common assumption that
a content can be always successfully transferred in one
contact, e.g., [3], [4], [8]. In this paper, we investigate a
more realistic case that a big content (e.g., podcasts) with
a size exceeding the maximum transmission volume of
a single contact is to be offloaded and disseminated over
a VDTN.
In addition to GPS-equipped personal cars, public
transportation systems [7], [23]–[26], e.g., buses and taxis
have been also investigated for offloading. The benefit of
such network is that their trajectories are relatively more
stable and predictable. In this paper, we also consider
buses, taxis or other public transportation vehicles as
potential helpers.
3 SYSTEM MODEL AND PROBLEM STATEMEN-
T
3.1 Network Model
We consider a VDTN model with a number of vehicles
Nas shown in Fig. 1. Our model applies to many
scenarios where a group of users have subscribed ser-
vices of common interests [3], [22], [27], e.g., receiv-
ing enhanced urban map, real-time traffic information,
and high-quality commercials from content server when
entering its covered region. The users interested in a
content are defined as subscribers D(D⊆N), and those
that are willing to forward traffic as relays are defined
as helpers H(H⊆N). Without loss of generality, we
consider all subscribers capable of both cellular commu-
nication and V2V communication, while helpers capable
of V2V communication only. By cellular communication,
a subscriber can always directly retrieve data from the
cellular network. The V2V communications are conduct-
ed in a “carry-and-forward” manner, in which a trans-
mission opportunity between two mobile nodes happens
only when they move into the reciprocal communication
range of each other. Furthermore, all vehicles are GPS-
enabled such that the V2V transmission opportunities
can be predicted based on their current geographical
locations and speeds.
Let the size of a packet be the amount of data trans-
mitted between any two vehicles on each encountering
duration. A content with K(K > 1) packets needs to
be disseminated to all its subscribers within deadline T
since its generation. Although any subscriber can get the
whole content via cellular communication, this is not
desired since it is at the expense of high communica-
tion cost and may incur heavy burden on the cellular
network. Therefore, the cellular communication will be
conducted only when necessary, while the offloaded
traffic will be disseminated to subscribers that are not
done yet with their full reception of the content by
cooperative V2V communications.
3.2 Network Coding based Cooperative Transmis-
sion
For multiple-packet dissemination in opportunistic net-
works, it has been well known that it is difficult to
make a right forwarding decision due to the unexpected
network dynamics [16], [28]. A wrong decision not only
wastes the precious transmission opportunity but also
may incur the coupon collector problem [29] that it takes
a quite long time for a subscriber to download the whole
content. Fortunately, the random linear network coding
(RLNC) is a compelling technique to address this issue
by making all data packets with equal importance [28],
[30]–[32]. Specially, a number of linearly independent
coded packets are generated on the server for the content
with Knative packets pn
1, pn
2,··· , pn
Kin the form pc=
PK
i=1 αipn
i, where αiis the coding coefficient randomly
chosen from Galois Field. Some subscribers are first
offloaded with certain coded packets, which are then
forwarded to the other subscribers by V2V cooperative
transmissions. When two vehicles meet, each of them
will generate and forward a by encoding all the coded
ones it possesses. Whenever a subscriber collects Klin-
early independent coded packets, it recovers its desired
content using Gaussian elimination.
3.3 Problem Statement
By offloading cellular traffic to a VDTN, our goal is
to minimize the cellular communication cost in terms
of the number of coded packets retrieved from the
Internet directly under the condition that all subscribers
are able to get their desired content before it expires.
With GPS-enabled mobile nodes, previous studies have
shown that it is possible to predict the future contact
events in VDTNs by different means [33]. Intuitively,
if an offloaded coded packet can be shared by more
subscribers via V2V communications, less offloading
traffic is required. It is straightforward to determine the
shared subscribers in a two-hop forwarding paradig-
m, i.e., cellular network→offloading point→subscriber.
However, when we consider the general multi-hop for-
warding paradigm, where helpers are introduced, this
4
issue becomes challenging due to a large number of com-
binatorial contact events that are possible to disseminate
a packet to subscribers. To the best of our knowledge,
no solution for such problem have been proposed in the
literature. In this case, our major problem is to determine
the set of subscribers as offloading points and how many
packets to be offloaded for each.
4 ONLINE FLOW-BASED PACKET OFFLOAD-
ING SCHEME
In this section, we propose a two-phase flow-based
packet offloading algorithm based on the dynamic con-
struction of DTCG, which is always updated by a cen-
tralized server. The system works in a way similar to
SDN (software defined networking), where the server
in the control plane locates at the 3G base station and
communications among vehicles in the data plane are
controlled by the server via a dedicated 3G channel.
Furthermore, in order to make the proposed algorithm
applicable, we assume that there is a separated control
network between vehicles and the centralized server.
When a dynamic event occurs on a vehicle, e.g., overtak-
ing others, entering or leaving the communication region
of a base station, it will report this information to the
centralized server. Accordingly, the server will deliver
partial offloading solution to the involved vehicles via
3G communication channels. All such interactions are
made by short messages. However, we notice that the
dynamic events happen frequently in the real world
and would incur significant computation overhead if
only the primitive LP based offloading algorithm is
applied. Therefore, we further propose a Lazy-control
mechanism, in which a Cumulative Hurting Indicator
(CHI) is defined to indicate the damage degree of the
previous offloading solutions caused by dynamic events.
Once the CHI exceeds a designated threshold, the DTCG
will be updated. Then, a new suite of solution will be
obtained by resolving our offloading algorithm. Finally,
only the relevant partial solution is delivered to the
involved vehicles by base station.
4.1 Phase-I: Initial Offloading
4.1.1 Predicted DTCG Construction
The number of subscribers that a coded packet can
benefit is determined by the available V2V transmis-
sion opportunities. With the current geo-information,
the predicted future contact events are modeled by a
predicted DTCG, in which a contact denoted by a tuple
hu, v, tiwith an encountering pair (u, v), u, v ∈Nand
their meeting time, is represented by a vertex. Let the
content generation time be 0. Only the contacts that
happen within Tare regarded as effective and will be
included in the predicted DTCG G= (V,E), where node
set Vincludes all such contact events and edge set Eis
constructed as follows.
If two contact events αand β(α, β ∈V) share a
common node and αhappens before β, a directed edge
Algorithm 1 Construction of a Predicted DTCG
Input: A sequence of predicted contact events
{hu, v, ti}, u, v ∈N, t ≤T
Output: A corresponding predicted DTCG G= (V,E)
1: G ← (V,E),V← ∅,E← ∅
2: Create and add vertexes for all contact events into V
3: Sort Vin an ascending order of the contact time
4: for all hui, vi, tii ∈ Vdo
5: for all huj, vj, tji ∈ V, tj> tido
6: if (ui== ujand vi6=vj)
or (ui6=ujand vi== vj)then
7: Add an edge connected from hui, vi, tiito
huj, vj, tjiwith weight 1 into E
8: end if
9: end for
10: end for
11: Add the source vertex hs, s, −1iinto V
12: for all di∈Ddo
13: Add the source-subscriber vertex hsi, di,0iinto V
14: Add an edge connected from hs, s, −1ito hsi, di,0i
with weight ∞into E
15: Add the subscriber vertex hdi, di, T iinto V
16: Add edges connected from each hu, di, ti ∈ Vto
hdi, di, T iwith weight ∞into E
17: end for
eαβ ∈Eis created with a weight cαβ representing the
maximum number of packets that can be transferred
over the two contact events. As the contact bandwidth is
normalized as one, we have cαβ = 1 for any two regular
contact events αand βif eαβ ∈E. For example, we
suppose α=hu1, v, t1i,β=hu2, v, t2iand t1< t2,. The
edge eαβ with cαβ = 1 represents that a data packet can
be transferred from u1to vat t1and then from vto u2
at t2.
In addition, we introduce three types of auxiliary
vertexes in the predicted DTCG as follows.
To incorporate the cellular communication together
with the V2V communication using a uniform graph
description, the source vertex ¯s=hs, s, −1iand source-
subscriber vertexes hs, d, 0i(d∈D) are created. The former
represents the content origin, i.e., in the cellular network,
with a meeting time attribute set to -1 to make sure that
the data ¯scan flow to any other connected ones in the
graph. The latter represents that any subscriber can di-
rectly retrieve data from the content server at the content
generation time 0 via the cellular network. Furthermore,
we connect the source vertex to each source-subscriber
vertex by an edge with weight ∞, indicating that the full
content can be offloaded directly.
For each subscriber d∈D, an auxiliary destination
vertex ¯
d=hd, d, T iis also created. Furthermore, any
vertex α=hu, d, ti ∈ V) needs to be connected to ¯
dwith
weight ∞.
Summarizing all the issues discussed above, we
present our predicted DTCG construction algorithm in
Algorithm 1. An example is given in Fig. 2.
5
h0,d0
h0,h1
h0,d1
h1,d1
d0,d1
s,s,-1
s,d0,0
h0,d0,5
h0,h1,6
h0,d1,7
s,d1,0
h1,d1,13
d0,d0,T
d0,d1,14
d1,d1,T
Fig. 2. An example of constructing predicted DTCG from
a sequence of predicted contact events
Remark 1: The DTCG construction algorithm is of time
complexity OV2, where Vis the total number of
predicted contact events within T.
4.1.2 Formulation of Flow-based Offloading
By V2V communication, a packet can be delivered
from its offloading point to a subscriber divia
multiple V2V transmission opportunities, e.g., P=
{hs, dj,0i,hdj, h1, t1i,hh1, h2, t2i,··· ,hhn, di, tni},0<
t1< t2<··· < tn≤T, where dj(dj6=di) is the initial
packet offloading point. Each edge in DTCG can be
regarded as one two-hop relay process. For example,
the edge between hh0, h1, t1i, and hh1, h2, t2iindicates
that a packet is first forwarded from h0to h1at time t1
and then relayed to h2via h1at t2. From this point of
view, a path denoted by a sequence of contact events
is equivalent to a flow in the DTCG. This motivates
us to build a flow-based model to describe the content
dissemination process.
Let b
f¯sβ and fdi
αβ denote the flow from source-vertex
¯sto vertex β, and from vertex αto βfor subscriber di,
respectively. The minimum offloading problem is then
formulated in linear programming (LP) format based on
a flow-based model on the constructed predicted DTCG
as follows:
min
b
f¯sβ
:C=X
β∈Vb
f¯sβ (1)
s.t.
fdi
αβ ≤cαβ ,∀α, β ∈V, di∈D,(2)
X
α∈V
fdi
αβ =K, where β=hdi, di, T i,∀di∈D,(3)
X
α∈V
fdi
αβ =X
γ∈V
fdi
βγ ,∀β∈V, di∈D,(4)
b
f¯sβ ≥fdi
¯sβ ,∀β∈V, di∈D.(5)
Recall that the edge weight denotes the maximum
number of packets that can be relayed during the two
contacts and therefore imposes capacity constraints on the
flow, as indicated by (2).
According to the RLNC decoding requirement, any
subscriber di∈Dcan decode a content with Kpackets
if and only if there is a K-flow from the vertex hs, s, −1i
to hdi, di, T iin DTCG. Note that for a subscriber di∈D,
the flow injected to the corresponding subscriber ver-
tex hdi, di, T icomes from two parts, direct offloading
from the content server and the remaining, if any, from
other vehicles via V2V communication, respectively. In
summary, the flow conservation constraints (3) and (4) are
given for the destination-vertexes and regular vertexes,
respectively.
Note that fdi
αβ represents the flow only destined to
subscriber di. The actual flow from αto vertex βshall
be the maximum one over all subscribers, as given in
(5) due to the property of RLNC. The total cost can
be obtained by summing up injected flows from ¯s, i.e.,
C=Pβ∈Vb
f¯sβ , as given in (1).
As soon as a content is generated, the content serv-
er can construct a predicted DTCG according to the
available geo-information. Then, by solving LP problem
(1), we are able to obtain the initial offloading solution,
which specifies all offloading points and the traffic vol-
ume to be offloaded to each of them.
4.2 Phase II: Dynamic Offloading
4.2.1 DTCG Update
The initial offloading solution is obtained through a
predicted DTCG. The resulting packet offloading and
transmission path shall be followed only if all vehicles
move just as we predicted. However, the predictions
may fail due to some unexpected events. For example,
a car serving as a helper may exit from the network
unexpectedly or a car join in accidentally creating some
new V2V transmission opportunities in the middle of the
content distribution process. We categorize these events
into three types: relevant to irrelevant transition (R2IR),
irrelevant to relevant transition (IR2R) and overtaking
event. Their corresponding DTCG update procedures are
described as follows.
R2IR event: When a predicted contact disappears
(i.e., it becomes irrelevant to the new DTCG), its related
vertexes and edges shall be removed from the current
DTCG. This happens when a vehicle turns off the road.
For example, after node h1becomes irrelevant at time
10, vertex hh1, d1,13i, and all its incoming and outgoing
edges, are removed from the original DTCG, as shown
in Fig. 3.
IR2R event: When a new contact opportunity e-
merges, new vertex and related edges shall be added.
For example, Fig. 4 shows the DTCG after a new vehicle
h2gets into the network and h2with predicted meeting
time to d0and d1at time 11 and 12, respectively. There-
fore, vertexes hh2, d0,11iand hh2, d1,12iare added in.
The incoming and outgoing edges to both vertexes are
then generated according to the rule discussed in Section
4.1.
Overtaking event: An overtaking event usually re-
sults in the update of meeting time of a predicted
contact. It happens when the speed change of a vehicle is
6
h0,d0
h0,h1
h0,d1
h1,d1
d0,d1
s,s,-1
s,d0,0
h0,d0,5
h0,h1,6
h0,d1,7
s,d1,0
h1,d1,13
d0,d0,T
d0,d1,14
d1,d1,T
Fig. 3. New predicted DTCG after h1becomes irrelevant
at time 10
h0,d0
h0,h1
h0,d1
h2,d0
d0,d1
s,s,-1
s,d0,0
h0,d0,5
h0,h1,6
h0,d1,7
s,d1,0
h2,d1,12
d0,d0,T
d0,d1,14
d1,d1,T
h2,d0,11
h2,d1
Fig. 4. New predicted DTCG after h2becomes relevant
at time 11
detected. For example, consider a scenario where vehicle
h2slows down after meeting d1. A delayed meeting with
h2is expected at time 15. Accordingly, vertex hh2, d0,11i
is replaced by hh2, d0,15i(the red vertex in Fig. 5). Some
related edges which violate the time sequence, i.e., the
red dashed arrows in the figure, are removed. Finally, a
new predicted DTCG is obtained.
4.2.2 Formulation Update
In addition to the above updates, all the vertexes for
the happened contacts and their related edges shall
be removed from the updated DTCG. After that, new
offloadings may be required to complement the loss
of vanished transmission opportunities or explore the
newly emerging ones. To this end, we build a new model
to describe the new offloadings as follows by taking the
offloaded traffic into consideration.
Recall that the contact prediction would fail in the
middle of content dissemination process. In other words,
a subscriber may have already received the content
partially. As a result, different from (3) where each
subscriber must receive Kpackets, we only consider the
residual coded packets that a subscriber requires for a
full decoding. Constraint (3) shall be rewritten as
X
α∈V
fdi
αβ +ϕ(di) = K, where β=hdi, di,∞i,∀di∈D,(6)
where ϕ(di)denotes the number of linearly independent
coded packets that have been scheduled for subscriber
dibut not affected by dynamic events.
h0,d0
h0,h1
h0,d1
h2,d0
d0,d1
s,s,-1
s,d0,0
h0,d0,5
h0,h1,6
h0,d1,7
s,d1,0
h2,d1,12
d0,d0,T
d0,d1,14
d1,d1,T
h2,d0,15
h2,d1
h2,d0
Fig. 5. New predicted DTCG after an overtaking event
On the other hand, for a regular vertex, the coded
packets that have been received shall be reutilized to
potentially reduce the offloading traffic. Let φ(α, β)de-
note the number of linearly independent coded packets
that have been received by their common node by the
dynamic event time. For example, suppose α=hu, v , tuvi
and β=hx, u, txuiand node uhave received two coded
packet. We have φ(α, β) = 2. The already received
coded packet can be also forwarded to an encountered
node and therefore the flow reservation constraints for
a regular vertex shall be updated to
X
α∈V
(fdi
αβ +φ(α, β)) = X
γ∈V
(fdi
βγ +φ(β, γ)) ∀β∈V,∀di∈D.
(7)
Overall, by summarizing all the issues as discussed
above, we are able to build a new flow model to deal
with dynamic events as
min
f¯sβ
:C=X
β∈V
f¯sβ .
s.t.(2),(5),(6),(7).
(8)
4.3 Dynamic Offloading with Constrained Updates
If new solutions are required based on updated DTCG
once each dynamic event is detected, the computation
overhead of the centralized server would be high, es-
pecially when dynamic events occur frequently. In this
section, we propose an update-constrained approach
for the dynamic offloading problem by introducing a
threshold-control mechanism.
Let ~denote the CHI of a dynamic event, defined as
the ratio of total number of encoded packets in the links
becoming invalid to the total number of encoded packets
in the previous solution, i.e.,
~=Peeφ(ee)
Pe∈EPdifdi
e
,∀di∈D,(9)
where eedenotes any invalid link caused by a dynamic
event. In the threshold-control mechanism, our LP based
algorithm will be launched for an updated solution only
when the CHI exceeds a given threshold value χ0as
summarized in Alg. 2.
In this algorithm, the CHI shall be updated accord-
ing to line 4 when the DTCG is updated according to
7
Algorithm 2 Lazy Control of Packets Offloading Algo-
rithm
Input: Initial DTCG
1: while deadline Tdose not expire do
2: if a dynamic event occurs then
3: update the DTCG relying on current geo-
information of all vehicles
4: update CHI by: ~←~+Peeφ(ee)
Pe∈EPdifdi
e
,∀di∈D
5: if ~≥χ0then
6: solve (8)
7: if new offloading links arise then
8: server allocates new coded packets to the
relevant vehicles
9: end if
10: ~←0
11: end if
12: end if
13: end while
a dynamic event. If the CHI achieves the designated
threshold χ0, formulation (8) will be solved and a new
suite of solutions will be obtained. Note that, only when
the updated solutions identify new offloading links o-
riginating from the server, the new coded packets are
allocated to the involved vehicles.
5 OFFLINE OPTIMAL OFFLOADING ANALYSIS
In this section, we present an analysis on the mini-
mum offloading problem based on the global contact
events trace that is collected during the content dis-
semination process. It provides an evaluation measure
to our proposed online algorithm. Different from the
speculative contact events used in online algorithm, it
is possible that two mobile nodes encounter multiple
times in the deterministic trace. We define the contacts
with the same pair of nodes but with different contact
time as repetitive contact (RC). The RC vertexes in the form
{hu, v, tii}, i = 1,2,··· ,0< t1< t2<· · · ≤ Tcompose
an RC vertex group, where any two neighbor elements,
i.e., hu, v, tiiand hu, v, ti+1i, are called consecutive RC
vertexes. Given a contact trace, a deterministic DTCG
will be constructed by applying the new rules developed
below for connection of intra- and inter-RC vertex group.
Rule 1: Connect any two consecutive RC vertexes α=
hu, v, tiiand β=hu, v, ti+1iby an edge α→βwith weight
∞. The weight ∞on the edge is to ensure that any
received packets are always maintained in the buffer.
For example, a direct edge from hu, v, t1ito hu, v, t2iis
created in Fig. 6(a). Note that other connections between
non-consecutive repetitive vertexes, e.g., the edge from
hu, v, t1ito hu, v, t3iin Fig. 6(a) are redundant and there-
fore could be omitted.
Now, we consider the edges between two RC vertex
groups. Without loss of generality, we consider two
groups, A={hu, v, ∗i} and B={hx, u, ∗i}, sharing
a common node u. To construct the inter-group edges
between Aand B, we apply the following rule.
Rule 2: Sort A∪Bin an ascending order of the meeting time
and connect any two adjacent vertexes αand βvia an edge
α→βwith unit weight if they are not in the same group.
Consider an example shown in Fig. 6(b) where t1<
t2<···< t7. While the link from hx, u, t1ito hu, v, t3iis
allowed, it could be removed based on Rule 2 because
it is already implied by edges hx, u, t1i → hx, u, t2iand
hx, u, t2i → hu, v, t3i. Similarly, hu, v, t5iis connected to
hx, u, t6ias they are adjacent and in different groups. In
this way, many redundant edges could be omitted.
(a) Multiple consecutive RC vertexes
x,u,t1x,u,t2x,u,t6
u,v,t3u,v,t4u,v,t5u,v,t7
A
B
(b) Connections between two RC vertex groups
Fig. 6. Connection between RC vertex groups
By incorporating the above two rules into Algorithm
1 to deal with the connections involving RC vertex
groups, a deterministic DTCG can be constructed from a
trace collected after content dissemination. For example,
a deterministic trace and its corresponding DTCG are
shown in Fig. 7. Based on such deterministic DTCG, we
can then apply exactly the same flow-based method as
discussed in Section 4.1 and build a LP formulation to
derive the optimal offloading solution and the minimum
cellular communication cost.
6 PERFORMANCE EVALUATION
In this section, we present the performance evaluation
results of our online algorithm for the minimum offload-
ing problem.
6.1 Simulation Settings
We have implemented a discrete-event simulator in C++
simulating the dissemination of a deadline-constrained
content to a group of subscribers according to the system
model described in Section 3. Particularly, the LP based
algorithm is implemented in Python and solved by the
commercial optimizer Gurobi [34] with a version 5.60.
The architecture of our simulator in Fig. 8 illustrates the
principle of handling dynamic events and how our LP
based algorithms are embedded in the simulation. In the
Data Plane, the Movement and Dynamic Events Emula-
tor module deals with NC-based packet dissemination in
a VDTN, and the Vehicles Status Monitor module detects
various events and reports them to the DTCG Update
module in the Control Plane. After updating the DTCG,
8
h0,d0
h0,h1
h1,d1
h0,h1
h0,h1
s,s,-1
s,d0,0
h0,d0,5
h0,h1,6
h1,d1,7
s,d1,0
h0,h1,10
d0,d0,T
h0,h1,14
d1,d1,T
h0,d0,13
h0,d0
d0,d1
h1,d1
h1,d1,8
d0,d1,14
Fig. 7. A deterministic DTCG example with intra-vertexes
Fig. 8. The architecture of simulator
the Lazy-control module will trigger the LP solver for
the new solution to be returned to the involved vehicles
if the CHI exceeds the designated threshold.
In simulations, a number of helpers and subscribers
are randomly distributed on a highway road when the
content is generated. The road is 5000 meters long,
within which there are 4 intersections in total. To show
the efficiency of our online algorithm, we also implement
a GPS-Density algorithm by adapting a Connectivity-
based single-packet offloading algorithm presented in
[18] to be able to deal with bulk traffic offloading in
VDTNs. In GPS-Density based algorithm, the subscribers
are sorted according to the number of subscribers to
meet within content deadline Tand the ones with the
most number of neighbors are selected as the offload-
ing targets. In our experiments, we consider two cases
where only the top 20% (GPS-Density 20%) and 80%
(GPS-Density 80%) subscribers are allocated with coded
packets, respectively. Additionally, other algorithms Base
and Random in [3] are considered. The Base algorithm
simply shows the total cellular cost without offloading.
Because its performance is always the worst compared
with others, we only show it in Fig. 12(d). Similar to
the GPS-Density algorithm, the Random algorithm only
selects a top percentage of subscribers at random to be
allocated with coded packets at the beginning. After
deadline expires, those unsatisfied subscribers will be
directly allocated packets through cellular network.
Both our algorithm (LP based) and these existing
algorithms are extensively investigated under various
simulation settings with different values of content size
K, dynamic event probability pT, content deadline T, the
number of helpers |H|and the number of subscribers |D|.
For each setting, we conduct 100 simulation instances
to obtain the average cost in terms of the number of
coded packets sent via cellular communications. In each
instance, we log all the contact events after the simu-
lation and apply the trace to Algorithm 1 to construct
a deterministic DTCG with all actual contacting events,
which is then imported into Gurobi Solver to get the
optimal value by solving (1) or (8).
6.2 Evaluation Results under Different Threshold
Values
We first study the effect of threshold values by varying
χ0in {0, 0.01, 0.05, 0.1, 0.2, 0.3}with a fixed deadline
T=60. The number of subscribers (|D|) and helpers (|H|)
are both set to 30.
Fig. 9 shows the cumulative distribution function (CD-
F) of computation overhead (in terms of the total number
of LP problems to be solved) and the cellular allocation
cost when pTis set to 0.5 and 0.8. From Fig. 9(a) and Fig.
9(c), we observe that the threshold value significantly
affects the computation overhead under dynamic sce-
narios. For example, under the setting pT= 0.5, in the
top 75% instances with low computation overhead, the
total number of generated LP problems reaches as high
as 30 when χ0= 0, but this number drops to 5 when
χ0equals to 0.01. Further, the computation overhead
becomes much lower when χ0is bigger than 0.05. The
similar observations are made in Fig. 9(c). On the other
hand, the CDF of cellular cost shows an increasing
function of χ0in Fig. 9(b) and Fig. 9(d). This is because
a larger threshold leads to less packets delivered via
VDTN before deadline and thus more encoded packets
are needed to be directly downloaded to subscribers
from cellular network.
With the same simulation settings, Fig. 10(a) shows
the average cellular cost v.s. χ0under different settings
of pT∈ {0.2,0.5,0.8}. It can be observed that the cellular
cost grows firstly when χ0increases and then converges
after χ0= 0.1 for each pT. This is because once the
threshold exceeds 0.1, the frequency of running the LP-
based algorithm becomes very low, and the content
server supplements almost the same number of encoded
packets for the unsatisfied subscribers after the deadline.
Due to this reason, we assign χ0=0.1 by default in all
the remaining simulations unless a different value is
declared explicitly. On the other hand, Fig. 10(b) illus-
trates the average cellular cost v.s. pTunder different
settings of χ0∈ {0,0.01,0.1}. We see the gap of the
average cellular costs corresponding to their thresholds
grows under each pT. This is because a larger transition
probability incurs a more dynamic VDTN scenario and
makes the cellular cost grow.
9
0 10 20 30 40 50
0
0.2
0.4
0.6
0.8
1
CDF
Computaion overhead
χ0=0
χ0=0.01
χ0=0.05
χ0=0.10
χ0=0.20
χ0=0.30
(a) CDF of the computation overhead, pT=0.5
0 50 100 150 200
0
0.2
0.4
0.6
0.8
1
CDF
Cost
χ0 = 0
χ0 = 0.01
χ0 = 0.05
χ0 = 0.20
(b) CDF of the cellular cost, pT=0.5
0 10 20 30 40 50 60
0
0.2
0.4
0.6
0.8
1
CDF
Computation overhead
χ0=0
χ0=0.01
χ0=0.05
χ0=0.10
χ0=0.20
χ0=0.30
(c) CDF of the computation overhead, pT=0.8
0 50 100 150 200
0
0.2
0.4
0.6
0.8
1
CDF
Cost
χ0 = 0
χ0 = 0.01
χ0 = 0.05
χ0 = 0.20
(d) CDF of the cellular cost, pT=0.8
Fig. 9. Performance evaluation under different threshold values
0 0.05 0.1 0.15 0.2 0.25 0.3
50
100
150
200
χ0
Average cost
pT = 0.2 pT = 0.5 pT = 0.8
(a) Average cellular cost v.s. χ0with different pT
0.2 0.5 0.8
50
100
150
200
Probability of transition
Average cost
χ0 = 0
χ0 = 0.01
χ0 = 0.10
(b) Average cellular cost v.s. pTwith different χ0
Fig. 10. Average cellular cost v.s. χ0and PT
6.3 Evaluation Results under Different Content
Sizes
We first investigate the relationship between the cellular
communication cost and the content size by varying the
Kfrom 1 to 10 under the settings T=60, |D|=30 and
|H|=30. Fig. 11(a) shows the performance comparisons
between other work and our algorithm when dynamic
event probability pT= 0.05. Obviously, we shall first
notice that our LP-based algorithm outperforms both
GPS-Density and Random variants under any content
size. In particular, the advantage even increases with
the content size. For example, when the content size
becomes 10, GPS-Density and Random algorithms cost
more than two fold of ours. Our LP-based algorithm can
always approach the optimal one. In addition, we can
also see that the average cellular cost is an increasing
function of content size. This is further validated by Fig.
11(b) where the evaluation results of LP-based algorithm
under different dynamic event probabilities and content
sizes are shown. Such relationship is due to the intuitive
reason that larger content requires more packets offload-
ed via cellular communication.
10
1 3 5 7 9 10
0
50
100
150
200
Content size
Average cost
GPS−Density 20%
GPS−Density 80%
Random 20%
Random 80%
LP based pT=0.05
Optimal pT=0.05
(a) Performance comparison when pT=0.05
5 6 7 8 9 10
0
20
40
60
80
100
Content size
Average cost
LP based pT=0
Optimal pT=0
LP based pT=0.15
Optimal pT=0.15
LP based pT=0.3
Optimal pT=0.3
(b) Average cellular cost v.s. content size
Fig. 11. Performance evaluation under different content sizes
0 50 100 150 200
0
0.2
0.4
0.6
0.8
1
CDF
Cost
GPS−Density 20%
GPS−Density 80%
Random 20%
Random 80%
LP based
Optimal
(a) CDF of the cellular cost when pT= 0
0 50 100 150 200
0
0.2
0.4
0.6
0.8
1
CDF
Cost
GPS−Density 20%
GPS−Density 80%
Random 20%
Random 80%
LP based
Optimal
(b) CDF of the cellular cost when pT= 0.1
0 50 100 150 200
0
0.2
0.4
0.6
0.8
1
CDF
Cost
GPS−Density 20%
GPS−Density 80%
Random 20%
Random 80%
LP based
Optimal
(c) CDF of the cellular cost when pT= 0.2
0 0.1
0
50
100
150
200
250
300
350
Probability of transition
Average cost
Base
GPS−Density 20%
GPS−Density 80%
Random 20%
Random 80%
LP based
Optimal
(d) Average cellular cost under different values
of pT
Fig. 12. Performance evaluation under different dynamic event probabilities
6.4 Evaluation Results under Different Dynamic
Event Probabilities
From Fig. 11, one shall notice that the dynamic event
probability has a deep influence on the cellular commu-
nication cost of our LP-based algorithm. Therefore, in
this section, we specially study this issue by varying the
dynamic event probabilities pT={0,0.1,0.2}and fixing
T,|H|and |D|as 60, 30 and 30, respectively. Fig. 12 first
shows the CDF of cellular cost under different values of
pT. When no dynamic event happens during the content
dissemination process, i.e., pT= 0, as shown in Fig.
12(a), we can see that our LP-based algorithm tightly
match with the optimal solution because the speculative
contact events truly reflect the actual ones. When pT
increases, as shown in Fig. 12(b) and Fig. 12(c), our
LP-based algorithm starts to deviate from the optimal
solution. This is attributed to the fact that more packets
may be required to be offloaded to compensate the
errors in the prediction of contact events. Nevertheless,
our LP-based algorithm can always outperform GPS-
Density and Random-based ones. For example, when
pT= 0.1, there are 95% cases that our algorithm can
11
30 60 90 120 150
0
50
100
150
200
250
Content deadline
Average cost
GPS−Density 20%
GPS−Density 80%
Random 20%
Random 80%
LP based pT=0.3
Optimal pT=0.3
(a) Performance comparison when pT=0.1
30 60 90 120 150 180
0
50
100
150
Content deadline
Average cost
LP based pT=0
Optimal pT=0
LP based pT=0.15
Optimal pT=0.15
LP based pT=0.3
Optimal pT=0.3
(b) Average cellular cost v.s. content deadline
Fig. 13. Performance evaluation under different content deadlines
2 4 6 8 10
100
150
200
250
300
Number of helpers
Average cost
GPS−Density 20%
GPS−Density 80%
Random 20%
Random 80%
LP based pT=0.05
Optimal pT=0.05
(a) Performance comparison when pT=0.05
2 4 6 8 10
100
120
140
160
180
Number of helpers
Average cost
LP based pT=0
Optimal pT=0
LP based pT=0.05
Optimal pT=0.05
LP based pT=0.1
Optimal pT=0.1
(b) Average cellular cost v.s. the number of
helpers
Fig. 14. Performance evaluation under different number of helpers
achieve a cost less than 80, while this percentage is only
85% to “GPS-Density 80%”, 78% to “Random 80%”, and
even worse to “GPS-Density 20%” and “Random 20%”.
Fig. 12(d) further validates this by showing the average
cost under different dynamic event probabilities. The
Base algorithm always has the worst performance, and
Random-based algorithm performs not so bad and close
to GPS-Density based algorithm. The most important
observation is that our proposed LP based algorithm
outperforms all other heuristics and approaches to the
optimal performance.
6.5 Evaluation Results under Different Deadline
Next, we study how the content deadline influences the
cost. In this group of experiments, we set |H|,|D|,Kset
as 30, 30 and 10, respectively. The content deadline Tis
varied from 30 to 180 with a step of 30. The evaluation
results are shown in Fig. 13. In the comparisons with
GPS-Density and Random in Fig. 13(a), we notice that
the average cellular cost in all the algorithms shows
as a decreasing function of the content deadline. This
phenomenon can be also observed in Fig. 13(b). This
is because larger deadline makes each subscriber have
more opportunities to collect coded packets and thus a
coded packet is able to benefit more subscribers. The
number of coded packets needed is reduced and the cost
decreases. However, with further increasing of content
deadline, we notice that the decreasing on the cost
becomes marginal. This is attributed to the fact that
although larger deadline can bring up more potential
V2V transmission opportunities, some of them become
redundant as the content has been decoded by most
subscribers.
6.6 Evaluation Results under Different Network Size
Finally, we investigate the effect of network size to the
cellular cost by varying the number of helpers |H|and
the number of subscribers |D|, respectively. Fig. 14 shows
the average cellular cost as a function of the number of
helpers in the settings T= 60,K= 10 and |D|= 30.
One may also first observe that the cost obtained by any
algorithm slightly deceases with the increasing of the
number of helpers. This is because a coded packet can
potentially benefit more subscribers via the relaying of
helpers if there are more helpers. When the number of
helpers is fixed, as shown in Fig. 15, we can see that
the cellular cost shows as an increasing function of the
number of subscribers. Without doubt, more subscribers
12
8 12 16 20
50
100
150
200
Number of subscriber
Average cost
GPS−Density 20%
GPS−Density 80%
Random 20%
Random 80%
LP based pT=0.05
Optimal pT=0.05
(a) Performance comparison when pT=0.05
2 4 6 8 10 16 20 30
0
50
100
150
Number of subscribers
Average cost
LP based pT=0
Optimal pT=0
LP based pT=0.05
Optimal pT=0.05
LP based pT=0.1
Optimal pT=0.1
(b) Average cellular cost v.s. the number of sub-
scribers
Fig. 15. Performance evaluation under different number of subscribers
potentially requires more packets and the cellular cost
is thus increased. However, one shall notice that the
increasing on the cellular cost do not linearly increases
with the number of subscribers. This is because more
subscribers makes a coded packet be reused more, po-
tentially lessening the burden on the requirement of the
coded packet.
7 CONCLUSION
In this paper, we study a minimum offloading prob-
lem on offloading cellular traffic to VDTNs by leverag-
ing V2V communication capability of on-road vehicles
where cooperative V2V transmissions are utilized to
help dissemination of a deadline-constrained content to
its subscribers. To minimize the cellular communica-
tion cost while ensuring that all subscribers fully get
the desired content, we first build an LP formulation
using flow-model on a predicted DTCG to obtain the
initial offloading solution. To deal with various unex-
pected network dynamics during the content dissemi-
nation process, an online dynamic offloading algorithm
is designed. Finally, an offline LP model is built on a
deterministic DTCG constructed from the contact events
trace to derive the optimal offloading solution. Through
extensive simulation studies, we validate the high ef-
ficiency of our online algorithm as it approaches the
results of the offline solution.
ACKNOWLEDGEMENTS
The work was partially supported by Fundamental
Research Funds for National University, China Uni-
versity of Geosciences, Wuhan (Grant No. CUG14065,
CUGL150830). S. Guo is the corresponding author.
REFERENCES
[1] “Cisco Visual Networking Index: Global Mo-
bile Data Traffic Forecast Update, 2012-2017,”
http://www.cisco.com/en/US/solutions/collateral/ns341/ns525/
ns537/ns705/ns827/white paper c11-520862.html.
[2] A. R. Mishra, Fundamentals of Cellular Network Planning and Opti-
misation: 2G/2.5 G/3G... Evolution to 4G. Wiley, 2004.
[3] B. Han, P. Hui, V. Kumar, M. V. Marathe, G. Pei, and A. Srinivasan,
“Cellular Traffic Offloading through Opportunistic Communica-
tions: a Case Study,” in Proceedings of the 5th ACM Workshop on
Challenged Networks (CHANTS). ACM, 2010, pp. 31–38.
[4] Y. Li, G. Su, P. Hui, D. Jin, L. Su, and L. Zeng, “Multiple Mobile
Data Offloading through Delay Tolerant Networks,” in Proceedings
of the 6th ACM Workshop on Challenged Networks (CHANTS). New
York, NY, USA: ACM, 2011, pp. 43–48.
[5] Y. Go, Y. Moon, G. Nam, and K. Park, “A Disruption-tolerant
Transmission Protocol for Practical Mobile Data Ofoading,” in
Proceedings of the third ACM international workshop on Mobile Op-
portunistic Networks (MobiOpp). New York, NY, USA: ACM, 2012,
pp. 61–68.
[6] B. Han, P. Hui, V. Kumar, M. Marathe, J. Shao, and A. Srinivasan,
“Mobile Data Offloading through Opportunistic Communications
and Social Participation,” IEEE Transactions on Mobile Computing,
vol. 11, no. 5, pp. 821–834, May 2012.
[7] S. Kashihara, M. Sanadidi, and M. Gerla, “Mobile, Personal Data
Offloading to Public Transport Vehicles,” in Proceedings of the 2012
IEEE International Conference on Mobile Computing and Ubiquitous
Networking. IEEE, May 2012.
[8] Y.-J. Chuang and K.-J. Lin, “Cellular Traffic Offloading through
Community-based Opportunistic Dissemination,” in Proceedings
of the 2012 IEEE Wireless Communications and Networking Conference
(WCNC). IEEE, Apr. 2012, pp. 3188–3193.
[9] J. Burgess, B. Gallagher, D. Jensen, and B. Levine, “MaxProp:
Routing for Vehicle-Based Disruption-Tolerant Networks,” in Pro-
ceedings of the 25th Conference on Information Communications (IN-
FOCOM), vol. 6, Barcelona, Spain. IEEE, Apr. 2006, pp. 1–11.
[10] R. Lu, X. Lin, and X. Shen, “Spring: A Social-based Privacy-
Preserving Packet Forwarding Protocol for Vehicular Delay Toler-
ant Networks,” in Proceedings of the 29th Conference on Information
Communications (INFOCOM). IEEE, Mar. 2010, pp. 1–9.
[11] V. Soares, F. Farahmand, and J. Rodrigues, “Improving Vehicular
Delay-Tolerant Network Performance with Relay Nodes,” in Pro-
ceedings of the 2009 Conference on Next Generation Internet Networks
(NGI). IEEE, July 2009, pp. 1–5.
[12] T. Yan, W. Zhang, and G. Wang, “DOVE: Data Dissemination to
a Fixed Number of Receivers in VANET,” in Proceedings of the 9th
Annual Conference on Sensor, Mesh and Ad Hoc Communications and
Networks (SECON). IEEE, Jun. 2012, pp. 272–280.
[13] Q. Xu, T. Mak, J. Ko, and R. Sengupta, “Vehicle-to-Vehicle Safety
Messaging in DSRC,” in Proceedings of the 1st ACM International
Workshop on Vehicular Ad hoc Networks (VANET). New York, NY,
USA: ACM, 2004, pp. 19–28.
[14] D. Jiang and L. Delgrossi, “IEEE 802.11p: Towards an Internation-
al Standard for Wireless Access in Vehicular Environments,” in
Proceedings of the 2008 IEEE Vehicular Technology Conference (VTC-
Spring ’08). IEEE, May 2008, pp. 2036–2040.
[15] B. B. Chen and M. C. Chan, “MobTorrent: A Framework for
Mobile Internet Access from Vehicles,” in Proceedings of the 28th
13
Conference on Information Communications (INFOCOM). IEEE, Apr.
2009, pp. 1404–1412.
[16] M. Sardari, F. Hendessi, and F. Fekri, “DMRC: Dissemination
of Multimedia in Vehicular Networks using Rateless Codes,” in
Proceedings of the 28th IEEE International Conference on Computer
Communications Workshops (INFOCOMW). Piscataway, NJ, USA:
IEEE, April 2009, pp. 182–187.
[17] M. Li, Z. Yang, and W. Lou, “CodeOn: Cooperative Popular
Content Distribution for Vehicular Networks using Symbol Level
Network Coding,” IEEE Journal on Selected Areas in Communica-
tions, vol. 29, no. 1, pp. 223–235, Jan. 2011.
[18] J. Whitbeck, Y. Lopez, J. Leguay, V. Conan, and M. D. de Amorim,
“Push-and-Track: Saving Infrastructure Bandwidth Through Op-
portunistic Forwarding,” Pervasive and Mobile Computing, vol. 8,
no. 5, pp. 682–697, 2012.
[19] M. Barbera, J. Stefa, A. Viana, M. de Amorim, and M. Boc, “VIP
delegation: Enabling VIPs to Offload Data in Wireless Social
Mobile Networks,” in Proceedings of the 2011 IEEE International
Conference on Distributed Computing in Sensor Systems and Work-
shops (DCOSS). IEEE, Jun. 2011, pp. 1–8.
[20] J. Whitbeck, M. Amorim, Y. Lopez, J. Leguay, and V. Conan, “Re-
lieving the Wireless Infrastructure: When Opportunistic Networks
Meet Guaranteed Delays,” in Proceedings of the 2011 IEEE Inter-
national Symposium on a World of Wireless, Mobile and Multimedia
Networks (WoWMoM). IEEE, Jun. 2011, pp. 1–10.
[21] S. Dimatteo, P. Hui, B. Han, and V. O. Li, “Cellular traffic
offloading through wifi networks,” in 2011 IEEE 8th International
Conference on Mobile Adhoc and Sensor Systems (MASS). IEEE,
2011, pp. 192–201.
[22] L. Xiaofeng, H. Pan, and P. Lio, “Offloading mobile data from
cellular networks through peer-to-peer wifi communication: A
subscribe-and-send architecture,” Communications, China, vol. 10,
no. 6, pp. 35–46, 2013.
[23] X. Zhang, J. Kurose, B. N. Levine, D. Towsley, and H. Zhang,
“Study of a bus-based disruption-tolerant network: mobility mod-
eling and impact on routing,” in Proceedings of the 13th annual
ACM international conference on Mobile computing and networking.
ACM, 2007, pp. 195–206.
[24] M. Doering, T. P¨ogel, W.-B. P ¨ottner, and L. Wolf, “A new mobility
trace for realistic large-scale simulation of bus-based dtns,” in
Proceedings of the 5th ACM workshop on Challenged networks. ACM,
2010, pp. 71–74.
[25] J. Luo, X. Gu, T. Zhao, and W. Yan, “Mi-vanet: a new mobile
infrastructure based vanet architecture for urban environment,”
in Vehicular Technology Conference Fall (VTC 2010-Fall), 2010 IEEE
72nd. IEEE, 2010, pp. 1–5.
[26] H.-s. Park and J.-d. Kim, “Wi-fi-based modeling and hybrid
routing scheme for delay-tolerant public bus network,” Wireless
Communications and Mobile Computing, 2013.
[27] P. Hui, J. Crowcroft, and E. Yoneki, “Bubble rap: Social-based for-
warding in delay-tolerant networks,” IEEE Transactions on Mobile
Computing, vol. 10, no. 11, pp. 1576–1589, 2011.
[28] J.-S. Park, M. Gerla, D. Lun, Y. Yi, and M. Medard, “CodeCast: a
Network-Coding-based Ad hoc Multicast Protocol,” IEEE Wireless
Communications, vol. 13, no. 5, pp. 76–81, Oct. 2006.
[29] L. Baum and P. Billingsley, “Asymptotic Distributions for the
Coupon Collector’s Problem,” The Annals of Mathematical Statis-
tics, pp. 1835–1839, 1965.
[30] S. Deb, M. M ´edard, and C. Choute, “Algebraic Gossip: A Network
Coding Approach to Optimal Multiple Rumor Mongering,” IEEE
Transactions on Information Theory, vol. 52, no. 6, pp. 2486–2507,
2006.
[31] C. Fragouli, J. Widmer, and J. Le Boudec, “On the Benefits of
Network Coding for Wireless Applications,” in Proceedings of the
4th International Symposium on Modeling and Optimization in Mobile
Ad Hoc and Wireless Networks (WiOpt). IEEE, 2006, pp. 1–6.
[32] ——, “Efficient Broadcasting using Network Coding,” IEEE/ACM
Transactions on Networking, vol. 16, no. 2, pp. 450–463, 2008.
[33] D. Zhang and C. K. Yeo, “Enabling Efficient WiFi-Based Vehicular
Content Distribution,” IEEE Transactions on Parallel and Distributed
Systems, vol. 24, no. 3, pp. 479–492, 2013.
[34] I. Gurobi Optimization, “Gurobi Optimizer Reference Manual,”
2013. [Online]. Available: http://www.gurobi.com
Hong Yao received the B.E. degree in Computer
and Applications from Wuhan Technical Univer-
sity of Surveying and Mapping in 1998, the M.E.
degree in Cartography and Geographic Informa-
tion Engineering from China University of Geo-
sciences, Wuhan, and the Ph.D. degree in Com-
puter Science and Technology from Huazhong
University of Science and Technology in 2010.
He is currently an associate professor in the
School of Computer Science at China University
of Geosciences, Wuhan. His research interests
include wireless and mobile network, delay tolerant networks, overlay
networks, and mobile cloud computing. He is a member of the IEEE.
Deze Zeng received his Ph.D. and M.S. degrees
in computer science from University of Aizu,
Aizu-Wakamatsu, Japan, in 2013 and 2009, re-
spectively. He received his B.S. degree from
School of Computer Science and Technology,
Huazhong University of Science and Technolo-
gy, China in 2007. He is currently an associate
professor in School of Computer Science, China
University of Geosciences (Wuhan), China. His
current research interests include: cloud com-
puting, software-defined sensor networks, data
center networking, networking protocol design and analysis. He is a
member of IEEE.
Huawei Huang received the Master degree in
computer science from the China University of
Geoscience (Wuhan) in 2013. He is currently a
PhD candidate at School of Computer Science
and Engineering, The University of Aizu, Japan.
His research interests are mainly in the area of
software defined networking and wireless net-
works.
Song Guo (M’02-SM’11) received the PhD de-
gree in computer science from the University
of Ottawa, Canada in 2005. He is currently a
Full Professor at School of Computer Science
and Engineering, the University of Aizu, Japan.
His research interests are mainly in the areas
of protocol design and performance analysis
for reliable, energy-efficient, and cost effective
communications in wireless networks. He has
published over 250 papers in refereed journals
and conferences in these areas and received
three IEEE/ACM best paper awards. Dr. Guo currently serves as Asso-
ciate Editor of IEEE Transactions on Parallel and Distributed Systems,
Associate Editor of IEEE Transactions on Emerging Topics in Computing
with duties on emerging paradigms in computational communication
systems, and on editorial boards of many others. He has also been in
organizing and technical committees of numerous international confer-
ences. Dr. Guo is a senior member of the IEEE and the ACM.
14
Ahmed Barnawi is an associate professor at
the Faculty of Computing and IT, King Abdu-
laziz University, Jeddah, Saudi Arabia, where
he works since 2007. He received PhD at the
University of Bradford, UK in 2006. He was
visiting professor at the University of Calgary in
2009. His research areas are cellular and mo-
bile communications, mobile ad hoc and sensor
networks, cognitive radio networks and security.
He received three strategic research grants and
registered two patents in the US. He is IEEE
member.
Ivan Stojmenovic was editor-in-chief of IEEE
Transactions on Parallel and Distributed System-
s (2010-3), is Associate Editor-in-Chief of Ts-
inghua Journal of Science and Technology, and
is founder of four journals (Journal of Multiple
Valued Logic and Soft Computing, Ad Hoc &
Sensor Wireless Networks, International Jour-
nal of Parallel, Emergent and Distributed Sys-
tems, Cyber Physical Systems). He is editor of
IEEE Transactions on Computers, IEEE Net-
work, IEEE Transactions on Cloud Computing
etc. Stojmenovic has top h-index in Canada for mathematics and statis-
tics, and has >16000 citations and h=63. He received five best paper
awards. He is Fellow of IEEE, Canadian Academy of Engineering and
Academia Europaea. He received Humboldt Research Award in Ger-
many and Royal Society Research Merit Award in UK. He is Tsinghua
1000 Plan Distinguished Professor.
