multiple content dissemination in roadside-unit-aided vehicular opportunistic networks

0018-9545 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. Seehttp://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI10.1109/TVT.2014.2308149, IEEE Transactions on Vehicular Technology

1

Multiple Content Dissemination in Roadside UnitAided Vehicular Opportunistic Networks

Yong Li, Member, IEEE, Xiangming Zhu, Depeng Jin, Member, IEEE, and Dapeng (Oliver) Wu, Fellow, IEEE

Abstract—Roadside Units (RSUs), which enable the vehicles-to-infrastructure communication, are deployed along roadsidesto handle the growing communication demands as the num-ber of vehicles increases. The current opportunistic RSU-aidedcontent dissemination schemes, however, do not address theheterogeneity networks in terms of data items and users. Weestablish a mathematical framework to study the problem ofmultiple content dissemination under realistic RSU-aided oppor-tunistic network assumptions, where 1) mobile content itemsare heterogeneous in terms of size and lifetime; 2) vehicles’interests are different to different data; and 3) the RSU’ storagesfor content dissemination are limited in size. We formulatethe maximum data dissemination as a submodular functionmaximisation problem with multiple linear constraints of limitedstorage. Then, we propose an efficient heuristic algorithm to solvethis NP-hard problem. Finally, we demonstrate the effectivenessof our algorithm through extensive simulations using realisticvehicular traces. The simulation results show that our proposedlow-complexity heuristic algorithm performs much better thanthe existing feasible solutions, and achieves similar performanceas the most accurate algorithm currently available, whose com-putational complexity is unacceptable in practice.

Index Terms—Mobile content dissemination, vehicular oppor-tunistic networks, roadside unit, submodular function maximiza-tion.

I. INTRODUCTION

With the ever-increasing number of vehicles on roads, trafficjams and accidents have become a serious and widespreadproblem [1]. To alleviate this serious problem, recently thereis a strong interest in developing vehicular networks to en-able wireless communications between vehicles and RoadsideUnit (RSU) infrastructure equipments, known as vehicles toinfrastructures (V2I), and between vehicles, known as vehiclesto vehicles (V2V) [2]. Federal Communications Commis-sion in USA has allocated 75 MHz of spectrum for suchdedicated short-range communications in vehicular networks[3], and IEEE is also working on standard specifications for

Copyright (c) 2013 IEEE. Personal use of this material is permitted.However, permission to use this material for any other purposes must beobtained from the IEEE by sending a request to [email protected].

Y. Li, X. Zhu and D. Jin are with Tsinghua National Laboratory forInformation Science and Technology (TNLIST), Department of Electron-ic Engineering, Tsinghua University, Beijing 100084, China (E-mails: [email protected]).

D. O. Wu is with the Department of Electrical and Computer Engineering,University of Florida, Gainesville, FL 32611-6130, USA.

This work is supported by National Basic Research Program of China(973 Program) (No. 2013CB329001), National Nature Science Foundationof China (No. 61301080, No. 61171065 and No. 61273214), National HighTechnology Research and Development Program (No. 2013AA013501 andNo. 2013AA013505), Chinese National Major Scientific and TechnologicalSpecialized Project (No. 2013ZX03002001), and Chinas Next GenerationInternet (No. CNGI-12-03-007).

vehicular networks. With an increasing number of vehiclesequipped with devices to provide vehicular communicationcapacities, large scale RSU-aided networks are expected tobe available [2]. Then, many related applications will emerge,which include safety applications, such as automatic collisionwarning, remote vehicle diagnostics, emergency management,assistance for safely driving, etc., and other applications, suchas automobile high speed Internet access, vehicle tracking,multimedia content sharing, and so on.

Content dissemination is useful in many vehicular appli-cations, including event notification, transportation systemstatus update, and content publishing for safety informationand entertainment data [4]–[7]. To achieve these applications,the V2I communications between vehicles and RSUs playan important role since usually V2V communications arevery opportunistic and stochastic [8]. In RSU-aided vehicularnetworks, i.e., RSUs deployed along the roadside [9], vehicleswill have more steady connections with the RSUs to transmitdata. In such kind of networks, data traffic demands initiatedfrom vehicles are random and bursty by nature. RSUs act asthe gateways to the Internet and to other infrastructure systemssuch as Intelligent Transport System (ITS). Vehicles transmittheir Internet access requests and information to RSUs, andRSUs then send responses to the Internet for querying thedata and information on behalf of the vehicles [8]. Usually, thecommunication coverage of one RSU is limited to a certainarea due to the short wireless communication range. On theother hand, vehicles are highly mobile and sometimes sparseby nature. Therefore, it is difficult to maintain a connectednetwork between the RSUs and vehicles to communicate.Thus, opportunistic contacts between the RSUs and vehicles,by contrast, are capable of providing high-bandwidth commu-nication capacity for data transmission, which is known asone type of vehicular opportunistic networks [7], [10], [11],[12]. With this kind of mechanism, when the vehicles are notavailable for the RSUs to transmit content, the RSUs willstore the content in their local buffers, and forward them toother appropriate vehicles when a transmission opportunityis available along their movement, which is referred to as acommunication contact [13].

Many challenging and open problems exist in designingRSU-aided vehicular opportunistic networks [14], and current-ly many consortia and standardization bodies are actively de-veloping technologies and protocols for efficient data transmis-sion [14], [15]. Recent works have focused on how to deployRSU infrastructure to handle the growing communication de-mands as the number of vehicles increases, and have proposedoptimal RSU placement schemes with the consideration of the



2

vehicular traffic and city structures [8], [14], [16]–[18]. Withan optimal RSU deployment, which dramatically enhances theperformance of the opportunistic vehicular networks in termsof data transmission delay and ratio [14], one of the majorremaining problems is how to efficiently utilise the RSUs toimprove the data dissemination performance. In the RSU-aidedvehicular opportunistic networks, data dissemination efficiencydepends on how the RSUs replicate the mobile data and,furthermore, the heterogeneous mobile content and vehicularmobility critically influence the opportunistic data transmis-sion. Therefore, new and different schemes are required forefficient mobile content dissemination.

In this paper, we establish a mathematical framework tostudy the content dissemination in the realistic opportunisticRSU-aided network environment with multiple data items.This problem is challenging for several reasons. Firstly, con-tent items in the opportunistic network are very different,in terms of their delay-sensitivity, content size, etc. On theother hand, a RSU-aided network’s resources are limited inpractice, e.g. limited storage of RSUs, and the communicationcontacts are opportunistic in nature. Therefore, it is difficult toefficiently disseminate these heterogeneous content items withsuch limited network resources. To model a realistic inter-vehicular network environment, we considers the followingnetwork settings: 1) the network contains heterogeneous vehi-cles, in terms of content preference, 2) the content items aremulti-types of different delay sensitivity and size, and 3) theRSUs’ storages are limited in size. These realistic conditionswere not taken into account in the previous works [7], [9] forsimplicity reasons. Our novel contributions are summarised asfollows:

• We formulate the optimal content dissemination withheterogeneous data items and RSUs of limited storageas a submodular function maximisation problem underlinear constraints.

• We prove that this optimal content dissemination problemis NP-hard, and we propose a heuristic algorithm to solveit.

• Through extensive real trace-driven simulations, wedemonstrate that our algorithm achieves good systemperformance in challenging opportunistic RSU-aided net-work environments, which performs much better than theexisting feasible solutions in terms of content dissemi-nation amount and ratio, and achieves almost the sameperformance by reducing the computational time by about500x in comparison with the most accurate algorithmcurrently available.

The rest of the paper is organized as follows. In Section II,we describe the communication contact aware mobile data dis-semination system and formulate the associated optimizationproblem. In Section III, we specify the optimization problemand design the heuristic algorithm to solved the formulatedproblem. In Section IV, we introduce the experimental en-vironment for performance evaluation and provide extensivesimulation results. Finally, we conclude the paper in Section V.

Internet

RSU Device

Coverage Area

Content ServerCentral Controller

Wired Link

Opportunistic Contact

Fig. 1. Illustration of content dissemination system for the RSU-aidedvehicular opportunistic communications.

II. SYSTEM OVERVIEW AND PROBLEM FORMULATION

A. System Overview and Network Modelling

The envisaged mobile content dissemination system is illus-trated in Fig. 1, where vehicles travel around the city roads,and the RSUs-based infrastructure network provides coverageover some regions. The RSUs are connected to the contentservers in the Internet through wired-line links. The mobilecontent are first delivered from the corresponding contentservers to the RSUs under the guidance of the content storagepolicy. The RSUs will further disseminate the mobile contentsto the vehicles that are interested in the contents, through op-portunistic communication which occurs when vehicles moveinto the communication range of RSUs.

A central controller is deployed for the mobile contentdissemination system, and it is tasked to make the contentstorage decisions for the content servers, based on the sys-tem parameters, such as the storage sizes of RSUs, mobilecontent interests, and the contact rate between vehicles andRSUs, so that the content servers can distribute the contentsappropriately into the buffers of RSUs. Note that the centralcontroller can communicate with the content servers, as it isconnected to them directly via wired-links, while it can alsocommunicate with vehicles via RSUs connecting to the wiredInternet. Thus, the central controller can collect the requiredinformation of the nodes’ contact rates, the storage sizes ofRSUs and the content interests. In other words, althoughthe required information are distributed in the opportunisticRSUs-aided network, these information can be obtained bythe central controller in order to make the storage allocationdecisions so as to optimize the system performance of the



3

opportunistic RSUs-aided network.In such an opportunistic content dissemination system, the

service provider transmit content to the deployed RSUs. RSUsthen further propagate the data to vehicles that subscribe thedata by opportunistic communication. As depicted in Fig. 2, aRSU can store more than one content item, depending on thebuffer and content sizes, and a vehicle may also be interestedin different content items. Then vehicles can obtain the contentfrom these deployed RSUs according to their subscribingrelationship to the content by the opportunistic communicationparadigm.

In general, there are V vehicles in the system, labeled asv ∈ {1, 2, · · · , V }. For the mobile content, we model themas I kinds of mobile data with different delay-sensitivities,content sizes, etc., which are labelled as I. For any i ∈ I,its data length is li. Also, there are R RSUs in the system,labeled as r ∈ {1, 2, · · · , R}. In the RSUs, the system requirestheir storages to buffer the mobile content, which may includemultimedia content of very large size, such as movies. Eventhough RSU devices may have large storage, it is impossibleto ask a RSU to contribute all its storage solely for themobile data dissemination purpose. Therefore, we should takethe storage that each RSU is willing to share as one of ourconstraints, which directly influences the number of data itemsthat can be stored. Considering this realistic condition, weassume that and RSU r can at most buffer Lr size of dataitems. We use V to denote the vehicles, and R for the RSUs,where |V| = V and |R| = R. Vehicles can communicate withRSUs only when they move to within the transmission rangeof the deployed RSUs, which is referred to as a communicationcontact. We assume that the communication contact betweenvehicle v and RSU r obeys the Poisson process with a contactrate γvr. Poisson distributed contact rate has been validated tofit well to real vehicular traces and is widely used to modelopportunistic vehicular systems [19]–[21].

Any vehicles in V may be interested in a data item,

Share data by

opportunistic contact

RSU Device

Vehicular User

Mobile Data

Internet

Download and Replicate

data into RSU

Fig. 2. Overview of the multiple data dissemination in RSU-aided oppor-tunistic vehicular network with the heterogeneous data of different length andlife time marked by the color.

and obtains it through opportunistic communication from thedeployed RSUs. In a system with multiple content items, thevehicles as the content subscribers will have different interestsin different items. Moreover, some data items are popular datathat are of interest to many subscribers, while some other dataitems are not popular data which may only be interesting toa very small number of subscribers. We describe the vehicle’sinterests to different mobile data by a subscriber profile, andmodel the popularity of mobile data by an interest distribution.In this way, we model the steady state of the subscribers’interests in different data items.

Specifically, for all the mobile data, the system has Hkeywords, denoted by the set H, to describe them. Any itemi ∈ I is described by a subset of keywords, denoted byHi ⊆ H, and the associated weights ςhi , which indicates theimportance of keyword hi ∈ Hi. In this way, we can define thepopularity of mobile data items. We assume

∑hi∈Hi

ςhi = 1.To model the interests of different subscribers on differentdata, we define ϑh

v as the degree of vehicle v ∈ V beeninterested in keyword h ∈ H. In this way, we can compare theinterests of vehicle v to two different keywords h1, h2 ∈ H byϑh1v and ϑh2

v . Thus, the interest profile of vehicle v is definedby the set {ϑh

v : h ∈ H}. We assume∑

h∈H ϑhv = 1. The

interest probability of vehicle v ∈ V in mobile data i ∈ I,defined by wvi, can be obtained as follows

wvi =∑

hi∈Hi

ςhiϑhiv . (1)

Thus, finally, we define the vehicle’s interest as the probabilitythat it may get interested on each data item, where vehicle vis associated with a vector wv = [wv1 wv2 · · ·wvI ]

T, wherewvi defines the degree of the vehicle’s interest in data item i.

B. Problem Formulation

Denote X = (xri), r ∈ R and i ∈ I, as the storageallocation policy, in which xri ∈ {0, 1} and xri = 1 indicatesthat RSU r stores item i in its buffer. A lifetime Ti is assignedto each data item i, and all the RSUs will discard this data atdeadline Ti. If vehicles do not receive a required item from thedeployed RSUs after the lifetime is expired, the disseminationof this data item fails. Therefore, the system should maximisethe expectation of the disseminated data size in all the vehicles,and this objective function can be expressed as

J(X) =∑i∈I

li∑v∈V

div, (2)

where div is the probability that vehicle v receives data i beforedeadline Ti.

Since more than one RSU may store item i, we considerthe dissemination opportunity metric for r ∈ R, v ∈ V andi ∈ I, denoted as tirv , which is the probability that vehiclev obtains content i from RSU r. Because the contact ratebetween r and v follows the Poisson distribution with rate γvrand the contact event is independent to the vehicle interests,we can model the dissemination opportunity as the Poissonprocess with rate γvrwvi. If RSU r stores content i in itsbuffer, we denote the probability that vehicle v obtains content



4

i from r before the lifetime Ti as tirv = 1 − e−xriγvrwviTi .Therefore, the probability that vehicle v cannot obtain contenti from all possible r ∈ R can be expressed as

∏r∈R

(1− tirv

).

It then becomes obvious that the probability of vehicle vreceiving data i before deadline Ti, namely div, is given bydiv = 1−

∏r∈R

(1−tirv

). substituting div into (2), the expectation

of the total disseminated data size before lifetime, J(X), canbe written as:

J(X) =∑i∈I

li∑v∈V

(1− exp{−

∑r∈R

xriγvrwviTi}). (3)

For the simplicity of analysis, we further define the utilityfunction J(X) over the subset of R×I. For A ⊆ R×I, wedefine the storage allocation policy X as

X = F (A), s.t. xri = 1 if (r, i) ∈ A and xri = 0 if (r, i) /∈ A.

Obviously, F (A) is a bijection, and we have⌢

J(A) = J(F (A)) =∑i∈I

li∑v∈V

(1−exp{−wviTi

∑r:(r,i)∈A

γvr}).

(4)Thus, maximising the system’s expected disseminated data

size for all the items and over all the vehicles can be specifiedas the following optimisation problem

max⌢

J(A)s.t. xri ∈ {0, 1}, ∀r ∈ R, i ∈ I,∑

i∈Ilixri ≤ Lr, ∀r ∈ R,

(5)

where∑

i∈I xrili ≤ Lr is the buffer size constraint of RSUr.

III. MULTIPLE DATA DISSEMINATION ALGORITHM

We first prove that the problem (5) is a submodular functionmaximisation (SFM) under multiple linear constraints (MLCs),which is NP-hard, and then design a heuristic algorithm tosolve it efficiently.

A. Utility Function Analysis

Submodularity has long been studied in various problems[22]–[24]. A function f defined on subsets of the universeC is called submodular, if and only if f(A ∪ x) − f(A) ≥f(B ∪ x) − f(B) holds for ∀A ⊆ B ⊆ C and ∀x ∈ C\B.For A ⊆ B ⊆ 2R×I and (r0, i0) ∈ 2R×I\B, throughsimple derivation, we can have

(⌢

J(A∪{(r0, i0)})−⌢

J(A))−(

⌢

J(B ∪ {(r0, i0)})−⌢

J(B))≥ 0, which shows that

⌢

J(A) isa submodular function on 2R×I . About the complexity of theformulated problem, we have the following Theorem.

Theorem 1: The formulated problem (5) is NP-complete.Proof: To prove that the problem (5) is NP-complete, we

use the technique of reduction [25]. In (5), let

|V| = |R| = 1, |I| = i, L1 = W, γ1,1 = n,

w1,i =1

nand Ti = ln

( lili − pi

), 1 ≤ i ≤ n.

Then, the problem (5) is reduced to the following problem,

maxn∑

i=1

li

(1−

(li−pi

li

)x1i)

s.t. x1i ∈ {0, 1}, ∀1 ≤ i ≤ n, andn∑

i=1

lix1i ≤W.

Furthermore, we haven∑

i=1

li

(1−

(li−pi

li

)x1i)

=n∑

j=1

pjzj ,

and the problem (5) is reduced to the following 0-1 KnapsackProblem

maxn∑

j=1

pjzj

s.t. zj ∈ {0, 1}, ∀1 ≤ j ≤ n, andn∑

j=1

ljzj ≤W,

which is NP-complete. Therefore, the problem (5) must beNP-complete.

The above analysis shows that the problem (5) is a SFMwith MLCs, which is widely studied by the computer sciencecommunity [22], [26]. In [22], an algorithm is proposed tosolve this problem by the (1 − 1/e − Θ(ε)) approximation,but it has a very high computational complexity. Specifically,the algorithm proposed in [22] consists of two procedures.The first one, the rounding procedure enumerates all thepossible combinations of items I and RSUs R, and it has acomputational complexity in the order of R! + I!, denoted asO(R!+I!), while the second procedure has a polynomial-timecomplexity which is negligible in comparison with O(R!+I!).Considering the problem (5) with R = 5 RSUs and I = 10data items, for example, the computational complexity of thealgorithm [22] is R! + I! > 36 × 105, which is clearlyunacceptable in practice.

B. Our Algorithm

Greedy approach is employed as a heuristic solution to thisSFM problem in the way that storage allocation is determinedone by one. When one more copy of an item is stored in aRSU, which is in accordance with the constraints, the objectivefunction will be enhanced. The gain of the objective functionis generally different for different choices of item and RSU.As our first greedy strategy, we can select the items and RSUsthat maximise the gain on the objective function at each stage,that is, select (r0, i0) as

(r0, i0) = arg max(r,i)∈P

(⌢

J(A′)−⌢

J(A)), (6)

where A′ = (A∪ (r, i)) is the set of data items and RSUs wehave chosen, and P is the set of possible solutions that satisfythe storage constraint.

The length of each data item is also important. This isbecause, although an item may offer a large gain, it may alsohave a huge length such that other items cannot be stored.Consequently, our second greedy strategy is to calculate thegain per unit data length for each choice of item and RSU andselect the pair (r0, i0) that maximises this per-unit-length gain

(r0, i0) = arg max(r,i)∈Z

(⌢

J(B′)−⌢

J(B))/li, (7)



5

Algorithm 1 Heuristic algorithm for the RSU-aided multiplecontent dissemination.

1: Initialize j ← 0, A0 ← ∅, j ← 0, B0 ← ∅, P, and Z2: for P = ∅ do3: j = j + 1, A′

j−1 = Aj−1 ∪ {(r, i)}4: (rj , ij) = arg max

(r,i)∈P

(⌢

J(A′j−1)−

⌢

J(Aj−1))

5: Aj = Aj−1 ∪ {(rj , ij)}6: Update P according to Lr and li, r ∈ R, i ∈ I7: end for8: for Z = ∅ do9: k = k + 1, B′

k−1 = Bk−1 ∪ {(r, i)}10: (rk, ik) = arg max

(r,i)∈Z

(⌢

J(B′k−1)−

⌢

J(Bk−1))/li

11: Bk = Bk−1 ∪ {(rk, ik)}12: Update Z according to Lr and li, r ∈ R, i ∈ I13: end for14: Return OPT∗ = max

(⌢

J(Aj),⌢

J(Bk))

where B′ = B ∪ (r, i) is the set of data items and RSUs wehave chosen, and Z is the set of possible solutions that satisfythe storage constraint.

Each of these two greedy strategies has its own drawbacks.Thus, a combination of these two strategies is used to enhancethe overall performance. In our algorithm, these two strategiesare both performed and we choose the better result fromthe two solutions. We present this heuristic algorithm inAlgorithm 1, which is a pseudo-polynomial-time algorithmwith the computational complexity O(R3I2V ). This is wellwith the computational capacity of the central controller. Onthe other hand, the performance bound of this algorithm isguaranteed, which will be demonstrated by the simulation.Besides, we have the following three comments about ourproposed scheme.

Remark 1: Our proposed heuristic algorithm is a centralisedalgorithm which is run at the central controller. The centralcontroller requires global information of the content lengthsli, the content lifetimes Ti, the storage sizes of RSUs Lr,the vehicles’ content interests wv , and the contact ratesγvr. Note that these information are not collected via theopportunistic RSUs-aided network. Since the central controlleris connected to the content servers via the wired Internet, itcan obtain the content-related parameters of li and Ti withease, and the required communication overhead is negligiblecompared with the bandwidth of the wired Internet. Also,the buffer size of RSUs Lr can be obtained via the wiredInternet since each RSU is linked to the Internet. On the otherhand, vehicles are equipped with communication interfaces toconnect to the RSUs. Therefore, they are able to send thevehicle-related parameters of wv, and γvr through the uplinkchannel of the RSU-based infrastructure wireless network tothe central controller, and this signalling overhead is alsoinsignificant compared with the bandwidth of the wirelessinfrastructure network. Basically, there already exist certainamount of signalling between a vehicle and the infrastructurenetwork, and the required vehicle-related parameters may take“piggybacking’ in these existing signallings. Moreover, li, Ti,

Lr and wi are statistic parameters, which are well-known bythe content servers or the central controller in advance. Manyvehicles travel on predetermined routes and schedules, andexamples include city buses and people traveling by cars to andfrom work. Therefore, daily mobility patterns exhibit certainregularity, and the contact rates between many vehicles andRSUs are often quasi-statistic, which can often be obtained bythe central controller in advance with high accuracy. Moreover,any new contact information can be collected regularly fromvehicles.

Remark 2: It is worth pointing out that all the big cellularnetwork providers are actively pursuing this “dual-mode” strat-egy to offload large amount of non-realtime mobile data fromtheir cellular networks to networks like WiFi or opportunisticRSUs-aided network. Mobile Internet access is getting increas-ingly popular for providing various services and applicationsincluding video, audio and images. According to the latestCisco forecasts [27], global mobile traffic will increase 18-foldbetween 2011 and 2016, and monthly global mobile data trafficwill surpass 10 exabytes in 2016. Mobile cellular networksprovide the most popular method of mobile access today. Withthe increase of mobile services and user demands, however,cellular networks will very likely be overloaded and congestedin the near future. By exploiting the delay-tolerant nature ofnon-realtime applications, service providers can shift the datatransmission to the opportunistic RSUs-aided network. Thiskind of traffic offloading is the quickest way at the smallestcost to support the exponential growth of mobile data whichotherwise could not be supported even by the 4G cellularnetwork [27].

Remark 3: In a generic opportunistic RSUs-aided contentdissemination system, the most important system performancemetrics are the amount of contents disseminated and thecontent dissemination delay. In the context of mobile dataoffloading, however, the content dissemination delay becomesa secondary and non-critical issue. In our investigation, weconsider the delay as the system constraint by imposing therequirement of delivering each mobile data item i beforeits deadline Ti. Therefore, we mainly focus on the systemperformance of the total amount of disseminated data items,which is clearly the optimisation goal in our formulatedproblem (5). For the generic opportunistic vehicular mobiledata dissemination system, it would be highly desired tosimultaneously maximise the total amount of disseminateddata items and minimise the expected dissemination delay.However, this is an extremely challenging multi-objectiveoptimisation, and the solution of which is unsolved. In thesimulation investigation, we will also evaluate the proposedalgorithm by observing the resulting dissemination delay.

IV. PERFORMANCE EVALUATION

In order to evaluate the performance of our proposed heuris-tic scheme for multiple mobile data dissemination in oppor-tunistic RSUs-aided networks, we compared the performanceof our Heuristic Algorithm with the following schemes:1) Approximation algorithm, in which we assume that the



6

lifetime Ti is small, and convert this problem into a knapsackproblem by approximating the objective function. Then it canbe solved by knapsack arithmetic [29].2) Average Algorithm, in which we deal with all items insequence in each turn to find whether the item can be storedin a new RSU;3) SFM Algorithm [22], which uses some approximationalgorithms to maximise a submodular set function subject toMLCs.

It is worth emphasising that the main contribution of thiswork is to consider the realistic heterogeneous conditions ofthe opportunistic RSUs-aided content dissemination system,which were not taken into account in the previous works[7], [9], and the three benchmark schemes adopted reflectthe whole spectrum of the currently available techniques formultiple mobile data dissemination in opportunistic RSUs-aided networks. SFM Algorithm is the best available schemein terms of performance but it has a very high computationalcomplexity. In the context of opportunistic RSUs-aided contentdissemination, SFM Algorithm is a centralized algorithmrequiring exactly the same set of the system’s parameters asour Heuristic Algorithm.

A. Evaluation System Setup

Our evaluation was conducted on the two real vehicularmobility traces, Shanghai trace [13] and Beijing trace, whichrecorded the positions of vehicles carrying GPS devices.Specifically, about 2100 operational taxis were running forthe whole month of February 2007 in Shanghai city to collectShanghai trace [13]. In collecting Beijing trace, we usedmobility track logs obtained from 27,000 participating Beijingtaxis carrying GPS receivers during the whole May month in2010. More specifically, we utilized the GPS devices to collectthe taxi locations and timestamps and GPRS modules to reportthe records every one minute for moving taxis.

In simulating the data dissemination process, the commu-nication range between vehicles was set to 100m, whichis about half of the typical setting of 200m for using thecurrent dedicated short range communication technology [30],in order to ensure reliable communication. Specifically, weassume that, when a vehicle, traveling according to the GPStrajectories recorded in the trace, move into the communicationrange of a RSU of 100m, the data dissemination occurssuccessfully if the vehicle is interested in any data stored inthe RSU’s buffer. In reality, we recognise that even when avehicle come into the communication range of a RSU, theymay still fail to exchange data reliably, as there are manyother factors that influence the physical layer link. However,since our aim is to compare the total amounts of data itemsdisseminated by a group of schemes, it is sufficient to assumezero error rate during data exchange. In fact, multiplying theresults of different algorithms by a same packet error rateobviously does not change the comparative performance ofdifferent algorithms.

To reflect the true contact rates as recorded in the trace, weused the first half of the trace to obtain the contact rates ofeach node pairs, while the second half of the trace was used

to simulate the data dissemination process. For the data items,we considered multimedia application and set the number ofmultimedia data items as 10. The sizes of the multimediadata items were generated randomly and uniformly in therange of [0MB, 200MB], unless otherwise specifically stated,while the data lifetimes followed the uniform distributionin [0, 2Ta s], where Ta was the average data lifetime. TheRSU buffer sizes were randomly and uniformly generated in[0, 2la MB], where la was the average buffer size. Vehicleinterests to different data items is also generated randomly.Besides the amount of data disseminated, the data dissemi-nation delay is also an important metric for opportunistic ve-hicular mobile data dissemination systems. Therefore, we alsocompared the dissemination delay of the proposed algorithmwith those of the three benchmarks. The data disseminationratio, defined as

DR =∑

∀v∈V,∀i∈I

wvipvi, (8)

where

pvi =

{1, vehicles v receives data i before deadline Ti,0, otherwise,

(9)specifies the average probability of all the vehicles receivingtheir requested data. We also evaluated this data disseminationratio for each of the four algorithms tested.

B. Amount of Data Disseminated

100 200 300 400 500 600 700 800 9002

4

6

8

10

12

14x 10

4

Average Buffer Length(la) (MB)

Dis

sem

inate

d D

ata

Am

ount

(MB

)

Heuristic

Approximation

Average

SFM

Fig. 3. The amounts of data disseminated by different algorithms forShanghai trace with the fixed average data lifetime of 10000 s and the variableaverage buffer size.

The results of the amount of data disseminated obtained bythe four algorithms for Shanghai trace are shown in Fig. 3and Fig. 4. Fig. 3 shows that, with the increase of the RSU’sbuffer size, the total amount of disseminated items increasessignificantly, as expected. Our Heuristic algorithm achievesalmost the same performance of the SFM algorithm, andit outperforms the Average and Approximation algorithmsconsiderably, especially when the buffer is small. The results



7

1 1.5 2 2.5 3 3.5 4 4.5 5

x 104

2

3

4

5

6

7

8

9

10

11x 10

4

Average Data Lifetime(Ta) (s)

Dis

sem

inate

d D

ata

Am

ount

(MB

)Heuristic

Approximation

Average

SFM

Fig. 4. The amounts of data disseminated by different algorithms forShanghai trace with the fixed buffer size of 100 MB and the variable averagedata lifetime.

under the different average data lifetimes are shown in Fig. 4.In general, more data items are disseminated to the vehicleswhen the data lifetime is larger, again as expected. Fig. 4 alsoshows that our Heuristic algorithm dramatically outperformsthe Average and Approximation algorithms. In particular,when the average data lifetime is 50000 s, our approachachieves about 2 and 3 times higher data amounts on average,respectively, over the Approximation and Average algorithms.

100 200 300 400 500 600 700 800 9000

0.5

1

1.5

2

2.5x 10

5


Dis

sem

inat

ed D

ata

Am

ount

(M

B)

HeuristicApproximationAverageSFM

Fig. 5. The amounts of data disseminated by different algorithms for Beijingtrace with the fixed average data lifetime of 50000 s and the variable averagebuffer size.

The results of the amount of data disseminated by thefour algorithms for Beijing trace are depicted in Fig. 5 andFig. 6. Similar observations from Shanghai trace can be drawn.Specifically, our Heuristic algorithm achieves almost the sameperformance of the SFM algorithm, and it outperforms theAverage and Approximation algorithms considerably, especial-ly when the average buffer size is small and average datalifetime is large. In particular, for Beijing trace, our Heuristic

1 1.5 2 2.5 3 3.5 4 4.5 5

x 104

1

1.2

1.4

1.6

1.8

2

2.2

2.4x 10

5

Average Data Lifetime(Ta) (s)

Dis

sem

inat

ed D

ata

Am

ount

(M

B)

HeuristicApproximationAverageSFM

Fig. 6. The amounts of data disseminated by different algorithms for Beijingtrace with the fixed average buffer size of 500 MB and the variable averagedata lifetime.

algorithm achieves a slightly smaller improvement comparedwith the results shown in Fig. 4 under the Shanghai trace.The reason is that in Beijing trace, the heterogenous of thecontact rates between the vehicular and RSUs is relativelysmaller than Shanghai since the traffic congestions are usuallyhappened. Thus, Approximation algorithm achieves relativelybetter results even when the lifetime is large. However, fromall these results, we can observe that our Heuristic algorithmachieves almost the same performance of the SFM algorithm,which clearly demonstrates the effectiveness of our proposedalgorithm.

C. Data Dissemination Latency and Ratio

We then evaluate the achievable data dissemination delayand the data dissemination ratio. Fig. 7 shows the simulationresults of the average data dissemination latency obtained bythe four algorithms with the fixed average data lifetime of20000 s and the variable average buffer size. Observe that theApproximation algorithms require much larger average datadissemination delays than our Heuristic algorithm and theSFM algorithm. The Average algorithm require respectivelysmall delays when buffer is small, and the difference will besmaller when buffer increases. Compared with the SFM algo-rithm, our Heuristic algorithm, our scheme achieves almost thesame performance of average data dissemination latency as theSFM algorithm, as can be seen from Fig. 7. Thus, our Heuristiccan obtain the same dissemination latency performance as theSFM algorithm, while benefiting from a significant saving incomputational complexity.

We next investigated the achievable average probability ofall the vehicles receiving their requested data before deadlines.Fig. 8 depicts the dissemination ratios obtained by the fouralgorithms with the fixed average data lifetime of 50000 s andthe variable average buffer size. Observe from Fig. 8 that ourHeuristic algorithm achieves a much higher dissemination ratiothan the Average and Approximation algorithms, especiallywhen buffer is small. Compared with the SFM algorithm, our



8

100 200 300 400 500 600 700 800 9001.15

1.2

1.25

1.3

1.35

1.4

1.45

1.5x 10

4


Dis

sem

ination L

ate

ncy (

s)

SFM

Heuristic

Approximation

Average

Fig. 7. Simulation Results of average data dissemination latency by differentalgorithms for Shanghai trace with the fixed average data lifetime of 20000 sand the variable average buffer size.

100 200 300 400 500 600 700 800 900

0.35

0.4

0.45

0.5

0.55

0.6

0.65

0.7

0.75

0.8


Data

Dis

sem

ination R

atio

SFM

Heuristic

Approximation

Average

Fig. 8. Simulation Results of data dissemination ratio by different algorithmsfor Shanghai trace with the fixed average data lifetime of 50000 s and thevariable average buffer size.

algorithm has almost achieve the same high dissemination ratiofor all the buffer length.

D. Influence of Mobile Data Size

In all the above simulation experiments, the sizes of themobile data were randomly and uniformly generated in therange of [0 MB, 200 MB]. To further investigate the impactof data sizes on the achievable performance, we designedan additional experiment by letting the mobile data sizes tofollow the uniform distribution in

[(ds−50), (ds+50)

]MB,

while varying the “average mobile data size” ds from 50 MBto 450 MB. This created various experimental conditions,where the ranges of the uniform distribution for the datasizes, changed from as small as [0, 100] MB to as large as[400, 500] MB. Fig. 9 shows the dissemination ratios obtainedby the four algorithms under the different average mobile datasizes ds, with the average data lifetime set to 50000 s and the

50 100 150 200 250 300 350 400 4500.2

0.3

0.4

0.5

0.6

0.7

0.8

Average Mobile Data Size (MB)

Data

dis

sem

ination R

atio

Heuristic

Approximation

Average

SFM

Fig. 9. Simulation Results of data dissemination ratio by different algorithmsfor Shanghai trace with the average data lifetime of 50000 s and the averagebuffer size of 900 MB, where the sizes of mobile data are randomly anduniformly generated from the range of

[(ds − 50), (ds +50)

]MB with the

given average mobile data size ds.

average buffer size to 900 MB. With the fixed average datalifetime and average buffer size, a higher data disseminationratio can be achieved when the average mobile data size dsis reduced, as is clearly demonstrated in Fig. 9. Also fromFig. 9, we can see that our Heuristic algorithm significantlyoutperforms the Average and Approximation algorithms interms of achievable data dissemination ratio if only data size isnot too small. Compared with SFM algorithm, our algorithmis almost the same.

E. Discussions

5 10 15 20 2510

0

101

102

103

104

The Number of Mobile Data Items

Runnin

g T

ime (

s)

HeuristicSFM

Fig. 10. Running time of SFM and Heuristic algorithms for Shanghai tracewith the average data lifetime of 50000 s, the average buffer size of 900 MB,the average data length is 200 MB, and variable number of mobile data items.

The above extensive simulation results confirm that ourHeuristic algorithm performs much better than the Approx-imation algorithm as well as the Average algorithm. These



9

two benchmark algorithms represent the most widely usedschemes in mobile data dissemination. Most significantly, ourHeuristic algorithm achieves almost the same performancein comparison with the SFM algorithm, which is the mostaccurate algorithm currently available to solve the optimisationproblem (5) of heterogeneous mobile data dissemination butit is not very practical due to its very high computationalcomplexity. In order to numerically show the computationalcomplexity, we present the running time comparison in Fig. 10with the variation of the number of data items, where the y-axisis plotted in log scale. From the results, we can observe thatwhen the number of data items increases from 5 to 25, whichenlargers the scale of the problem, the running time of SFMalgorithm is increased by about 5x. However, the running timeof our Heuristic algorithm is only increased from 3.2s to 6.0s.On average, in the above simulated scenarios, the running timeof the SFM algorithm is about 0.81 hours, while our Heuristicscheme only takes a few seconds to obtain the results, whichreduces the computational time by about 500x. This clearlydemonstrates the effectiveness of our proposed algorithm.

V. CONCLUSIONS

We have investigated the problem of multiple mobile datadissemination in opportunistic RSUs-aided networks. We havestudied this problem in a realistic environment, where thenetwork is heterogeneous, in terms of the disseminated databeing multi-types with different delay sensitivities and lengthsas well as the network dissemination RSUs’ storages beinglimited with difference sizes. By formulating this challengingproblem as a submodular function maximisation, we havedesigned an efficient heuristic algorithm to allocate the buffer.Extensive simulation results have demonstrated that ourlow-complexity algorithm achieves similar performance asthe very high-complexity SFM algorithm, traditionally usedto solve this type of challenging problems.

REFERENCES

[1] M. Khabazian, S. Aissa, and M. Mehmet-Ali, “Performance modelingof message dissemination in vehicular ad hoc networks with priority,”IEEE J. Selected Areas in Communications, vol. 29, no. 1, pp. 61–71,Jan. 2011.

[2] J. J. Blum, A. Eskandarian, and L. J. Hoffman, “Challenges of interve-hicle ad hoc networks,” IEEE Trans. Intelligent Transportation Systems,vol. 5, no. 4, pp. 347–351, Dec. 2004.

[3] J. Zhao and G. Cao, “VADD: Vehicle-assisted data delivery in vehicularad hoc networks,” in Proc. 25th IEEE INFOCOM (Barcelona, Spain),April 23-29, 2006, pp. 1–12.

[4] L. Wischhof, A. Ebner, and H. Rohling, “Information dissemination inself-organizing intervehicle networks,” IEEE Trans. Intelligent Trans-portation Systems, vol. 6, no. 1, pp. 90–101, March 2005.

[5] M. Saito, J. Tsukamoto, T. Umedu, and T. Higashino, “Design andevaluation of intervehicle dissemination protocol for propagation ofpreceding traffic information,” IEEE Trans. Intelligent TransportationSystems, vol. 8, no. 3, pp. 379–390, Sept. 2007.

[6] A. Skordylis and N. Trigoni, “Efficient data propagation in traffic-monitoring vehicular networks,” IEEE Trans. Intelligent TransportationSystems, vol. 12, no. 3, pp. 680–694, Sept. 2011.

[7] M. Johnson, L. De Nardis, and K. Ramchandran, “Collaborative contentdistribution for vehicular ad hoc networks,” in Proc. 44th Allerton Conf.Communication, Control, and Computing (Monticello, IL), Sept. 27-29,2006, pp. 751–760.

[8] A. Abdrabou and W. Zhuang, “Probabilistic delay control and roadside unit placement for vehicular ad hoc networks with disruptedconnectivity,” IEEE Journal on Selected Areas in Communications,vol. 29, no. 1, pp. 129–139, 2011.

[9] X. Lin, R. Lu, X. Liang, and X. Shen, “Stap: A social-tier-assisted packetforwarding protocol for achieving receiver-location privacy preservationin vanets,” In IEEE INFOCOM, 2011.

[10] D. Camara, N. Frangiadakis, F. Filali, and C. Bonnet, “Vehicular delaytolerant networks,” in: M. M. Cruz-Cunha and F. Moreira, Eds., Hand-book of Research on Mobility and Computing: Evolving Technologiesand Ubiquitous Impacts. IGI Global, 2011, pp. 356–367.

[11] X. Zhuo, Q. Li, W. Gao, G. Cao, and Y. Dai, “Contact duration awaredata replication in delay tolerant networks,” in Proc. 19th IEEE ICNP(Vancouver, BC Canada), Oct. 17-20, 2011, pp. 236–245.

[12] M. Doering, W. Pottner, T. Pogel, and L. Wolf,, “Impact of radiorange on contact characteristics in bus-based delay tolerant networks,” inProc. 8th Int. Conf. Wireless On-Demand Network Systems and Services(Bardonecchia, Italy), Jan. 26-28, 2011, pp. 195–202

[13] H. Zhu, L. Fu, G. Xue, Y. Zhu, M. Li, and L. Ni, “Recognizing expo-nential inter-contact time in VANETs,” in Proc. 29th IEEE INFOCOM(San Diego, CA), March 14-19, 2010, pp. 1–5.

[14] J. Gozalvez, M. Sepulcre, and R. Bauza, “IEEE 802.11p vehicle toinfrastructure communications in urban environments,” IEEE Commu-nications Magazine, vol. 50, no. 5, pp. 176–183, May 2012.

[15] G. Dimitrakopoulos and P. Demestichas, “Intelligent transportation sys-tems,” IEEE Vehicular Technology Magazine, vol. 5, no. 1, pp. 77–84,March 2010.

[16] P.-C. Lin, “Optimal roadside unit deployment in vehicle-to-infrastructurecommunications,” in Proc. 12th IEEE ITST (Taipei, Taiwan, China),Nov. 5-8, 2012, pp. 796–800.

[17] I. Filippini, F. Malandrino, G. Dan, M. Cesana, C. Casetti, and I. Marsh,“Non-cooperative RSU deployment in vehicular networks,” in Proc. 9thIEEE Annual Conf. WONS (Courmayeur, Italy), Jan. 9-11, 2012. pp. 79–82.

[18] S.-I. Sou, “A power-saving model for roadside unit deployment invehicular networks,” IEEE Communications Letters, vol. 14, no. 7, pp.623–625, July 2010.

[19] K. Lee, Y. Yi, J. Jeong, H. Won, I. Rhee, and S. Chong, “Max-contribution: On optimal resource allocation in delay tolerant networks,”in Proc. 29th IEEE INFOCOM (San Diego, CA), March 14-19, 2010,pp. 1–9.

[20] W. Gao and G. Cao, “User-centric data dissemination in disruptiontolerant networks,” in Proc. 30th IEEE INFOCOM (Shanghai, China),April 10-15, 2011, pp. 3119–3127.

[21] M. H. R. Khouzani, S. Sarkar, and E. Altman, “Maximum damagemalware attack in mobile wireless networks,” in Proc. 29th IEEEINFOCOM (San Diego, CA), March 14-19, 2010, pp. 1–9.

[22] A. Kulik, H. Shachnai, and T. Tamir, “Maximizing submodular setfunctions subject to multiple linear constraints,” in Proc. 20th AnnualACM-SIAM Symp. Discrete Algorithms (New York), Jan. 4-6, 2009,pp. 545–554.

[23] J. Leskovec, A. Krause, C. Guestrin, C. Faloutsos, J. VanBriesen, andN. Glance, “Cost-effective outbreak detection in networks,” in Proc. 13thACM SIGKDD Int. Conf. Knowledge Discovery and Data Mining (SanJose, CA), Aug. 12-15, 2007, pp. 420–429.

[24] M. Sviridenko, “A note on maximizing a submodular set function subjectto a knapsack constraint,” Operations Research Letters, vol. 32, no. 1,pp. 41–43, Jan. 2004.

[25] D. S. Hochbaum, ed., Approximation Algorithms for NP-Hard Problems’PWS Publishing Co.: Boston, MA, 1997.

[26] J. Vondrak, “Optimal approximation for the submodular welfare problemin the value oracle model,” in Proc. 40th Annual ACM Symp. Theory ofComputing (Victoria, Canada), May 17-20, 2008, pp. 67–74.

[27] Cisco visual networking index: Global mobile da-ta traffic forecast update, 2009-2014, February 2010.http://www.cisco.com/en/US/solutions/collateral/ns341 /n-s525/ns537/ns705/ns827/white paper c11-520862.html.

[28] J. Reich and A. Chaintreau, “The age of impatience: Optimal replicationschemes for opportunistic networks,” in Proc. 5th Int. Conf. EmergingNetworking Experiments and Technologies (Rome, Italy), Dec. 1-4,2009, pp. 85–96.

[29] Li. Y, Su. G, Hui. P, et al, ”Multiple mobile data offloading throughdelay tolerant networks,” in Proceedings of the 6th ACM workshop onChallenged networks, 2011, pp. 43–48.

[30] M. Ikawa, Y. Goto, Y. Igarashi, H. Kumazawa, K. Koizumi, and K. Oka,“DSRC local communication platform and its application to information



10

push service,” in Proc. 2004 IEEE Intelligent Vehicles Symposium(Parma, Italy), June 14-17, 2004, pp. 105–110.

Yong Li (S’09–M’13) received his B.S. degreein electronics and information engineering fromHuazhong University of Science and Technology,Wuhan, China, in 2007, and his Ph.D. degree inelectronic engineering from Tsinghua University,Beijing, China, in 2012.He is now a postdoctoral researcher at TsinghuaUniversity. He serves as a paper reviewer for in-ternational conferences of IEEE ICC, VTC, ICOIN,PIMRC, APCC and many others. His research fieldsinclude mobile delay tolerant networks, topics in-

cluding forwarding polices design, buffer management design and perfor-mance evaluation; mobility modeling; and mobility management in nextgeneration wireless IP networks, topics including Mobile IP, SIP, Proxy mobileIP, cross-layer design, multicast mobility, modeling for mobility performanceevaluation, enhancing handoff performance and proposing mobility manage-ment architecture.

Depeng Jin received the B.S. and Ph.D. degreesfrom Tsinghua University, Beijing, China, in 1995and 1999 respectively both in electronics engineer-ing.He is an associate professor at Tsinghua Universityand vice chair of Department of Electronic Engi-neering. Dr. Jin was awarded National Scientific andTechnological Innovation Prize (Second Class) in2002. His research fields include telecommunica-tions, high-speed networks, ASIC design and futureInternet architecture.

Dapeng Oliver Wu Dapeng Oliver Wu (S’98–M’04–SM06–F’13) received B.E. in Electrical En-gineering from Huazhong University of Science andTechnology, Wuhan, China, in 1990, M.E. in Elec-trical Engineering from Beijing University of Postsand Telecommunications, Beijing, China, in 1997,and Ph.D. in Electrical and Computer Engineeringfrom Carnegie Mellon University, Pittsburgh, PA, in2003.

He is a professor at the Department of Electricaland Computer Engineering, University of Florida,

Gainesville, FL. His research interests are in the areas of networking, com-munications, signal processing, computer vision, and machine learning. Hereceived University of Florida Research Foundation Professorship Award in2009, AFOSR Young Investigator Program (YIP) Award in 2009, ONR YoungInvestigator Program (YIP) Award in 2008, NSF CAREER award in 2007, theIEEE Circuits and Systems for Video Technology (CSVT) Transactions BestPaper Award for Year 2001, and the Best Paper Awards in IEEE GLOBECOM2011 and International Conference on Quality of Service in HeterogeneousWired/Wireless Networks (QShine) 2006.

Currently, he serves as an Associate Editor for IEEE Transactions onCircuits and Systems for Video Technology, Journal of Visual Communi-cation and Image Representation, and International Journal of Ad Hoc andUbiquitous Computing. He is the founder of IEEE Transactions on NetworkScience and Engineering. He was the founding Editor-in-Chief of Journal ofAdvances in Multimedia between 2006 and 2008, and an Associate Editorfor IEEE Transactions on Wireless Communications and IEEE Transactionson Vehicular Technology between 2004 and 2007. He is also a guest-editorfor IEEE Journal on Selected Areas in Communications (JSAC), SpecialIssue on Cross-layer Optimized Wireless Multimedia Communications. He hasserved as Technical Program Committee (TPC) Chair for IEEE INFOCOM2012, and TPC chair for IEEE International Conference on Communications(ICC 2008), Signal Processing for Communications Symposium, and as amember of executive committee and/or technical program committee of over80 conferences. He has served as Chair for the Award Committee, andChair of Mobile and wireless multimedia Interest Group (MobIG), TechnicalCommittee on Multimedia Communications, IEEE Communications Society.He was a member of Multimedia Signal Processing Technical Committee,IEEE Signal Processing Society from Jan. 1, 2009 to Dec. 31, 2012. He isan IEEE Fellow.

multiple content dissemination in roadside-unit-aided vehicular opportunistic networks

Documents