Balancing Wireless Data Broadcasting and Information Hovering for Efficient Information...

66 IEEE TRANSACTIONS ON BROADCASTING, VOL. 58, NO. 1, MARCH 2012

Balancing Wireless Data Broadcastingand Information Hovering for Efficient

Information DisseminationChristos Liaskos, Graduate Student Member, IEEE, Andreas Xeros,

Georgios I. Papadimitriou, Senior Member, IEEE, Marios Lestas, Member, IEEE, andAndreas Pitsillides, Senior Member, IEEE

Abstract—Wireless data broadcasting is an efficient, bandwidthpreserving way of data dissemination. However, as the amountof data increases, the waiting time of the clients becomes un-acceptably high. The present paper proposes the combinationof optimal wireless broadcasting and information hovering, asan effective means of performance improvement in vehicularnetworks with locality of demand. While state-of-the-art worksexploit user collaboration only as a means of wireless coverageextension, the proposed scheme proposes parallel disseminationthrough broadcasting and user networking over the whole studiedarea. Optimal, periodic broadcast scheduling is adopted at thehighest tier for data dissemination. At the lowest tier, users canexploit information hovering around selected anchoring points, inorder to retrieve data faster than their next scheduled broadcast.The issue of sharing the dissemination load optimally between thebroadcasting and the hovering subsystems is mapped to the classicpull-push balancing problem. Through analysis, the long-standing“optimal cut-off point” balancing method is shown to be subop-timal, and a new method is proposed which achieves lower clientserving time in any of the cases. Simulation results in realisticVANETs show that the proposed dissemination scheme surpassesstate-of-the-art works in terms of efficiency and client satisfaction.

Index Terms—Information hovering, vehicular networks, wire-less broadcast scheduling.

I. INTRODUCTION

U BIQUITOUS inter-networking has enabled access to anever-growing amount of data, by a steadily increasing

number of clients. As the client-base expands, similarities inuser preferences become more noticeable. Data broadcastingoffers a comprehensive way of exploiting this fact, enablingthe realization of efficient, bandwidth preserving disseminationschemes. Broadcast scheduling, the process of serializing the

Manuscript received April 01, 2011; revised July 11, 2011; accepted July 12,2011. Date of publication September 06, 2011; date of current version February23, 2012. This work was supported by the European Union (European SocialFund–ESF) and Greek national funds through the Operational Program “Edu-cation and Lifelong Learning” of the National Strategic Reference Framework(NSRF) - Research Funding Program: Heracleitus II. Investing in knowledgesociety through the European Social Fund.

C. Liaskos and G. I. Papadimitriou are with the Department of In-formatics, Aristotle University, 54124, Thessaloniki, Greece (e-mail:[email protected]; [email protected]).

A. Xeros, M. Lestas, and A. Pitsillides are with the Department of ComputerScience, Cyprus University, Nicosia, Cyprus (e-mail: [email protected];[email protected]; [email protected]).

Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/TBC.2011.2163449

data delivery in an optimal way, constitutes a well-founded fieldwith every-day applications, such as television, radio and tele-text scheduling. While broadcasting favors wireless transmis-sion, it has been incorporated successfully in wired scenariosas well. Well known applications include broadcasting for dataavailability advertisement and cache scheduling in named con-tent networking [1].

Wireless data broadcast scheduling typically assumes a singlefrequency or cellular network, which covers a densely popu-lated area. The clients therein are assumed to be interested ina common set of discrete data items. Each item is associatedwith a request probability and a size measured in bytes. Whileeach item’s size is static and given, its request probability varieswith time. In order to approximate the probability distribution ofthe items, the clients are usually required to provide some light-weight feedback regarding their preferences, either explicitly orindirectly. In [2] for example, the clients are required to emita single pulse upon reception of a wanted item. The aggregatereceived power level is then mapped to the probability distri-bution by a learning automaton [3]. Once the item probabilitiesbecome known, a central authority creates a broadcast schedulewhich optimizes a given criterion. Examples are the minimiza-tion of the mean waiting time in [4] and of the mean incurredimpatience in [5]. The schedule is transmitted over the coveredarea, and the clients retrieve needed data items in a streamingfashion.

While wireless broadcasting offers perfect scalability in termsof served clients, its performance does not scale as well withthe increase of the amount of data. Furthermore, broadcast sys-tems with large spatial coverage are typically assigned narrowbands in low frequencies. According to [6] for example, channel2 for national TV broadcast in the U.S. is assigned a 6 MHzband starting at 54 MHz, while the amateur zone at 2.4 GHz isassigned a band of 17 MHz. In addition, the achievable meanwaiting time in a broadcasting system is known to have an an-alytical lower bound [4]. The emerging challenge can be ex-pressed as follows:

Is it possible to improve the performance of a wirelessbroadcast system via architectural modifications or clientnetworking?

Architectural modifications include the deployment of mul-tiple, spatially distributed broadcast systems, each one handlinga subset of the original data [7]. In a similar fashion, data subsetscan be assigned to different frequency channels, as in [8]. Clientnetworking has been proposed as another alternative. In this

0018-9316/$26.00 © 2011 IEEE

LIASKOS et al.: BALANCING WIRELESS DATA BROADCASTING AND INFORMATION HOVERING 67

case data items are retrieved via collaborative caching and P2Pnetworking [9]–[11]. Architectural modifications and client net-working typically imply backbone and client device upgrades,extra bandwidth or strict locality of demand. Recently however,Information Hovering [12] was proposed as a lightweight col-laborative networking of end-users in Mobile Ad Hoc Networks(MANETs). The information remains in the vicinity of an an-choring geographical location, e.g. by hopping probabilisticallyamong clients [13]. Users that enter or are present in this loca-tion have rapid access to the hovered data. Due to its simplicity,information hovering constitutes a viable, cost effective alterna-tive over classic P2P networking.

The present work proposes a novel collaborative schemebetween centralized wireless broadcasting and user networking(information hovering). Previous state-of-the-art works havetreated user networking simply as a means of extending thecoverage of wireless access points [11]. In [9] this conceptis repeated, and network coding at symbol level is employedfor boosting the throughput of the users’ network. In contrast,the present proposes parallel data dissemination through cen-tralized, optimal broadcasting and through user networking(information hovering) at the same time. Thus, a two tier archi-tecture is formed: at tier 1 (broadcasting), data is disseminatedthrough optimal periodic scheduling. The scheduling considerstime-variant data content and user preferences. At tier 2 datawith locality of demand are hovered around time-variant an-choring points. A user inside the hovering zone can then retrievethe needed information before its next scheduled broadcast(tier 1). Data dissemination load can be freely shared betweenthe centralized broadcasting and the hovering, following apull-push balancing model. As an additional contribution,the classic balancing method of “optimal cut-off point” [14]is shown to produce suboptimal results, and the analyticallyoptimal alternative is presented.

The remainder of this paper is organized as follows: Section IIpresents the related work on broadcast scheduling and infor-mation hovering. Section III introduces the proposed collab-orative scheme. Validation through simulation takes place inSection IV, and the conclusion is given in Section V.

II. RELATED WORK

A. Information Hovering

The term Information Hovering was outlined qualitatively in[15], and was formally defined later, in [16]. A more detaileddescription is provided in [17] with a brief description of per-formance metrics of the concept. The term of Information Hov-ering involves decoupling of the hovering information from itshost and promotes coupling it directly with a specific geograph-ical location which is called the anchor location. In this re-gard, the hovering information stays “attached” to a specific ge-ographical area (called the anchor area). The information hoversfrom one mobile device to another, in a quest to remain withina specific vicinity and avail itself to users currently present orentering its anchoring geographical location. The InformationHovering concept can be used in a variety of applications inMobile Ad hoc Networks with characteristic usage examplesprovided in [12].

A relevant concept in the networks literature is geocasting,which addresses the problem of information dissemination ingeographical target regions. protocols which have appeared inthe literature [18] are intended to forward the relevant messagein the target region and distribute the message once, to all de-vices residing in the region. However, this one time message de-livery is not relevant in the ’information hovering’ paradigm, asthe information needs to remain in the target area for a specificamount of time. This concept has appeared in the literature asabiding or time-stable geocasting [19]. In [19] authors presentthree reasonable approaches for abiding geocast in ad-hoc net-works. However, none of the three approaches cope with lownode density which may result in periods with zero nodes in thegeocast area.

Flooding based schemes, such as epidemic routing [20] canbe applied to the entire network to effectively serve such a par-adigm, achieving high percentage of nodes in the hovering areareceiving the relevant message (high reachability). However,this is done at the expense of a large number of redundant mes-sages which strain the communication channel and lead to ex-tensive contention and large latency of message delivery.

In order to make efficient use of the available resources andsubstantially reduce the number of exchanged messages, onemay choose to apply epidemic routing merely in the hoveringarea. However, this approach may also reduce the achievedreachability in cases of low traffic density. Low traffic den-sity in the hovering area may cause the network to becomeintermittently connected. This implies that sections of thenetwork within the hovering area which are partitioned fromthe information sources may never receive the messages, thusdecreasing the achieved reachability. This demonstrates that thedesign of information hovering protocols is highly challengedby the intermittently connected nature of the network in casesof low traffic densities. One approach to address this challengeis to allow controlled exchange of messages outside the hov-ering area. The reasoning behind this approach is that informedvehicles outside the hovering area can serve as informationbridges towards partitioned uninformed areas withing the hov-ering site, thus increasing message reachability. In addition,since ’epidemic routing’ outside the hovering area is avoided,the number of exchanged messages is reduced. The authorsin [12] follow this reasoning to allow epidemic disseminationof messages in an extended area beyond the hovering area.They demonstrate through simulations that such an approachcan increase the recorded reachability in cases of low trafficdensity.

B. Push-Based Broadcast Scheduling

Push-based scheduling refers to the procedure of serializingdata items for broadcasting, based on their request probabilitiesonly. On the other hand, on-demand scheduling [21] operateson explicit client queries, as well as in shorter time scales. Thepresent work studies push-based scheduling.

Research on wireless broadcast systems initially focused onthe minimization of the clients’ mean waiting time in the contextof Teletext systems. This problem was solved by authors in [22],under the assumption of equally-sized data items. The problemwas revisited in [4], studying items with small variation in theirsizes. It was proven that the mean waiting time depends on data


item attributes (i.e. item request probabilities and sizes), and noton the number of clients. A lower bound for the mean waitingtime was given, as well as scheduling algorithms that achievedit at the expense of increased complexity ( , beingthe number of data items and the number of scheduledbroadcasts). These algorithms worked for nearly-equally sizeditems only. Authors in [23] presented an analysis-derived peri-odic scheduler that achieved the same performance withcomplexity, for any item sizes.

Heuristic, low complexity scheduling methods were intro-duced in [24] with the introduction of the Broadcast Disksmodel. According to it, items are grouped by popularity,forming virtual disks rotating around a common axis. Imagi-nary heads read and serialize data from the disks, producingthe final schedule. In [25], the authors applied clustering tech-niques for performing the data grouping. In [26] the groupingof items was analytically optimized. The analytical results wereexploited in [27] for producing minimal complexity schedulers.All these studies focused on the minimization of the clients’mean waiting time.

According to the analysis of [4], any increase in the amount ofdata to be broadcasted leads inevitably to higher mean waitingtimes. Authors in [8] propose the use of additional broadcastbandwidth, and enabling the users to listen to multiple channelsconcurrently. The number of channels and their respective band-width become extra static parameters, which are inserted in theformula of the lowest achievable mean waiting time. The orig-inal lower bound of [4] is essentially diminished by times,

being the number of channels. Another solution [7] proposesthe use of multiple antennas and corresponding scheduling au-thorities, each covering a sub-region of the total area. Underthe assumptions of strict locality of demand and uniform spa-tial distribution of clients, the lower bound of [4] is once againdiminished by times, being the number of scheduling au-thorities. Finally, another research direction focuses on P2P usernetworking, such as the collaborative caching scheme of [10].However, the users are required to support ad hoc network for-mation and communications protocols, implying equipment up-grades.

The present work differentiates from the mentioned researchdirections: regarding the central infrastructure, no changes orextra bandwidth are required. This also implies incorporationto multi-channel approaches. Possible locality of demand isexploited through information hovering, but it is not critical tothe functionality of the proposed scheme. Finally, the straight-forward, probabilistic nature of routing in information hovering[28] enables direct implementation via trivial, existing userfunctionality (e.g. Bluetooth or WiFi connectivity).

III. PROPOSED COLLABORATIVE SCHEME

The goal of the present study is the collaborative data dis-semination through broadcasting and information hovering. Themethodology consists of the following steps:

• Represent the collaboration as a push/pull balancingproblem. Then, define analytically the load of data to behandled by the hovering system, minimizing the meanwaiting time of the broadcasting system. It is assumedthat the time required to answer a query via pull is trivialcompared to the push alternative.

• The hovering system must be notified of the data it shouldhandle. Therefore, a signaling protocol between the broad-casting and the hovering systems is presented.

These steps are studied in Section III-B (Load Sharing) andSection III-C (Collaboration between subsystems) respectively.

Notation and Standard Assumptions: We assume a set ofdata items arbitrarily indexed by . Each item isassociated with its size (in ) and its request probability

. Therefore, it holds that . No assumptions aremade concerning the nature of a data item during the analysis.In accordance with the related work on scheduling [2], [4], [5],[7], [8], [21]–[27], an item is simply a piece of information thata client may acquire through a single query. During the sim-ulations of Section IV, items are specialized to carry informa-tion about traffic, parking availability and road conditions in aVANET covering an urban area. It is clarified that in push-basedbroadcast scheduling, the term “client query” does not implyposting a request to a server, but rather waiting for the broad-cast of a specific item.

According to [4], the minimum mean waiting time of a stand-alone, push-based broadcast system is given by:

(1)

Notice that does not depend on the number of clients [4].This bound is achievable in practice, as shown in [4], [23].

Finally, it is noted that broadcast scheduling and informationhovering are application-layer techniques, independent of theunderlying physical implementation parameters. This is a stan-dard assumption, in accordance with [2], [4], [5], [7], [8], [12],[21]–[28]. The reader is referred to [29] and [30] for exemplarystudies regarding physical aspects of digital content dissemina-tion via broadcasting and P2P networking respectively.

A. Load Sharing

Load sharing between hovering and broadcasting implies aprobabilistic approach. Push-based broadcasting assumes peritem request probabilities (see (1)). In addition, a hoveringscheme can disseminate a data item with a certain success prob-ability, which depends on the number of clients participating inthe hovering process [13]. A query posed by a mobile client willbe answered by listening to the broadcast schedule, unless thehovering supplies the needed item sooner. This configurationcalls for a probabilistic pull/push balancing model, which isillustrated in Fig. 1. The goal is to define the optimal percentageof queries that should be answered successfully by the hoveringscheme.

The optimality refers to minimizing the lower bound ex-pressed by (1), for a given strain on the hovering system.

The minimization of (1) expresses the fact that since certainitems may be available through external sources (i.e. hovering),their broadcast frequency can be lowered. Thus, more “air time”becomes available to other items. Notice that (1) regards item re-quest probabilities and sizes only. This means that the broadcastscheduling authority must alter the original probability distribu-tion by offloading items to the hovering system, in order forthe mean waiting time to be lowered.


Fig. 1. Illustration of the proposed load sharing between the broadcasting andthe hovering systems. In a total of � queries, � �� refer to item �. A percentage� � �� of these are addressed by the hovering system, while the remainingare answered via the broadcasting.

The proposed load sharing cannot occur regardless of theimposed strain on the hovering system. Indeed, one could re-quire that all clients hover all available data items with per-fect efficiency. This would effectively nullify (1) in theory, butwould imply that each mobile client had vast memory, pro-cessing power and available bandwidth. It is therefore logical tomake the load sharing tunable in terms of imposed strain. Sincea query for item receives a reply of size , the strain can beexpressed as the mean amount of pulled data :

(2)

where denotes the percentage of queries for itemthat are addressed by the hovering system. An expression in-volving only the rate of pull queries is also introduced:

(3)

The goal of the study can be defined as follows: given a desiredtotal strain , where should hovering occur and with what effi-ciency (i.e. define optimal values for , , )in order to minimize the mean waiting time of the clients (1)?

We assume that hovering occurs instantaneously, and inter-mediate transient phenomena are negligible. With these remarksin mind, the analysis is initiated by assuming two discrete states:

STATE 1: The starting state, in which the item requestprobabilities are unaltered: . This state cor-responds to total absence of hovering, which has not beeninitiated yet. In this phase the broadcast system assumesthe total data dissemination load.STATE 2: The ending state, in which:

(4)

In other words, this state represents the extreme case where hov-ering has assumed most of the dissemination load, handling alldata items but item . Notice that one could also consideredthe total absence of broadcasting as the ending state. However,this would entail that the central broadcasting infrastructure has

stopped functioning, for no apparent reason. In other words, thefinal load attributed to broadcasting can be minimal, but a nullvalue would yield inefficient use of resources and therefore badoverall design. However, the results of the analysis remain un-altered in any case.

We proceed to introduce a set of continuous variationalflows , , , which fulfill the conditions:

(5)

In this sense, expression (1) can also be rewritten as a functionof the flow control variable , by simply substituting with

. The same applies to (2) and (3) when substituting withand with . The notation , and cor-

responds to this expression. The goal is to define the flows ,which minimize for any given .

Lemma 1: The continuous flows , whichminimize , with regard to the chosen item of STATE 2,are linear functions of :

(6)

(7)

where ,

and . denotes the ordering of all itemsbut by ascending ratio, up to (but without) item .denotes the corresponding inclusive ordering.

Proof: See Appendix.Notice that the flows defined in Lemma 1 can be directly ex-

pressed as functions of only, discarding the redundant vari-able . This was expected, since was introduced only for facil-itating the analysis. For the remainder of this work, the expres-sion of flows will follow this convention.

Lemma 1 refers to flows with respect to the chosen item ofSTATE 2. For a given strain we examine all possible valuesof and keep the one which produces the lowestmean waiting time (1). This optimal value, , corresponds tooptimal flows, which will be denoted as . Concerningthe optimal rate of queries that should be an-swered via pulling (see Fig. 1), it holds that:

(8)

Equation (8) essentially states that for each item , the queriesanswered via the hovering are equal to the total number ofqueries for item , minus the ones answered via the centralizedbroadcasting. Finally, solving (8) for yields:

(9)

which expresses the required hovering efficiency for data items(and, assuming locality of demand, at site

) which minimizes the client mean serving time, givena desired strain .


As a closing remark, lemma 1 states that the normalized strainaffects the quantities of (7) according to the or-

dering in a serial manner. Starting from STATE 1, the initialincrease of will affect the first (e.g. -indexed) item of theordering until , then the second item, and so on.From (2) and (3) it is easily seen that for any item that is thesole one currently affected in the ordering, it holds:

(10)

Equation (10) states that the relation between and isstrictly increasing, and therefore 1 1. Any value of can thusbe converted to one unique, corresponding value and vice-versa. The conversion is straightforward, since the quantities ,

of (2) and (3) are static and known, and are calculatedthrough (9). Therefore, the real strain, , and its normalizedversion, , are interchangeable in all equations. For the sakeof proper presentation we will assume the format, since it isalways bounded in [0,1] regardless of the data item sizes.

B. Collaboration Between Subsystems

The preceding analysis considered the generic case of hybrid(pull/push) data dissemination. The conclusions of Lemma 1and (9) are adapted to the case of collaboration between broad-casting and hovering as follows.

When initiating operation, the broadcast scheduling authorityis unaware of the per item preferences. The request probabilitydistribution of the items is approximated through an adaptationprocedure, which has been analysed in detail in [2] and [31].As discussed in [2], upon reception of a desired item, a clientemits a single pulse designating content approval. The learningautomaton[3] of [2] processes the aggregate feedback from allclients in real-time and re-approximates the popularity distribu-tion of the data items. There are no distinct cycles of adapta-tion and subsequent rescheduling. The employed scheme of [2]schedules the broadcast of the items based on their attributes(size , approximate popularity ) and the time interval passedsince their last broadcast . Before each broadcast the fol-lowing array is constructed:

(11)

The item that produces the maximum entry is broad-casted, and the current timeslot is marked as its last broadcasttime moment. The values are renewed by the learning au-tomaton upon the reception of new feedback. Notice that nei-ther the number of items nor their content is static. As in [2],when new items are added, their corresponding values areset to the current time slot, while their initially assumed requestprobability is set to the median of the existing probability dis-tribution. Normalization is applied to ensure that the sum of theapproximate probability distribution is equal to one. The sched-uling authority may also update the content referring to the geo-graphical site , changing its size , or remove an item entirely.The procedure is repeated, and the automaton converges to thereal probability distribution.

The number of users and the influence of its variations on theconvergence of the adaptation has been studied in [4]. It wasproven that for large user sets, all clients can be collectively

handled as a simple Gaussian process. This outcome, whichis expected due to the central limit theorem, typically holds inVANETs, since they typically contain a large number of users.Consequently, [4] showed that their actual number is irrelevant,as long as the popularity of each data item becomes known. Thelatter is the case of the present study.

The scheduling authority initiates or terminates informationhovering through broadcasted signals. Upon successful conver-gence of the adaptation process the scheduling authority pro-ceeds to calculate the and values via (9) andLemma 1, for a given strain . The value expressesthe efficiency with which the hovering subsystem should handlequeries for item . As shown in [13] however, this efficiency isa function of the percentage of clients contributing in thehovering process:

(12)

As in [13], the form of the function depends on the cov-ered area, the hovering range, the employed routing protocoland the mean density of vehicles. Since these parameters canbe assumed to be known, the scheduling authority can calculatethe optimal percentage of clients, that should contributeto the hovering of item , in order to achieve an efficiency of

. The values , are then broadcastedin a single package to all clients. For each entry , a client flipsa weighted coin with probability of contributing to thehovering of item . The successful clients initiate the hoveringprocess. Hovering termination occurs with the broadcast of aspecific message from the central authority, at any appropriatesituation (e.g. item replenishment or invalidation). Finally, thescheduling authority re-adapts the broadcast schedule to complywith the probabilities.

Notice that precise knowledge of (12) is not crucial to thesystem operation. Hovering can easily achieve 100% efficiency[13] in almost any case. If (12) is imprecise, the learning au-tomaton will converge to a point that is either below or abovethe expected value of the mean waiting time (1). The schedulingauthority can then simply increase or decrease the strain in orderto get the desired result.

IV. SIMULATION

The procedure described in Section III is validated throughsimulation in a VANET environment. The proposed scheme iscompared to the CodeON [9] and CodeTorrent [32] approaches,in terms of mean client waiting time and user satisfaction [11].The transient phenomena of P2P network establishment andefficiency, as well as the adaptive capabilities of the feedbackmechanism are addressed. All simulations were performedusing the VISSIM software package [33]. The implementationof the network coding technique of [9] and [32] was based onthe freely available source code of [34].

A. Simulation Setup and Results

The simulated VANET refers to an area of Bellevue and Red-mond in Washington, depicted in Fig. 2. The dimensions of thearea are 4.5 2.5 Km. At any given moment, approximately3,000 vehicles (clients) roam the arterial streets and the freeway,their exact number being time-variant. The speed of each vehicle


Fig. 2. Simulated VANET area of Bellevue and Redmond, Washington, as depicted in Google Maps (Coordinates: 47.626413,�122.149944). A number of events(dots) regarding traffic, road condition, and parking availability are selected. Information on these events is disseminated via broadcasting (global coverage) andhovering around the corresponding locations.

varies indicatively from 20 Km/h to 80 Km/h. The relation be-tween the and parameters is obtained statistically from[13] as:

(13)

In all tests, a central broadcasting system is assumed to coverthe studied area, enabling every vehicle to retrieve data fromthe broadcast stream with a guaranteed rate of 25 KBps. Noticethat typical choices for physical implementation of the broad-cast system (3G, DVB/H) supply bitrates well beyond 2–5 Mbps[35]. However, we choose this particularly low value to makeup for the global coverage and bitrate guarantee assumptions.Finally, the probability of successful retrieval of data from thebroadcast stream and correct client feedback reception is set to0.95.

Concerning the physical attributes of the P2P networking, weassume that each vehicle has a varying transmission range of180 40 m. The range is randomized prior to any transmission.Inside this range, and assuming MAC-layer clearance, the prob-ability of successful data reception is 0.95. In all cases we adoptthe MAC protocol of the 802.11 specification. Carrier sensingsimulation is based on the two-ray ground propagation model.The reader is directed to the NS-2 [36] implementation of the802.11 specification for the definition of additional attributes.The Rx/Tx rate of each client is 11 Mbps.

Regarding the nature of the data items and their geographicalsites, we assume random anchoring points on the terrain ofFig. 2. P2P networking occurs inside a radius of 500 m aroundeach point. Each site (arbitrarily indexed) defined inthis way, is associated with a piece of information with size

. This emulates a typical web page which containstweets on all local events of interest. Each site , and thereforeeach associated data item, has a time-variant popularity among

the clients, , which is defined as the portion of thetotal clients interested in site . For simplicity, each client mayhave only one request pending at any given time. Therefore,

.In order to emulate locality of demand conditions, the data

requests are assigned to the clients in a way that followsa Gaussian distribution around the anchoring point. When

, clients on the terrain are interested in site , regardlessof their current position. When , 99.7% of the interestedclients (three standard deviations) are already inside the P2Psite. Intermediate values are set proportionally. In order togenerate requests and assign them to clients, we set the sitepopularity distribution according to the ZIPF distribution[37]. The skewness of this distribution is regulated through asingle parameter, . When , the distributionis flat and . Therefore,corresponds to high locality of demand. As increases, thepopularity of several sites becomes higher, and the locality oftheir demand drops. Consequently, the manipulation of isused for testing various cases of locality of demand. Finally,each simulation typically considers 600,000 client queriesgenerated according to the distribution in use.

The popularity distribution is initially not known to thebroadcast scheduling authority. In the experiment of Fig. 3 weexamine the efficiency of the employed adaptation technique. Inrandom time moments designated by the thin vertical lines, thedata set in question changes in cardinality , content popu-larity and content size (mean item size given in Kbytes).We assume the worst case, with the data set being completelyrenewed each time. Likewise, the server has no initial, roughinfo on the form of the distribution, and assumes a flat onein any case. The dotted line expresses the time-variant lowerbound of mean waiting time (1), should the distribution wasknown. The solid line represents the time-variant product of theautomaton, based on the current approximation. Even in this


Fig. 3. Performance of the learning automaton-based adaptation technique,which is used for approximating the time-variant popularity of a dynamicallyformed data set. Random changes are set to occur in the cardinality of the studieddata set and the popularity and size of the contained items. Even in the worst-case scenario (no content correlation between successive data set updates), theautomaton adapts optimally in 50–250 sec.

Fig. 4. Plots of the requested strain,� , versus the achieved mean waiting time� and the actual imposed strain. The plots reflect the case of � � ��, i.e., aslightly skewed ZIPF distribution of the items’ popularity. The mean waitingtime is reduced by 10 times, when compared to the case of stand-alone broad-casting �� . While the reduction follows the general theoretical form,some degree of divergence is present. Results are identical for other values of �.

worst case approach, the automaton proves its efficiency, as in[2], [3]. The convergence time for the current setup is shown tobe 50–250 sec. Note that this effort refers to measuring the popu-larity of geographical sites in a city, which are known to changeinsignificantly over time (months or years are required for theconstruction of new main roads, malls, parking points and othersignificant points of interest). Therefore, for the remainder ofthe simulations it will be assumed that the 1–5 minutes requiredfor convergence have passed, and the study will focus on thebehavior of the compared approaches in the steady state. Thenumber of sites/data items will be kept constant at .

In Fig. 4 we examine the validity of the preceding analysisregarding the pull-push load balancing optimization. Startingwith zero client strain (no hovering initiated yet) the broad-casting authority begins to signal the clients to initiate thehovering process, gradually increasing their load. The achievedmean waiting time in the steady state is then logged and plottedversus the corresponding strain. Three observations can bemade: firstly, the experimental results follow the general formof theoretical expectations. Secondly, the average waiting timeis never nullified, even for maximum strain. This is expected,

Fig. 5. Time required for the initialization of the compared P2P schemes in anarea of interest. Hovering requires a trivial amount of time before being estab-lished. CodeOn and CodeTorrent may be viable, requiring approximately 5 secfor establishment in the best case.

since the analysis assumes ability to handle the total load.However, the hovering process has a 500 m of activity radius.Consequently, there will always exist clients that require anitem but are outside the corresponding P2P zone. Thirdly, therequested strain is not upheld. The finally incurred strain is lessthan the expectation. This phenomenon is attributed to the im-precision of (13) which is proven to be optimistic: a requestedstrain value causes less real strain. This issue also highlightsthe fact that the performance of the system does not dependon the precision of the formula (13). The central schedulingauthority is aware of both the theoretical curves (through directdefinition) and the experimental curves (through the clientfeedback mechanism) of Fig. 4. Therefore, the divergence canbe directly quantified, and the central authority may requestless or more strain than the expectation of (13), achieving thedesired result. In other words, (13) simply serves as an initialguide.

The proposed approach offers tun-ability in terms of imposedclient strain but CodeOn and CodeTorrent do not. Therefore, itis required to define a standard strain for fair comparison. Thus,for the remainder of the simulations, 95 out of 100 items (byascending ratio) will be deterministically assigned for fulldissemination via P2P networking, while the remaining 5 itemswill be handled explicitly via the central broadcasting.

The proper initiation of the P2P networking is of crucialimportance, since the highly dynamic VANET environmentmay hinder the employed schemes from operating efficiently.In Fig. 5 we consider 210 vehicles (typically present in a site)in a rectangular formation. We assume immobility, and that arandomly selected client initiates the P2P networking, seekingto disseminate a single item to the whole client set. The Hov-ering, CodeON and CodeTorrent P2P schemes are comparedin terms of time required for the task. The hovering technique,relying simply on probabilistic flooding, achieves minimaltimes. The number of seeding users increases exponentially,yielding overall completion times of even less than 1 sec in theemployed setup. The CodeOn and CodeTorrent approaches onthe other hand, fall back due to the requirement for explicitpacket routing. The tested approaches include blind forwarding(BF) and selective forwarding (SF) [34]. Network coding(NC) is used for maximizing the P2P network throughput, atthe expense of added computational load on the clients. Bothapproaches present viable results in the case of selective for-warding. It is notable however, that the use of network coding


Fig. 6. Mean waiting time achieved by all compared approaches in a full sim-ulation. Steady state is assumed, with the initialization and adaptation phaseshaving been completed. The hovering-broadcasting combination yields the bestresults in the tested cases of locality of demand.

Fig. 7. Average time required to obtain a data item, if the client is inside thecorresponding area of interest. Hovering achieves minimal time, with CodeOnand CodeTorrent following. The results also imply that network coding may notbe more beneficial than an efficient routing algorithm.

does not provide a significant boost to the overall performance.This is a concern due to the high computational cost, especiallyfor CodeON, which employs symbol-layer network coding,producing large amounts of data segments to be processed persingle packet.

Based on the requirements for initialization, we allow fora 30 sec interval to pass before posing client queries. In thefull-scale simulation of Fig. 6 the achieved mean waiting timeover 600,000 client queries is presented. The locality in the de-mands of the clients is gradually decreased, and the performanceof each opposing scheme is logged. The broadcasting-hoveringcombination achieves the best results in all cases. All schemesare affected by the loss of demand locality, since a user may befar beyond the reach of the P2P networking around the area ofinterest. CodeON performs better than CodeTorrent, since it en-forces smaller data segment sizes, limiting the impact of packetlosses. However, it is once shown that network coding is not ac-tually more beneficial than an efficient routing algorithm in ahighly dynamic VANET. This fact is more apparent in Fig. 7,where we measure the mean waiting time for clients inside theirarea of interest. Hovering-achieved times are tantamount to datapulling, as assumed. Network coding however, produces ben-efits only compared to blind packet forwarding. Selective for-warding is not positively affected by the added data segmenta-tion overhead imposed by the network coding technique.

The results of Fig. 6 can also be expressed through the metricof user satisfaction [11]. The satisfaction metrics are simple

Fig. 8. Alternative representation of the performance of the opposing schemesin terms of user satisfaction (try-best metric of [11]). The hovering-broadcastingcombination achieves the best results, for all items � � � � �� chosen fordissemination through P2P networking.

transformations which receive the mean waiting time as an inputand magnify the importance of certain ranges of values. Fig. 8refers to the try-best metric, which is valid for items that shouldbe disseminated preferably within 5–10 sec. It is defined as

, being the waiting time and the user satisfac-tion. We examine the average satisfaction for each data item(and therefore each site) separately. As expected from the pre-ceding results, the proposed hovering-broadcasting combina-tion achieves higher user satisfaction than any version of theopposing schemes. This conclusion holds for all data items

that were chosen to be disseminated through P2Pschemes.

V. CONCLUSIONS AND FUTURE WORK

A novel collaborative scheme between data broadcastingand information hovering has been proposed. The clients ina realistic VANET had their mean waiting times lowered,provided that they participated in the hovering of informationin the vicinity of a given area. This result demonstrates thatlightweight user collaboration is sufficient for significant im-provement of the system’s overall performance. Future stepstarget the comparison with high-end solutions (such as broad-cast network redeployment and exquisite P2P user networking)in terms of achieved cost-to-performance improvement ratio.

APPENDIX

Proof of Lemma 1

Assume an appropriately small interval for whichit holds that

(14)For brevity, the notation will be employed.Since , it is derived from (14) that

(15)


Furthermore, assume that in this first infinitesimal step, the chosen flow (referring to the sole data item in

STATE 2) increases while the rest decrease:

and (16)

Up to this point the flow control parameter was an auxiliaryvariable bounded in an arbitrary range and deprived ofany physical meaning. We will now proceed to assign a spe-cific physical context to variable . From this point and on,will express the normalized strain imposed on the hovering sub-system and be a synonym to the metric. Therefore,

expresses the increase of the total rate of pulled items. Con-sequently:

(17)

In other words, the increase in the rate of pulled items is equal(by absolute value) to the decrease in the rate of pushed items.From (15) and (17) we deduce that:

(18)

Next, we calculate , (1) with regard to (14):

(19)

The goal is to make reduce as fast as possible, i.e. to find the, values which minimize (19). Consider the ordering

of all items by ascending ratio, an arbitrary item in theordering for which it holds:

(20)

Let , , . The derivative (19) then yieldsthe value:

(21)

We will prove through reduction that , for any ,values abiding by the statement (16):

(22)

(23)

Fig. 9. Brief comparison between the flow optimization scheme ofSection III-A, the optimal cut-off approach of [14] and brute force re-sults. The proposed scheme yields more efficient exploitation of the imposedstrain than the cut-off approach, and nearly coincides with the brute forceresults. The cut-off approach operates up to � � ��.

which holds always, because of (16) and (20). Therefore,for any , , QED.

Notice that the best possible values ( and ,, , ) do not present any dependence on , reminding

that expresses the normalized hovering strain. Therefore, weproceed to the next infinitesimal step and repeat the process(continuity assumption). Observe that the place of in theordering remains unchanged. Repeat the process until ,i.e., until . At that point, consider that the original setof items has been reduced to , i.e. the now improbableitem has been removed. Name this state as STATE 1 and repeatthe proof recursively until only item remains. This concludesthe proof of Lemma 1.

Further Validation of the Proposed Load-Sharing Scheme

In this paragraph we perform numerical calculations for val-idating the complete load sharing procedure of Section III-B,including Lemma 1.

In Fig. 9, the proposed procedure is briefly compared to re-sults extracted via brute force, and the optimal cut-off pointapproach [14]. The latter states that items with large ratioshould be assigned deterministically to the pull system, until thetargeted strain has been reached. The setup consists ofitems (a restriction imposed by the brute force procedure), with

following the ZIPF p.d.f [37] with skewness parameter, and (random). The brute force

searches all possible combinations of, for every . The optimal cut-off approach,

being an empirical rule, does not always achieve optimal ex-ploitation of the strain. The proposed scheme achieves near-op-timal results in any case. Furthermore, better tun-ability in termsof maximum exploitable strain is achieved. Future steps targetthe study of non-continuous optimal flows.

REFERENCES

[1] V. Jacobson, D. K. Smetters, J. D. Thornton, M. F. Plass, N. H. Briggs,and R. L. Braynard, “Networking named content,” in Proc. 5th ACMConf. Emerging Netw. Exp. Technol. (CoNEXT), Rome, Italy, Dec.2009, pp. 1–12.

[2] P. Nicopolitidis, G. Papadimitriou, and A. Pomportsis, “Continuousflow wireless data broadcasting for high-speed environments,” IEEETrans. Broadcast., vol. 55, no. 2, pp. 260–269, 2009.


[3] G. Papadimitriou, M. Sklira, and A. Pomportsis, “A new class of e-op-timal learning automata,” IEEE Trans. Syst., Man, Cybern., Part B, vol.34, no. 1, pp. 246–254, 2004.

[4] N. H. Vaidya and S. Hameed, “Scheduling data broadcast in asym-metric communication environments,” Wireless Netw., vol. 5, no. 3, pp.171–182, 1999.

[5] M. Raissi-Dehkordi and J. S. Baras, “Broadcast scheduling fortime-constrained information delivery,” in Proc. IEEE GlobalTelecommun. Conf. (GLOBECOM), Washington, DC, Nov. 2007, pp.5298–5303.

[6] “U.S. frequency allocation chart,” U.S. Department of Commerce,2003 [Online]. Available: http://www.ntia.doc.gov/osmhome/al-lochrt.html

[7] P. Nicopolitidis, G. Papadimitriou, and A. Pomportsis, “Exploiting lo-cality of demand to improve the performance of wireless data broad-casting,” IEEE Trans. Veh. Technol., vol. 55, no. 4, pp. 1347–1361,2006.

[8] B. Zheng, X. Wu, X. Jin, and D. L. Lee, “TOSA: A near-optimal sched-uling algorithm for multi-channel data broadcast,” in Proc. 6th Int.Conf. Mobile Data Manage. (MDM), Ayia Napa, Cyprus, May 2005,pp. 29–37.

[9] M. Li, Z. Yang, and W. Lou, “CodeOn: Cooperative popular con-tent distribution for vehicular networks using symbol level networkcoding,” IEEE J. Sel. Areas Commun., vol. 29, no. 1, pp. 223–235,2011.

[10] Y. Ma and A. Jamalipour, “A cooperative cache-based content deliveryframework for intermittently connected mobile ad hoc networks,” IEEETrans. Wireless Commun., vol. 9, no. 1, pp. 366–373, 2010.

[11] L. Zhou, Y. Zhang, K. Song, W. Jing, and A. V. Vasilakos, “Distributedmedia services in P2P-based vehicular networks,” IEEE Trans. Veh.Technol., vol. 60, no. 2, pp. 692–703, 2011.

[12] D. Konstantas and A. Villalba, “Hovering information: A paradigmfor sharing location-bound information,” in Proc. Int. Conf. Intell.Syst. Comput.: Theory Appl. (ISYC), Ayia Napa, Cyprus, Jul. 2006,pp. 116–122.

[13] A. Xeros, M. Andreou, A. Pitsillides, and M. Lestas, “Information hov-ering in vehicular AdHoc networks,” in Proc. IEEE Workshop Intell.Veh. Infrastructures (NiVi), Co-located with IEEE Global Commun.Conf., (GLOBECOM), Honolulu, Hawaii, Nov. 2009, pp. 1–6.

[14] Y. Guo, S. K. Das, and C. M. Pinotti, “A new hybrid broadcast sched-uling algorithm for asymmetric communication systems,” in Proc. 4thACM Int. Workshop Model., Anal. Simulation Wireless Mobile Syst.(MSWIM), Co-Located With MOBICOM, Rome, Italy, Jul. 2001, pp.123–130.

[15] A. Villalba and D. Konstantas, “Towards hovering information,” in Lec-ture Notes in Computer Science. Heidelberg, Berlin: Springer, 2006,pp. 250–254.

[16] A. Di Villalba, G. Marzo Serugendo, and D. Konstantas, “Hovering in-formation—Self-organising information that finds its own storage,” inProc. IEEE Int. Conf. Sensor Netw., Ubiquitous Trustworthy Comput.(SUTC ), Newport Beach, CA, Jul. 2008, pp. 193–200.

[17] G. Di Marzo Serugendo, A. Villalba Castro, and D. Konstantas, “De-pendable requirements for hovering information,” in Suppl. Proc. 37thAnnu. IEEE/IFIP Int. Conf. Dependable Syst. Netw. (DSN), Edinburgh,Scotland, Jul. 2007.

[18] S. Hermann, C. Michl, and A. Wolisz, “Time-stable geocast in intermit-tently connected IEEE 802.11 MANETs,” in Proc. Veh. Technol. Conf.(VTC), Dublin, Ireland, Apr. 2007, pp. 1922–1926.

[19] C. Maihofer, T. Leinmuller, and E. Schoch, “Abiding geocast: Time-stable geocast for ad hoc networks,” in Proc. 2nd ACM Int. WorkshopVeh. Ad hoc Netw.(VANET ), Co-located with MOBICOM, Cologne,Germany, Sep. 2005, pp. 20–29.

[20] A. Vahdat and D. Becker, “Epidemic routing for partially connected adhoc networks,” Duke University, Tech. Rep. CS-200006, 2000.

[21] J. Xu, X. Tang, and W.-C. Lee, “Time-critical on-demand data broad-cast: Algorithms, analysis, and performance evaluation,” IEEE Trans.Parallel Distrib. Syst., vol. 17, no. 1, pp. 3–14, 2006.

[22] J. Gecsei, The Architecture of Videotex Systems. Englewood Cliffs,NJ: Prentice-Hall, 1983.

[23] C. Liaskos, S. Petridou, and G. Papadimitriou, “Towards realizable,low-cost broadcast systems for dynamic environments,” IEEE/ACMTrans. Netw., vol. 19, no. 2, pp. 383–392, 2011.

[24] S. Acharya, R. Alonso, M. Franklin, and S. Zdonik, “Broadcast disks,”ACM SIGMOD Rec., vol. 24, no. 2, pp. 199–210, 1995.

[25] C. Liaskos, S. Petridou, G. Papadimitriou, P. Nicopolitidis, M.Obaidat, and A. Pomportsis, “Clustering-driven wireless data broad-casting,” IEEE Wireless Commun. Mag., vol. 16, pp. 80–87, 2009.

[26] C. Liaskos, S. Petridou, G. Papadimitriou, P. Nicopolitidis, and A.Pomportsis, “On the analytical performance optimization of wirelessdata broadcasting,” IEEE Trans. Veh. Technol., vol. 59, pp. 884–895,2010.

[27] C. Liaskos, S. Petridou, and G. Papadimitriou, “Cost-aware wirelessdata broadcasting,” IEEE Trans. Broadcast., vol. 56, no. 1, pp. 66–76,Mar. 2010.

[28] A. Xeros, M. Andreou, A. Pitsillides, and M. Lestas, “Adaptiveprobabilistic flooding for information hovering in VANETs,” inProc. IEEE Veh. Netw. Conf. (VNC), Jersey City, NJ, Dec. 2010, pp.121–127.

[29] Y. Wu, E. Pliszka, B. Caron, P. Bouchard, and G. Chouinard, “Com-parison of terrestrial DTV transmission systems: The ATSC 8-VSB, theDVB-T COFDM, and the ISDB-T BST-OFDM,” IEEE Trans. Broad-cast., vol. 46, no. 2, pp. 101–113, Jun. 2000.

[30] M.-F. Leung and S.-H. G. Chan, “Broadcast-based peer-to-peer collab-orative video streaming among mobiles,” IEEE Trans. Broadcast., vol.53, no. 1, pp. 350–361, Mar. 2007.

[31] C. Liaskos and G. Papadimitriou, “Ultra lightweight adaptation pro-cesses for scheduling servers in push-based systems,” in Proc. 17thIEEE Symp. Commun. Veh. Technol. (SCVT), Enschede, Netherlands,Oct. 2010, pp. 1–6.

[32] S.-H. Lee, U. Lee, K.-W. Lee, and M. Gerla, “Content distributionin VANETs using network coding: The effect of disk I/O and pro-cessing O/H,” in Proc. 5th Annu. IEEE Conf. Sensor, Mesh and AdHoc Commun. and Netw. (SECON), San Francisco, CA, Jul. 2008, pp.117–125.

[33] PTV Vision, The VISSIM Traffic Simulator 2011 [Online]. Available:http://www.vissim.de

[34] A. Hamra, C. Barakat, and T. Turletti, “Network coding for wirelessmesh networks: A case study,” in IEEE Int. Symp. World Wireless, Mo-bile Multimedia Netw. (WoWMoM), 2006, pp. 103–114.

[35] M. Kornfeld, “DVB-H—The emerging standard for mobile data com-munication,” in IEEE Int. Symp. Consum. Electron., 2004, 2004, pp.193–198, IEEE.

[36] The Network Simulator 2. California, USA: I. S. I. University of Cal-ifornia, 2011 [Online]. Available: http://www.isi.edu/nsnam/ns/

[37] L. Pietronero, E. Tosatti, V. Tosatti, and A. Vespignani, “Explainingthe uneven distribution of numbers in nature: The laws of Benford andZipf,” Physica A: Statistical Mechanics and its Appl., vol. 293, no. 1/2,pp. 297–304, 2001.

Christos Liaskos received the Diploma degree inelectrical and computer engineering in 2004 and theM.S. degree in medical informatics in 2008 from theAristotle University of Thessaloniki, Thessaloniki,Greece, where he is currently pursuing the Ph.D.degree in communication networks.

Andreas Xeros received the Bachelor degree fromthe Computer Engineering and Informatics Depart-ment from University of Patras in 2002. In 2004 hereceived his Master’s Degree in computer sciencewith specialization in multimedia and creative tech-nologies from the University of Southern California(fellowship). He is now a Ph.D. candidate at theUniversity of Cyprus. His research interests includeInformation Dissemination in Vehicular Ad HocNetworks (VANETs), Vehicular Traffic modelingand Graph Theory.


Georgios I. Papadimitriou (M’89–SM’02) re-ceived the Diploma and Ph.D. degrees in computerengineering from the University of Patras, Greece in1989 and 1994, respectively.

In 1997 he joined the faculty of the Departmentof Informatics, Aristotle University of Thessaloniki,Greece, where he is currently serving as an AssociateProfessor. His main research interests include opticalnetworks and wireless networks. Prof. Papadimitriouis Associate Editor of five IEEE journals. He is co-au-thor of three international books published by Wiley.

He is author or coauthor of 94 journal and 102 conference papers. He is a SeniorMember of the IEEE.

Marios Lestas (S’00–M’10) received the B.A.and M.Eng. degrees in electrical and informationengineering from the University of Cambridge,Cambridge, U.K., in 2000, and the Ph.D. degreein electrical engineering from the University ofSouthern California, Los Angeles, in 2005.

He is currently an Assistant Professor with Fred-erick University, Nicosia, Cyprus. His research inter-ests include application of nonlinear control theoryand optimization methods in computer networks.

Andreas Pitsillides (M’89–SM’05) received theB.Sc. (Honors) degree from the University ofManchester Institute of Science and Technology(UMIST), Manchester, U.K., in 1980, and the Ph.D.degree from Swinburne University of Technology,Melbourne, Australia, in 1993, both in electricalengineering.

He is a Professor with the Department of ComputerScience, University of Cyprus, Nicosia, Cyprus, andheads the Networks Research Laboratory (NetRL).He has particular interest in adapting tools from var-

ious fields of applied mathematics, such as adaptive nonlinear control theory,computational intelligence, and recently nature-inspired techniques, to addressproblems in communication networks.

Balancing Wireless Data Broadcasting and Information Hovering for Efficient Information...

Documents

Transcript of Balancing Wireless Data Broadcasting and Information Hovering for Efficient Information...