DAWN: A Density Adaptive Routing for Deadline …anrg.usc.edu/www/papers/DAWN_Fu.pdfQiao Fu et al.: DAWN: A Density Adaptive Routing for Deadline-Based Data Collection in Vehicular

TSINGHUA SCIENCE AND TECHNOLOGYISSNll1007-0214ll03/11llpp230-241Volume 18, Number 3, June 2013

DAWN: A Density Adaptive Routing for Deadline-Based DataCollection in Vehicular Delay Tolerant Networks

Qiao Fu, Bhaskar Krishnamachari, and Lin Zhang�

Abstract: Vehicular Delay Tolerant Networks (DTN) use moving vehicles to sample and relay sensory data for

urban areas, making it a promising low-cost solution for the urban sensing and infotainment applications. However,

routing in the DTN in real vehicle fleet is a great challenge due to uneven and fluctuant node density caused by

vehicle mobility patterns. Moreover, the high vehicle density in urban areas makes the wireless channel capacity

an impactful factor to network performance. In this paper, we propose a local capacity constrained density adaptive

routing algorithm for large scale vehicular DTN in urban areas which targets to increase the packet delivery ratio

within deadline, namely Density Adaptive routing With Node deadline awareness (DAWN). DAWN enables the

mobile nodes awareness of their neighbor density, to which the nodes’ transmission manners are adapted so as

to better utilize the limited capacity and increase the data delivery probability within delay constraint based only

on local information. Through simulations on Manhattan Grid Mobility Model and the real GPS traces of 4960 taxi

cabs for 30 days in the Beijing city, DAWN is demonstrated to outperform other classical DTN routing schemes

in performance of delivery ratio and coverage within delay constraint. These simulations suggest that DAWN is

practically useful for the vehicular DTN in urban areas.

Key words: delay tolerant networks; node density adaptive routing; deadline-based data collection; channel

capacity

1 Introduction

The Delay Tolerant Networks (DTN) are networkswhere contemporaneous end-to-end paths are unstableor unlikely. Routing in such networks is differentfrom those in wired or well-connected wirelessnetworks, because of its characters of extremelylong delay, frequent failure of link and opportunistic

�Qiao Fu and Lin Zhang are with the Department of ElectronicEngineering, Tsinghua University, Beijing 100084, China. E-mail: [email protected]; [email protected].�Bhaskar Krishnamachari is with the Ming Hsieh Department of

Electrical Engineering, University of Southern California, LosAngeles, CA 90089, USA. E-mail: [email protected].�To whom correspondence should be addressed.

Manuscript received: 2013-03-24; accepted: 2013-04-17

connections[1]. Therefore, researchers have proposedthe DTN routing to exploit the opportunisticconductivities amongst a group of mobile wirelessnodes and to deliver data in a “store-carry-forward”manner[2-8]. By tolerating certain amount of delay,such routing schemes make data transmission in DTNpossible and increase the throughput. Therefore, DTNrouting technologies have been widely used in theintermittently connected and decentralized scenarios,such as the disaster rescue, wildlife monitoring, fieldwork, etc.

Recently, with the maturation of the Vehicle-to-Vehicle (V2V) communication technologies and thestandardization and upcoming enforcement of theIEEE 802.11p in the automotive industry, vehicularnetwork becomes a promising technology. Applicationsin vehicular networks include not only the “active

Qiao Fu et al.: DAWN: A Density Adaptive Routing for Deadline-Based Data Collection in Vehicular � � � 231

safety services”, but also the infotainment, whichprovides useful information and on-board entertainmentapplications[9]. Large scale urban sensing applicationshave also drew people’s attention these days, includingremote metering, environmental sensing, etc.[10]

Although these services can be supported by awireless infrastructure (e.g., 3G), the cost of doingso is high and may not be possible when such aninfrastructure does not exist or when the service isintermittent[11]. Since immediate message delivery isnot a necessity for infotainment and urban sensing,these tasks can be supported by DTN. By toleratingcertain amount of delay, the vehicular DTN enjoyvery low deployment and operational costs. Moreover,in case of disaster, the wireless infrastructure maybe damaged, whereas vehicular DTN can be usedto provide important traffic, rescue, and evacuationinformation to the users.

The vehicular network enjoys several advantagesin supporting DTN routing. First, since vehicles canprovide substantial electricity supplies and transportbulky hardware, neither the power nor the storagewill create system design issues. Furthermore, themovement of the vehicles is statistically predictablebecause of the road constraint and the human behavioralhabits. Therefore, the routing algorithms can utilizethe vehicle’s geographic location information to makeprecise data forwarding or duplication decisions so asto achieve good performance.

Nevertheless, in DTN deployed in urban areas, thechallenges for routing algorithms are still huge. Inurban areas, the density of vehicles fluctuates fiercelyin time and location. For example, the vehicle densityat intersections or during the traffic jams could behigh enough to cause packet congestions in thechannel, while in suburbs or at night, vehicles canhardly encounter or communicate with others[12]. Theheterogeneity of the vehicle density makes the volumeof data exchanged upon vehicle encounters at differentgeographic locations and time highly diversified, whichwill cause a channel utilization problem. In the areasor time of low car density, part of the channelcapacity could be wasted, while in the areas or timeof high car density, the network will suffer packetloss due to excessive channel contention. Therefore,the constraint of wireless channel capacity should becarefully considered as an impact factor in the routingalgorithm design in large scale urban DTN. But, to the

best of our knowledge, so far there is still not a practicaland implementable routing algorithm for DTN whichtakes channel constraint into consideration.

Moreover, in most infotaiment and urban sensingapplications, there exists a deadline, beyond whichpackets are considered obsolete. For example, in realtime urban environment monitoring applications, datagenerated many hours ago may be considered uselesssince the environment may have changed greatlyduring that time. Therefore, such a deadline restrictionshould be taken into consideration when evaluatingthe DTN performance. However, conventional DTNrouting algorithms normally seek for either higherthroughput or smaller delay, which may not be anaccurate measurement for the network performance inlarge scale urban DTN.

In this paper, we propose a local capacityconstraint density adaptive DTN routing algorithm,namely Density Adaptive routing With Node deadlineawareness (DAWN), to remedy the impact of nodedensity heterogeneity, so as to improve the deadline-based network performance. Unlike traditional DTNmodels, DAWN considers the local capacity limitationof the wireless channel, especially for the case in whichthe throughput is restricted by congestion. Moreover,a heuristic measurement of network performance isproposed to measure the packet delivery ratio withindeadline. Unlike other conventional DTN routings, suchmeasurement takes both the network throughput and thedelay of the packet into consideration. Because of thedeadline restriction characteristic of most urban sensingand infotaiment applications, such deadline-based datacollection measurement is rather practical and accuratein realistic deployments. In DAWN, a deadline-basedutility function for each sensory data packet is defined,which indicates the probability that this packet can bedelivered successfully to the base stations within thedelay constraint. The replication procedure is densityadaptive and each node chooses to replicate the packetswith highest utility gain estimation under capacityconstraint. Note that even though we propose DAWNin a field-gathering model, where the destination ofthe packets is only the base stations, DAWN can beeasily extended to scenarios where destinations can beany node in the network. Simulation results based onManhattan Grid model and Beijing taxi trajectories[13],show that DAWN outperforms other traditional DTNrouting schemes in the performance of delivery ratio

232 Tsinghua Science and Technology, June 2013, 18(3): 230-241

and coverage area within delay constraint.

2 Related Work

As the key component of DTN, a large number ofrouting protocols have been proposed based on variousassumptions regarding the connectivity and mobilitypatterns. Vahdat and Becker[14] propose the first routingprotocol for DTN, namely the Epidemic RoutingProtocol. Its idea is almost the same as the floodingrouting scheme in traditional ad hoc networks. When anintermediate node receives a message, it will broadcastto all its neighbors, and thus information disseminatesin a flooding manner. It is intuitive to consider theEpidemic routing as an unconstrained optimization tomaximize the performance of the throughput. However,in real world deployments, especially in large scale,resource constraint scenarios like the one we arefacing in the urban DTN, careful consideration of thecomplicated tradeoff between the performance and theresource cost, and adaptation to the heterogeneity of thenetwork is a must for sophisticated protocols.

In Ref. [2], a vehicular based DTN routingprotocol called MaxProp is proposed to address theresource constraint scenarios. MaxProp uses severalmechanisms to define the order in which packetsare transmitted and deleted in order to better utilizethe network resource. However, it is not explicitlyaddressed how to optimize a specific routing metricusing MaxProp. In Ref. [15], a resource allocationprotocol called RAPID is introduced. RAPID isdesigned to explicitly optimize an administrator-specific routing metric, such as minimizing averagedelay, minimizing missed deadline, or minimizingmaximum delay. RAPID translates the routing metricto per-packet utility and determines at every transferopportunity if the marginal utility of replicating a packetjustifies the resource used. However, RAPID requiresthe flooding of information about all the replicas ofa given message in the queues of all nodes in thenetworks in order to derive the utility. In scalablenetworks such as the urban DTN, such informationis difficult to achieve and information flooding alsorequires tremendous network resource.

Besides simply maximizing the performance, severalapproaches have been introduced to control the copy ofpackets based on Epidemic algorithm. One successfulcase is the Spray and Wait (S&W)[16], in which eachpacket is only transmitted L times by the source

node. Packets are then held by the relay nodes anddelivered to the destination until they meet. Thisapproach obviously reduces the overhead, and achievesacceptable network performance under the randomwaypoint model. However, S&W manifests apparentdisadvantage in large scale network considering thefairness of the coverage. Packets generated near thebase stations have more chance to be collected thanthose generated far away.

Some protocols assume the network topology canbe predicted and intermediate nodes can estimate thechance of reaching the destination. Based on thisestimation, the intermediate nodes decide whetherto store the packet or send it to other nodeswho enjoy better chance to reach the destination(see Refs. [17-22]). In Ref. [17], the authorspropose a routing scheme called Probabilistic RoutingProtocol Using History of Encounters and Transitivity(PROPHET). PROPHET estimates a probabilisticmetric called delivery predictability at every node a,for each known destination b. This metric indicates howlikely the node will be able to deliver a message to thedestination. When two nodes meet, they compare eachother’s delivery predictability for each destination, andtransfer packets to the party that has higher probabilityof reaching the destination. Since nodes make choicesbased on a priori information, the average cost perpacket delivery of PROPHET is lower than that ofEpidemic. Nevertheless, such approach strongly relieson mobility prediction. Performance can be improvedwhen more precise prediction is made, and a specificlink quality estimation for urban DTN should beproposed.

There are algorithms successfully optimize thenetwork performance to some extend. The queue-differential backpressure routing scheme is shown byTassiulas and Ephremides[23] to be throughput optimalin terms of being able to stabilize the network underany feasible traffic rate vector. The basic idea ofsuch scheme is to prioritize transmissions over linksthat have the highest queue differentials, so thatpackets are flowing in the network as if pulled bygravity to the destination. Further research showsthat backpressure scheme performs well under hightraffic condition. On the contrary, when the trafficis low, backpressure scheme becomes inefficient interms of packet delay, since many other nodes apartfrom the destination may also have a small or 0queue size. In Ref. [24], the authors propose an


adaptive redundancy backpressure scheme to reducedelay under low load conditions, while at the same timepreserving the performance and benefits of traditionalbackpressure routing under high traffic conditions. Suchadvantage is achieved by creating copies of packets ina new duplicate buffer upon an encounter, when thetransmitters queue occupancy is low. These duplicatepackets are transmitted only when the original queue isempty. In this way, the algorithm builds up gradientstowards the destinations faster and reduces packetlooping. However, such scheme does not guaranteedeterministic delay for the packets.

The routing algorithms stated above ignore the effectsof network contention on performance arguing thatits effect is small in sparse intermittently connectednetworks. Nevertheless, some recent work has shownthrough simulations that ignoring capacity constraintleads to inaccurate and misleading results. In Ref.[25], the authors propose an analytical framework tomodel network contention to analyze the performanceof any given DTN routing scheme for any givenmobility and channel model. They then calculatethe expected delay for representative DTN routingschemes, namely direct transmission, Epidemic, andS&W, with capacity constraint in the network forthe random direction, the random waypoint, and themore realistic community-based mobility model. Theypoint out through these delay expressions that ignoringcapacity constraint can lead to suboptimal or evenerroneous decision. The authors also point out somedirections to modify the existing S&W scheme inorder to decrease the delay when capacity constraintis taken consideration. However, no practical andimplementable heuristics is proposed in the ongoingwork.

Density adaptive routing schemes have also gainedinterests recently. In Ref. [26], the authors pointout that routing algorithms, such as Epidemic andS&W, perform poorly in density dynamic networks,such as the RollerNet examined in the paper. Basedon observation of the RollerNet, which is obtainedfrom a rollerblading tour that consists of 15,000people, the authors propose the Density-Aware Sprayand Wait (DA-SW), a measurement-oriented variantof the S&W algorithm that tunes, in a dynamicfashion, the number of message copies to bedisseminated in the networks. In Ref. [4], an Epidemic-based density adaptive routing scheme called ECAMis proposed. ECAM limits the maximum times

of message transmission among nodes to controlflooding. Transmission times decrease exponentiallywith hop counts so that less transmissions wouldbe allowed as packets disseminate. Simulation resultsshow that ECAM successfully decreases overheadwithout delivery rate reduction. All these algorithmsrealize the influence of density to networks and theimportance of adjusting message copies with networkdensity, but none of them aims at urban DTN, whereextreme heterogeneity presents. How to detect localdensity, and to tune message replication with networkdensity in large scale urban DTN are problems remainunsolved. The characteristics of routing algorithms forDTN are summarized in Table 1.

In this paper, we consider both local channel capacityconstraint and heterogeneous node density for the DTN,and propose a capacity constrained and density adaptiverouting algorithm to improve the delivery ratio withinthe delay constraint. In our algorithm, the local capacityis better utilized to improve the packet delivery ratiowithin deadline, and the impact of fluctuant nodesdensity has been taken into account as well.

3 DAWN

In this section, we will first propose a network modelfor DTN taking consideration of network capacityconstraint. The notation “First Passage Time” is thenexplained addressing the time for a mobile node tofirst meet the base stations, which will be used in theestimation of utility value in DAWN. The DTN routingscheme DAWN is then proposed in detail.

3.1 Network model

The following algorithm and simulations are consideredunder the data collection network model stated asfollows, in which packets are generated by themobiles, and are intended to be uploaded to thebase stations. Such network model can be utilized fordata collection and distribution in urban sensing andinfotainment applications. Nevertheless, applicationof DAWN is not limited to such data collectionmodel. DAWN can be easily extended to supportvehicle to vehicle data routing like any other DTNrouting algorithms by simply replacing the utilityfunction to corresponding scenarios.

The data collection network model consists of Mmobile nodes and one or more static data gathering basestations (See Fig. 1). TheM mobile nodes move withinthe field following a certain mobility pattern. Each


Table 1 DTN routing algorithms.

Routing Flooding Resource Copy Link Density Design Performance Deadlinedissemination constraint control estimation adaptive for VSN optimization aware

Epidemic[14] X � � � � � � �

MaxProp[2] X X � � � X � �

RAPID15 X X � � � � X �

Spray and Wait[16] X � X � � � � �

PROPHET[17] � � � X � � � �

BWAR[24] X � � � � � X �

ECAM[4] X � X � X � � �

DA-SW[26] X � X � X � � �

Fig. 1 Network model.

mobile generates fix-length data packets at a certainrate while traveling. Each packet can be denoted bythe tuple of its source node and the generation time,denoted by .s; t0/, and it shall be delivered to the basestations before a deadline TMAX. The data gatheringbase stations can be located in arbitrary spots in the fieldand collect the sensory data packets from mobile nodes.

Both the mobile nodes and the base stations areequipped with short range communication devices, thelimited bandwidth of which constrains the volume ofdata deliveries over wireless link when multiple nodesare in each others’ transmission range and competingthe channel. We assume that in this case, only anumber of K packets can get through in unit timedue to channel capacity constraint. The constraint Kis determined by the Media Access Control (MAC)and physical layer. For example, when IEEE 802:11p

standard is utilized for mobiles communication, thereexists a maximum number of packet input, inputlarger than which will cause collision and decrease thenumber of packets getting through the network and

being received by the destinations[27]. Such optimalnumber of input can be defined as the capacityconstraint K in our model. The local network densityis defined as the number of nodes within transmissionrange. Such definition is rational since only the nodeswithin transmission range can compete the channel andreplicate packets.

3.2 First passage time

Definition 1 (First Passage Time) The First PassageTime (FPT) is defined as the interval between the time amobile node starts from a certain spot and the time thatit first meets the base stations.

The FPT can be defined for any mobilitymodel. Intuitively, the distribution of FPT can berelated to the location of the starting spot, the size ofthe map, and the velocity of the mobile nodes. However,there is not a close form solution for the ComulativeDistribution Function (CDF) of FPT in such datacollection model even for the simple mobility modelssuch as random walk on the torus.

We could achieve the CDF of FPT from eithernumerical simulations (for mobility models such asthe Manhattan Grids) or empirical data (for realworld vehicle mobility patterns) as a reference for thefollowing calculations. Figure 2 shows the CDF of FPTfor the Manhattan Grid Model and the Beijing taxitrajectories with single base station located.

The Manhattan Grid Model uses a grid road topologythat fully manifests the movements of vehicles in urbanareas, where the streets are in an organized manner. Itis often used in vehicular network evaluations becauseof its similarity to the urban street networks. In aManhattan Grids, the mobile nodes move in horizontalor vertical direction with a velocity of one grid eachtime slot on the urban map, and choose directions withidentical probability at each grid point. The borders of


Fig. 2 The CDF of FPT for Manhattan Grid Model andBeijing taxi trajectories.

the Manhattan Grids are cyclic, that is, the mobile nodesgetting out of the network through one side of the borderwill enter the network through the other side. Suchassumption guarantees that the following evaluationswill not be influenced by the border effects. Moreover,this Manhattan Grid Model reflects the situation in realdeployments. In large cities with wide area coverage,the area will be divided into multiple zones, eachof which is equipped with a base station, and thebase stations are connected by a wide-band backbonenetwork. In such scenarios, with negligible bordereffects at the edge of the whole coverage area, eachbase station and its zone is equivalent, and theirborders can be considered cyclic since vehicles arefree to move in and out of the borders. Therefore, theManhattan Grids provide a pretty good approximationof the real world multi-base-station situation. Eachcurve for the Manhattan Grid Model in the figure isthe results of 108 simulation experiments. Distanceis the Manhattan distance, i.e., the sum of horizontaland vertical distance in number of grids, between thestart and end points. In each experiment, mobile nodestarting from a certain distance moves according tothe Manhattan Grid Mobility Model. The time it takesto first meet the base station, which can be addressedas an FPT sample, is then counted. We calculate thefrequency of the FPT samples and derive the statisticCDF for FPT.

The curves for Beijing taxi trajectories are obtainedfrom the Beijing taxi trajectory data set of 27; 848

taxis traveling for 15 days from May 16, 2009 tillMay 30, 2009[13]. The data set includes taxi trajectoriesranging between 39:759ıN and 40:023ıN latitude, and116:209ıE and 116:544ıE longitude, the area insidethe fifth-ring road of Beijing. The curves are derivedusing the same methods as in Ref. [28], in which FPTfor seabird traveling is studied. We divide the area of26:3 km � 33.5 km into 26 � 33 coordinates with eachcoordinate of 1 km � 1 km. The base station is locatedat coordinate .15; 19/ with a communication range of500 m. We then calculate the frequency of the FPTsamples of taxis starting from each coordinate to thebase station to derive the statistic CDF of FPT for theBeijing taxi trajectories. The statistics of such CDF ofFPT in Beijing is relatively coarse grained in space,this is because we used enough data from each cell toguarantee the validity. CDF of FPT for scenarios wheremore than one base station located can also be drawnwith the same method.

Note that FPT for vehicle to vehicle data routing canalso be defined as the time it takes for a mobile node tofirst meet the destination mobile. The CDF of FPT formobiles following the random walk model can be foundin Ref. [29], and can also be utilized in the followingutility calculation of DAWN. In this case, DAWN canbe easily extended to scenarios where destinations aremobiles rather than static base stations.

3.3 DAWN

DAWN models DTN routing as a utility-based resourceallocation problem. At a specific time, all the copiesof a specific packet compose a packet companyp.s;t0/. Upon any of the copies in the data companyarriving at the base stations, the packet is considereddelivered. The DAWN algorithm executes when nodesare within each others’ transmission range. Nodes thenmake local decisions to replicate packets to all theother nodes within range. DAWN remedies the impactof density heterogeneity and capacity constraint andimproves the packet delivery probability by takingcareful considerations on the following two aspects foreach node within each others’ transmission range ateach time slot, namely the quota to replicate, and theselection of replication.

3.3.1 How many packets to replicateSince the capacity within transmission range isconstrained, each node is assigned with a certain quotato transmit packets according to the node density withinrange. As stated above, the optimal input K is decided


by the channel. Given K, each node in the range is

assigned withK

�l.t/input, in which �l.t/ indicates the

density, that is the number of the mobile nodes withintransmission range of mobile node l at time t . Since themobile node density is slowly varying with time, thisnumber can be estimated by the history encounters ofthe node. Each node may learn �l.t/ by counting thesource nodes it has communicated with. Local mobilenode number is estimated by each node as the average ofthis number in history encounters. In order to guaranteethe accuracy of density estimation, time window for theestimation should be carefully chosen according to thenetwork density variation feature. Such time windowcan be optimized at each communication opportunityby minimizing the error between estimation and actualdensity.

When local density is low and the node is assignedwith enough input, all the transfer requirements can besatisfied. In this case, the forward strategy is almostthe same as Epidemic, which has been proved to beoptimal in none capacity constraint situation. Whenlocal density increases and the assigned input is notenough for the node, each node will make decisionson which packets to replicate in order to improve theoverall network performance.

3.3.2 Which packets to replicateFor most urban DTN applications, there exists a delayconstraint TMAX, beyond which packets are consideredwasted. Therefore, the performance measurement toincrease the delivery ratio of packets within deadlineis more practical and fundamental. To make thereplication decisions, DAWN introduces a utility valuefor each packet indicating the delivery probability of thepacket within the delay constraint. And nodes alwaysreplicate the packets with highest utility gains in orderto better utilize the limited capacity.

Definition 2 (Packet Utility) The utility value of thepacket company p.s;t0/ at time t , denoted by U t

.s;t0/, is

defined as the probability that after time t , if no furtherreplication happens, at least one of the copies will reachthe destination before the deadline.

Due to packet replication, the packet utility valueis increasing with time. If at time � , packet p.s;t0/ isreplicated only by mobile node l , then the utility gain ofpacket p.s;t0/ due to this replication, defined as�U �;l

.s;t0/,

is as follows:

�U�;l.s;t0/

D 1 � .1 � U �.s;t0//˚ l.�/�l .�/ � U �.s;t0/

D .1 � U �.s;t0//.1 � ˚ l.�/�l .�// (1)

where ˚ l.t/ D 1 � ˚l.t/ is the ComplementaryCumulative Distribution Function (CCDF) of FPTfor the location of mobile node l at time � , and� D TMAX � � C t0 is the remaining time. Notice thatthe function ˚l.t/ here should be the correspondingCDF of FPT that address the scenario, which meansdifferent ˚l.t/ should be used if the number andpositions of the base stations change. DAWN can alsobe extended to the situation where the destinations aremobile nodes rather than static base stations given theCDF of FPT of the relays and the destinations.U t.s;t0/

is a global parameter indicating the utilityvalue of packet company p.s;t0/ at time t . In practice,such global parameter is difficult to derive since itneeds to spread information of each packet to theentire network. So DAWN introduces a distributed wayto estimate the utility value of each packet. At eachreplication, the packet utility value is updated by boththe source node and the intermediate nodes as theoriginal packet utility value OU t

.s;t0/(which is known

by the source node) pluses the �U t;l.s;t0/

. Because localestimation loses the global picture, the estimated utilityvalue OU t

.s;t0/is upper bounded by U t

.s;t0/.

When network is dense and only part of the packetscan be replicated, packets that own higher utility gainwill be transmitted with higher priority. Note that suchutility gain estimation is not precise since we do notconsider the global packet replication scheme. Howeverit claims low communication and computing cost sinceit does not require any global information transmission.

Since it is adaptive to density changes in networks,DAWN can increase the delivery probability ofpackets within delay constraint in dynamic networkswhere capacity is constrained by successfully avoidingnetwork congestion. Moreover, this algorithm iscompletely distributed. Neither center node norinformation exchange is needed in the network tocollect density information. Nodes only have tocalculate the number of nodes in its communicationrange to estimate local node density and make forwarddecision. The pseudo code of DAWN is shown inTable 2. Simulations on Manhattan Grid Model andBeijing taxi trajectories show that DAWN outperformsall other traditional DTN routing protocols in theperformance of packet delivery ratio and coveragewithin delay constraint.


Table 2 Pseudo Code of DAWN.

Algorithm DAWNFor each node1: update density estimation O�l .t/2: calculate the assigned input Kassigned D

KO�l .t/

3: calculate self-required input Krequire

4: rank packets according to �U t;l.s;t0/

in decreasing order,5: IF Krequire < Kassigned

6: all input requirements be satisfied7: ELSE8: transfer Kassigned packets successively9: END10: update utility value U t

.s;t0/for each packet

11: END

4 Simulations

4.1 The Manhattan Grid simulations

To analyze the performance of DAWN, we first simulateon a 25 � 25 Manhattan Grids, which is considered asa proper abstraction for the metropolitan areas. In theManhattan Grid Model, each mobile generates packetsaccording to a certain rate �s D 0:05 packet per timeslot. The time constraint of each packet is 100 slots. Abase station locates in the center of the network, andthe communication range is assumed to be zero, whichmeans nodes can only communicate when they areat the same grid point. In our simulation, four othertraditional DTN routing algorithms, namely Epidemic,S&W, PROPHET and ECAM, are compared withthe DAWN algorithm. Furthermore, in DAWN, twoapproaches which respectively achieve packet utilityvalue in a global manner and a distributed estimationmanner are both considered.

Figure 3 compares the delivery ratio within timeconstraint of different routing algorithms in termsof node density changing. It shows that DAWNoutperforms all other protocols with considering oflocal network capacity constraint. As is shown, thedelivery ratio traces of all algorithms increase whennode density increases, and the delivery ratio of DAWNis about 20% higher than that of Epidemic, PROPHET,ECAM and S&W when node density is low. DAWNperforms even better with high node density, and about98% collected packets can be delivered to the basestation. It suggests that DAWN takes full advantage ofnetwork capacity and also efficiently avoids networkcongestion as network becomes dense. Meanwhile,

Fig. 3 Delivery ratio v.s. node density in Manhattan Gridsimulation.

the performance of distributed global utility estimationmethod is compared with that of the global utilityaware method, and the results show that they haveapproximate performance. It means that our distributedestimation method is feasible to estimate the utilityvalue.

To achieve a global view of the coverage of differentrouting schemes, we calculate the delivery ratio of thepackets generated at each grid point and count thepercentage of grids that enjoy a delivery ratio of morethan 50%, as is shown in Fig. 4. From these resultswe notice that DAWN not only enjoys higher deliveryratio, but also covers larger areas with high deliveryratio than the other routing schemes. By introducingFPT into the utility function, DAWN gives packetsgenerated far from the base station higher probabilityto be replicated in the network than other schemes. Thismakes sense because comparing with packets generatednear by, packets generated far away may need more timeto disseminate in the network and finally reach the basestation.

4.2 Taxi trajectory based simulations

Based on the Beijing taxi trajectory database, whichcontains 30-day GPS records from May 1st, 2009till May 30th, 2009 of 27 848 taxis in Beijing,DAWN is simulated to compare with other routingalgorithms. The trajectories from May 16th till May30th are utilized to study the FPT of the taxis. And thefollowing simulations are conducted using trajectoriesfrom May 1st till May 15th, after invalid traces being


Fig. 4 Percentage of grids that attain delivery ratio > 50%in Manhattan Grid simulation.

removed, and the area within the fifth-ring road ofBeijing is considered.

In the simulation, a base station is set at the locationof the center of the Beijing city. The communicationrange for each taxi is 100 meters, and that ofthe base station is 1000 meters, typical for V2Vcommunication. Taxis generate packets according to acertain rate �s D 0:05 packet per minute. The timeconstraint of each packet is 250 minutes.

As is shown in Fig. 5, the delivery ratio of allrouting algorithms in terms of the number of taxisincreases when the number of taxis increases, andDAWN shows apparent advantage over other routingschemes. Its delivery ratio is about 15% higher thanthat of Epidemic and PROPHET, and about 40% higherthan that of S&W and ECAM. Since this simulationutilizes authentic trajectories, PROPHET obtains betterperformance than the simulations on Manhattan Gridsas the link prediction comes into play. Meanwhile,ECAM suffers performance decay as the networkbecomes scalable and packets can hardly disseminatesufficiently before the exponential rule in ECAM stopsthem from transferring. It also shows that the deliveryratio of DAWN can reach about 70% when using only3000 taxis in a 881 km2 city area. Such results suggestthat using DAWN can effectively collect and transmitdata in urban DTN.

To show the coverage property of DAWN, we dividethe area of 26:3 km � 33:5 km into 263 � 335 blockswith each block of 100m � 100m. Figure 6 showsthe percentage of blocks that enjoy a delivery ratio

Fig. 5 Delivery ratio v.s. taxi number in taxi trajectorysimulation (1 base station).

Fig. 6 Percentage of blocks that attain delivery ratio> 50%(1 base station).

of more than 50%. From the results we can see thatDAWN enjoys high delivery ratio in more than 60%of the area of Beijing when using only 3000 taxis,while Epidemic and PROPHET mainly collect packetsnear the base station and S&W and ECAM achievelow delivery ratio in the whole area of Beijing. Suchsuperiority guarantees the large portion of the urbanarea can enjoy acceptable delivery ratio of the messagescollected.

In practical applications, more than one base stationwill be deployed in the network to increase the deliveryratio and decrease delay. DAWN can be easily extendedto the situation where more than one base station is


located by simply utilizing the CDF of FPT for multiplebase stations when calculating the utility incrementalvalue. We further deploy 4 base stations, respectivelylocated at .70; 80/, .70; 240/, .200; 80/, .200; 240/, anddecrease the communication range of each base stationto 500 m, which is half of that of the simulations withonly one base station. Figures 7 and 8 respectivelyshow the delivery ratio within time constraint in termsof taxi numbers and the percentage of blocks thatenjoy a delivery ratio of more than 50%. From thesimulation results, we can see that the delivery ratioof all the routing algorithms increase because of theextra deployment of base stations, and DAWN stilloutperforms all other schemes. The delivery ratioreaches 75% with only 3000 taxis and almost 80% with4960 taxis, which are nearly 20% higher than that ofEpidemic and PROPHET and 45% higher than that ofS&W and ECAM. DAWN also enjoys high deliveryratio in nearly 70% of the area of Beijing when usingonly 3000 taxis, when other routing schemes mainlycover the areas around the base stations. With betterallocation of the base stations, for example, to locatethe base stations at the spots where more taxis pass by,performance can be further improved.

Figures 9 and 10 respectively show the deliveryratio within time constraint in terms of taxi numbersand the percentage of blocks that enjoy a deliveryratio of more than 50% when 4 base stations arelocated at the top 4 points with the highest taxi passingrate, namely .128; 160/, .138; 160/, .139; 181/, and.128; 170/. From the results we can see that the deliveryratio of DAWN is more than 80% with 3000 taxis,and the coverage of DAWN improves to more than

Fig. 7 Delivery ratio v.s. taxi number in taxi trajectorysimulation (4 base stations).

Fig. 8 Percentage of blocks that attain delivery ratio> 50%(4 base stations).

Fig. 9 Delivery ratio v.s. taxi number in taxi trajectorysimulation (with better base station locations).

70%. Such results further prove the effectiveness ofDAWN in large scale DTN.

5 Conclusions

In this paper, we firstly propose a model thatfully manifests the special features of vehicularDTN with consideration of the limited local channelcapacity caused by wireless interference. We furtherpropose a heuristic and accurate network performancemeasurement for the large scale DTN which evaluatesthe network throughput within deadline. A routingscheme for DTN, called DAWN is then introducedto improve the packet delivery ratio within deadlinewhere network capacity is constraint. The core ideaof DAWN is to adjust forward strategy and protocolparameters based on local node density with theawareness of local channel capacity. In DAWN, a utility


Fig. 10 Percentage of blocks that attain delivery ratio>50% (with better base station locations).

value suggesting the delivery probability of the packetswithin time constraint is maintained for each packet byeach mobile node. Mobile nodes in each cell decidehow many packets to broadcast and which packet tobroadcast independently based on information of localdensity and the packet utility gain. Simulations on bothManhattan Grids and Beijing taxi trajectory databaseshow that DAWN outperforms all other protocols interms of delivery ratio within deadline. The deliveryratio of DAWN is about 35% higher than those ofother routing protocols in Manhattan Grids, and about20% higher in Beijing taxi trajectory-based simulationwith only one base station. The delivery ratio furtherincreases when more base stations are deployed andbetter located. In simulations based on Beijing taxitrajectories, the delivery ratio of DAWN reaches 70%with only 3000 taxis and one base station, and 75%with 4 base stations, and reaches 80% when the 4 basestations are carefully located at the spots where moretaxis pass by. Simulation results also show that DAWNenjoys better fairness across the coverage area, whileother routing algorithms only attain acceptable deliveryratio near the base stations. The percentage of blocksthat reach a delivery ratio of over 50% is over 60%when using only 3000 taxis and one base station, thepercentage boosts to 68% when 4 base stations aredeployed, and 72% when the base stations are locatedat spots where more taxis pass by. Such performancemetrics suggest that DAWN may perform quite well inlarge scale vehicular DTN deployments.

References

[1] Z. Zhang, Routing in intermittently connected mobile adhoc networks and delay tolerant networks: Overview andchallenges, IEEE Communication Surveys and Tutorials,vol. 8, no. 1, pp. 24-37, 2006.

[2] J. Burgess, B. Gallagher, D. Jensen, and B. Levine,Maxprop: Routing for vehicle-based disruption-tolerantnetworking, in Proc. IEEE INFOCOM, Barcelona, Spain,2006, pp. 1-11.

[3] T. Spyropoulos, K. Psounis, and C. Raghavendra, Efficientrouting in intermittently connected mobile networks: Themulti-copy case, IEEE/ACM Trans. Networking, vol. 16,no. 1, pp. 77-90, 2008.

[4] Z. Jin, J. Wang, S. Zhang, and Y. Shu, Epidemic-based controlled flooding and adaptive multicast for delaytolerant networks, in Proc. 7th International Conferenceon Ubiquitous Intelligence and Computing and 7thInternational Conference on Autonomic and TrustedComputing (UIC/ATC), Xi’an, China, 2010, pp. 191-194.

[5] M. Mcmahon and S. Farrell, Delay- and disruption-tolerantnetworking, IEEE Internet Computing, vol. 13, no. 6,pp. 82-87, 2009.

[6] H. Zhu, X. Lin, R. Lu, X. Shen, D. Xing, and Z. Cao,An opportunistic batch bundle authentication scheme forenergy constrained DTNs, in Proc. IEEE INFOCOM,Shanghai, China, 2010, pp. 1-9.

[7] B. J. Choi and X. S. Shen, Adaptive asynchronous sleepscheduling protocols for delay tolerant networks, IEEETrans. Mobile Computing, vol. 10, no. 9, pp. 1-9, 2011.

[8] W. Wang, M. Motani, and V. Srinivasan, Opportunisticenergy-efficient contact probing in delay-tolerantapplications, IEEE/ACM Trans. Networking, vol. 17,no. 5, pp. 1592-1605, 2009.

[9] H. Hartenstain and K. P. Laberteaux, A tutorial surveyon vehicular ad hoc networks, IEEE CommunicationMagazine, vol. 46, no. 6, pp. 164-171, 2008.

[10] A. Campbell, S. B. Eisenman, N. Lane, E. Miluzzo,and R. A. Peterson, People centric urban sensing, inProc. 2nd annual international workshop on WirelessInternet, Boston, USA, 2006.

[11] J. Zhao and G. Gao, VADD: Vehicular-assisteddata delivery in vehicular ad hoc networks, IEEETrans. Vehicular Technology, vol. 57, no. 3, pp. 1910-1922, 2008.

[12] Y. Toor, P. Muhlethaler, and A. Laouiti, Vehicular adhoc networks: Applications and related technical issues,IEEE Communication Surveys and Tutorials, vol. 10, no. 3,pp. 74-88, 2008.

[13] Beijing Taxi Trajectory Datasets, http://sensor.ee.tsinghua.edu.cn, 2011.

[14] A. Vahdat and D. Becher, Epidemic routing for partiallyconnected ad hoc networks, in Tech. Rep. CS-200006,Dept. Computer Science, Duke University, Durham, USA,2000.

[15] A. Balasubramanian, B. Neil, and A. Venkataramani,DTN routing as a resource allocation problem, inProc. SIGCOMM’07, Kyoto, Japan, 2007, pp. 373-384.


[16] T. Spyropoulos, K. Psounis, and C. S. Raghavendra,Spray and wait: An efficient routing scheme forintermittently connected mobile networks, in Proc. 2005ACM SIGCOMM workshop on Delay-tolerant networking,Philadelphia, USA, 2005, pp. 252-259.

[17] A. Lindgren, A. Doria, and O. Schelen, Probabilisticrouting in intermittently connected networks, ACMSIGMOBILE Mobile Computing and CommunicationsReview, vol. 7, no. 3, pp. 19-20, 2003.

[18] M. Musolesi, S. Hailes, and C. Mascolo, Adaptive routingfor intermittently connected mobile ad hoc networks, inProc. 6th IEEE International Symposium on a Worldof Wireless Mobile and Multimedia Networks, Taormina,Italy, 2005, pp. 183-189.

[19] C. Shen, G. Borkar, S. Rajagopalan, and C. Jarkaeo,Interrogation-based relay routing for ad hoc satellitenetworks, in Proc. IEEE Global TelecommunicationsConference, Taipei, China, 2002, pp. 2920-2924.

[20] B. Burns, O. Brock, and B. N. Levine, MV routingand capacity building in disruption tolerant networks, inProc. IEEE INFOCOM, Miami, USA, 2005, pp. 398-408.

[21] K. Tan, Q. Zhang, and W. Zhu, Shortest path routingin partially connected ad hoc networks, in Proc. IEEEGlobal Telecommunications Conference, San Francisco,USA, 2003, pp. 1038-1042.

[22] E. Jones, L. Li, J. K. Schmidtke, and P. A. S. Ward,Practical routing for delay tolerant networks, in Proc. 2005ACM SIGCOMM workshop on Delay-tolerant networking,Philadelphia, USA, 2005, pp. 237-243.

[23] L. Tassiulas and A. Ephremides, Stability properties ofconstrained queueing systems and scheduling policies formaximum throughput in multihop radio networks, IEEETrans. Automatic Control, vol. 37, no. 12, pp. 1936-1948,1992.

[24] M. Alresaini, M. Sathiamoorthy, B. Krishnamachari, andM. J. Neely, Backpressure with Adaptive Redundancy(BWAR), in Proc. IEEE INFOCOM, Orlando, USA, 2012,pp. 2300-2308.

[25] A. Jindal and K. Psounis, Contention-aware performanceanalysis of mobility-assisted routing, IEEE Trans. MobileComputing, vol. 8, no. 2, pp. 145-161, 2009.

[26] P. Tournoux, J. Leguay, and F. Benbadis, Density-awarerouting in highly dynamic DTNs: the rollernet case, IEEETrans. Mobile Computing, vol. 10, no. 12, pp. 1755-1768,2011.

[27] S. Eichler, Performance evaluation of the IEEE 802.11pwave communication standard, in Proc. IEEE 66thVehicular Technology Conference, Baltimore, USA, 2007,pp. 2199-2203.

[28] P. Fauchald and T. Tveraa, Using first-passage time inthe analysis of area-restricted search and habitat selection,Ecology, vol. 84, no. 2, pp. 282-288, 2003.

[29] R. Metzler and J. Klafter, The restaurant at the end of therandom walk: recent developments in the description ofanomalous transport by fractional dynamics, Journal ofPhysics A: Mathematical and General, vol. 37, no. 31,pp. R161-R208, 2004.

Qiao Fu received her BS degree inElectronic Engineering from TsinghuaUniversity in 2010. She is currentlya Master candidate at Departmentof Electronic Engineering, TsinghuaUniversity. Her research interests includeMAC and network layer algorithms inmobile wireless sensor networks.

Bhaskar Krishnamachari received hisBEng in Electrical Engineering at TheCooper Union, New York, in 1998,and his MS and PhD degrees fromCornell University in 1999 and 2002respectively. He is currently an AssociateProfessor and a Ming Hsieh FacultyFellow in the Department of Electrical

Engineering at the University of Southern California’s ViterbiSchool of Engineering. His primary research interest is inthe design and analysis of algorithms and protocols for next-generation wireless networks.

Lin Zhang received his BS degree in1998, MS degree in 2001, PhD degree in2006 all from Tsinghua University. Heis currently an associate professor atTsinghua University. His current researchfocuses on sensor networks, data andknowledge mining, and informationtheory. He is a co-author of more than

40 peer-reviewed technical papers and five U.S. or Chinesepatents applications. Lin and his team were also the winnerof IEEE/ACM SenSys 2010 Best Demo Awards. In 2010, hereceived Excellent Teaching Awards from Tsinghua University.

Documents

DAWN: A Density Adaptive Routing for Deadline …anrg.usc.edu/www/papers/DAWN_Fu.pdfQiao Fu et al.: DAWN: A Density Adaptive Routing for Deadline-Based Data Collection in Vehicular