Video Streaming Distribution in VANETs

Video Streaming Distribution in VANETsFabio Soldo, Member, IEEE, Claudio Casetti, Member, IEEE,

Carla-Fabiana Chiasserini, Senior Member, IEEE, and Pedro Alonso Chaparro

Abstract—Streaming applications will rapidly develop and contribute a significant amount of traffic in the near future. A problem,

scarcely addressed so far, is how to distribute video streaming traffic from one source to all nodes in an urban vehicular network. This

problem significantly differs from previous work on broadcast and multicast in ad hoc networks because of the highly dynamic topology of

vehicular networks and the strict delay requirements of streaming applications. We present a solution for intervehicular communications,

called Streaming Urban Video (SUV), that 1) is fully distributed and dynamically adapts to topology changes, and 2) leverages the

characteristics of streaming applications to yield a highly efficient, cross-layer solution.

Index Terms—Vehicular networks, streaming video, traffic forwarding.

Ç

1 INTRODUCTION

IT is commonly acknowledged that Vehicular Ad HocNetworks (VANETs) are ill-suited to support streaming

media traffic. Low bandwidth, fleeting connectivity, andhighly dynamic, unpredictable topology are the main short-comings hindering the support of multimedia applications.The network characteristics, along with the variable bit rate(VBR) nature of the traffic and the strict delay constraints,making no allowance for store-and-forward, pose a differentproblem from the ones previously addressed in ad hocnetworks [1], [2]. Even in the context of VANETs, theliterature carries few solutions dealing with channel accessand traffic forwarding for the support of streaming media [3],[4], [5], and the majority of them considers a highway scenarioor does not leverage the application characteristics (a moredetailed discussion of related works can be found in [6]).

In this work, we propose a fully distributed solutioncalled Streaming Urban Video (SUV) that efficiently dissemi-nates streaming video to all vehicles in a city VANET. SUVcompletely relies on intervehicular communication: a videostream, generated in a point in space (e.g., at a roadsideaccess point), is fed to SUV nodes and disseminated acrossthe VANET through a distribution structure, which is laidover the physical topology of mobile nodes. We refer to thenodes that belong to the distribution structure and areresponsible for the forwarding of the streaming video asrelay nodes. Streaming video distribution in SUV thereforeoccurs through a mix of local broadcasting from a relay

node to its neighboring nodes and MAC-layer multicastingfrom a relay node to its next-hop relay nodes.

Ideally, we would like the distribution structure to be agrid, so as to minimize the number of relay nodes requiredto cover a network area while providing a good level ofconnectivity [7]. To this aim, in SUV, each relay node notonly forwards the streaming traffic but also exploits a built-in positional device (e.g., a GPS) and the received powerlevel, to dynamically select its next-hop relays, so that thedistribution structure always approximates a grid as closelyas possible. We stress that due to the time-varying networktopology, the gateway can select different relays for eachnewly generated packet, and any relay node selects its next-hop relays according to the current status of its neighbor-hood. It follows that the set of relays thus identifiedrepresents a distribution structure that continuously adaptsto the network topology changes.

The GPS signal is also used to synchronize all relaynodes to a structured TDMA transmission, as done inseveral other works, e.g., in [8]. To efficiently schedule thetransmission of relay nodes, and thus minimize the chanceof collisions, we derive some results from graph-coloringtheory and apply them to the distribution structure. Unlikethe work in [9], we do not make any assumption on thenodes density and define a distance-k coloring, with k beingany positive integer, which is optimal and has the sameasymptotic complexity as the algorithm in [9].

Clearly, due to collisions stemming from nonideal gridselection, the scheduling algorithm may sometimes fail andresult in bandwidth waste. To overcome this, SUV nodes arecapable of 1) detecting a collision by means of passiveacknowledgments at the MAC layer, 2) reclaiming part of thewasted bandwidth to solve the contention, and 3) salvaging afailed transmission by sending a reduced amount of informa-tion. The last approach (hereinafter called “opportunisticaccess”) is particularly amenable to cross-layer interactionsbetween the application and the MAC layer, such as the use ofmultiple-description video streams allowing the selectivediscarding of descriptors. In addition, when VBR video istransmitted, each stream can be seen as a succession of burstsat peak rate, interspersed with periods of little or no activity.

IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 22, NO. 7, JULY 2011 1085

. F. Soldo is with the Henry Samueli School of Engineering, University ofCalifornia, Irvine, Irvine, CA 92697-2625. E-mail: [email protected].

. C. Casetti and C.-F. Chiasserini are with the Dipartimento di Elettronica,Politecnico di Torino, C.so Duca degli Abruzzi 24, 10129 Torino, Italy.E-mail: {casetti, chiasserini}@polito.it.

. P.A. Chaparro is with the Ciudad Politecnica de la Innovacion,iTEAM—Mobile Communications Group, Polytechnic University ofValencia, C\Camino de Vera S/N, Edificio 8G, 46022 Valencia, Spain,and the Telematics Department, Technical University of Catalonia,Castelldefels, Barcelona, Spain. E-mail: [email protected].

Manuscript received 29 May 2009; revised 27 Nov. 2009; accepted 25 May2010; published online 6 Oct. 2010.Recommended for acceptance by C. Pinotti.For information on obtaining reprints of this article, please send e-mail to:[email protected], and reference IEEECS Log Number TPDS-2009-05-0245.Digital Object Identifier no. 10.1109/TPDS.2010.173.

1045-9219/11/$26.00 � 2011 IEEE Published by the IEEE Computer Society

SUV leverages the traffic characteristics, and hence, theleftover bandwidth at the MAC layer, to support best-effortdata traffic via a contention-based access, thus enhancingradio resource efficiency.

The remainder of the paper is organized as follows: Weintroduce the system model in Section 2. The SUV scheme isdetailed in Section 3 as a fully distributed solution spanningseveral layers from the application to the MAC layer. InSection 4, by using graph coloring theory, we derive ascheduling algorithm that 1) is specifically designed for ourscenario and 2) aims at maximizing the distance betweenthe closest pair of nodes that simultaneously access thesame radio resources. We also mention that a study of theperformance of SUV against theoretical results for broadcastcapacity in multihop networks, as well as of its suitability tosupport video streaming in a realistic vehicular scenario,can be found in [6].

2 NETWORK SCENARIO AND SYSTEM MODEL

Consider a VANET deployed in an urban environment. Wemake no assumptions on the vehicle density since, as shownin [6], SUV can achieve a good performance even with spotty,volatile vehicular connectivity. We assume that one or moregateway nodes, either fixed or mobile, provide streamingvideo to car passengers. Examples of streaming videoinclude news, tourist information, commercial advertise-ments, football games, or music video clips. Distribution ofmultimedia content relies on intervehicular communication;in addition, vehicles may wish to exchange best-effort datatraffic in a peer-to-peer fashion: news summaries, publictransportation timetables, traffic warnings, and so on.

As in [10], we define the node transmission range as themaximum distance at which the expected packet error rateis still acceptable, namely, equal to 0.08 as in the 802.11standard, and we denote the corresponding received powerlevel by P ðthÞrx . This power level, measured in dBm, dependson the node wireless interface and data rate. Also, allnetwork nodes are supposed to be equipped with apositioning system, such as GPS, so that they are awareof their location and accurately synchronized in time. Eachvehicle periodically broadcasts an in-band HELLO signalingmessage, which carries the sender’s ID and GPS position. Avehicle can therefore keep 1) an updated list of its 1-hopneighbors, i.e., the nodes from which it receives a HELLO

with power level that is equal to or greater than P ðthÞrx , and2) the position and the power level received from each of its1-hop neighbors.

Streaming video and best-effort traffic (the latter includ-ing HELLO messages) are transmitted over a data channel,which is organized according to a TDMA structure. Asshown in Fig. 1, the data channel is structured in fixed-length time frames of duration TF . Each time frame isfurther divided into S identical slots, where the value of S,as well as the subset of relay nodes that transmit in eachtime slot, is determined using the graph-coloring algorithmdescribed in Section 4.

The multimedia content is assumed to be a videosequence. Note that various video coding techniques havebeen defined to allow streaming video to withstand thepotentially harsh conditions of wireless networks; examples

are multiple descriptions and layered coding [11]. Here, weconsider the video to be encoded into three descriptorsalthough other techniques could be considered as well. Eachdescriptor is composed of several video frames (e.g., I, B, orP ) of different size. The node protocol stack includes aSegmentation and Reassembly (SAR) layer, such that, at thetransmitter, each video frame is segmented (if needed) andformatted into a packet that will cover up to one third of aMAC payload of maximum size. In other words, every MACpacket to be transmitted in a time slot carries three (or parts ofthree) video frames, each corresponding to a differentdescriptor. The SAR header also carries the video framesequence number. At the receiver, the SAR layer reassembles(if both necessary and possible) different parts of the samevideo frame and sends the video frame to the upper layers.We highlight that due to the VBR nature of video traffic, Iframes are typically very large, while P frames are small.This implies that when an I frame needs to be transmitted,one or more slots will be filled up; when, instead, P framesare sent, a large portion of the slot will remain free and canthus be reused for best-effort traffic (see Fig. 1).

3 THE SUV PROTOCOL

The tenets of the solution we propose are as follows:

1. selection of relay nodes so as to maximize thecoverage area,

2. scheduling of relay nodes in TDMA fashion,3. scheduled access for streaming video,4. opportunistic access for streaming video,5. contention-based access for best-effort traffic.

In the rest of this section, we provide a detailed descriptionof the SUV solution.

3.1 Dynamic Children Selection and SchedulingAlgorithm

In our scheme, for every packet generated at the MAC layer,the gateway undertakes a set of actions with the purpose ofidentifying up to four relay nodes. The same set of actions isthen undertaken by each of the relays thus selected with thepurpose of selecting up to three more relays. Each newlyselected relay repeats the procedure until either a relay nodecannot find any schedulable neighbors, or the relay is askedto forward a slot content with a video frame sequence numberthat is smaller or equal to the one previously received. The setof relays thus selected represents the distribution structurethat is in charge of forwarding the content.

1086 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 22, NO. 7, JULY 2011

Fig. 1. Data channel access: time frame structure and an example of slotusage with three video descriptors (MD1, MD2, and MD3).

Next, we describe the children selection proceduretaking on the point of view of a generic relay node. Wewill refer to the relay node scheduling the tagged relay nodeand feeding traffic to it as parent node, while we will refer tothe relay nodes being scheduled and fed traffic by thetagged relay node as children nodes. The parent node andchildren nodes are expected to behave as described for thetagged relay node in their own turn.

The scheduling of children nodes by the relay noderequires the following: 1) the identity of its parent, 2) theposition and the received power level received from its 1-hopneighbors, and 3) the current scheduling status of the timeframe (i.e., the slots already scheduled by its parent node).The list of children nodes scheduled by the relay changeswith time; therefore, their choice is a dynamic process that isperiodically repeated. The length of the refresh period ischosen as twice the positioning system update period (e.g., 1 sin GPS). As will become clearer in the following, therefreshing of children node identities does not entailadditional overhead to the system.

The relay node will therefore partition the surroundingspace in four identical sectors at 90 degrees of each other;ideally, near the center of one of them its parent sits (parentsector). Near the center of the remaining three (childrensectors), the relay node will look for nodes who are eligibleto be children. The “sector center” in our case is a point onthe bisectrix of the angle formed by sector borders, and at adistance that satisfies two conflicting requirements: 1) closeenough that radio reception is not impaired and 2) farenough from the relay node to minimize cochannelinterference with other nodes scheduled in the same slot,while maximizing stream spacial advancement.

The orientation of the four-sector space may be chosen inseveral ways; one of the most efficient choices is thefollowing: the orientation of each sector is deterministicallychosen so that its bisectrix points toward one of the fourcardinal points. By selecting this orientation, the resultingdistribution topology formed by all relays has a grid-likestructure, where each relay ideally sits in between fourother relays (save for possibly missing relay nodes). Fig. 2shows one possible orientation as seen by a relay node (X)

in a sample topology. Sectors are tagged using theircardinal orientation (N, E, S, and W) with respect to X.X’s task is to select three children, one for each childrensector (W, N, and E since sector S is already taken up by itsparent node P ). We emphasize that, in general, in urbanvehicular scenarios the resulting topology is not a regulargrid structure due to the nodes mobility and road topology.Instead, in our dissemination system, a grid-like distribu-tion topology is created in the sense that every relay node isconnected to its parent and its (up to) three children todistribute the streaming traffic. Also, the children selectionand scheduling process is periodically updated to accountfor the varying topology.

The children selection leverages the availability of thefollowing information stored at each node: the list of 1-hopneighbors, their position, and the most recent value ofpower level received from them. In particular, similarly to[12], we define a range of preferred values of received signalpower as ðP ðthÞrx þ fout; P ðthÞrx þ fin�, where fout and fin arepositive values measured in dB. We set fout and fin inaccordance with the studies in [12], [10] so that thepreferred values correspond to the signal power oftransmissions that a node would receive from neighborsthat are neither too far away, so as to have good linkquality, nor too close, so as to let streaming traffic advanceby at least half the radio range. In particular, a relaymeasures the signal power received from its neighborsusing the HELLO messages they send to it.

Then, for each children sector, X chooses its (relay) childnode among its neighbors, according to the followingprocedure:

. select the node closer to the bisectrix and such thatthe associated received power level falls in the rangeof preferred values (shaded area in Fig. 2); if morethan one node satisfies the above condition, selectthe one associated with the lowest received signalpower, OR

. if no such node can be found, select the node closerto the bisectrix whose associated received powerlevel is the lowest among those that are greater thanP ðthÞrx þ fin (white area in Fig. 2), OR

. if no such node can be found, declare the sectorunschedulable.

Fig. 2 shows the outcome in the sample topology, i.e., thechoice of the three children nodes, each tagged as C. Blacknodes are neighboring nodes that were unsuitable, orsecond-choice, to become relays according to the aboveprocedure. For clarity, the width of the sectors in the figureonly depends on the distance from X (i.e., the receivedpower level isotropically decreases with the distance). Westress that our approach has several advantages: it max-imizes the coverage area while using few, reliable nodes; itallows for efficient, fully distributed scheduling algorithms;and it provides multiple redundant paths (since every relayschedules up to three children) that help coping with thevolatile connectivity typical of VANETs.

If no children sector is schedulable, or the relay is asked toforward a slot content whose video frame sequence numberis smaller or equal to one it has already received, the relaywill just transmit for the benefit of its own neighboringnodes, but will refrain from scheduling any children. If oneor more children are selected, the relay assigns their

SOLDO ET AL.: VIDEO STREAMING DISTRIBUTION IN VANETS 1087

Fig. 2. Orientation of the four-sector space. The received power fromnodes in the middle ring falls in the preferred range; the power receivedfrom nodes in the inner (white) area and the outer ring is, respectively,greater than P ðthÞrx þ fin and less than or equal to P ðthÞrx þ fout.

transmissions to a specific slot according to the “gridcoloring” scheme explained in Section 4. Then, within itsown transmission, it notifies the scheduling to the children.

3.2 Channel Access Rules: Scheduled Access forStreaming Video

When transmitting within its allocated slot, the relay nodeappends the following information to the MAC-layer header:

1. number of bits of each descriptor,2. identity of the three (or fewer) children,3. slot number in which each children is scheduled.

No explicit acknowledgment (ACK) is expected fromchildren. Rather, a passive ACK is obtained at the relaynode by monitoring transmissions in the slots where itschildren are supposed to be relaying the content. If childrentransmissions, relaying the correct content, are heard, thetraffic relaying is considered to be successful (see Fig. 3, firstrow). Otherwise, one of the following situations may arise:

1. No transmission is heard during the time slot: in thiscase, it is assumed that the child did not receive thecontent or its scheduling instructions; therefore, therelay node will try an “on-the-fly” rescheduling ofthe missing child within the same slot (see below,opportunistic access and second row in Fig. 3).

2. A garbled transmission is received: the relay nodemakes the assumption that the child has beencorrectly scheduled and has relayed the content.Indeed, the relay node assumes it did not receive itbecause of concurrent neighboring transmissionsthat only affected its own reception; the relaying isconsidered successful and no rescheduling isdeemed necessary.

3. A correct transmission is decoded, but its source is not thescheduled child: again, the relay node cannot makeany definite assumption since the scheduled childmay have successfully received and then lost achildren conflict, as explained below in case 1, or it

may not have received the relay node’s transmissionat all. No action is taken since the slot is occupied.

4. A correct transmission by the scheduled child is decoded,but the slot content has a higher video frame sequencenumber than the one transmitted by the relay: thissituation may arise if the child is scheduled by twoor more relay nodes at different levels in thedistribution tree (hence relaying just the slot contentwith the newer video stream). In this case, the relaynode swaps its own parent role with the child’s ifscheduled by it, thus forwarding the more recentstream feed (see the example in Fig. 4). The morerecent copy of the stream will then worm its wayoutward at the expense of the older copy of the stream.

If two or more relay nodes schedule two or more different

children within the same sector, their transmissions will

overlap. The following two children conflicts may arise:

1. the children are aware of the conflict (they heartransmissions from their own parent as well as fromother children’s parents). The conflict is thenresolved using the children node IDs: only the childwith the lowest ID will transmit, while the othersrefrain from it. This behavior will trigger one of thefour situations outlined above since one (or more) ofthe relay nodes will assume that its child has notreceived the transmission;

2. if the children are too far apart to hear each other’sparent, they will collide and, again, one of the fourabove situations will occur.

We finally remark that if a nonrelay node does not

receive a transmission due to collision (i.e., overlapping

transmissions by neighboring relays), no recovery action is

foreseen. However, the probability that this event occurs is

minimized using the grid coloring scheme in Section 4, as

shown by the good performance of SUV in [6].


Fig. 3. Children selection: F and s denote time frame and slot number,respectively; boxed letters indicate the scheduled child and arrows thetransmissions. At F ¼ 1, A transmits to B, and then, B to D. At F ¼ 10(s ¼ 0 and s ¼ 3), nodes have moved and B’s transmission to D fails.Since B does not hear D resending the slot content, it reschedules thetransmission in s ¼ 5 (opportunistic access), but it fails again. At F ¼ 11,B schedules C.

Fig. 4. F and s denote time frame and slot number, respectively; Pktindicates the content ID; dashed and solid arrows indicate a transmissionfrom older and newer flows, respectively; boxed letters indicate the nodescheduled by two parents. In the top scheme, D is scheduled by C, duringslot 7 of time frame 0, to transmit Pkt ¼ 11, in s ¼ 0 (of the following timeframe). However, E has a fresher copy of the stream and schedules D toforward Pkt in F ¼ 2, s ¼ 0. In the middle scheme, D receives Pkt ¼ 12from C (F ¼ 1, s ¼ 7), but it does not forward it, having already receivedPkt ¼ 13 from E. D forwards Pkt ¼ 13 in F ¼ 2, s ¼ 0. In the bottomscheme, C receives Pkt ¼ 13 from B (F ¼ 2, s ¼ 6) and forwards it for thebenefit of its own neighbors only, however C does not schedule any child,having already heard D send Pkt ¼ 13 (in F ¼ 2, s ¼ 0).

3.3 Channel Access Rules: Opportunistic Accessfor Streaming Video

Opportunistic access within a time slot occurs in thefollowing situation: a relay node does not hear one of itschildren using the slot where it was scheduled fortransmission. Likely, a node close to the silent child hascollided with the relay node transmission. The relay nodethen will try and start a contention procedure to claim theslot and, at the same time, to salvage its latest transmissionby sending one or two out of the three (parts of the)multiple-description video frames.

The time slot is therefore used as follows:

. a grace period (similar to IEEE 802.11 DIFS) neededto declare the medium as idle,

. an RTS/CTS exchange between the relay node andthe silent child to reclaim the slot,

. a leftover period carrying the transmission of one ortwo (depending on the length of the RTS/CTSexchange) of the multiple descriptors.

The RTS/CTS exchange provides for contending relaynodes to claim the leftover slot through a slotted Alohaprocedure [13]. After sending the RTS, if the contending relaynode hears a CTS addressed to itself, it will use the leftoverslot time to transmit two out of three of the descriptors.

If no CTS is heard, the contending relay node will start overthe RTS/CTS procedure one last time, in the hope of sendingat least a single descriptor. Clearly, two or more contendingnodes will keep competing for the same slot in every timeframe, till one of them moves out of range (the latter is afrequent occurrence due to the highly mobile environment).

3.4 Channel Access Rules: Contention-BasedAccess for Best-Effort Traffic

If a time slot is underutilized, i.e., a relay node cannot fill upthe time slot with the streaming video available at the timeof transmission, the residual slot time is allocated tocontention-based access. The sender is assumed to advertisethe slot portion that it is about to use through a MACheader field. All nodes within the radio range of thetransmitter are entitled to contend for the residual slot time,provided they correctly decode the slot occupancy indica-tion. A node willing to transmit must first perform apostbackoff as in standard 802.11 DCF and, if the residualslot time at the end of postbackoff is sufficient for thetransmission of at least one minimum-size data packet, itaccesses the medium as in DCF.

4 GRID COLORING

We now describe the procedure to dimension the schedulingof relay nodes in SUV. We draw on the fact that the networktopology composed of relay nodes has a grid-like structure.We first consider a regular grid topology, i.e., every node hastwo neighbors along each spatial dimension and thatneighboring nodes are all at the same distance R from eachother. Under such a network scenario, we formulate thescheduling problem and provide a solution that is proven tobe optimal for all cases of practical interest. The obtainedscheduling scheme is then applied to a realistic VANET,where, in general, the distance between relay nodes is shorterthan R and relays form an irregular grid.

As a first step, we model the network of relay nodes as an

undirected graph G ¼ ðV ;EÞ, where V and E are the set of

vertices and edges, respectively. We associate each node with

a vertex, and we say that an edge between any pairs of nodes,

u; v 2 V , exists if their distance is less than or equal toR. Also,

let the set of radio resources available in the system (namely,

the number of time slots within a time frame) be represented

as a set of colors C, and let C be the set cardinality. Then, a

proper distance-k C-coloring (or briefly ðC; kÞ-coloring) of

graphG is a mapping � from the set V into the set of available

colors C s.t. �ðuÞ 6¼ �ðvÞ, 8u, v 2 V connected by a shortest

path of at most k hops, with k > 0.Based on the above considerations, we formulate our

scheduling problem as follows:P1. Given an undirected graph G ¼ ðV ;EÞ and a set of

colors C, find a mapping � : V ! C, s.t.

max�

nmin

u;v2V ;�ðuÞ¼�ðvÞlu;v

o; ð1Þ

where lu;v is the length of the shortest path connecting nodes

u and v, in number of hops.For arbitrary topologies and values of C and k of practical

interest, ðC; kÞ-coloring problems, and hence, P1 are known

to be NP-complete [14]. However, in the following, we show

that, in the case of a regular grid topology, P1 can be reduced

to a problem solvable in polynomial time.Consider a regular grid in the x,y-plane. Let us denote each

color with a non-negative integer (i.e., C ¼ f0; 1; . . . ; C � 1g);also, let ZC be the set of non-negative integers modulo C.

Assume that the grid vertices are assigned colors according to

the following method that we name constant-step coloring.

Given two adjacent nodes on the x-dimension, ux and vx,

we have:

�ðvxÞ ¼ �ðuxÞ þ aðmod CÞ;

while for any pair of adjacent nodes on the y-dimension, uyand vy, we have:

�ðvyÞ ¼ �ðuyÞ þ bðmod CÞ;

with a; b 2 ZC . As an example, in Fig. 5, we show a grid-

structure topology whose nodes have been colored using

the constant-step method with C ¼ 5, a ¼ 1, and b ¼ 2.It can be proven that the following lemma holds (the

proof can be found in [15]):

Lemma 4.1. Given C and a constant-step coloring with steps

a; b 2 ZC , for any pair of nodes, u; v 2 V , we have

�ðvÞ ¼ �ðuÞ þ ðasþ btÞðmod CÞ;

where jsj þ jtj is the length of the shortest path connecting u

and v, expressed in number of hops.


Fig. 5. Example of constant-step coloring, with C ¼ 5, a ¼ 1 and b ¼ 2.

This lemma tells us how to assign colors to the gridvertices, and it allows us to prove the theorem below.

Theorem 4.1. Under the assumption of constant-step coloring and

of a regular grid, P1 is equivalent to the following problem:P2. Given a set of colors, C, with cardinality C 2 N, find

integers a; b; s; t s.t.

maxa;b2ZC

mins;t2Zðjsj þ jtjÞ

� �; with ð2Þ

asþ bt ¼ 0 ðmod CÞ: ð3Þ

The proof can be found in [15]. Observe that, for fixed a and b,(3) is a Diophantine equation1 in the indeterminates s and t.

It can be shown that (3) admits integer solutions if andonly if C ¼ 0 ðmod gcdða; bÞÞ, where gcdða; bÞ is the greatestcommon divisor between a and b, and that infinite integersolutions exist. A single solution ðs0; t0Þ can be obtained,using, for instance, the Extended Euclidean algorithm [16];then, all other solutions can be derived through a lineartransformation on ðs0; t0Þ.

Based on these observations, we present the followingsimple, but very efficient, algorithm to solve P2.

Given a fixed number of colors, C; do

1. for every a; b ¼ f1; 2; . . . ; C � 1g

a. find an integer solution ðs0; t0Þ (if any) to (3)(e.g., using the Extended Euclidean algorithm);

b. if a solution exists, find the other solutions ðs; tÞwith the linear transformation

s ¼ s0 � bH; H 2 Z;

t ¼ t0 þ aH;

c. among all solutions obtained for the consideredða; bÞ, choose ð�s; �tÞ that minimizes k ¼ jsj þ jtj.

2. Select the pair of integers ða; bÞ for which k ismaximized.

Hence, we obtain the values of the steps ða; bÞ that allowto color all grid vertices, while maximizing the shortest path(expressed in number of hops) between any two nodesassigned with the same color.

We point out that, at least for distance values k of practicalinterest (namely, k ¼ 1; . . . ; 4), the proposed constant-stepcoloring optimally solves the distance-k coloring problemover a grid topology, that is, it finds the minimum number ofcolors needed to color the grid. This can be easily proven

through geometrical considerations (see [15]). Table 1 reportsthe minimum values of C that yield the distance k alongwith the euclidean distance between any two nodes assignedwith the same color. It can be shown (see [15]) that,considering a; b < C, the overall running time complexityof the algorithm is upper bounded by OðC2 logCÞ, which isvery limited considering the values of C of practical interest.

Now, let us apply the above results to the systemdimensioning and the scheduling problem in our VANET.Consider that the number of colors represents the number ofslots, S, within a single time frame and that relay nodesassigned with the same color are scheduled for transmissionin the same slot. Clearly, it is of crucial importance to thenetwork performance to maximize the spatial separationbetween vehicles transmitting simultaneously. In particular,recall that the leftover part of a slot used by node v for videotransmission can be accessed by v’s neighbors for best-efforttraffic. Other relay nodes assigned to the same slot as v mayinstead fill up the slot with their streaming transmission. Inthis case, to avoid collisions, the distance between v and anyother relay node scheduled in the same time slot within thesame time frame should be strictly greater than three timesthe maximum radio range (i.e., k ¼ 3). Theoretically, thiswould ensure that, even if a 1-hop neighbor of thescheduled relay uses the leftover room for sending best-effort traffic, it will not interfere with other relay nodesscheduled in the same time slot.

Looking at Table 1, we have that the minimum number ofslots (colors) needed to achieve such a separation is equal toeight. We therefore consider that each time frame includes atleast S ¼ 8 slots and each relay node schedules its childrenaccording to the constant-step coloring, as in Lemma 4.1.

We point out that since the vertices of the forwarders gridare chosen at variable distance depending on propagationconditions and node availability, the number of colors neededto achieve separation may be higher. However, in oursimulations [6], S ¼ 8 proved to be an excellent trade-offbetween interference mitigation and radio resource efficiencyshowing that the theoretical results derived under ouradmittedly simplistic assumptions yield a practical payoff.

An example of constant-step coloring applied to aVANET with S ¼ 8 is shown in Fig. 6. In a perfect grid, ithas been proved that the constant-step coloring algorithmachieves maximum spatial separation between concurrent


TABLE 1Parameters Characterizing a Grid-Like Topology Coloring Using

the Constant-Step Algorithm

1. A general Diophantine problem consists in a system of polynomialequations with more unknowns than equations, with the additionalconstraint that variables can take only integer values.

Fig. 6. Example of scheduling in the VANET: constant-step coloring witheight colors. The optimal solution is a ¼ 1, b ¼ 5, see Fig. 1. Below eachvehicle, the assigned slot (i.e., color) is indicated. In this example, fourhops separate two vehicles scheduled for slot 0; indeed, (3) can besolved by setting s ¼ t ¼ 2 (i.e., jsj þ jtj ¼ 4 and 2 � 1� 2 � 5 ¼ 0;mod 8);also, no solution exists for an s; t pair such that jsj þ jtj < 4.

nodes (i.e., nodes assigned to the same slot). As illustrated

in the example, though, even in an irregular grid-like

structure, the algorithm still exhibits large separation

between vehicles that are assigned the same slot.

5 CONCLUSION

We addressed the issue of on-board live video streaming in

VANETs by providing a comprehensive solution, SUV,

spanning several architectural layers. The content distribu-

tion is achieved through a fully distributed, dynamic

selection of forwarders which, in turn, perform local

broadcasting. Graph coloring is exploited to solve the

scheduling problem and optimize the spacial separation of

concurrent streaming transmissions. Whenever a collision

occurs, SUV leverages the properties of video coding to

design a collision-resolution mechanism and the character-

istics of VBR traffic to efficiently exploit radio resources.

ACKNOWLEDGMENTS

This was supported by the European Union through the

“myMED” project funded through the ALCOTRA action.

REFERENCES

[1] F.J. Ros, P.M. Ruiz, and I. Stojmenovic, “Reliable and EfficientBroadcasting in Vehicular Ad Hoc Networks,” Proc. IEEEVehicular Technology Conf. (VTC) Spring 2009, pp. 1-5, Apr. 2009.

[2] W. Chen, R.K. Guha, T.J. Kwon, J. Lee, and Y.-Y. H. Wirel, “ASurvey and Challenges in Routing and Data Dissemination inVehicular Ad Hoc Networks,” Wireless Communications and MobileComputing, Oct. 2009.

[3] M. Guo, M.H. Ammar, and E.W. Zegura, “V3: A Vehicle-to-Vehicle Live Video Streaming Architecture,” Proc. IEEE Int’lConf. Pervasive Computing and Comm. (PerCom), pp. 171-180,Mar. 2005.

[4] M. Bonuccelli, G. Giunta, F. Lonetti, and F. Martelli, “Real-TimeVideo Transmission in Vehicular Networks,” Proc. IEEE MobileNetworking for Vehicular Environments (MOVE Workshop), pp. 115-120, May 2007.

[5] Y.-C. Chu and N.-F. Huang, “Delivering of Live Video Streamingfor Vehicular Communication Using Peer-to-Peer Approach,”Proc. IEEE Mobile Networking for Vehicular Environments (MOVEWorkshop), pp. 1-6, May 2007.

[6] F. Soldo, C. Casetti, C.-F. Chiasserini, and P. Chaparro, “SUV:Related Work and Performance Evaluation,” technical report,Politecnico di Torino, available at http://www.telematica.polito.it/casetti/TechRep_SUV_Performance.pdf, Mar. 2010.

[7] X. Bai, D. Xuan, Z. Yun, T.H. Lai, and W. Jia, “Complete OptimalDeployment Patterns for Full-Coverage and k-Connectivity(k � 6) Wireless Sensor Networks,” Proc. ACM MobiHoc, pp. 401-410, May 2008.

[8] R. Mangharam, R. Rajkumar, M. Hamilton, P. Mudalige, and F.Bai, “Bounded-Latency Alerts in Vehicular Networks,” Proc. IEEEMobile Networking for Vehicular Environments (MOVE Workshop),pp. 55-60, May 2007.

[9] V. Naik, A. Arora, P. Sinha, and H. Zhang, “Sprinkler: A Reliableand Energy Efficient Data Dissemination Service for Extreme ScaleWireless Networks of Embedded Devices,” IEEE Trans. MobileComputing, vol. 6, no. 7, pp. 777-789, July 2007.

[10] P. Barsocchi, G. Oligeri, and F. Potortı̀, “Frame Error Model inRural Wi-Fi Networks,” Proc. IEEE Int’l Symp. Modeling andOptimization (WiOpt), WiNMee/WiTMeMo Workshop, pp. 41-46,Apr. 2007.

[11] F.H.P. Fitzek, B. Can, R. Prasad, and M. Katz, “Traffic Analysisand Video Quality Evaluation of Multiple Description CodedVideo Services for Fourth Generation Wireless IP Networks,”Wireless Personal Comm., vol. 35, nos. 1/2, pp. 187-200, 2005.

[12] V. Naumov, R. Baumann, and T.R. Gross, “An Evaluation of Inter-Vehicle Ad Hoc Networks based on Realistic Vehicular Traces,”Proc. ACM MobiHoc, pp. 108-119, May 2006.

[13] D.P. Bertsekas and R.G. Gallagher, Data Networks. LongmanHigher Education, 1987.

[14] A. Sharp, “Distance Coloring,” Proc. 15th Ann. European Conf.Algorithms, 2007.

[15] F. Soldo, C. Casetti, and C.-F. Chiasserini, “Constant-Step Color-ing,” technical report, Politecnico di Torino, available at http://www.telematica.polito.it/casetti/TechRep_Scheduling.pdf, Mar.2010.

[16] E. Bach, Algorithmic Number Theory. The MIT Press, 1996.

Fabio Soldo received the BS degree in mathe-matics from Politecnico di Torino, Italy, in 2004,the MS degree in mathematical engineeringfrom Politecnico di Torino and Politecnico diMilano, Italy, in 2006. He is currently workingtoward the PhD degree in the NetworkedSystems Program from UC Irvine. He hadinternships with Telefonica Research, NTTDocomo Labs, and Google. His research inter-ests are in the areas of design and analysis of

network algorithms and network protocols, defense mechanisms againstmalicious traffic, and wireless networks. He is a member of the IEEE.

Claudio Casetti (M’05) received the graduateddegree in electrical engineering in 1992 fromPolitecnico di Torino, where he received the PhDdegree in electronic engineering, in 1997. He isan assistant professor in the Dipartimento diElettronica e Telecomunicazioni, Politecnico diTorino. He has coauthored more than 120 journaland conference papers in the fields of networkingand holds three patents. His interests focus on adhoc wireless networks and vehicular networks.

He is a member of the IEEE.

Carla-Fabiana Chiasserini (M’98-SM’09) re-ceived the electrical engineering degree (summacum laude) from the University of Florence, Italy,in 1996, and the PhD degree from the Politecni-co di Torino, Italy, in 1999. Since then, she hasbeen with the Department of Electrical Engineer-ing, Politecnico di Torino, where she is currentlyan associate professor. From 1998 to 2003, shehas worked as a visiting researcher at theUniversity of California, San Diego. Her research

interests include architectures, protocols, and performance analysis ofwireless networks for integrated multimedia services. She has publishedmore than 170 papers in prestigious journals and leading internationalconferences, and she is the coauthor of three patents. She serves as anassociate editor of Ad Hoc Networks Journal (Elsevier), ComputerCommunications (Elsevier), IEEE Wireless Communications Magazine,IEEE Transactions on Wireless Communications, and IEEE Commu-nications Letters. She is a senior member of the IEEE.

Pedro Alonso Chaparro received the telecom-munications engineer degree from Santo TomasUniversity, Bucaramanga, Colombia, in 2002,and currently he is working toward the PhDdegree in telematics engineering from thePolytechnic University of Catalonia. In 2007, heworked in the Electronic Department at Politec-nico di Torino, Italy. In 2008, he worked in theComputer Network Group at the PolytechnicUniversity of Valencia, Spain. Since November

2008, he has been with the Telecommunications and MultimediaApplications Institute (iTEAM) at the Polytechnic University of Valencia.His research interests are in video transmissions over mobile networksand fourth generation wireless networks.

. For more information on this or any other computing topic,please visit our Digital Library at www.computer.org/publications/dlib.


Documents

Video Streaming Distribution in VANETs