On Shooting a Moving Vehicle with Data Flows

Transcription

1 On Shooting a Moving Vehicle with Data Flows Seokhoon Yoon, Hung Q. Ngo, and Chunming Qiao Compute Science and Engineeing State Univesity of New Yok at Buffalo {syoon4,hungngo,qiao}@cse.buffalo.edu Abstact This pape poposes an ITS system that uses not only integated cellula and ad hoc elaying technologies, but also uses mobility pofiles. It focuses on a subsystem whee cellula bandwidth is used mostly fo contol signals, and an ad hoc distibution netwok (ADN) is used fo file downloading and video steaming. The pape fomulates an optimization poblem with the objective being to maximize the amount of data to be deliveed to a moving vehicles via the ADN, and theeby minimizing the usage of the costly cellula bandwidth fo data tansfe. Thee appoaches based on non-linea and linea pogams ae descibed and compaed, and simulation esults indicate that a satisfying pefomance fo file download applications can be achieved. I. INTRODUCTION We envision an achitectue fo ITS in metopolitan aeas that integates the cellula and ad hoc elaying technologies. The poposed achitectue, called icar-mopa 1, will povide vaious sevices to vehicles, which include file downloading, stoed o eal-time video steaming, and customized taffic alet as well as adaptive navigation. In this wok, we focus on a subsystem fo the file download sevices, but the concepts and techniques discussed can be adapted fo othe types of applications such as video steaming. In addition to exploiting the integated cellula and ad hoc elaying technologies, anothe unique featue of the poposed achitectue is that it incopoates knowledge of mobility pofiles of vehicles. Heeafte, we efe to the subsystem unde consideation as mobility pofile aided downloading systems o icar-mopads. In the envisioned icar-mopads (and icar-mopa), vehicles ae equipped with both cellula (e.g., 3G WCDMA) and ad hoc elaying (e.g., 82.11) intefaces as well as GPS. One o moe seve stations (SS) at distibuted locations have connections to a high-bandwidth and elatively low-cost-pebit backbone (e.g., a wieline netwok). The cellula netwok is also integated with the backbone so the files on the SS ae accessible though the cellula channel. In addition, many access points based on low-cost wieless technologies (e.g ) equipped with stoage, which ae simila to the infostations in [1], [2], ae placed at stategic locations such as majo intesections thoughout the metopolitan aeas. Unlike most of the poposed system, these APs ae not diectly connected to the same wieline backbone as 1 whee icar and MOPA stand fo Integated Cellula and Ad Hoc Relaying Systems and Mobility Pofiled Assistance espectively SS AP Wieles Mesh Netwok ICMAN Fig. 1. To the Backbone Cellula Towe An Ad hoc Distibution Netwok the SS to avoid the expected high cost. Instead, some of the APs can be connected with each othe (and also with the SS) via a wieless mesh netwok. Othe APs may fom a pat of an intemittently connected mobile ad hoc netwok (ICMAN [3]) whee data fom one AP (o SS) to anothe AP must be caied by vehicles moving fom the fome towad the latte. Futhemoe, with mobility pofile infomation of the subscibe, a vehicle can also download data fo anothe vehicle by acting as a mobile AP deliveing the data at thei encounte. In eithe case, deliveing data though such a netwok, heeafte efeed to as the ad hoc distibution netwok o ADN, is expected to povide a highe bandwidth at a lowe cost pe bit than using a cellula channels. Fig. 1 shows an example of an ADN whee a vehicle downloads data via and/o the cellula channel. A basic idea behind the envisioned icar-mopads is that vehicles (o in geneal othe mobiles) can download a file stoed on the SS o cached at some APs when the vehicle passes by an AP using the ad hoc (e.g., based) inteface, instead of the cellula inteface. Vehicles can also download a file using thei cellula intefaces but the cost pe bit using a cellula channel is expected to be much highe and thus the usage of the cellula channel fo file downloading is to be minimized. Nevetheless, the APs and vehicles will use cellula channels fo the tansmission and eception of contol signals which equie a elatively low bandwidth. When a icar-mopads sevice subscibe stats a tip, he/she may equest via a cellula channel fo a file (a new piece of softwae o a video fo example). Ideally, icar-mopads can lean, based on diect input o mobility pofile of the use,

2 the oute to be taken by the use and calculate its estimated time to aival (ETA) at each and evey majo intesection along the oute 2. icar-mopads detemines the appopiate amount of data to be pushed fom one o moe SS to the APs at those intesections so that the vehicle can download the data when it aives (and passes by). Fo a given AP i and a vehicle that passes by the AP, the SS detemines the optimal amount of data, denoted by x i, to be pushed so that the expected amount of data downloaded by the vehicle is maximized. The expected amount of data that the vehicle could download at the AP depends on the pobability of the vehicle passing by the coesponding intesection, the expected duation of AP association time and the expected download ate of the AP. Since the data has to aive at an AP befoe the vehicle does, x i also depends on the status of the ADN, including its link tansmission speed and latency, cuent queuing delays and infomation about othe vehicles (including thei equests fo file download and outes). If the vehicle fails to download all the data of a file fom an AP oiginally planned by icar-mopads, the cellula channel bandwidth (in addition to that devoted to contol signals) is used to delive the est of the data as needed. To simplify ou pesentation in this pape, we will focus on the poblem of deliveing a maximum amount of data to a vehicle via the ADN (wieless mesh netwoks of APs and ICMAN), theeby minimizing the cost associated with the use of cellula channel fo data. Moeove, we will only conside a simple outing stategy in ADN whee SS uses a fixed, pecomputed path to oute data to each AP, obviously leaving many open poblems elated to outing in a wieless mesh netwok o via elaying by othe vehicles (as needed in an ICMAN netwok) fo futue woks. The simplified vesion of the poblem is thus essentially analogous to the poblem of shooting a moving vehicle to maximize the numbe of bullets hitting it. We believe that this is the fist time that such a poblem is being investigated. The est of the pape is oganized as follows. We fist fomally descibe the poblem in Section II. We then popose thee appoaches to fomulate the optimization poblem in Section III. Section IV pesents ou simulation esults, and Sections V and VI discuss elated studies and concludes the pape espectively. II. PROBLEM DESCRIPTION Conside a vehicle whose oute includes APs v 1, v 2,..., v m in that ode which can be estimated fom the mobility pofile of the vehicle. Assume that, at time t=, the vehicle makes a equest fo a file of size F bytes stoed at SS via a cellula channel. Let T i be the (estimated) time that the vehicle will be at v i. Note that in pactice the numbe of bytes the vehicle can download fom an AP vaies, and has a cetain uppe-bound say b i. The basic idea of the envisioned icar-mopads is to patition the file into seveal pats, x i <= b i, whee i = 2 This is staightfowad if the icar-mopa also povides an adaptive navigation sevice that the use subscibes to. 1, 2, m and m x i F, and ty to push x i bytes fom SS to v i befoe the vehicle aives at T i. Poblem 1 (Shooting a Moving Vehicle SMV). Find file segmentation and tansmission schedule at SS that maximizes the expected amount of data that a vehicle will eceive on its tip. Let X i be the andom vaiable denoting the time (stating fom ) at which AP v i eceives the x i bytes sent fom SS. Note that X i not only is a function of x 1,, x m, but also depends on the ode in which SS sends these messages out (since the outes these messages take may ovelap). Note also that depending on vaious pactical conditions (e.g. taffic light, diving speed in the AP coveage zone including the likelyhood that the vehicle will stop at a ed-light), a vehicle may only be able to download data available fo it fom v i with pobability p i. Let B (a andom vaiable) be the total numbe of bytes the vehicle eceives on this tip, then ou objective function is E[B] = x i p i Pob[X i T i ], (1) subject to x i b i fo all i, and m x i F. III. APPROACHES We pesent thee appoaches fo fomulating the afoementioned SMV poblem with non-linea and linea pogams (NLP and LP). To simplify the poblem, we assume that thee is only one seve station SS which has the data equested by uses and outes data to each of the n APs using a fixed path. The union of the outing paths foms a diected tee T ooted at SS. In addition, in all thee appoaches, we will assume that the tansmissions of messages fom SS ae in the ode x 1,..., x m (though this assumption may not lead to optimal solutions). Appoach 1. Solve (1) diectly with an NLP optimization softwae. In ode to ealize this appoach, an explicit fomula fo Pob[X i T i ] has to be deived. As a fist appoximation, we assume that each AP opeates as an M/M/1 queue whose sevice time is exponential with mean 1/µ. In addition, data geneated at SS in esponse to downloading equests fom vehicles is Poissonian with ate λ. At equilibium, we assume that a file segment coming into a node v has equal pobabilities of going out on any outgoing edge of v. The entie netwok is thus an open Jackson netwok. The data aival ate on each edge is thus Poissonian and can be ecusively computed using flow consevation equations. In paticula, if the incoming data to node v (though edge e v ) is Poissonian with ate λ v, then the flow though each outgoing edge of v is also Poissonian with ate λ v /out-deg(v). Let P i be the path in the outing tee fom SS to v i. Let k i be P i s length. Denote the nodes (i.e. APs) on this path as follows: SS = a i, a i 1,..., a i k i = v i. Fo each node a i j along P i, let λ i j be the mean incoming flow ate fo its queue, and ρ i j = λi j /µ be the taffic intensity, both of which depend on

3 x i ; also let Wj i be the andom vaiable denoting the queuing time at the node. Then, it is standad that Pob[W i j t] = 1 ρ i je (µ λi j )t, t, (2) E[W i j ] = Va [ Wj i ] ρ i j µ λ i, (3) j = (2 ρi j )ρi j (µ λ i. (4) j )2 Also, fo 1 j k i, let d i j be the popagation delay fom a i j 1 to ai j. Obviously, we must have k i j=1 di j < T i in ode fo the data to have any chance to aive at v i in time. When ki j=1 di j >= T i, icar-mopads will simply set x i =, namely no data will be sent to v i fo this equest. Let be the aveage tansmission speed of links along P i. Then, X i = x ik i + d i j + Wj i (5) j=1 Note that the fist tem xiki is simply the sum of the tansmission times of all nodes befoe v i. Howeve, X i will be longe if thee ae othe data to some othe node v j sent ealie by the SS which also use the fist few hops along P i (i.e. when P i and P j shae a pefix). Even if one finds a way to take into account the effect of peviously tansmitted data x j (j < i) on X i, in ode to compute Pob[X i T i ], we need to compute the distibution of the sum k i W j i. This is computable in closed fom, but the fomula is faily complex, involving the sum of vaious exponential functions on the x i. Accodingly, the poblem has an NLP fomulation. Since solving this poblem may be costly in time, we pesent two othe simple appoaches as follows. Appoach 2. To avoid the costly computation needed to optimize the complicated, non-linea objective function in (1), in this appoach we ty to appoximate the optimization poblem by putting a theshold δ on each failue pobability Pob[X i > T i ], and then optimize the expected numbe of bytes eceived by the vehicle. This appoach is quite consevative. We expect it to wok bette when the delays fluctuate widely. In this case, the optimization poblem becomes max k p i x i subject to Pob[X i > T i ] δ, i 1 i m (6) x i b i, i 1 i m x i F. This is not yet a linea pogam since Pob[X i > T i ] is non linea on the x i. To tun it into an LP, we can uppebound Pob[X i > T i ] using the one-sided Chebyshev s inequality, and then enfoce the theshold δ indiectly on the uppebound instead of diectly on Pob[X i > T i ]. This is doable since X i is a linea function of the waiting times in the queue along the path to v i, and since ou queueing netwok is an acyclic open Jackson netwok [4] [6], implying the waiting times ae independent. Thus we can easily compute both E[X i ] and Va [X i ] using (2), (3), and (4). The uppebound of Pob[X i > T i ] will then be a linea combination of the vaiables x 1,..., x m. As an example of how to find an uppebound fo Pob[X i > T i ], conside a path P i which shaes no pefix with any othe P j, implying X i can be expessed as in (5). We have, ν i := Va µ i := E W i j W i j = = (2 ρ i j )ρi j (µ λ i j )2 ρ i j µ λ i j We will only conside the case when T i µ i + k i j=1 di j, othewise we set x i =. By Chebyshev s inequality, we have Pob[X i > T i ] = Pob Wj i > T i x ik i ν i j=1 d i j ( ν i + T i x ik i k i j=1 di j µ i Consequently, Pob[X i > T i ] δ will hold if x i k i T i d i j µ i ν i ( 1 k i δ 1) (7) j=1 This constaint can then be used to substitute the constaint in (6) (as an appoximation), which esults in an LP fomulation (whose detailed desciption is omitted due to space limitation). Appoach 3. A potential poblem with Appoach 2 is that the Chebyshev inequality used to uppe bound Pob[X i > T i ] makes it quite consevative. In this Appoach 3, we think of the waiting times at all queues as constants. Moe specifically, we assume that each node a i j along the fixed outing path fom SS to AP v i peiodically sends its queuing delay to the SS using a cellula channel. The SS uses the mean qj i of the epoted queuing delays fo node a i j to appoximate the actual queuing delay that packets will expeience. This way, the SS can estimate the appoximate value of X i as a linea function ) 2

4 of the x i. Then, we can fomulate a linea pogam as follows: max p i x i subject to k i (d i j + qj 1 i + x i ) + j=1 (8) k i 1 i 1 x u I(j, i, u) T i δ, 1 i m (9) u= { 1 if a i I(j, i, u) = j = a u j ; (1) othewise x i b i, 1 i m (11) x i F (12) i= whee δ is a system paamete which povides a safety magin to ensue data aives ealie than the vehicle, and F is total amount of data equested by a use. In the constaints of the linea pogam fomulation, constaint (9) states that the data x i should be deliveed to v i befoe the estimated aival time of the vehicle at the AP minus the safety magin δ. Note that if the actual delay time distibutions ae highly concentated with a small tail, a small value of δ is sufficient to mimic the theshold constaint in (6) without having to become consevative as in Appoach 2. The fist tem of (9), i.e. tem (8), is the sum of the popagation, tansmission, and queueing delay along the path to v i. The next tem of (9) appoximates the total additional queueing delay fo x i caused by the pevious tansmission of data, {x 1,, x i 1 }. Note that, I(j, i, u) in (9) and (1) indicates whethe o not P i and P u shae an AP. The constaint (11) keeps the SS fom tansmitting data moe than the maximum amount of data a vehicle can download at v i. Also, the sum of x i can not exceed the total amount of data equested as epesented in (12). One may be able to send moe data than equested (with some sot of coding/eplication algoithm such as easue codes) to incease the expected amount of deliveed data. This appoach is an open eseach diection. IV. NUMERICAL RESULTS In this section, we evaluate Appoach 3 (selected mainly fo its feasibility) using simulation. The basic idea is that fo each equest, we solve the LP in Appoach 3 to detemine the optimal x i (1 i m). The esulting x 1,..., x m ae then fed into a simulato, along with othe input paametes, to obtain pefomance esults. The basic simulation set-up is as follows. A total of 25 APs ae deployed foming a 5-by-5 gid with a distance of 75m between hoizontally and vetically adjacent APs to mimic a 3km 3km section of cental Manhattan. They ae connected to the SS located at the cente of the egion with a tee topology (diffeent tee topologies ae simulated), and in each of such tee topologies, only two geogaphically neighboing APs in the gid may be connected with a tee banch. In all the tee topologies simulated, SS is connected to fou close-by APs with one tansmitte (and queue) pe % of each equested file sent by SS % of each equested file aived in time Numbe of Requests within 2 minutes 2 (a) 4 6 Numbe of Requests within 2 minutes (b) Fig. 2. Pecentage of the equested file sent to and eceived by v 1 v 5 AP. Each AP may be connected to to 3 othe downsteam APs but has only one tansmitte. We assume that an AP can eceive (fom upsteam) and tansmit (to downsteam) data simultaneously using two sepaate channels. Note that although the simulation set-up models the ADN mainly with a wieless mesh (with diectional antenna), it can also be adapted, by changing the assumptions on the popagation delay and link tansmission speed, to simulate othe ADNs, e.g., an ICMAN whee data fom one AP can be elayed to anothe AP in a multi-hop fashion using othe vehicles. In this study, the time-vaying tansmission ate (in Mbps) of each AP anges fom high, medium and low, modeled using a nomal distibution with (µ, σ) = {(3, 5), (2, 3.4), (1, 1.7)} espectively. The popagation delay of each link anges fom low ( ms epesenting the time fo a signal to tavel between two APs at light speed), medium (1ms on aveage, taking into consideation the delay caused by 5-hop stoe-and-fowad elay among vehicles between two APs that ae located 75m apat, assuming each stoe-andfowad delay is 2ms) and high (25ms on aveage, taking into consideation about 12-hop elay). Within a total simulation time of 2 minutes, a numbe of equests (up to 8), one fom each vehicle, wee geneated at andom intevals. Each vehicle tavels along a shotest path that includes exactly 5 APs (v 1,..., v 5 ), whee v 1 and v 5 ae selected abitaily, and making a andom decision at each possible intesection (e.g., a vehicle going Noth- East diagonally will andomly choose Noth o East at an 1st AP 2nd AP 3d AP 4th AP 5th AP 8 1st AP 2nd AP 3d AP 4th AP 5th AP 8

5 % of each equested file deliveed via ADN pd:~ms, t:3mbps pd:1ms, t:2mbps pd:25ms, t:1mbps Numbe of equests within 2 minutes Avg. Delivey Rate via ADN pe tip (kbits/sec) pd:~ms, t:3mbps pd:1ms, t:2mbps pd:25ms, t:1mbps Numbe of equests within 2 minutes Total numbe of bytes via ADN (MBytes) pd:~ms, t:3mbps pd:1ms, t:2mbps pd:25ms, t:1mbps Numbe of equests within 2 minutes (a) Pecent of equested data deliveed via ADN (b) Effective delivey ate via ADN (c) Total numbe of bytes deliveed via ADN Fig. 3. Simulation Results intesection as long as it will get close to the destination). The moving speed between two APs (in metes pe second) follows a nomal distibution with (µ, σ) = (1, 3), and the sojoun time of a vehicle at an AP (in seconds) also follows a nomal distibution with (µ, σ) = (3, 3). This means that on aveage, a tip will take 525 seconds. We also assume that each vehicle equests fo a file whose size is exactly F = 4Mbytes, and that, fo each equest, SS will send at most b i = 2Mbytes to each AP (v i ) to be visited by the vehicle in an attempt to avoid oveloading cetain links and APs in the ADN. As fo pefomance metics, we ae inteested in knowing the following values aveaged ove all vehicles and all simulation uns: (a) the pecentage of each equested file sent by SS to each v i (accoding to the LP solve), which is x i ove 4Mbytes; (b) the pecentage of each equested file eceived by each v i ; (c) the pecentage of each equested file deliveed to all the five APs in time and deliveed to a vehicle via the ADN; (d) the effective delivey ate and the total numbe of bytes downloaded to a vehicle via ADN. Fig. 2 shows the esults obtained unde high tansmission ate (3Mbps) and low link popagation delay ( seconds). Fom Fig. 2(a), we can see that at a low load (e.g., 5 equests within 2 minutes), SS evenly distibutes the amount of data (x i ) to APs, because SS has enough time to tansmit data to all APs including APs close to the cuent location of the vehicle. Also, almost all data seems to get deliveed by the time a vehicle eaches its fifth AP. These esults have positive implications fo video steaming applications (as a efeence point, 4Mbytes of video data can be played fo 64 seconds at the playing speed of 5kbps which gives a decent pictue quality/size). As the load inceases (e.g. 4 equests), howeve, SS is foced to send moe data to the APs fathe along the path to be taken by a vehicle due to the inceased delivey delay (mainly esulted by the high queuing delay). At a vey high load (8 equests), SS sends most data to the last AP of the vehicle s jouney. Also note that if SS is not esticted to send data in the ode of x 1,..., x 5, moe data may be deliveed to the fifth and fouth APs along a vehicle s paths, and the total amount of data deliveed may also incease. As shown in Fig 2 (b), the distibution of the eceived data by v 1 v 5 is almost same as that of sent data fom SS. This indicates that most data sent by SS aive at APs befoe vehicles do, which enables the vehicles to download the data. Fig. 3 compaes the pefomance of ADN with diffeent combinations of the popagation delay and tansmission ate of the links. As shown in Fig. 3, the system has acceptable pefomance with less than 4 equests within 2 minutes at a low to medium load, high to medium tansmission ate, and low to medium popagation delay. Fo example, Fig. 3(a) shows moe than 8% of each equested file can be deliveed though the ADN when the load is low to medium and the popagation delay and tansmission ate ae 1ms and 2Mbps espectively. As the load inceases, the pecent deceases due to a longe queuing delay within the ADN fom the netwok congestion. As shown in Fig. 3(b), the file delivey ate via ADN is moe than 5 kbits/sec up to 4 uses when (popagation delay, tansmission ate) = (1ms, 2Mbps) o bette, which is enough fo most Intenet video steaming sevices. Fig. 3(c) shows that the total numbe of bytes vehicles eceive via ADN. At a low to medium load, the numbe of bytes a vehicle can download duing a 525 seconds tip is moe than 3 MBytes on aveage when tansmission ate and popagation delay ae 2Mbps and 1ms o bette. Howeve, as the load gows, the total numbe of bytes deceases due to the congestion. V. RELATED WORKS Existing achitectues fo ITS can oughly be categoized into thee kinds: puely infastuctue-based, puely ad hoc, o heteogeneous. The fist kind is mainly fo vehicle-to- Infa-stuctue o infa-stuctue-to-vehicle communications, and typically involves cellula towes o base stations (BS), o Access Points (APs) that ae connected to the Intenet. Vehicles communicate indiectly with seves o othe vehicles via BSs using medium-to-long ange wieless cellula intefaces (e.g., 3G o WiMAX), o via APs using shot-ange wieless intefaces (e.g., based). [7] [9] In the second kind of achitectues, which ae mainly fo vehicle-to-vehicle communications, vehicles ae typically equipped with vaious sensos and communicate mainly with

6 each othe using shot-ange wieless intefaces [1] [13]. Communications between two emote vehicles may be via elays by othe vehicles and oad-side access points (APs). (Note that unlike the BSs o APs in the fist kind of achitectues, these APs, when used in this context, ae not connected to the Intenet.) In the thid kind of achitectues, heteogeneous wieless technologies such as cellula, WiMAX, and DSRC [14] ae used. Thee ae vaious appoaches with diffeent eseach foci. Fo example, each vehicle may download a small piece of the same file fom a oad-side AP (connected to the Intenet) duing the limited peiod of AP association time, and then shae with othe vehicles [15]. The eseach focus hee is to emulate the Bit-Toent in a vehicula envionment. In anothe appoach, a vehicle in an aea coveed by multiple BS and APs may choose one of seveal possible connections with diffeent costs and QoS pefomance chaacteistics [16]. The eseach focus hee is to minimize the cost subject to cetain QoS constaint. Yet anothe appoach is to enable a vehicle, fa fom any BS o AP, to use an ad hoc inteface to communicate with anothe vehicle acting as a elay o gateway to a BS/AP (and then the Intenet). The poposed icar-mopa achitectue is somewhat simila to those heteogeneous achitectues (i.e., the thid kind), especially the wok in [16], but with significant diffeences fom the existing appoaches. Fo instance, the subsystem icar-mopads descibed in this pape allows a vehicle to download most data via the Ad hoc Distibution Netwok (ADN) and use the cellula channels to send and eceive contol signals and a minimal amount of data only if it is necessay. This is to addess high cost and low bandwidth issues associated with cellula based achitectue (the fist kind) and the limited scalability issues associated with the puely ad hoc achitectue (the second kind), and allow icar- MOPADS to povide quality and scalable sevices (to lage numbe of potential subscibes) at affodable pice. In icar- MOPADS, unlike in [16] and othe existing appoaches, the content seve(s) is (ae) a pat of the ADN, instead of esiding in the backbone. Accodingly, ou eseach focus is on how to patition a equested file into diffeent segments and oute them though the multi-hop ADN to diffeent APs to be visited by a vehicle unde cetain delay and bandwidth constaints such that a maximum amount of data can be deliveed via the ADN. VI. CONCLUDING REMARKS In this wok, we have poposed a new achitectue called icar-mopa that integates these heteogeneous wieless/mobile technologies by utilizing both a cellula system and an ad hoc distibution netwok (ADN), and in addition, takes advantage of uses mobility pofiles in ode to povide good scalability, QoS suppot, and cost-efficiency. We have investigated a subsystem that povides file download sevices called icar-mopads, fomulated an optimization poblem called SMV (Shooting a Moving Vehicle) using NLP and LP, evaluated a simple LP appoach via simulations, and obtained encouaging esults. The wok will be extended to developing othe subsystems poviding sevices such as video steaming, customized eal-time taffic alets, and adaptive navigation unde the poposed achitectue. ACKNOWLEDGEMENT The wok of Hung Q. Ngo was suppoted in pat by NSF CAREER Awad CCF The wok of S. Yoon and C. Qiao was suppoted in pat by NSF SGER gant CNS REFERENCES [1] T. Small and Z. Haas, The shaed wieless infostation model - a new ad hoc netwoking paadigm (o whee thee is a whale, thee is a way), in In Poceedings of IEEE MobiHoc 23, 23. [2] R. Fenkiel, B. Badinath, J. Boas, and R. Yates, The infostations challenge: Balancing cost and ubiquity in deliveing wieless data, EEE Pesonal Communications, vol. 7, no. 2, pp , 2. [3] T. Spyopoulos, K. Psounis, and C. Raghavenda, Spay and Wait: An efficient outing scheme fo intemittently connected mobile netwoks, in Poceeding of the 25 ACM SIGCOMM wokshop on Delaytoleant netwoking, August 25. [4] B. Simon and R. D. Foley, Some esults on sojoun times in acyclic Jackson netwoks, Management Sci., vol. 25, no. 1, pp (198), [5] A. J. Lemoine, Netwoks of queues a suvey of equilibium analysis, Management Sci., vol. 24, no. 4, pp , 1977/78. [6] E. Reich, Note on queues in tandem, Ann. Math. Statist., vol. 34, pp , [7] N. Imai, H. Moikawa, and T. Aoyama, Pefetching achitectue fo hot-spotted netwoks, in IEEE Intenational Confeence on Communications (ICC 21), vol. 7, Helsinki, Finland, June 21. [8] B. Hull, V. Bychkovsky, K. Chen, M. Goaczko, A. Miu, E. Shih, Y. Zhang, H. Balakishnan, and S. Madden, Catel: A distibuted mobile senso computing system, in Poceedings of the 4th ACM Confeence on Embedded Netwoked Senso Systems (SenSys), Coloado, USA, 26. [9] J. Ott and D. Kutsche, Dive-thu intenet: Ieee 82.11b fo automobile uses, in Poceedings of IEEE INFOCOM 24, 24. [1] Q. Xu, T. Mak, J. Ko, and R. Sengupta, Vehicle-to-vehicle safety messaging in dsc, in Poceedings of the Fist ACM Wokshop on Vehicula Ad Hoc Netwoks (VANET 24), PA, USA, Octobe 24. [11] L. Biesemeiste, L. Schafes, and G. Hommel, Disseminating messages among highly mobile hosts based on inte-vehicle communication, in Poceedings of the IEEE Intelligent Vehicles Symposium, MI, USA, Octobe 2. [12] J. Zhao and G. Cao, VADD: Vehicle-assisted data delivey in vehicula ad hoc netwoks, in Poceedings of IEEE INFOCOM, 26. [13] X. Yang, J. Liu, F. Zhao, and N. Vaidya, Vehicle-to-vehicle communication potocol fo coopeative collision waning, in Poceedings of MobiQuitous 24, August 24. [14] Standad specification fo telecommunications and infomation exchange between oadside and vehicle systems - 5ghz band dedicated shot ange communications (dsc) medium access contol (mac) and physical laye (phy) specifications, in ASTM E2213-3, SEPTEMBER 23. [15] A. Nandan, S. Das, G. P. M. Sanadidi, and M. Gela, Coopeative downloading in vehicula ad hoc wieless netwoks, in Poceedings of IEEE/IFIP Intenational Confeence on Wieless On demand Netwok Systems and Sevices, St. Moitz, Switzeland, Januay 25, pp [16] H. Wu, C. Qiao, S. De,, and O. Tonguz, Integated cellula and ad hoc elaying systems: ica, IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, vol. 19, no. 1, pp , OCTOBER 21.