Wi-Fi Hotspot at Signalized Intersection: Cost …n5cheng/Publication/Wi-Fi...LU et al.: Wi-Fi...

3506 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 65, NO. 5, MAY 2016

Wi-Fi Hotspot at Signalized Intersection:Cost-Effectiveness for Vehicular Internet Access

Ning Lu, Student Member, IEEE, Nan Cheng, Student Member, IEEE, Ning Zhang, Student Member, IEEE,Xuemin (Sherman) Shen, Fellow, IEEE, Jon W. Mark, Life Fellow, IEEE, and Fan Bai, Senior Member, IEEE

Abstract—In this paper, we investigate the cost-effectivenessof a Wi-Fi solution for vehicular Internet access. We define thecost-effectiveness as the cost saving by deploying and operatinga low-cost Wi-Fi infrastructure instead of a costly benchmarkcellular network. To characterize the service quality of Wi-Fideployment, we also define the normalized service delay (NSD),which is the service time to fulfill a data application via the Wi-Finetwork normalized by that via the cellular network. To derivethe service time, we analyze the average throughput capacity ofa generic vehicle in the Wi-Fi network and the average down-link capacity in the cellular network. In particular, we proposedeploying Wi-Fi access point (AP) at signalized intersection andstudy the fundamental influence of traffic signals (which yield aninterrupted vehicle traffic) on Wi-Fi access. Then, we examinethe tradeoff between cost-effectiveness and NSD by identifyinginterplays between controllable (e.g., the density of Wi-Fi de-ployment and user’s satisfaction) and uncontrollable parameters(e.g., vehicle traffic statistics). Our results are very useful for net-work operators to make strategic planning of Wi-Fi deploymentfor vehicular Internet Access.

Index Terms—Cellular networks, cost-effectiveness analysis,drive-thru Internet, vehicle traffic modeling, vehicular Internetaccess, Wi-Fi networks.

I. INTRODUCTION

W ITH growing awareness of road safety and the ever-increasing demand for high-speed mobile Internet

services, Internet connectivity is becoming a must-have fea-ture of modern vehicles. The telecommunication industry hasresponded promptly by using off-the-shelf wireless technolo-gies to establish a huge mass market of Internet-connectedcars, which is expected to reach $131.9 billion by 2019 [1].Not surprisingly, cellular technology, such as third-generation(3G) (UMTS, HSPA, and HSPA+) and fourth-generation (4G)(LTE), is the top choice for delivering Internet service to carsdue to its prominent role in providing reliable and ubiquitousmobile Internet access. Very recently, via AT&T’s LTE net-work, General Motors’s LTE-connected car has begun hitting

Manuscript received January 3, 2015; revised April 13, 2015; acceptedMay 20, 2015. Date of publication June 10, 2015; date of current versionMay 12, 2016. This work was supported in part by a Joint Research Grant fromthe Natural Sciences and Engineering Research Canada and General Motors.The review of this paper was coordinated by Prof. Y. Qian.

N. Lu, N. Cheng, N. Zhang, X. Shen, and J. W. Mark are with the Departmentof Electrical and Computer Engineering, University of Waterloo, Waterloo,ON N2L 3G1, Canada (e-mail: [email protected]; [email protected];[email protected]; [email protected]; [email protected]).

F. Bai is with General Motors Corporation, ECI Lab, Warren, MI 48092 USA(e-mail: [email protected]).

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

Digital Object Identifier 10.1109/TVT.2015.2443808

the roads, powering many automotive telematics applications(e.g., emergency services and online diagnostics) [2]. However,the cellular network nowadays faces an uphill battle against theexplosive growth of mobile data traffic, which has been report-edly doubling each year over the last few years. Exploitingcomplementary spectrum for vehicular Internet access isthereby an immediate need, which is also part of the solutionto the so-called 1000x data challenge [3].

Operating in unlicensed frequency bands, Wi-Fi is astonish-ingly popular with millions of hotspots deployed all over theworld for public Internet access. Due to its low per-bit cost andthe feasibility of serving outdoor users at vehicular mobility[4], Wi-Fi is expected to be an attractive and complementarytool to deliver broadband services to moving cars—the built-inWi-Fi radio or Wi-Fi-enabled mobile devices on board canaccess the Internet when vehicles drive-thru the coverage ofWi-Fi hotspots. Recent advances in Passpoint/Hotspot 2.0,powered by Wi-Fi Alliance, make Wi-Fi more capable ofproviding secure connectivity than before. Moreover, with sub-scriber identity module (SIM)-based authentication, seamlessroaming between Wi-Fi and cellular networks is also enabled.

Compared with cellular access with wide availability, theWi-Fi infrastructure has limited utility of the service offeringfor vehicles over intermittent network connectivity, as observedin real-world tests, e.g., [5]. Typically, vehicle users on theroad have to experience a number of drive-thrus/connectionsto fulfill a mobile application (e.g., buffering a video clip of100 MB from the Internet), implying a large service delaythat degrades the user’s satisfaction. While the Wi-Fi footprintcan be enlarged by deploying more Wi-Fi access points (APs)[6], the network cost or TCO1 would be increased as well. Inparticular, a solution that tries to achieve ubiquitous coverage ascellular networks is prohibitive and not practical [7]. Therefore,great uncertainty remains as to whether it makes economicsense to deploy Wi-Fi networks for highly mobile vehicle users.Due to the new-generation Wi-Fi hotspot, many mobile networkoperators (MNOs), such as AT&T and China Mobile, haveshown strong interest in the Wi-Fi solution. However, the realbenefit of Wi-Fi solution should be validated through cost-effectiveness analysis considering user’s satisfaction so thatMNOs would be fully convinced to turn the strong interestto strong commitment to deploy large-scale outdoor Wi-Finetworks in favor of vehicle users.

1Total cost of ownership, including one-time cost component (capital ex-penditures: CAPEX) and recurring cost component (operational expenditures:OPEX).

0018-9545 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

LU et al.: Wi-Fi HOTSPOT AT SIGNALIZED INTERSECTION 3507

As an effort to that end, in this paper, we study cost-effectiveness of vehicular Internet access through Wi-Fihotspots. The cost-effectiveness is defined as the TCO savingby deploying and operating drive-thru Wi-Fi networks insteadof the cellular network for vehicular Internet access. To es-tablish the relationship between cost-effectiveness and networkperformance, we analyze the maximum average throughput ofindividual vehicle under any given density of deployed Wi-FiAPs (different deployment yields different cost-effectiveness),which can be used to determine the service delay once the totalthroughput required for fulfilling a data application is given.For apple-to-apple comparison, the network performance ofbenchmark cellular network is also analyzed so that the servicedelay experienced in Wi-Fi network can be normalized by thatin cellular network. This normalized service delay (NSD) isthen able to reflect the service quality degradation because ofusing Wi-Fi networks characterized by the corresponding cost-effectiveness. We examine tradeoff between cost-effectivenessand NSD and then demonstrate the benefit of Wi-Fi solutionsquantitatively.

The main contributions of this paper are as follows.

• We deploy Wi-Fi APs at signalized intersections andstudy the fundamental influence of traffic signals ondrive-thru Internet access, which has attracted significantresearch interest. Deploying AP at intersection can re-duce the need for non-line-of-sight (NLOS) transmissionsalong the road to provide better service coverage [8].One example of real-world deployment is Wickedly FastWi-Fi network, which covers the downtown area ofSan Jose, CA, USA [9]. To facilitate the network perfor-mance analysis of urban drive-thru Wi-Fi networks, wedevelop a simple yet effective traffic modeling, which tothe best of our knowledge, represents the first model toincorporate the influence of traffic lights at intersectionson the throughput performance of a vehicle in a drive-thru Internet system. Interestingly, for the first time, weshow a significant throughput gain for vehicles stoppingat intersections due to red signals. Our modeling andanalysis are validated via extensive SUMO and NS-3simulations.

• We propose a framework for cost-effectiveness analysisof Wi-Fi network, in which the cost-effectiveness (howmuch TCO can be saved for the MNO) and NSD (howmuch service degradation the vehicle user will tolerate)are mathematically defined and the explicit relation be-tween these two metrics are established. Quantitatively,we show that the TCO of Wi-Fi can be traded withuser’s satisfaction, which could aid MNOs in strategicdecision-making for Wi-Fi deployment. Our frameworkalso lays a foundation for helping in understanding cost-effectiveness of other complementary wireless technolo-gies for vehicle users, such as small cells in the contextof cellular networks and super Wi-Fi operating in TVwhite space.

The remainder of this paper is organized as follows.Section II surveys the related works. Section III presents the

problem formulation. We analyze the time-average through-put capacity of drive-thru Wi-Fi and the average downlinkcapacity of cellular network in Sections IV and V, respec-tively. Section VI presents the cost-effectiveness analysis.Section VII provides concluding remarks.

II. RELATED WORK

A. Drive-Thru Internet Access

There have been extensive research efforts on drive-thruInternet access. Due to high vehicle mobility and intermittentconnectivity, network performance of drive-thru Wi-Fi is quitedifferent from that of a normal Wi-Fi network. Many researchefforts focus on performance evaluation through real-worldmeasurement campaign in different scenarios, including urbanarea (e.g., [4]), highway (e.g., [5]), and traffic-free road (e.g.,[10]). All these measurements show a limited throughput ofper drive-thru access. To improve the performance of drive-thruWi-Fi, a number of solutions have been proposed in theliterature (refer to a survey [11] for details).

The first effort on analytical modeling of drive-thru Internetsystems is presented in [12], which provides the closed-formexpression of average number of bytes downloaded by a genericvehicle during its sojourn in an AP’s coverage range consider-ing both traffic flow theory and wireless network features. Inthis paper, we also propose an analytical model of drive-thruInternet access. Different from [12] and only for uninterruptedtraffic flows, we consider the fundamental influence of trafficsignals, leading to an interrupted vehicle traffic Moreover, ouranalysis is for the entire period during which vehicles travelthrough multiple Wi-Fi APs, whereas in [12], the throughput ofone typical drive-thru is the focus.

B. Cost-Effectiveness Analysis

Research on cost-effectiveness of Wi-Fi solution for vehicu-lar Internet access is quite limited. An apple-to-apple compar-ison of the performance characteristics of a 3G network anda metro-scale commercial Wi-Fi network (aiming to serve no-madic users) is performed in [13], and a hybrid network designis suggested to enhance the network performance. Fundamentalrelations between network throughput and infrastructure TCOare established in [14] for cellular deployment and Wi-Fi-based deployment, respectively. The Wi-Fi-based solution issuggested for providing a cost-effective data pipe to vehicles,however, without considering service delay/user’s satisfactionand fundamental influence of traffic signals in urban scenarios.For vehicular ad hoc networks (VANETs), Banerjee et al. in[15] first examined the cost-performance tradeoffs for VANETsby considering three infrastructure enhancement alternatives:base stations (BSs), meshes, and relays. They demonstratedthat, if the average packet delay can be reduced by a factor oftwo by adding X BSs, the same reduction needs 2X meshesor 5X relays. They argued that relays or meshes can be a morecost-effective enhancement due to the high TCO of BSs. Notethat the objective of their work is to improve network delay byaugmenting mobile ad hoc networks with infrastructure, whichis different from ours on cost-effectiveness of vehicular Internet


access. Moreover, our methodology for modeling and analysisis also different from that adopted in [15]. Some other works(e.g., [16]) only consider cost issues of deploying a particulartype of infrastructure.

III. PROBLEM FORMULATION

We study the cost-effectiveness of drive-thru Wi-Fi accessin a city area Ω with a set of moving vehicles. Two differ-ent networking scenarios are considered. In the first scenario(Wi-Fi), the Internet gateways are sparsely deployed Wi-FiAPs (at intersections), and vehicles have only opportunisticdrive-thru access. We use the second scenario (Cellular)with cellular macrocell BSs providing full service coverage asa benchmark for performance comparison.

We denote by NW the number of Wi-Fi APs deployed inthe Wi-Fi scenario and by NC the number of macrocell BSsdeployed in the Cellular scenario. Under both scenarios,vehicles consume Internet data services as long as Internet con-nectivity is available. The maximum average data throughputachieved by individual vehicles are denoted by �W and �C ,respectively, for Wi-Fi and Cellular. To achieve a targetaggregate throughput G (considering a mobile application ofdownloading a file of size G), the average time required (servicedelay) in the Wi-Fi scenario and in the Cellular scenarioare denoted by DW and DC , respectively.

Definition 1: The NSD of a Wi-Fi deployment is defined as

α = DW /DC .

Straightforwardly, the NSD characterizes the service qualitydegradation if the vehicle uses the Wi-Fi network instead of thecellular network. For a fixed deployment of cellular networkNC , α merely depends on the deployment of Wi-Fi APs (NW ).From users’ standpoint, a lower NSD is desirable as the delayincurred to achieve a target throughput by using drive-thruInternet access would be easier to be tolerant instead of usingfast but costly cellular services.

We denote by eW the TCO (including CAPEX and OPEX)of deploying and operating one Wi-Fi AP. The total cost ofa Wi-Fi deployment is thereby EW = eWNW . Similarly, wehave eC for deploying and operating one macrocell BS, andEC = eCNC .

Definition 2: Cost-effectiveness of a Wi-Fi deployment isdefined as

η = 1 − EW /EC . (1)

The cost-effectiveness is used to characterize the cost sav-ing for deploying and operating a Wi-Fi infrastructure in amodel city Ω. Intuitively, a lower NSD yields lower cost-effectiveness. The NSD cannot be reduced while increasing thecost-effectiveness. The main objective of this paper is to studythe tradeoff between α and η. A summary of the mathematicalnotations used in this paper is given in Table I.

IV. TIME-AVERAGE THROUGHPUT CAPACITY OF WI-FI

To obtain the service delay DW , we first derive the time-average throughput capacity �W of a Wi-Fi network. In anurban environment, vehicle mobility is regulated by traffic

TABLE IKEY NOTATIONS

signals at intersection, which imposes a significant impact ondrive-thru Wi-Fi access: vehicle stopping at the intersectionprolongs the connection time with the Wi-Fi AP to potentiallyincrease the data volume downloaded/uploaded. To facilitatethe throughput analysis, a simple yet effective modeling ofvehicle flow regulated by traffic signals is developed.

A. Modeling Vehicle Flow With Fixed Signals

Appropriate modeling of vehicle flow regulated by trafficsignals at intersections is the prerequisite to analyze the networkperformance of urban drive-thru Wi-Fi networks. However, itturns out to be a challenging task as it is difficult to determinehow many details of vehicle mobility and road network shouldbe incorporated into the modeling. Traffic models developedin transportation engineering, such as car following models,depend on many details/factors that increase the accuracy butcould make the network analysis intractable or tedious. Hence,we develop a simple yet effective modeling with adequatedetails, which can capture the main characteristics of vehicletraffic at signalized intersections. To the best of our knowledge,our model represents the first model to incorporate the influenceof traffic lights at intersections on the throughput performanceof a vehicle in a drive-thru Internet system.


Fig. 1. Wi-Fi hotspot at signalized intersections.

Road Network: The road network of the target city area Ωis considered a regular grid, which is a common street patternin many cities, such as Houston, TX, USA, and Portland, OR,USA [17]. In specific, we consider a two-way traffic on eachroad and one single lane for each way. The intersection ofany two roads is signalized, i.e., having traffic signal control.We denote by L the distance between any two neighboringintersections. Further, to facilitate our analysis, we definethe intersection area (IArea) for each intersection. It is asquare centering at one intersection and consisting four lanes(eastbound, westbound, northbound, and southbound), each ofwhich is of length L, as shown in Fig. 1. By doing so, the cityarea is partitioned into distinct IAreas.

Wi-Fi Deployment: In Wi-Fi scenario, Wi-Fi APs are de-ployed in the target city area for vehicular Internet access. Weconsider the following random deployment strategy2: Wi-FiAPs are deployed only at intersections, and an intersectionhas an AP deployed with probability pap. Deploying APs atintersections can reduce the need for NLOS transmissionsalong the road to provide better service coverage [8], whichis also considered in literature, such as the theoretical study[7]. Moreover, placing APs at intersections mathematicallyfacilitates the investigation on the impact of traffic signals dueto the introduced symmetry. We denote by R the radius of APcoverage and define the area of interest (AreaOI) for eachintersection, which is a disk centering at the intersection withradius R. The AreaOI thereby consists four lanes (eastbound,westbound, northbound, and southbound), each of which isof length 2R, as shown in Fig. 1. It is obvious that if oneintersection has an AP deployed, the AreaOIwill be the Wi-Ficoverage region.3

Stop-and-Go Flow: The objective of this section is to analyzethe throughput of drive-thru Wi-Fi network, considering theimpact of road intersections controlled by traffic signals.To this end, we theoretically derive the vehicle connectiontime with AP (depending on the time the vehicle arrives atthe intersection) and how many vehicles share the wirelessresources within the AP coverage (changing over time duringthe connection period). Under the control of traffic signals,vehicles arriving at the intersection during the red period stopand form a queue, whereas vehicles arriving at the intersectionduring the green period keep going without delay if the vehicle

2Although Wi-Fi APs may be regularly deployed at intersections, the en-counter of next AP for a vehicle is still random due to the randomness inmovement when we observe the vehicle. Therefore, we consider a randomWi-Fi deployment.

3We only consider the case in which R < L/2 so that there is no overlappedWi-Fi coverage regions, which also makes economic sense.

queue is completely dissolved. In transportation engineering,many modeling approaches for signalized intersection focus onthe development of delay and queue models [18]. The main ob-jective is to analyze the signal delay a vehicle experiences at theintersection (i.e., the extra waiting time due to signal operationand the vehicle queue), and the dynamics and the stochasticbehavior of the overflow queue (i.e., the vehicle queue at the endof a green period). For example, one of the best-studied modelsis the fixed-cycle traffic light queue, where the traffic signalalternates between fixed green and red periods, and vehiclesqueued at the intersection are assumed to depart during thegreen period at equal time intervals [19]. However, the existingmodeling approaches from traffic engineering cannot bedirectly applied to solve our problem. The reasons are twofold:1) The existing approaches focus on the steady-state or time-dependent analysis of delay and overflow queue length (themean and the distribution), which are unable to characterize thedynamics of vehicle flow in the AreaOI; and 2) the existingapproaches are for the analysis of one-way traffic interrupted bytraffic signals (the case of multiple lanes may be considered),which are too complex to apply to the scenario where thewhole IArea (including four lanes of different directions) isconsidered. In wireless networking research area, a stochastictraffic model is proposed for VANETs in signalized urban roadsystems [20], which can describe the average vehicle densityand the random interactions among vehicles. The difficulty ofapplying this model to our scenario is also the high complexity.To describe the main behaviors of vehicle flow in the AreaOI,we have to ignore the minor details, such as the randombehavior of individual vehicles. Thus, we develop a stop-and-govehicle flow model controlled by fixed traffic signals. Theproposed model is described as follows.

As a common practice, we simplify the three signal periods(i.e., green, amber, and red) into two periods, effective green pe-riod (EGP) and effective red period (ERP). We define the trafficsignal cycle (TSC) ([0, τ)). For the eastbound and westboundlanes, an EGP ([0, τg)) is followed by an ERP ([τg, τg + τr))during a TSC, whereas for the northbound and southboundlanes, an ERP ([0, τg)) is followed by an EGP ([τg, τg + τr)).The length of one TSC is hence denoted by τ = τg + τr. Weconsider a deterministic vehicle flow, i.e., vehicles arrive at theIArea on each lane with arrival rate λ(t) = λ. Every vehiclekeeps moving at the same and constant speed v unless it stopsand joins a vehicle queue due to red signals. During the EGP,4

each stopped vehicle departs the intersection at the speed vafter a short delay Δ of being head of line. We introduce Δto consider the time loss due to vehicle acceleration in reality.We also consider that vehicles do not change the lane at theintersection (i.e., no left, right, or U turns) for simplicity. Thus,vehicles flow into the IArea with a density ρ in vehicles permeter (vehicle/m), and according to [21]

ρ = λ/v. (2)

Further, we denote the jam density (maximum density) byρmax vehicle/m, which is the density of vehicles stopping and

4We make statement for the eastbound and westbound lanes, unless other-wise specified.


Fig. 2. Single-lane stop-and-go vehicle flow.

queueing at the intersection due to signal operations (a typicalrange of ρmax is 0.116–0.156 vehicle/m [22]).

We analyze the performance of the urban drive-thru Wi-Finetwork by observing a tagged vehicle. The tagged vehiclemoves along a fixed lane (the eastbound lane in this paper) andtraverses IAreas in sequence, as shown in Fig. 1. To simplifyour analysis, we have the following two assumptions.5

• Unsaturation: We consider the case that the length ofvehicle queue is less than R, so that if the tagged vehiclestops at the intersection, it will always be in the AreaOI,i.e., the AreaOI is always unsaturated. In addition, weassume that there is no overflow queue, i.e., all queuedvehicles can pass through the intersection during the EGP.This assumption is often valid under a regular traffic loadcondition and for a typical value of R.

• Independence: The IAreas are treated independently,i.e., traffic signals at different intersections are not coordi-nated, and the vehicle arrival process for one IArea doesnot depend on the upstream traffic. We further assume thatthe tagged vehicle arrives at the IArea equally likely forany instant τa during the cycle [0, τ), i.e., τa is uniformlydistributed over the interval [0, τ).

We denote by N(t) the number of vehicles in an AreaOIin a TSC, where t ∈ [0, τ). Moreover, we denote by S(τa) thesojourn time the tagged vehicle stays in an AreaOI, wherethe arrival time τa ∈ [0, τ). The sojourn time is equal to theconnection time if there is an AP deployed in the intersection.Next, we derive the analytic expressions of N(t) and S(τa).

B. Vehicle Dynamics

As shown in Fig. 2, we first focus on the vehicle flowregulated by traffic signals on the eastbound lane in a genericIArea, and denote the number of vehicles at time t in theAreaOI (only the eastbound lane is considered) by Ne(t),where t ∈ [0, τ). We immediately have Ne(0) = λ(τr +R/v),which is the number of vehicles that arrive in the AreaOIbut do not pass through the intersection during the previousTSC. Note that all these vehicles are located in [−R, 0], i.e.,Ne(0) should be less than ρmaxR, according to the unsaturationassumption. Hence, vehicles arrive in the AreaOI at the rateλ(t) = λ during [0, τ). As the arrival traffic would be regu-lated by the traffic signals, the vehicle departure rate from theAreaOI is the key to characterize Ne(t). Let μ(t) denote thisdeparture rate. Ne(t) can be characterized as follows:

Ne(t) = Ne(0) +

t∫0

λ(t̂)dt̂−t∫

0

μ(t̂)dt̂, t ∈ [0, τ). (3)

5Relaxation of these two assumptions would be considered for future work.

The departure rate function μ(t) is given by Lemma 1.Lemma 1: Vehicles (on the eastbound lane) depart from the

AreaOI at the rate μ(t), t ∈ [0, τ), where

μ(t) =

⎧⎪⎪⎪⎨⎪⎪⎪⎩

0, t ∈[0, R

v +Δ)

ρ∗v, t ∈[Rv +Δ, R

v +Δ+ t∗)

λ, t ∈[Rv +Δ+ t∗, τg +

Rv

)0, t ∈

[τg +

Rv , τ

)where ρ∗=ρmax/(1+ρmaxΔv), and t∗=λ(τr+Δ)/(ρ∗v−λ).

Proof: Note that the time interval [0, τg) is the EGP, and[τg, τ) is the ERP. At t = Δ, the head-of-line vehicle queued inthe previous ERP starts to pass through the intersection at thespeed v. Since the head-of-line vehicle has to move a distanceof R to depart from the AreaOI, during [0, (R/v) + Δ), thereare no vehicle departures. At t = (R/v) + Δ, the vehicle queuewith the density ρ∗ starts to depart from the AreaOI, where ρ∗

can be easily determined by the following equality: (1/ρmax) +Δv = 1/ρ∗. Next, we determine the duration of this departure,which is denoted by t∗. Since, at t = (R/v) + Δ + t∗, ρ∗vt∗

vehicles have departed from the AreaOI and all the vehiclesin the AreaOI are located in [−R,R] with density ρ, we havethe following equation with respect to t∗ under the unsatura-tion assumption: λ((R/v) + τr + (R/v) + Δ+ t∗) = ρ∗vt∗ +ρ2R. From (2), we have t∗ = λ(τr +Δ)/ρ∗v − λ. Therefore,during [(R/v) + Δ, (R/v) + Δ+ t∗), the departure rate is ρ∗v.Following the vehicle queue, vehicles depart at the arrival rateλ until t = τg + (R/v). Note that, at t = τg , the traffic signalturns from green to red. Again, there is no vehicle departureduring [τg + (R/v), τ). �

Thus, from (3) and Lemma 1, we have

Ne(t)=

⎧⎪⎪⎪⎪⎪⎪⎨⎪⎪⎪⎪⎪⎪⎩

λ(τr +R/v + t), t ∈[0, Rv +Δ

)λ(τr +R/v + t)

−ρ∗v(t−R/v −Δ), t ∈[Rv +Δ, Rv +Δ+t∗

)2λR/v, t ∈

[Rv +Δ+t∗, τg+

Rv

)λ(R/v + t− τg), t ∈

[τg +

Rv , τ

).

(4)

Under the same control of traffic signals, the vehicle flow onthe westbound lane has the same behavior as that on the east-bound lane. Thus, Nw(t) = Ne(t), where Nw(t) is the numberof vehicles at time t on the westbound lane in the AreaOI.Similarly, Nn(t) and Ns(t) are denoted for the northbound andsouthbound lanes, respectively. Ns(t) = Nn(t) and Nn(t) canbe derived in the same way as Ne(t). The departure rate ofvehicles on the northbound lane is given as follows: Whent∈ [0,R/v), μ(t)=λ; when t∈ [R/v, τg+(R/v)+Δ), μ(t) =0; when t ∈ [τg + (R/v) + Δ, τg + (R/v) + Δ + (λ(τr +


Fig. 3. Comparison of vehicle dynamics in the AreaOI between our analytic results and simulations based on SUMO. ρmax = 0.12 vehicle/m, v = 14 m/s,λ = 0.17 vehicle/s, τr = τg = 40 s [23], Δ = 1.3 s, and R = 100 m. (a) Ne(t). (b) Nn(t). (c) N(t).

Δ)/ρ∗v − λ)), μ(t) = ρ∗v; and when t ∈ [τg + (R/v) + Δ+(λ(τr +Δ)/ρ∗v − λ), τ), μ(t) = λ. Therefore, we have

Nn(t) =

⎧⎪⎪⎪⎪⎪⎪⎨⎪⎪⎪⎪⎪⎪⎩

2λR/v, t ∈[0, Rv

)λ(R/v + t), t ∈

[Rv , τg +

Rv +Δ

)λ(R/v + t)− ρ∗v(t− τg −R/v −Δ),

t ∈ [τg +Rv +Δ, τg +

Rv +Δ+ λ(τr+Δ)

ρ∗v−λ )

2λR/v, t ∈ [τg +Rv +Δ+ λ(τr+Δ)

ρ∗v−λ , τ).

(5)

As N(t) is a summation of Ne(t), Nw(t), Nn(t), and Ns(t),we can immediately have the following result.

Lemma 2: Under the stop-and-go flow model and the unsat-uration assumption, the number of vehicles in the AreaOI attime t, t ∈ [0, τ), is given by

N(t)=

⎧⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩

2λ(τr + 3R/v + t), t ∈[0, R

v

)2λ(τr + 2R/v + 2t), t ∈

[Rv ,

Rv +Δ

)2λ(τr + 2R/v + 2t)− 2ρ∗v(t−R/v −Δ),

t ∈[Rv +Δ, R

v +Δ+ t∗)

2λ(3R/v + t), t ∈[Rv +Δ+ t∗, τg +

Rv

)2λ(2R/v + 2t− τg), t ∈

[τg+

Rv , τg+

Rv +Δ

)2λ(2R/v + 2t− τg)− 2ρ∗v(t− τg −R/v −Δ)

t ∈[τg +

Rv +Δ, τg +

Rv +Δ+ t∗

)2λ(3R/v+t−τg), t ∈

[τg +

Rv +Δ+ t∗, τ

).

To evaluate how our proposed stop-and-go flow model canreflect the major behavior of vehicle dynamics in the AreaOI,we perform simulations in an open-source traffic softwareSUMO6 in which the car following model developed byStefan Krauß [24] is used. We compare the obtained analyticresults with simulation results, as shown in Fig. 3. The sim-ulation data are averaged over 8000 s (i.e., 100 TSCs). Thefluctuation of simulation results is due to randomness of driverimperfection, which is a parameter controlled in the StefanKrauß car following model, as shown in Table II.

6SUMO is an open source, microscopic, and continuous road traffic simula-tor designed to handle large road networks.

C. Sojourn Time

The sojourn time S(τa) is the time duration the taggedvehicle stays in an AreaOI. Note that the sojourn time is equalto the connection time for intersections with an AP deployed.S(τa) depends on the arrival time τa, which is uniformlydistributed over the interval [0, τ). The analytic expression ofS(τa) is given in the following lemma.

Lemma 3: Under the stop-and-go flow model and the unsatu-ration assumption, the sojourn time of the tagged vehicle giventhe arrival time τa is as follows:

S(τa) =

⎧⎪⎪⎪⎪⎪⎪⎪⎨⎪⎪⎪⎪⎪⎪⎪⎩

λ(τr+Rv+τa)

ρ∗v + Rv +Δ− τa

τa ∈[0, t∗ +Δ− R

v

)2Rv , τa ∈

[t∗ +Δ− R

v , τg −Rv

)λ(τa−τg+

Rv )

ρ∗v + Rv +Δ+ τ − τa

τa ∈[τg − R

v , τ).

Proof: Note that the tagged vehicle arrives in the location−R at τa. It can be seen that the tagged vehicle will be thehead of line in the vehicle queue if τa = τg − (R/v). Whenτa ∈ [τ ′a, τg − (R/v)), the tagged vehicle passes through theintersection without stop (S(τa) = 2R/v), where τ ′a satisfiesthe following equality: λ(τ ′a + τr + (R/v)) = ρ∗vt∗. Hence,τ ′a = t∗ +Δ− (R/v). If τa ∈ [0, τ ′a), the tagged vehicle willstop and join the vehicle queue because the vehicle queueformed during the previous ERP has not been dissolved yet. Thesojourn time S(τa) is thereby (λ(τr + (R/v) + τa)/ρ

∗v) +(R/v) + Δ− τa. When τa ∈ [τg − (R/v), τ), the sojourn timeof the tagged vehicle can be expressed as τ − τa + τ ′′a . The firstpart denotes the time elapsed before the traffic signal turns fromred to green, and the second part denotes the time the tagged ve-hicle continues the movement until it departs from the AreaOI.τ ′′a can be obtained by τ ′′a = (λ(τa − τg + (R/v))/ρ∗v) +(R/v) + Δ. �

D. Throughput Capacity Per Drive-thru

Based on Lemmas 2 and 3, we next derive the throughputcapacity per drive-thru for the tagged vehicle.


TABLE IINS-3—SUMO SIMULATION PARAMETERS

Definition 3: Throughput capacity per drive-thru: the max-imum number of bits received by the tagged vehicle from theWi-Fi AP during one typical drive-thru of Wi-Fi coverage.

Wi-Fi transmission based on the IEEE 802.11 protocoladopts an adaptive modulation scheme with different trans-mission bit rates, depending on the communication distancefrom the AP. However, to reduce the complexity of our modelcomputation, we consider a nonadaptive scheme with constanttransmission bit rate in a fixed AP coverage range. The non-adaptive transmission rate is also analytically considered in[12]. In addition, the contention-based medium-access-control(MAC) protocol, i.e., IEEE 802.11 distributed coordinationfunction, is adopted to schedule parallel transmissions. Tocharacterize the protocol overhead (including the overhead ofphysical layer), we introduce an empirical efficiency factor ϕ ∈(0, 1), which can be obtained from real-world measurementsor through theocratical analysis. We consider that the taggedvehicle share the Wi-Fi resource equally with other vehicles inthe AP’s coverage range. Hence, the throughput capacity perdrive-thru achieved by the tagged vehicle is given by

�(τa) =

τa+S(τa)∫τa

ϕmaxC

N(t mod τ)dt (6)

where ϕmax is the maximum efficiency factor for a giventransmission rate, and C is the AP’s transmission rate. ϕmaxCthereby indicates the maximum bit rate that can be utilizedfor data transmission. For example, according to [25], ϕmax =5/11 for IEEE 802.11b and C = 11 Mb/s, as the theoreticalmaximum throughput is shown to be 5 Mb/s for a 11-Mb/stransmission rate. It is worth noting that the throughput capacityper drive-thru depends on the arrival time of the tagged vehicle.That is, with different arrival time to the AreaOI (with anAP deployed), the tagged vehicle achieves different throughputcapacity per drive-thru, reflecting the impact of traffic signals.In particular, there is a significant throughput gain for vehiclesstopping at intersections due to red signals, as shown in thefollowing simulation and analytic results, which is an importantfinding of this paper.

To validate our modeling and analysis, we conduct sim-ulations in the network simulator NS-3 and the road trafficsimulator SUMO. We first use SUMO to generate the mobilitytrace file of vehicle traffic in one IArea. Then, the trace file isused as an input for network simulations in NS-3. Simulation

TABLE IIISTOP-AND-GO FLOW MODEL PARAMETERS FOR ANALYTIC RESULTS

parameters and model parameters for analytic results are re-spectively given in Tables II and III. Fig. 4 presents the analyticand NS-3—SUMO simulation results on throughput capacityper drive-thru with respect to the arrival time. We adopt theIEEE 802.11g standard and adaptive data rates up to 54 Mb/s inthe simulation.7 As the results shown in Fig. 4, vehicles arrivingat the AreaOI with an arrival time around 31 s (τa ≈ 31) in aTSC, which thereby stop at the intersection due to the red signalto prolong the connection time with the AP, can achieve a muchhigher throughput capacity per drive-thru (approximately threetimes as high as the lowest one (τa ≈ 11) analytically), indicat-ing a significant throughput gain. In Fig. 4, it is shown that, evenwith the nonadaptive data rate and simplified MAC operation,our theoretical and simulation measurements still match wellin terms of general trends, which demonstrates the validity ofour proposed modeling approach on throughput analysis of avehicle driving through Wi-Fi coverage considering the impactof traffic signals.

Fig. 5 presents the maximum throughput capacity perdrive-thru, i.e., �max = maxτa∈[0,τ)�(τa), under differentconfigurations of TSC. It is intuitive that, with a longer EGPor ERP, vehicles achieve a larger �max due to the prolongedconnection time with the AP. However, the gain is not verysignificant since the increase in the EGP or ERP also incursa larger number of vehicles waiting at the intersection due tothe red signal to increase the number of vehicles contendingfor Wi-Fi resources. As shown in Fig. 5, the impact ofincreasing the connection time after all dominates the impactof increasing the number of vehicles in the AP’s coverage.Fig. 6 presents the throughput gain, which is the maximumvalue over the mean value of throughput capacity per drive-thru

7We consider Wi-Fi standard IEEE 802.11g in the simulation evaluation fortwo reasons. First, the IEEE 802.11g protocol is well developed in networksimulator NS-3, which is widely recognized and used in academia. Second,employing other versions of Wi-Fi standard may yield slightly different resultsin the simulation. However, our result considering the IEEE 802.11g can be abenchmark once the performance bias of a different Wi-Fi standard over theIEEE 802.11g is known.


Fig. 4. �(τa) versus τa [τg = τr = 40 s].

Fig. 5. max�(τa), τa ∈ [0, τ) versus τg .

with respect to τa (mathematically defined asmaxτa∈[0,τ)�(τa)/[(1/τ)

∫ τ

0 �(τa)dτa]), under differentconfigurations of TSC. It can be seen that, with a longerEGP or ERP, we have a higher throughput gain. For example,for τg = τr = 40 s, the maximum throughput capacity perdrive-thru (achieved when τa ≈ 31 s) is 1.9 times the meanvalue by simulation and 1.8 times by analysis, demonstrating asignificant impact of traffic signals.

E. Time-Average Throughput Capacity

The tagged vehicle moves along the road and experiencesalternate disconnected period and connected period (being in-side the AreaOI with an AP deployed). We are interested inthe average bit rate of the tagged vehicle received from the APsover a long time, e.g., the entire travel time. Mathematically, wepresent the definition of time-average throughput capacity.

Definition 4: Time-average throughput capacity: the max-imum average bit rate received by the tagged vehicle from

Fig. 6. Throughput gain (max over mean) versus τg .

the drive-thru Wi-Fi networks in a long term, which is givenmathematically by

�W = limt→∞

�̃(t)

t(7)

where �̃(t) is the total number of bits received by time t.We next derive the time-average throughput capacity. The

IAreas on the route of the tagged vehicle are indexed by1, 2, . . . , n, as shown in Fig. 1. Let Tn denote the time fromthe departure of the (n− 1)th AreaOI to the departure of thenth AreaOI. Thus, we have

Tn = (L − 2R)/v + S (τna ) (8)

where τna is the arrival time to the nth AreaOI. It can be seenthat T1, T2, . . . , Tn are independent and identically distributed(i.i.d.) random variables with a common distribution T underthe independence assumption. Further, we denote by �̃n thethroughput capacity achieved in the nth AreaOI. �̃n = �(τna )with probability pap and �̃n = 0 with probability 1 − pap. Thenumber of AreaOIs passed through by the tagged vehicle bytime t is denoted by {I(t), t ≥ 0}. Then, we have

I(t)∑n=1

�̃n ≤ �̃(t) <

I(t)+1∑n=1

�̃n. (9)

Since �̃n can be considered the reward earned during thetime period of Tn, we model {�̃(t); t > 0} as a renewal rewardprocess with interrenewal time {Tn;n ≥ 1}. The interrenewaltimes have a finite expectation E[T ] < ∞. �̃1, �̃2, . . . , �̃n arei.i.d. random variables with a common distribution �̃. Thefollowing lemma holds for renewal reward processes.

Lemma 4: Consider a renewal reward process {R(t); t > 0}with expected interrenewal time E[X ] = X < ∞ (See [26, Th.5.4.1]). If each Rn is a random variable with E[Rn] < ∞, thenwith probability 1, we have

limt→∞

R(t)

t=

E[Rn]

X. (10)


Fig. 7. �W versus pap.

Proposition 1: Under the independence assumption, withprobability 1, the time-average throughput capacity of thetagged vehicle is given by

�W =pap

∫ τ0 �(τa)dτa

τ(L − 2R)/v +∫ τ0 S(τa)dτa

. (11)

Proof: According to Lemma 4, �W =E[�̃]/E[T ] as, ob-viously, we have E[�̃] < ∞ and E[T ] < ∞. Specifically, fromLemma 3 and (8), E[T ] = ((L− 2R)/v) + E[S(τa)]=((L −2R)/v) + (1/τ)

∫ τ0 S(τa)dτa. From (12), E[�̃] = pap(1/τ)∫ τ

0 �(τa)dτa. Note that τa is uniformly distributed over [0, τ).We omit the tedious calculations of these two integrals. �

It can be seen that the time-average throughput capacity ofa Wi-Fi network is determined by the urban environment (L),Wi-Fi deployment (pap), Wi-Fi coverage R, and connectiontime depending on traffic signal operation (τg, τr) and vehicletraffic (λ, v). The analytic and simulation results of �W interms of deployment scale are shown in Fig. 7. It can beseen that the time-average throughput capacity increases witha larger Wi-Fi deployment. Note that the analytic result is quiteoptimistic. This is because we use an empirical efficiency factorto simplify the MAC-layer and physical-layer operations ofWi-Fi. Our analytic result can be considered an upper boundof Wi-Fi throughput performance.

V. BENCHMARK: CELLULAR MACROCELL SERVICE

The performance of cellular macrocell service is considereda benchmark in the study of cost-effectiveness of the Wi-Finetwork. Thus, we use the networking scenario with cellularmacrocell BSs providing full service coverage for performancecomparison. A macrocell enables cellular services relying ona high-power cellular BS [27]. For apple-to-apple comparison,we assume that the considered cellular network only servesvehicle users. We focus on the analysis of average downlinkcapacity achieved by the tagged vehicle.

Definition 5: Average downlink capacity: the maximum av-erage downlink data rate received from cellular BSs in a long

term, which is given mathematically by

�C = limt→∞

1t

t∫0

U(t)dt (12)

where U(t) is the instantaneous maximum data rate at time t.

A. Spectrum Efficiency

The maximum data rate U(t) can be determined by U(t) =b(t)Υ(t), where b(t) is acquired transmission bandwidth, andΥ(t) is the maximum spectrum efficiency (bits/s/Hz), at time t.The maximum spectrum efficiency is theoretically governedby the Shannon capacity. However, the Shannon capacity isnot achievable in reality due to limited coding block length,nonavoidable system overhead, etc. [28]. By following [29], weadopt a modified Shannon capacity formula, i.e.,

Υ = BWe · σ · log2(

1 +SINR

SINRe

)(13)

where BWe is the system bandwidth efficiency, SINRe is the ef-ficiency of signal-to-interference-plus-noise ratio (SINR), andσ is a correction factor due to the dependence between SINReand SINR. Equation (13) can be used to approximate themaximum spectrum efficiency of real-world cellular systemswith different settings. For example, for an LTE cellular systemwith single-antenna transmissions and round robin scheduling,BWeσ = 0.56, and SINRe = 2.0 [29].

B. Distribution of SINR

The SINR on the wireless link between cellular BS and thetagged vehicle is an important basis to determine the downlinkcapacity. The interference experienced by the vehicle comesfrom the transmission of other-cell BSs. In urban areas withdensely deployed BSs, other-cell interference is a major imped-iment to high spectrum efficiency [30]. To model the deploy-ment of macrocell BSs in the considered area Ω, we considera homogeneous Poisson point process (PPP) of density ξ, inwhich each point represents a location of the BS. Modeling BSlocation as a PPP is widely adopted in the literature, e.g., [31]and [32], which is able to characterize the variety of macrocellsize due to differences in transmission power, tower height, etc.We consider the same vehicle density ρ as in the drive-thruWi-Fi scenario. The impact of traffic signals on vehicle densityis not considered here since cellular macrocell is much largerthan Wi-Fi coverage and thereby not sensitive to such variationsof vehicle density. Considering Rayleigh fading on other-cellinterference, the complementary cumulative distribution func-tion of SINR is given in [31], i.e.,

Pr(SINR > Z) =1

1 + Z2β∫∞Z

− 2β

1

1+uβ2

du(14)

where β > 2 is called the path-loss exponent. Typically, wehave β = 4 for urban environments [33]. Thus, (14) can befurther simplified as follows:

Pr(SINR > Z) =1

1 +√Z(π/2 − arctan(1/

√Z)

) . (15)


The distribution of downlink SINR given in (15) is for theconfiguration of single transmit and single receive antenna.In addition, the thermal noise is ignored as the urban cellularnetwork is often interference limited.

C. Average Downlink Capacity

To derive the average downlink capacity of the tagged vehi-cle, we consider a simple bandwidth sharing model: an everyvehicle can obtain a constant bandwidth b0 from BSs, i.e.,b(t) = b0. Given that each BS provides a bandwidth of B,the total bandwidth resource in Ω is thus ξΩB. As the totalnumber of vehicles in Ω is given by Ω · 4Lρ/L2 (approximatelyΩ contains Ω/L2 IAreas and each IArea contains 4Lρvehicles), we have b0 = ξBL/4ρ. Considering the spectrumefficiency and distribution of SINR, the following result ofaverage downlink capacity is obtained.

Proposition 2: The average downlink capacity of the taggedvehicle is given by

�C = U1

∫r>0

1

1 +√U2

(π/2 − arctan(1/

√U2)

)dr (16)

where U1 = ξBLBWeσ/4ρ, and U2 = SINRe(2r − 1).Proof:

�C = limt→∞

1t

t∫0

b(t)Υ(t)dt = b0E[Υ]

=ξBL

4ρE

[BWeσ log2

(1 +

SINR

SINRe

)]

=ξBLBWeσ

4ρ

∫r>0

Pr(log2

(1 +

SINR

SINRe

)> r

)dr

=ξBLBWeσ

4ρ

∫r>0

Pr (SINR > SINRe(2r − 1)) dr.

The third equality holds due to E[X ] =∫x>0 Pr(X > x)dx

for a nonnegative random variable X . In (15), the propositionholds. �

In the analysis of average downlink capacity, the intercellinterference management techniques are not considered, suchas frequency reuse. However, these advanced techniques areindeed beneficial for improving the spectrum efficiency andavailable data rates. In particular, the performance of LTEsystem is more limited by other-cell interference comparedwith 3G cellular systems [34], implying that interference man-agement is also necessary for LTE systems. Hence, our resultgiven in Proposition 2 is conservative and can be considereda lower bound of average downlink capacity. In providing thebenchmark cellular service for comparison, small cell serviceon top of the existing macrocell service is not considered.Under the same comparison level with Fig. 7, the analyticand simulation results of �C in terms of deployment scaleare shown in Fig. 8. It can be seen that the average downlinkcapacity increases with a larger cellular deployment in terms ofξ. We can also notice the conservativeness of our analytic resultthat we have discussed.

Fig. 8. �C versus ξ.

VI. COST-EFFECTIVENESS ANALYSIS

We examine the cost-effectiveness of Wi-Fi deploymenthere. For service delay of Wi-Fi scenario, mathematically,DW = E[{min t0 : �̃(t0) > G}]. Since it is difficult to obtainthe distribution of �̃(t) and given that Pr(limt→∞(�̃(t)/t) =�W ) = 1, DW is approximated to be G/�W for a largeG, e.g., 100 MB. Similarly, DC ≈ (G/�C). Hence, the NSDα ≈ �C/�W .

The cost-effectiveness η depends on the TCO (includingCAPEX and OPEX) of infrastructure node and the numberof infrastructure node deployed in each scenario. The CAPEXincludes the cost of equipment, planning, installation, commis-sioning, etc., and the OPEX includes the cost of site rental,power, maintenance, etc. [35]. Based on the cost model pro-vided in [35], the ratio of the TCO of a Wi-Fi AP to the TCOof a macro three-sector LTE BS is around 12%, i.e., eW /eC ≈0.12. To evaluate the cost-effectiveness of a Wi-Fi scenario,we fix the cellular deployment (Ω, NC) and thereby the �C isdetermined. Given that the average number of APs deployed ina Wi-Fi scenario is papΩ/L

2, the explicit relation between ηand α is given by

η = 1 − eWeC

Ω�C

(τ(L − 2R)/v +

∫ τ

0 S(τa)dτa)

αNCL2∫ τ

0 �(τa)dτa. (17)

This is the main result of this paper, which presents thetradeoff between η (TCO savings) and α (service degradation).It can be seen that η depends on both controllable (e.g., Wi-Fideployment (pap, R, α)) and uncontrollable (e.g., the urbanenvironment (L, τ), vehicle traffic statistics (λ, v)) parameters.The analytic and simulation results on η are shown in Fig. 9.The gap between theory and simulation is due to the conserva-tive result on �C given in Proposition 2 (see Fig. 8). Throughthese results, the MNO is able to deploy the Wi-Fi networkaccording to the required level of service and to have theknowledge of the corresponding TCO savings (compared with acellular solution) immediately. For example, theoretically, fromFig. 9, almost 90% of the TCO can be saved if the averageservice delay of Wi-Fi deployment is three times larger than


Fig. 9. Tradeoff between η and α [τr = τg = 40 s].

that of cellular deployment, demonstrating the great potential ofWi-Fi solution for vehicle users in terms of cost-effectiveness.Fig. 10 analytically presents the tradeoff between η and α underdifferent configurations of TSC. We can clearly see the impactof traffic signals on the tradeoff between the cost-effectivenessof Wi-Fi deployment and the service degradation. However, theimpact is not significant particularly for a large NSD. However,for a single drive-thru, there exists a significant throughput gaindue to the impact of traffic signals.

In our study, we do not consider deploying cellular macrocellBSs and Wi-Fi APs in one scenario for the purpose of apple-to-apple comparison and explicitly showing the great potentialof Wi-Fi. For small-cell cellular BSs (e.g., LTE small cells),although it is a promising solution to the massive increase inmobile data demand, the feasibility of outdoor small cells formobile user at vehicular speed is not clear yet. However, webelieve that a heterogeneous network with cellular macrocellservice for coverage and cellular small cells and Wi-Fi forcapacity would be desired. Moreover, in this paper, althoughwe do not consider store-carry-and-forward communicationsbetween vehicles, which may incur additional cost/complexityfor opportunistic data exchange, vehicle-to-vehicle (V2V) com-munications can indeed benefit Internet traffic delivery. For ex-ample, the vehicles in the coverage of Wi-Fi hotspots can helpto relay the traffic to virtually extend the Wi-Fi service region.

VII. CONCLUSION

In this paper, we have investigated the cost-effectiveness ofWi-Fi solution for vehicular Internet access considering thetradeoff with the user’s satisfaction. In Wi-Fi scenario, wehave particularly studied the fundamental impact of traffic sig-nals at intersection on Wi-Fi access. By examining the tradeoffbetween cost-effectiveness and NSD, we have demonstratedthat Wi-Fi has great potential to serve vehicle users with amuch lower TCO. Our results provide a quick and efficientway of determining the Wi-Fi deployment strategy and thecorresponding TCO savings. Future work includes large-scale

Fig. 10. Tradeoff between η and α under different configurations of TSC.

simulations with real-world data set of vehicle mobility inurban scenarios, and cost-effectiveness analysis taking V2Vcommunications into consideration.

REFERENCES

[1] “Connected car market—Global industry analysis, size, share, growth,trends and forecast, 2013–2019,” Transparency Market Res., Albany, NY,USA, 2013.

[2] “OnStar LTE.” [Online]. Available: https://www.onstar.com/us/en/4glte/[3] Rising to meet the 1000x mobile data challenge, QUALCOMM Inc.,

San Diego, CA, USA, Whitepaper, 2013.[4] V. Bychkovsky et al., “A measurement study of vehicular Internet access

using in situ wi-fi networks,” in Proc. ACM MobiCom, 2006,pp. 50–61.

[5] J. Ott and D. Kutscher, “Drive-thru Internet: IEEE 802.11b for Automo-bile,” in Proc. IEEE INFOCOM, Hong Kong, China, Mar. 2004,pp. 1–12.

[6] T. Wang, W. Jia, G. Xing, and M. Li, “Exploiting statistical mobilitymodels for efficient Wi-Fi deployment,” IEEE Trans. Veh. Technol.,vol. 62, no. 1, pp. 360–373, Jan. 2013.

[7] Z. Zheng, P. Sinha, and S. Kumar, “Sparse wifi deployment for vehicularInternet access with bounded interconnection gap,” IEEE/ACM Trans.Netw., vol. 20, no. 3, pp. 956–969, Jun. 2012.

[8] T.Mangel, M.Michl, O.Klemp, and H.Hartenstein, “Real-world measure-ments of non-line-of-sight reception quality for 5.9 GHz IEEE802.11 p at intersections,” in Proc. Commun. Technol. Veh., 2011,pp. 189–202.

[9] “Wickedly Fast Wi-Fi.” [Online]. Available: http://www.wickedlyfastwifi.com/

[10] R. Gass, J. Scott, and C. Diot, “Measurements of in-motion 802.11networking,” in Proc. IEEE WMCSA, Semiahmoo Resort, WA, USA,Apr. 2005, pp. 69–74.

[11] N. Cheng, N. Lu, N. Zhang, X. Shen, and J. W. Mark, “Vehicular wifioffloading: Challenges and solutions,” Veh. Commun., vol. 1, no. 1,pp. 13–21, Jan. 2014.

[12] W. L. Tan, W. C. Lau, O. Yue, and T. H. Hui, “Analytical models andperformance evaluation of drive-thru internet systems,” IEEE J. Sel. AreasCommun., vol. 29, no. 1, pp. 207–222, Jan. 2011.

[13] P. Deshpande, X. Hou, and S. R. Das, “Performance comparison of 3G andmetro-scale WiFi for vehicular network access,” in Proc. ACM SIGCOMMConf. Internet Meas., Melbourne, Vic., Australia, Nov. 2010, pp. 301–307.

[14] N. Lu et al., “Vehicles meet infrastructure: Towards capacity-cost trade-offs for vehicular access networks,” IEEE Trans. Intell. Transp. Syst.,vol. 14, no. 3, pp. 1266–1277, Sep. 2013.

[15] N. Banerjee, M. Corner, D. Towsley, and B. Levine, “Relays, basestations, and meshes: Enhancing mobile networks with infrastructure,”in Proc. ACM MobiCom, 2008, pp. 81–91.

[16] H. Omar, W. Zhuang, and L. Li, “Gateway placement and packet routingfor multihop in-VehicleInternet access,” IEEE Trans. Emerging TopicsComput., vol. 3, no. 3, pp. 335–351, Sep. 2015.


[17] S. Kostof and R. Tobias, The City Shaped. London, U.K.: Thames andHudson, 1991.

[18] F. Viti, The Dynamics and the Uncertainty of Delays at Signals. Delft,The Netherlands: Netherlands TRAIL Res. School, 2006.

[19] J. S. van Leeuwaarden, “Delay analysis for the fixed-cycle traffic-lightqueue,” Transp. Sci., vol. 40, no. 2, pp. 189–199, May 2006.

[20] I. W.-H. Ho, K. K. Leung, and J. W. Polak, “Stochastic model and con-nectivity dynamics for VANETs in signalized road systems,” IEEE/ACMTrans. Netw., vol. 19, no. 1, pp. 195–208, Feb. 2011.

[21] J. D. Fricker and R. K. Whitford, Fundamentals of TransportationEngineering. Upper Saddle River, NJ, USA: Prentice-Hall, 2004.

[22] M. Van Aerde and H. Rakha, “Multivariate calibration of single regimespeed-flow-density relationships,” in Proc. 6th Veh. Navigat. Inf. Syst.Conf., 1995, pp. 334–341.

[23] J. A. Bonneson, S. R. Sunkari, and M. P. Pratt, “Traffic signal opera-tions handbook,” Texas Transp. Instit., Texas A&M Univ. Syst., CollegeStation, TX, USA, Tech. Rep., 2009.

[24] S. Krauß, “Microscopic modeling of traffic flow: Investigation of collisionfree vehicle dynamics,” Ph.D. dissertation, Dept. Comput. Sci., Univ.Cologne, Koln, Germany, 1998.

[25] J. Jun, P. Peddabachagari, and M. Sichitiu, “Theoretical maximumthroughput of ieee 802.11 and its applications,” in Proc. IEEE Int. Symp.Netw. Comput. Appl., 2003, pp. 249–256.

[26] R. G. Gallager, Stochastic Processes: Theory for Applications.Cambridge, U.K.: Cambridge Univ. Press 2013.

[27] J. H. Schiller, Mobile Communications. New Delhi, India: Pearson,2003.

[28] E. Dahlman, S. Parkvall, J. Skold, and P. Beming, 3G Evolution: HSPA andLTE for Mobile Broadband. Amsterdam, The Netherlands: Academic,2010.

[29] P. Mogensen et al., “LTE capacity compared to the shannon bound,” inProc. IEEE VTC–Spring, 2007, pp. 1234–1238.

[30] S. Mukherjee, “Distribution of downlink sinr in heterogeneous cellularnetworks,” IEEE J. Sel. Areas Commun., vol. 30, no. 3, pp. 575–585,Apr. 2012.

[31] J. G. Andrews, F. Baccelli, and R. K. Ganti, “A tractable approach tocoverage and rate in cellular networks,” IEEE Trans. Commun., vol. 59,no. 11, pp. 3122–3134, Nov. 2011.

[32] X. Lin, R. Ganti, P. Fleming, and J. Andrews, “Towards understandingthe fundamentals of mobility in cellular networks,” IEEE Trans. WirelessCommun., vol. 12, no. 4, pp. 1686–1698, Apr. 2013.

[33] H. Xia, “A simplified analytical model for predicting path loss in urbanand suburban environments,” IEEE Trans. Veh. Technol., vol. 46, no. 4,pp. 1040–1046, Nov. 1997.

[34] D. Astély et al., “LTE: The evolution of mobile broadband,” IEEECommun. Mag., vol. 47, no. 4, pp. 44–51, Apr. 2009.

[35] M. Paolini and S. Fili, “The economics of small cells and Wi-Fi offload,”SenzaFili Consult., Sammamish, WA, USA, 2012.

Ning Lu (S’12) received the B.Sc. and M.Sc. degreesfrom Tongji University, Shanghai, China, in 2007and 2010, respectively, and the Ph.D. degree fromthe University of Waterloo, Waterloo, ON, Canada,all in electrical engineering.

He is currently a Postdoctoral Fellow with theDepartment of Electrical and Computer Engineering,University of Waterloo. His research interests in-clude the fundamentals of wireless networking withspecial interests in connected vehicles.

Dr. Lu served as a Technical Program Committeemember for the 2012 IEEE International Symposium on Personal, Indoor, andMobile Radio Communications (PIMRC), the 2013 and 2014 InternationalConferences on Wireless Communications and Signal Processing, and the 2015International Conference on Computing, Networking, and Communications.

Nan Cheng (S’13) received the B.S. and M.S. de-grees from Tongji University, China, in 2009 and2012, respectively. He is currently working towardthe Ph.D. degree with the Department of Electricaland Computer Engineering, University of Waterloo,Waterloo, ON, Canada.

Since 2012, he has been a Research Assistant withthe Broadband Communication Research Group,Department of Electrical and Computer Engineering,the University of Waterloo. His research interests in-clude vehicular communication networks, cognitive

radio networks, and resource allocation in smart grid.

Ning Zhang (S’12) received the B.Sc. degreefrom Beijing Jiaotong University, Beijing, China,in 2007; the M.Sc. degree from Beijing Univer-sity of Posts and Telecommunications, in 2010; andthe Ph.D. degree from the University of Waterloo,Waterloo, ON, Canada, in 2015.

His current research interests include smart grid,fifth generation, and vehicular networks.

Xuemin (Sherman) Shen (M’97–SM’02–F’09) re-ceived the B.Sc. degree from Dalian Maritime Uni-versity, Dalian, China, in 1982 and the M.Sc. andPh.D. degrees from Rutgers University, Newark, NJ,USA, in 1987 and 1990, all in electrical engineering.

From 2004 to 2008, he was the AssociateChair for Graduate Studies with the University ofWaterloo, Waterloo, ON, Canada. He is currently aProfessor and the University Research Chair with theDepartment of Electrical and Computer Engineer-ing, University of Waterloo. His research interests

include resource management in interconnected wireless/wired networks, wire-less network security, social networks, smart grid, and vehicular ad hoc andsensor networks.

Dr. Shen served as the Technical Program Committee Chair/Cochair forthe 2010 Fall IEEE Vehicular Technology Conference (IEEE VTC) and the2014 IEEE International Conference on Computer Communications; SymposiaChair for the 2010 IEEE International Conference on Communications (IEEEICC); Tutorial Chair for the 2011 Spring IEEE VTC and 2008 IEEE ICC, theTechnical Program Committee Chair for the 2007 IEEE Global Communica-tions Conference; the General Cochair for the Annual International Conferenceon Mobile Computing and Networking, the 2007 International Conferenceon Communications and Networking in China, and the 2006 InternationalConference on Heterogeneous Networking for Quality, Reliability, Security andRobustness; and the Chair for the IEEE Communications Society TechnicalCommittee on Wireless Communications and Peer-to-Peer Communicationsand Networking. He serves/served as the Editor-in-Chief for IEEE NETWORK,PEER-TO-PEER NETWORKING AND APPLICATION and IET Communica-tions; a Founding Area Editor for the IEEE TRANSACTIONS ON WIRELESSCOMMUNICATIONS; an Associate Editor for the IEEE TRANSACTIONS ONVEHICULAR TECHNOLOGY, Computer Networks, and ACM/Wireless Net-works, etc.; and as a Guest Editor for the IEEE JOURNAL ON SELECTEDAREAS IN COMMUNICATIONS, IEEE WIRELESS COMMUNICATIONS, IEEECOMMUNICATIONS MAGAZINE, ACM Mobile Networks and Applications, etc.He received the Distinguished Performance Award in 2002 and 2007 from theFaculty of Engineering, University of Waterloo; the Premier’s Research Excel-lence Award from the Province of Ontario, Canada, in 2003; and the ExcellentGraduate Supervision Award in 2006 and the Outstanding Performance Awardin 2004, 2007, 2010, and 2014 from the University of Waterloo. He is an electedmember of the IEEE Communication Society Board of Governors and the Chairof the Distinguished Lecturer Selection Committee, as well as a DistinguishedLecturer of IEEE Vehicular Technology and Communications Societies. He isa Fellow of the Engineering Institute of Canada and the Canadian Academyof Engineering. He is a registered Professional Engineer in the province ofOntario, Canada.


Jon W. Mark (M’62–SM’80–F’88–LF’03) receivedthe Ph.D. degree in electrical engineering fromMcMaster University, Hamilton, ON, Canada, in 1970.

Since September 1970, he has been with theDepartment of Electrical and Computer Engineer-ing, University of Waterloo, Waterloo, ON, wherehe is currently a Distinguished Professor Emeritusand where he served as the Department Chair fromJuly 1984 to June 1990. In 1996, he established theCenter for Wireless Communications, University ofWaterloo, and is currently serving as its founding

Director. He had been on sabbatical leave at the following places: IBMThomas J. Watson Research Center, Yorktown Heights, NY, as a VisitingResearch Scientist (1976–1977); AT&T Bell Laboratories, Murray Hill, NJ,as a Resident Consultant (1982–1983): Laboratoire MASI, Universit Pierreet Marie Curie, Paris, France, as an Invited Professor (1990–1991); and Depart-ment of Electrical Engineering, National University of Singapore, as a VisitingProfessor (1994–1995). He has previously worked in the areas of adaptiveequalization, image and video coding, spread spectrum communications, com-puter communication networks, asynchronous-transfer-mode switch design,and traffic management. His current research interests include broadband wire-less communications, resource and mobility management, and cross-domaininterworking.

Dr. Mark served as an Editor for the IEEE TRANSACTIONS ON COM-MUNICATIONS from 1983 to 1990, a member of the Inter-Society SteeringCommittee of the IEEE/ACM TRANSACTIONS ON NETWORKING from 1992to 2003, a member of the IEEE Communications Society Awards Committeefrom 1995 to 1998, an Editor for Wireless Networks from 1993 to 2004,and an Associate Editor for Telecommunication Systems from 1994 to 2004.He received the 2000 Canadian Award for Telecommunications Research andthe 2000 Award of Merit of the Education Foundation of the Federation ofChinese Canadian Professionals. He is a Fellow of the Canadian Academy ofEngineering.

Fan Bai (SM’15) received the B.S. degree in au-tomation engineering from Tsinghua University, Bei-jing, China, in 1999 and the M.S.E.E. and Ph.D.degrees in electrical engineering, from the Univer-sity of Southern California, Los Angeles, CA, USA,in 2005.

Since September 2005, he has been a Senior Re-searcher with the Electrical–Control Integration Lab-oratory, Research and Development and Planning,General Motors Corporation, Warren, MI, USA. Heis also serving as a Ph.D. supervisory committee

member with Carnegie Mellon University, Pittsburgh, PA, USA, as well aswith the University of Illinois-Urbana Champaign, Urbana, IL, USA. He isthe author of about 40 book chapters and conference and journal papers,including papers presented at the ACM Annual International Conferenceon Mobile Computing and Networking; the IEEE International Confer-ence on Computer Communications; the ACM International Symposium onMobile Ad Hoc Networking and Computing; IEEE International Confer-ence on Sensing, Communication, and Networking; the IEEE InternationalConference on Communications; the IEEE Global Communications Con-ference; and the IEEE Wireless Communications and Networking Confer-ence. He has published in the IEEE JOURNAL ON SELECTED AREAS INCOMMUNICATIONS, IEEE WIRELESS COMMUNICATION MAGAZINE, IEEECOMMUNICATION MAGAZINE, and the Elsevier AdHoc Networks Journal.His current research interests include the discovery of fundamental princi-ples and the analysis and design of protocols/systems for next-generationvehicular ad hoc networks and for safety, telematics, and infotainmentapplications.

Dr. Bai has served as a Technical Program Cochair for the 2007 IEEEInternational Symposium on Wireless Vehicular Communications and the 2008IEEE International Workshop on Mobile Vehicular Networks. He serves asan Associate Editor for the IEEE TRANSACTIONS ON VEHICULAR TECH-NOLOGY and the IEEE TRANSACTIONS ON MOBILE COMPUTING and asa Guest Editor for IEEE WIRELESS COMMUNICATION MAGAZINE, IEEEVEHICULAR TECHNOLOGY MAGAZINE, and the Elsevier AdHoc NetworksJournal. He received the Charles L. McCuen Special Achievement Awardfrom General Motors Corporation in 2006, in recognition of extraordinaryaccomplishment in the area of vehicle-to-vehicle communications for driverassistance and safety.

Wi-Fi Hotspot at Signalized Intersection: Cost …n5cheng/Publication/Wi-Fi...LU et al.: Wi-Fi...

Documents

Transcript of Wi-Fi Hotspot at Signalized Intersection: Cost …n5cheng/Publication/Wi-Fi...LU et al.: Wi-Fi...