ABSTRACT. CARPENTER, SCOTT E. When evaluating safety in a vehicular ad hoc network, do not forget the obstacles.

Transcription

1 ABSTRACT CARPENTER, SCOTT E. When evaluating safety in a vehicular ad hoc network, do not forget the obstacles. Wireless communication between vehicles enables both safety applications, such as accident avoidance, and non-safety applications, such as traffic congestion alerts [1] with the intent of improving safety in driving conditions. Because the high costs of implementing limited test-bed environments constrain prototype testing, VANET researchers often turn instead to simulation toolsets from which a rich set of environmental scenarios are modeled. However, despite the availability of such tools, results are inconsistent. While VANET simulations allow investigators to select from a wide range of models for radio wave propagation, fading, and shadowing, they often fail to consider realistic road topologies and the presence of obstacles [2]. Safety performance assessments improve when models accurately reflect environmental conditions such as the fading effects of radio wave propagation through buildings and other obstacles. The goal of this research is to show quantitatively how accurate, deterministic obstacle fading models impact the performance assessment of VANET safety applications. An obstacle model and a fading model that uses it, along with a VANET simulation script, have been implemented and contributed to the open source ns-3 network simulation community. Results generated from simulation experiments that use these models show that deterministic obstacle fading greatly degrades Packet Delivery Ratio (PDR) and compares differently than stochastic fading models, such as Nakagami-m. Such network-level results can have dramatic impacts to safety performance assessment, such as awareness range. By considering the number of packets, time tolerance window, and awareness probability and

2 communications awareness range [3] can be used to assess application reliability. We can, in fact, assess the effectiveness of different applications by determining the likelihood that a target number of safety packets can be reliably received over a time window. Such results are insightful to the relative comparison of safety application performance. Failing to account for the effects of obstacles in safety assessment can therefore inaccurately or even greatly overstate the performance of VANET safety applications. Including realistic obstacle fading in VANET simulation modeling improves VANET assessment and strengthens safety, thus supporting one of the primary goals of connected vehicle systems. Copyright 2014 Scott E. Carpenter All Rights Reserved

3 When evaluating safety in a vehicular ad hoc network, do not forget the obstacles by Scott E. Carpenter A report for the written preliminary exam submitted to the Graduate Faculty of North Carolina State University in partial fulfillment of the requirements for the degree of Doctor of Philosophy Computer Science Raleigh, North Carolina 2014 APPROVED BY: Mihail L. Sichitiu, Ph. D. Committee Chair Insert committee member name Insert committee member name Insert committee member name

4 To God, who makes all knowledge possible. DEDICATION ii

5 BIOGRAPHY Scott Carpenter is an Information Analyst and Agile Project Manager at Houghton Associates, Inc., and a Doctoral student in the Department of Computer Science at North Carolina State University. In his current working roles, Mr. Carpenter analyzes a wide variety of technical problems in aviation and aerospace, much of which is conducted at the request of the Federal Aviation Administration (FAA) and involves cost-benefit analysis of complex systems. Prior to his recent position, Mr. Carpenter managed co-located and separated software development teams at Charles Schwab, Smith Breeden Associates, and Fidelity Investments. Furthermore, Mr. Carpenter has over 20 years of technical software engineering experience in computer graphics, data communications, protocol analysis, pharmaceutical, code analysis, finance, and aviation / aerospace. He specializes in software development using Java, C++, and SQL, and process management and improvement. He has also worked for IBM, Attachmate, and Quintiles. Mr. Carpenter achieved his Project Management Professional (PMP) certification in 2003, and is also a Certified Scrum Master (CSM). He has successfully managed numerous projects, using a variety of processes, tools, and techniques. Mr. Carpenter believes that strong team-based collaboration is very important when producing strong technical solutions. He currently favors agile practices, which he has championed and established within several organizations. Mr. Carpenter earned his BS degree in computer engineering from Case Western Reserve University in 1989, and his MS degree, also in computer engineering, from Case Western iii

6 Reserve University in His master s project involved an artificial intelligence (AI) based tutoring system. Mr. Scott E. Carpenter was named in honor of the American astronaut, Mr. Scott M. Carpenter, who was selected as one of the original seven Mercury Astronauts. Although they share no direct family history, Mr. Carpenter strongly admires his namesake and expresses his deepest thanks to his parents, as the family names being handed down to sons prior to his birth were primarily limited to Clarence, Eugene, and Earl. Considering that he may have thus been given one of those names, Scott likes very much the name that he was ultimately given. iv

7 TABLE OF CONTENTS Table of Contents Chapter Introduction... 1 Chapter Background and Related Work Vehicular Safety DSRC Standards and Operations Safety Assessment and Metrics Mobility Simulation Chapter Obstacle Models Radio Propagation Models Obstacle Modeling Methodology Chapter Experimental Setup and Results Experimental Setup Results and Discussion The Dinner Party Effect Safety Awareness Performance Threats to Validity Chapter Conclusions and Future Work Conclusions Future Work REFERENCES APPENDICES v

8 Appendix A Summary of Radio Propagation Models Appendix B Determining Awareness Range vi

9 LIST OF TABLES Table 1 DSRC device classes and operating parameters Table 2 Communications requirements for three CAMP/VSCC safety applications Table 3 Experimental parameters for path loss using (a) two-ray ground propagation loss only, (b) two-ray ground propagation loss and Nakagami-m fast fading, and (c) two-ray ground propagation loss and obstacle shadowing Table 4 Scenario topology parameters Table 5 Network simulation parameters Table 6 Awareness probabilities Table 7 Required PDR to achieve 95% awareness probabilities for 5, 7, and 9 messages out of 10 within a time window of 1 sec Table 8 Coefficients for (16) vii

10 LIST OF FIGURES Fig. 1 Total U.S. motor vehicle fatalities per 100,000 population, [4] Fig. 2 DSRC reference model Fig. 3 Representative VANET simulation tools. Adapted from: [5] Fig. 4 ns-3 reference model [6] Fig. 5 A downtown Raleigh, NC simulation scenario, in SUMO. Buildings (light shading) may significantly attenuate radio signal propagation between vehicles, affecting VANET application safety performance Fig. 6 Pseudo-code for Algorithm GETOBSTRUCTEDDISTANCEBETWEEN that determines the number of obstacle wall intersections and obstructed distance between two points Fig. 7 GoogleEarth view (a) and SUMO view (b) of the open highway scenario near Raleigh, NC USA Fig. 8 GoogleEarth view (a) and SUMO view (b) of the residential neighborhood scenario near Raleigh, NC USA Fig. 9 GoogleEarth view (a) and SUMO view (b) of the urban downtown scenario near Raleigh, NC USA Fig. 10 Process flow for experimental setup Fig. 11 PDR for residential neighborhood scenario with no fading, for vehicles as a function of the distance between transmitter and receiver Fig. 12 PDR for urban downtown scenario with 150 vehicles, for three different fading models as a function of the distance from the transmitter Fig. 13 PDR for three different building density scenarios as a function of distance from the transmission Fig. 14 PDR for three different fading models for (a) highway, (b) residential, and (c) downtown scenarios Fig. 15 Awareness probability vs. PDR for theoretical and simulation results for open highway scenario with 100 vehicles and no fading Fig. 16 Awareness range as a function of receiving n out of 10 safety messages per second vehicles using different fading models for 100 and 250 vehicles in (a) highway (b) residential, and (c) downtown scenarios Fig. 17 Awareness ranges for receiving n out of 10 safety beacons per second for the open highway scenario and different fading models for (a) n = 5, (b) n = 7, and (c) n = Fig. 18 Awareness ranges for receiving n out of 10 safety beacons per second for the neighborhood scenario and different fading models for (a) n = 5, (b) n = 7, and (c) n = Fig. 19 Awareness ranges for receiving n out of 10 safety beacons per second for the downtown scenario and different fading models for (a) n = 5, (b) n = 7, and (c) n = Fig. 20 The difference between awareness range (ΔR A ) for (a) open highway, (b) residential neighborhood, and (c) urban downtown scenarios Fig. 21 Average clock-time and confidence bars (minutes) of the sum of times to simulate 100 vehicles in the highway, residential, and downtown scenarios for different fading models viii

11 Fig. 22 Packet Delivery Ratio (PDR) for the downtown scenario with 150 vehicles and different fading models Fig. 23 Finding awareness range from Packet Delivery Ratio (PDR) for an obstacle fading scenario Fig. 24 Finding awareness range from Packet Delivery Ratio (PDR) for an obstacle fading scenario ix

12 Chapter 1 Introduction Collisions among moving vehicles lead all causes of traffic fatalities, injuries, and property damage [7]. By exchanging information using wireless communication, cars and trucks enable both safety applications, such as accident avoidance, and non-safety applications, such as traffic congestion alerts [1]. The US Intelligent Transportation Systems (ITS) Joint Program Office (JPO) suggests that vehicle safety applications and supporting technologies will prevent tens of thousands of automobile crashes every year [8]. The most promising technologies in the U.S. that will enable such a vehicular ad hoc network (VANET) are collectively referred to as Dedicated Short Range Communications (DSRC). While prototype testing thus far has been constrained to limited test-bed environments, numerous available simulation frameworks assist VANET researchers by providing toolsets with supporting models from which a rich set of environmental scenarios can be simulated. Despite the availability of such tools, results are inconsistent. While VANET simulations allow investigators to select from a wide range of models for radio wave propagation, the fading and shadowing models used are often entirely probabilistic and therefore fail to consider realistic road topologies and the presence of obstacles [2]. VANET modeling improves when a visibility scheme describes the topology as a configuration space and supports obstacle detection deterministically based on a realistic environmental representation. Model accuracy influences simulation results and directly impacts the performance assessment of VANET safety applications. 1

13 VANET applications differ by their communications requirements, which include [9]: Update Rate, Allowable Latency, Data to be Transmitted and/or Received, and Maximum Required Range of Communications. For example, a safety-critical application such as forward collision warning system (FCWS) requires rapid information dispersal of the SAE J2735 Basic Safety Message (BSM) to nearby traffic [10]. Other non-safety, broadcastoriented applications, such as traffic-jam alert or vehicle-to-vehicle video chat, require efficient information delivery to large or targeted areas with reduced reliability and delay requirements as compared to safety applications [10]. By considering the number of packets, time tolerance windows, and awareness probability, a communications awareness range [3] can be used to assess application reliability. We can, in fact, assess the effectiveness of different applications by determining the likelihood that a target number of safety packets can be reliably received over a time window. While such results are insightful to the relative comparison of safety application performance, they do not allow us to answer the general, open question, How much safety can be achieved by Intelligent Transportation Systems? The goal of this research is to show quantitatively how accurate, deterministic obstacle fading models impact the performance assessment of VANET safety applications. Our investigations are guided by the following research questions (RQ): RQ1: Can fast fading and shadowing effects of obstacles, such as buildings, vehicles and trees, be modeled and efficiently simulated in ns-3 [6]? 2

14 Because model realism improves simulation results, a deterministically real fast fading and shadowing model that accounts for radio wave propagation through obstacles will improve the usefulness of existing detailed network simulation tools, such as ns-3. RQ2: How does an obstacle model effect packet delivery ratio as compared to other VANET fading and shadowing models? By implementing a deterministic obstacle fading model and comparing simulation results to other stochastic shadowing models, a performance characterization of the obstacle fading model can be made. RQ3: What is the impact of a deterministic obstacle fading model to VANET application safety performance as compared to other fading and shadowing models? By using simulation results to assess safety performance, the safety impact of a deterministic obstacle fading model can be quantifiably compared to other fading and shadowing models. Our primary contributions include: 1) An obstacle model built using techniques from computational geometry and a fast-fading model that uses it, along with a VANET simulation script, were developed and contributed to the open source ns-3 network simulator community. 2) Simulation results of highway, residential, and downtown scenarios in the Raleigh, NC area compare quantitatively the effects of i) deterministic obstacle shadowing to ii) stochastic Nakagami-m fast fading and iii) no fading. 3

15 3) A resulting comparison that extends awareness range determination using simulation results and suggests the consequences of ignoring obstacles when evaluating VANET safety application performance. 4) Detailed analysis of results. This paper is organized as follows: Chapter 2 provides background information and discusses related work and Chapter 3 describes the modeling approach and details the methodology. Chapter 4 describes the experimental setup, results and analysis and Chapter 5 concludes the paper and discusses future work. Two appendices extend explanations; Appendix A details radio propagation models, and Appendix B gives an example of how safety awareness range is calculated. 4

16 Chapter 2 Background and Related Work 2.1 Vehicular Safety Collisions among moving vehicles lead all causes of traffic fatalities, injuries, and property damage. According to data from the National Highway Transportation Safety Administration (NHTSA), although approximately 5.3 million vehicle crashes in the US resulted in about 32,000 fatalities in 2011, motor-vehicle fatalities continue to decrease as safety measures, such as seal belts and airbags, have been mandated (see Fig. 1) [4]. Intending to further improve driving safety and reduce traffic-related fatalities, the United States Department of Transportation (USDOT) expects to push for regulations requiring vehicles to communicate cooperatively using wireless technologies [11]. The most promising technologies in the U.S. that will enable such a vehicular ad hoc network (VANET) are collectively referred to as Dedicated Short Range Communications (DSRC) National seat belt law passed Airbags requried Source: NHTSA Fig. 1 Total U.S. motor vehicle fatalities per 100,000 population, [4]. 5

17 Management Security By exchanging information using wireless communication, cars and trucks enable both safety applications, such as accident avoidance, and non-safety applications, such as traffic congestion alerts [1]. The US Intelligent Transportation Systems (ITS) Joint Program Office (JPO) suggests that vehicle safety applications and supporting technologies will prevent tens of thousands of automobile crashes every year [12]. 2.2 DSRC Standards and Operations In the U.S. Dedicated Short-Range Communications (DSRC) technologies support wireless communication within a vehicular ad hoc network (VANET). DSRC-related standards include IEEE Std [13], IEEE Std. 1609/WAVE [14] [15] [16] [17] [18], and SAE J2735 Message Set Dictionary [19]. Figure 2 shows the DSRC reference model. Application Applications DSRC Message Set Dictionary SAE J2735 Transport TCP UDP Internet IPv6 WSM Data Link Physical LLC MAC PHY (p) Fig. 2 DSRC reference model. IEEE Std [13] incorporates modifications to the MAC and PHY layers (formerly known as IEEE p), and IEEE Std. 1609/WAVE provides further 6

18 capabilities, such as: channel allocation and multi-channel access, priority queueing and channel routing, congestion control, security and privacy mechanisms, and an application programming interface (API) for messaging. IEEE Std specifies the MAC and PHY layers of the reference model and provides extensions intended to address the dynamic nature of potentially fast-moving vehicles. A primary enhancement allows a STA that is not a member of a Basic Service Set (BSS) to transmit data frames, allowing the PHY to operate outside the context of a BSS, (i.e., OCB) and thus defining a new type of communications. The PHY layer of DSRC uses Orthogonal Frequency Division Multiplexing (OFDM) at a 10 MHz channel bandwidth to achieve data rates from 3 to 27 Mbps. Applying forward error correction (FEC) reduces the effective bit rate but improves the probability of successful decoding [20]. The majority of DSRC testing in the U.S. has utilized the 6 Mbps configuration (i.e., Quadrature PSK with rate 1/2 coding), since it balances channel load and signal-to-noise requirements [21]. While the family is well-understood, unicasting is not well-suited to the VANET environment [22]. The IEEE Std supports safety beacons via broadcast transmissions (i.e., beaconcasting) using Carrier-Sense Multiple Access with Collision Avoidance (CSMA/CA). While sensing of the network delays a transmission which would otherwise cause a collision, CSMA/CA does not prevent all such collisions, especially in the hidden node scenario [23] in which one vehicle cannot sense that a transmission would collide with another nearby vehicle. Transmitters employ a back-off scheme when unicast 7

19 collisions occur to re-attempt a transmission at a randomly selected future time controlled using the Congestion Window (CW) parameter. The IEEE 1609/WAVE stack is built on top of the MAC and PHY layers. WAVE supports IPv6-based data transfers and also non-ip-based traffic through the WAVE Short Message Protocol (WSMP). IEEE specifies the WAVE Management Entity (WME) and corresponding network services, as well as WSMP. Channel coordination is a collection of enhancements to the IEEE MAC, and interacts with the IEEE LLC and IEEE PHY; IEEE describes multi-channel operations. WAVE security services are specified in IEEE The design of IEEE 1609/WAVE supports a single control channel (CCH) and six service channels (SCH), and it is generally assumed that the CCH will be mostly dedicated to safety applications [20]. Channels are defined in the 5.9 GHz range and typically occupy 10 MHz each, although the potential exists for 5 MHz and/or 20 MHz channels. While the WAVE standard does not preclude multiple channel-dedicated radios for continuous channel access, it also supports three additional options for periodic channel switching, commonly known as alternating, immediate, and extended channel access modes. Sync intervals split the channel at a rate of 10 Hz, which CCH and SCH intervals then typically split equally to support channel switching. A 4ms guard interval separates channel switches, resulting in a CCH interval of (½ x 1/10s) 4ms = 46ms. Vehicles operate autonomously to synchronize clocks using an external GPS or a WAVE-advertised timing service. The WAVE standard addresses message priority using four different Access Classes (AC) per (CCH or SCH) channel, AC0 (lowest priority) to AC3 (highest priority), with the 8

20 MAC layer maintaining separate queues and channel access for each AC [20]. The WAVE contention mechanism is similar to the one used in conventional Wireless Local Area Network (WLAN) and the IEEE e Enhanced Distributed Channel Access (EDCA) Quality of Service (QoS) enhancements [24]. During packet transmission selection, the four ACs first contend internally with the winning packet then contending for the channel externally using its contention parameters. Each AC specifies a number of Arbitration Inter- Frame Spacing (AIFS) and CW slots and each AC waits at least its AIFS slots, plus additional slots determined by the selected CW value. Using the SAE J2735 DSRC Message Set Dictionary [19], the VANET applications are developed over the 1609/WAVE stack. Data elements that enable many safety applications are described in the most important message in the J2935 standard [20], the Basic Safety Message (BSM), which every vehicle has to broadcast at a nominal rate of 10 Hz. The BSM is divided into i) Part I mandatory elements, e.g., position, motion, braking status, and vehicle size and ii) Part II optional elements, e.g., vehicle events, path history and path prediction. Additionally, SAE J2735 defines non-safety messages, e.g., for toll collection, and also provides guidelines on message prioritization among different message types. Applications may involve strictly vehicle-to-vehicle (V2V) messaging using an On-Board Unit (OBU), or may also involve the use of a Roadside Units (RSU) to support vehicle-toinfrastructure (V2I) communications. The Federal Communications Commission (FCC) defines four classes for DSRC device operations with desired communications zones ranging from 15 to 1000 meters, as shown in Table 1 [25] [20]. While vehicles and infrastructure expect to communicate reliably over these ranges, radio-blocking obstacles and other 9

21 interference that prevent message delivery challenge the safety effectiveness of all applications. RSU Class Maximum output power (dbm) Communications zone (meters) A 0 15 B C D Table 1 DSRC device classes and operating parameters. Many VANET applications have been proposed. The Crash Avoidance Metrics Partnership (CAMP) put forth a comprehensive list for the Vehicle Safety Communications Consortium (VSCC) [26] that was subsequently used within SAE J2735. Examples of purely V2V safety applications that assist with impending crash notifications include: Approaching Emergency Vehicle Warning, Wrong Way Driver Warning, Cooperative Forward Collision Warning, Emergency Electronic Brake Lights, Lane Change Warning, Blind Spot Warning, Highway Merge Assist, Cooperative Collision Warning, and Pre-Crash Sensing. 2.3 Safety Assessment and Metrics The plethora of envisioned safety applications creates confusion when attempting to evaluate and compare their performance. For example, what is the difference between Cooperative Forward Collision Warning (CFCW), Emergency Electronic Brake Lights (EEBL), and Pre-Crash Sensing (PCS)? VANET applications differ by their communications requirements, which include [9]: Transmission Mode, Update Rate, Allowable Latency, Data to be Transmitted and/or Received, and Maximum Required Range of Communications. For example, Table 2 summarizes the data frequency (i.e., update rate), latency, and range communications requirements for CFCW, EEBL, and PCS and shows that 10

22 FCW and EEBL have similar performance criteria, except that EEBL functions over a longer communication range (~300m) than FCW (~150m). Furthermore, PCW requires the highest update rate, the least latency, and the shortest communications range. Intuitively, we might say that PCW has stricter operational characteristics and thus may have a greater safety meaning to the driver. Cooperative Forward Collision Warning Emergency Electronic Brake Lights Pre-Crash Sensing Application (FCW) (EEBL) (PCS) Update Rate ~10 Hz ~10 Hz ~50 Hz Latency ~100 ms ~100 ms ~20 ms Range ~150 m ~300 m ~50 m Fig. 3 - Comparison of Communications Requirements for three CAMP/VSCC safety applications Table 2 Communications requirements for three CAMP/VSCC safety applications. Driving condition safety improvements from VANET technologies remain difficult to assess. Foremost, while numerous safety applications have been proposed, none are standardized [20]. Furthermore, prototype testing thus far has been constrained to limited test-bed environments where DSRC adoption rates have been relatively low, and even if comprehensive testing results were available, significant legal hurdles remain regarding risks and liabilities that could jeopardize successful deployment [27]. Put bluntly, if injuries still occur in operational conditions involving safety-certified VANET applications, then who is responsible? How, then, do we assess application safety in a VANET? The WAVE protocol stack of Fig. 2 shows a five-layer model top-down from Application to Physical layers and various metrics are used to assess the performance of each layer. Indeed, researchers [24] [28] [29] have studied in depth the physical layer performance of IEEE Std / WAVE VANETs, commonly using throughput, end-to-end delay, an collision probability as comparative 11

23 metrics. While network performance is often assessed using throughput and end-to-end delay to quantify Quality of Service (QoS), these metrics generally express the Link and MAC layer effectiveness. In a VANET, information dissemination of safety packets is critical, requiring a need to evaluate network-layer routing and delivery [30]; for accurate packet delivery we have to consider the packet delivery ratio (PDR) [31] and packet drop rate [32]. Additionally, routing overhead (i.e., number of routing bytes required by the routing protocol to construct and maintain its route [33]) and hop count from source to destination help assess routing performance [32]. Potential safety benefits of VANET applications can be assessed based on the estimated effectiveness of the application in terms of reduction of crash-related factors (e.g., functional years lost, vehicles crashed, and direct costs) [9]. While network level metrics have been historically important to the understanding and tuning of traditional Internet and Mobile Ad hoc Networking (MANET) communities [30], application performance depends instead on reliability metrics [30] [3] [31]. Performance measurements considered are often network- (or packet-) level metrics (e.g., PDR and Average Per-Packet Latency) and/or application-level reliability metrics (e.g., Application-level T-Window Reliability [30]). Packet Delivery Ratio PDR net (d) is the percentage of broadcast packets that are successfully received by all vehicles within each transmitting vehicle s coverage range, d [30] [31]. Receipt failures degrade PDR due to concurrent transmissions, hidden terminal problems and the impact from channel fading [31]. PDR is derived [31] with respect to two aspects: 1) PDR C (x), the impact of collisions due to concurrent transmissions, and 12

24 2) PDR F (x), the impact of channel fading and path loss. Taken together, it can be shown [31] that: ( ) ( ) ( ) (1) Average Per-packet latency,, is the duration of the time between the packet generation by a transmitting vehicle and its successful receipt at a receiving vehicle [30]. Application-level T-window reliability (TWR), P app (d), is the probability of receiving at least one packet out of multiple packets from a broadcasting vehicle at distances smaller than or equal to d, within a time interval, T. P app (d) describes the application-level reliability for safety applications, whereas PDR net (d) describes the wireless communication reliability at the packet level [30]. Application-level reliability can be expressed in terms of network-level reliability as given in (2) [30]: ( ) ( ( )), (2) where t is the beaconcasting interval and τ is the tolerance time window. Generalizing, the awareness probability, P A (x, n, T Tol ), as given in (3), is the probability of receiving at least n packets in the tolerance time window, T Tol [31] [3], where P s (x) = PDR net (x); awareness probability equals application-level T-window reliability when n = 1. ( ) ( ) ( ) ( ( )) (3) By considering the number of packets and time tolerance windows, awareness probability can be applied to different safety applications in order to assess application reliability. We can, in fact, assess the effectiveness of FCW and EEBL, both of which expect 10 messages per second. However, although an awareness probability can be calculated analytically using 13

25 network-level PDR, confusion remains over how to best use this application-level metric to evaluate application performance, as there are no standards that define suitable numbers for messages received or time tolerance windows and the requirements are hard to capture. For example, do we need to receive 1 message every 0.1 seconds? Or 5 messages every half a second? Or ten messages in one second? What if we receive 9 messages in one second, or 99 messages in 10 seconds would that be safe? The authors of [3] provide a methodology that extends the use of awareness probability to instead determine the awareness range, R A (P A )=d, which is the maximum distance, d, at which a threshold awareness probability, P A, is met or exceeded. By first fixing an awareness probability threshold (i.e., P A = 90%, 95%, 99.99%), the network-level PDR necessary to achieve P A can be determined and the effective awareness range of the PDR can further be calculated. The awareness range can then be compared to the vehicular operating conditions (e.g., braking distance, expected communications range, etc.) to determine if such conditions can be met. Intuitively, awareness range describes whether or not there is sufficient distance and time for a safety application to support user responses that maintain safe operating conditions. Therefore, this is conceptually much closer to a suitable measure of application-level performance. Nonetheless, the authors strongly caution that even they have failed to answer the general, open question, How much safety can be achieved by Intelligent Transportation Systems? [3]. 2.4 Mobility Vehicular movements are modelled using mobility simulators, which often support various mobility models, such as random waypoint (RWP), random direction (RD), and 14

26 Manhattan grid. The Simulator for Urban Mobility (SUMO) [34] is an open source macroscopic traffic simulation package [35] that represents traffic demands throughout road networks with the capability to produce vehicular movement trace files in a number of formats. SUMO can import road network data from various sources including TIGER, OpenStreetMap (OSM), RoboCup, and opendrive. Trips (i.e., routes) can be explicitly defined or generated randomly from pathways within the road network. SUMO supports vehicle variety in which vehicles behave as expected according to laws of physics, and can use car-following models, such as the Krauss model [36]. Lane-changing models may be employed. Vehicles obey traffic light signaling and other traffic laws. SUMO integrates with several network simulation tools, and provides a real-time interactive control interface, TRaCI. 2.5 Simulation Numerous available simulation frameworks assist VANET researchers by providing toolsets with supporting models from which a rich set of environmental scenarios are simulated. Often treated as separate architectural modules, simulators address mobility, networking, and radio propagation [37], while some tight coupling among these primary modules may be found in some over-arching VANET simulators [5]. For a good review of VANET simulator toolsets and models, see [5]. However, despite the availability of such tools, results are inconsistent. For example, when simulating the same environment and protocol, one investigation of VANET mobility generators shows performance deviations among them, making simulation results unconvincing and inconclusive [5]. Indeed, the simulation requirements of VANET environments presents numerous challenges to the 15

27 VANET researcher, such as constrained road topology, multi-path fading and roadside obstacles, traffic flow models, trip models, varying vehicular speed and mobility, traffic lights, traffic congestion, drivers behavior, etc. [5] The level of capability for addressing these challenges varies among current simulators. Most concerning to environmental accuracy, VANET simulations often fail to consider realistic road topologies and the presence of obstacles [2]. Figure 3 shows a classification of representative mobility generators and VANET and network simulators. Fig. 3 Representative VANET simulation tools. Adapted from: [5]. Because obstacles not only constrain vehicular movement but also interfere with radio transmissions, accurate environmental representation of obstacles is absolutely required in urban scenarios to benefit VANET simulation, although less critical in highway scenarios [37]. Previous simulation and experimental research studies the impact of radio obstacles, with obstacle classes (i.e., buildings or vehicles) often studied separately. For example, the 16

28 authors of [38] investigate OLSR in a simulation environment with and without large buildings and conclude that the presence of buildings creates communication holes resulting in disconnections. The authors of [39] use GloMoSim to evaluate under street environments the impact to routing protocols using a Manhattan-style grid environment and assuming all non-street areas are large buildings. The authors of [40] use VanetMobiSim and ns-2 to investigate the effects of four kinds of radio signal features (including one in which the signal can be diffracted up to 15 at the edge of obstacles), while assuming uniformly large rectangular buildings on 4 block x 4 block streets. Instrumenting vehicles to characterize IEEE b signal propagation between them in a variety of urban and suburban settings, the authors of [41] found that signal propagation varies especially between line-of-sight (LOS) and obstructed-line-of-sight (OLOS). The authors of [42] found similar effects occur when radio waves propagate Around-the-Corner (ATC). Due to the costs of vehicles, communications equipment, and road systems access and control, establishing a vehicular ad hoc network (VANET) test environment remains mostly cost-prohibitive in academic research, although collaborative organizations have developed some limited connected-vehicle test beds [43]. While simulation tools thus remain a highly useful tool set to the VANET modeler, realism continues to present challenges within the environmental models including: vehicular mobility, city-scape, and radio wave propagation loss and shadowing. Radio wave attenuation is often modeled deterministically based mainly on inter-vehicular line-of-sight (LOS) distance, while some modelers improve upon this with stochastic modeling intended to account for radio wave shadowing, which may be further characterized by environmental differences, such as open highway versus urban settings. 17

29 Recent models allow shadowing to be modeled more deterministically when interference includes radio wave propagation through buildings [42] [2] [44]. However, availability remains limited for such models within non-proprietary network simulation tools. For example, within OMNeT++ [45], the Simple Obstacle Model [42] is only available through Veins [46], making it useful only to vehicular network modeling. Within ns-3 [6], a buildings model exists, but limits buildings to cubes with linearly placed x- and y- coordinates and is implemented mainly for LTE modeling only. Figure 4 shows the ns-3 reference model. None of these models generically supports realistic building footprints for shadowing effects that are available to all simulation models (e.g., LTE, WiFi, Bluetooth, Zigbee, etc.). protocols applications internet Source code modules test helper devices network core propagation mobility Usage and examples High-level wrappers for everything else Aimed at scripting Mobility modules (static, random walk, etc.) Packets, Packet tags, Packet headers, Pcap/ASCII file writing Node class, NetDevice, Address types (IPv4, MAC, etc.), Queues, Sockets Smart pointers, Dynamic type system, Attributes, Callbacks, Tracing, Logging, Random variables, Events, Schedulers, Time arithmetic Fig. 4 ns-3 reference model [6]. 18

30 Chapter 3 Obstacle Models 3.1 Radio Propagation Models A Radio Propagation Model (RPM) handles the effects of signal attenuation due to distance, multipath signal fading due to reflectors, and shadowing as radio waves move through free space and obstacles, such as buildings. Modeling that uses empirical data to evaluate a vehicle-to-vehicle communications channel shows that path loss is primarily influenced by: i) LOS signal attenuation, ii) fast fading effects such as strong ground reflection, and iii) slow fading from random scatters and shadowing effects [47]. Many different RPMs have been proposed for VANET, ranging in their design complexity from simpler models such as unit-disk up through more complex ones that specifically address the interference that obstacles cause. Appendix A summarizes the RPMs included herein. The simplest model, often used in VANET simulation, is the unit-disk model, in which vehicles can communicate with each other if they are within a threshold distance and cannot communicate otherwise [48]. The authors of [49] conclude a complex shadow fading environment is well approximated by a simpler and more tractable unit-disk model. Another commonly simulated RPM in VANET is the two-ray model, which takes into account signal reflection from the road surface and sufficiently represents path loss in IVC [50]. Because signals between road-based vehicles are assumed present in at least direct 19

31 LOS and ground reflection, the two-ray model is more appropriate for VANET than the freespace model. The authors of [51] claim that the approximation of path loss given by the Two- Ray Interference model is more accurate than a simplified two-ray ground model used within most simulation tools. The authors of [48] conclude that the largely used unit disc model fails to realistically model a communication channel, while parameters of simplistic models like log normal can be adjusted to match the corresponding system metrics of more complex and hard to implement obstacle based models. For (small-scale) fast fading models, various stochastic distributions have been proposed, including Rice, Rayleigh, Nakagami-m and Weibull distributions [48]. The authors of [52] report that the large-scale fading in radio propagation channels at 5.3 GHz was found to be lognormally distributed, whereas the small-scale fading was characterized by the Weibull distribution. Channel models commonly used for multipath and shadow fading can be normalized by the general gamma distribution, as in (4) [53]: ( ) ( ) ( ), (4) where α is the fading parameter, β is the power-scaling parameter, v is the shape parameter, and Γ(α) is the gamma function. Distribution special cases include Nakagami-m (v=1), Weibull (α=1), and Rayleigh (v=α=1), while the lognormal distribution is a limiting case ( ). The Nakagami-m fading model determines signal power reception probabilistically dependent on model parameters that simulate fading levels. The authors of [54] assume 20

32 symmetric physical radio channels with Nakagami-m fading. The Nakagami-m fading model may be described as given by (5): ( ) ( ), (5) where m is the Nakagami parameter (i.e., shape parameter), Γ(m) is the gamma function, and Ω is the average power of multipath scatter field, which controls the spread of the distribution. The Nakagami-m distribution in (5) can be calculated using the general gamma distribution of (4) by setting v = 1, α=m, and β=ω/m. The stochastic nature of Nakagami-m fading fails to differentiate between line of sight (LOS) and obstacle-blocked vehicles and does not take inter-vehicle distance into account, thus resulting in the model failing to deterministically address building (i.e., obstacle) density differences between otherwise identical scenarios. Stochastic models determine the physical parameters of the vehicular channel in a completely probabilistic manner without considering underlying geometry [48]. Stochastic communications modeling could therefore deviate severely from realistic behavior, negatively impacting simulations of transmission-critical safety applications [42]. Obstacles play a part in radio signal propagation, with various modeling approaches being undertaken. Both Radio Propagation Model with Obstacles (RPMO) [55] and the Mahajan Model [56] behave like two-ray ground, adding the influence of obstacles and distance attenuation [57]. The impact of surrounding vehicles as obstacles is considered lognormally by [48] and using a multiple knife-edge model by [58]. The authors of [59] assert 21

33 that obstruction by a large building does not involve a pure shadowing but heavily reduces the received power. Applying sophisticated methods like ray-tracing achieves realistic representation of signal propagation but remains impractical for evaluating typically large-scale VANETs [41] since it requires site-specific propagation details [48]. A progressive series of models combining distance-based attenuation with building modeling has been proposed by the authors in [57] [60] [61]. The authors of [41] propose a hybrid solution that differentiates between LOS conditions using one set of settings for LOS conditions, and another for Around the Corner (ATC) settings, based on experimental measurements. The authors of [62] propose a similar approach, with different settings for LOS vs. NLOS, using a multi-ray propagation model and including the foliage effect in both NLOS and LOS models. Similarly the authors of [44] classify vehicle communications potential into one of three cases: LOS, NLOS1 (NLOS with one corner along the path) and NLOS2 (NLOS with two corners along the path). The authors of [59] also use two classifications of NLOS, but simulate only in a Manhattan grid scenario. Because radio waves can physically penetrate one or more buildings, models based on direct LOS alone are insufficient, leading the authors of [42] to collect empirical results showing the effects of building shadowing in urban environments. Gathering measurements using IEEE p devices, the authors of [42] present a realistic, yet computationally inexpensive path loss simulation model that uses the number of obstacle walls and distance between two vehicles to estimate the effect that buildings and other obstacles have on the 22

34 radio communications. However, in later work, the authors of [63] point out simulating path loss in (sub)urban environments to capture predictable shadowing effects seems to require more complex models than attenuation per wall or attenuation per meter of penetration approaches while noting that the computational complexity of the resulting required ray-tracing approaches makes this prohibitive. Similarly, the authors of [44] propose CORNER, a low computational cost yet accurate urban propagation model for mobile networks that uses information from urban digital maps to estimate building and obstacle presence along the signal path between two vehicles, classifying the propagation environment between them. 3.2 Obstacle Modeling VANET modeling improves when a visibility scheme describes the topology as a configuration space and supports obstacle detection. Most visibility schemes divide the configuration space into sub-areas by some criteria. For example, the Manhattan grid model assumes that all vehicles move only in streets arranged as a Manhattan-style grid and treats non-street areas as buildings. However, Manhattan layouts are uncommon in real scenarios [60]. The authors of [44] extend the areas of each roadway segment to include a proximity area to include locations within one half-width of the roadway. Figure 5 shows an example urban downtown scenario (Raleigh, NC USA) simulated in SUMO using buildings data from OpenStreetMap (OSM) [64]. 23

35 Fig. 5 A downtown Raleigh, NC simulation scenario, in SUMO. Buildings (light shading) may significantly attenuate radio signal propagation between vehicles, affecting VANET application safety performance. The authors of [60] propose Visibility Scheme for Real Maps, designed to be used in scenarios where streets are irregular. However, the model assumes a divide of the 24

36 configuration space into only streets and non-streets, with vehicles only traveling along street centerlines, resulting in an assumption that communications are only possible between vehicles mutually visible within streets. Open communication areas often extend beyond street edges in real urban scenarios, limiting the model s realism. The authors of [2] propose Topology-Based Visibility model, which extends their previous work of Building and Distance Attenuation Model (BDAM) [57] to take into account realistic road topologies, such as roundabouts, angled roads, merged-and-split roads, etc., that are not considered by other schemes. The approach converts the roadmap into an undirected graph where junctions are vertices and streets which connect them are edges. The resulting polygonal areas create realistic city profiles [2] as some of the polygons formed using the streets as sides are considered clear areas with little interference in the signal propagation (obstacle-free areas such as roundabouts, gardens, etc.), and the rest of the map is regarded as a set of buildings. In this scheme, radio signals can propagate through streets and clear areas but not buildings. Additionally, in certain situations, vehicles close to junctions are able to receive enough power to obtain the messages in non-line-of-sight (NLOS) conditions [2]. As such, the model accounts for message receipt potential for vehicles within 20 meters of a junction. Current approaches to detect obstacles within a configuration space employ different algorithms with varying complexity. Many come from computational geometry techniques including intersection problems [58] [42] and binary space partitions (BSP) [65] [42] that are applied to model vehicles and/or buildings as obstacles. 25

37 Geometric intersection problems deal with pairwise intersections between line segments in n-dimensional space [58]. In their study, the authors of [58] restrict their line segments of interest to LOS rays between vehicles and lines composing a bounding rectangle representing the vehicle, further restricting the class of intersection problems to so-called red-blue intersections. The problem is thus stated: Given a set of red line segments r and a set of blue line segments b, with a total of N = r + b segments, the goal is to report all K intersections between red and blue segments. The authors of [58] report that Agarwal presented an efficient algorithm with time-complexity using the randomized approach of O(N 4/3 log N + K), where K is the number of red-blue intersections, and with space complexity of O(N 4/3 ). The authors of [65] find V2V obstacle intersections using a Binary Space Partitions (BSP) algorithm, but only apply the approach to a Manhattan-style grid scenario with uniformly quadratic building sizes with edge length of 480 m. Obstacle edges are stored in a BSP structure. Using a bounding box derived from the LOS path between two vehicles, the potential obstructing buildings between them are identified. Lastly, obstacle edges are tested for intersection with the LOS path. The calculation of intersection between all lines of sight and all buildings is an expensive step [42]. However, finding the intersections can be done in O(n 2 log n) using caching and binary space partitioning approaches [65]. 3.3 Methodology This section elaborates the research methodology in terms of each request question (RQ). RQ1: Can fast fading and shadowing effects of obstacles, such as buildings, vehicles and trees, be modeled and efficiently simulated in ns-3? 26

38 An obstacle fading model was developed for the ns-3 network simulation toolset that uses computational geometry techniques (e.g., an obstacle model) to determine NLOS distances traversed and number of segment intersection (i.e., walls penetrated) in roadway scenarios using building footprint information from OSM. The models are tested using a VANET simulation script that was also developed for ns-3, and all code (obstacle model, obstacle fading model, and VANET simulation script) have been contributed to the ns-3 community. Three environments differentiate between obstacle density and illustrate the problem: i) an open highway setting, ii) iii) a residential neighborhood, and an urban downtown scenario. Street and building data derives from an open source, OpenStreetMap [64]. Multiple experiments inject 50 to 250 vehicles (in steps of 50) into each scenario to model vehicular mobility. Using SUMO, randomly-placed vehicles move realistically and follow routes that intend to obey expected traffic laws, lane changing, and traffic light signaling. Vehicles vary speed and braking using common car-following models, such as the Krauss Model [36]. Generated ns-2 movement files capture the resulting mobility traces. Modeling assumes a 100% DSRC equipage penetration rate and all vehicles use a single p continuous access 10 MHz channel operating in the 5.9 GHz range to broadcast a 200-byte BSM 10 times per second using WiFi and WAVE models from a network simulator, ns-3. As radio waves propagate through space (e.g., directly through free space, reflected off of surfaces, and/or through objects), the power density diminishes as a result of what is 27

39 commonly referred to as path loss. The transmitter and receiver power relates to path loss as follows: when the power at the input (i.e., transmitter) and output (i.e., receiver) are represented as and, respectively, then the power ratio, R, is given in (6): (6) Path loss is often expressed in units of decibel (i.e., db) as in (7): [ ] ( ) (7) where X represents a value to be converted to db, ( ) represents the base-10 logarithm, and [ ] is the converted value. In all equations that follow, path loss terms (i.e., L x ) are in db. The path loss in db, L, is defined as the power ratio, R, in decibels, which is calculated by substituting (6) into (7) to arrive at (8): ( ) (8) Total path loss is modeled as the cumulative effects of: i) signal attenuation due to distance (i.e., propagation loss), ii) iii) fast fading multipath such as ground reflections, and slow fading from shadowing due to buildings and other obstacles. Therefore, total path loss, L t, can be expressed as:, (9) where L p represents path loss due to propagation, L f represents path loss due to fading such as multipath interference, and L s represents path loss due to shadowing such as obstacles. Each of the terms in (9) is now considered in turn. 28

40 Path loss due to propagation (i.e., the L p term in (9) is simulated in all experiments using the two-ray ground propagation loss model, thusly following the works of [42] and [46]. The relation between received power after propagation loss, P r,prop_loss, and initial transmission power, P t, as a function of distance, d, can be approximated using the two-ray ground propagation model as given in (10), where G t and G r are the transmitter and receiver gains, respectively, and h b and h m are the heights of the receiver and transmitter, respectively: By rewriting (10), the power ratio, R p, as in (6) can be stated as (11): (10) (11) Two-ray ground path loss, L p,2-ray, can be calculated by substituting (11) and (6) into (8) and simplifying as in (12-15): ( ) (12) ( ) (13) ( ) ( ) (14) ( ) ( ) (15) Path loss due to fast fading (i.e., the L f term in (9)), L f,nak, is modeled assuming Nakagami-m fading and can be expressed as (16): (16) where I f,nak is the fading loss indicator function (i.e., 1 = include Nakagami-m fading, and 0 = exclude) and L Nak is the resulting path loss resulting from the Nakagami-m distribution in (5). Fast fading further reduces power received after propagation loss. Therefore, the fast 29

41 fading output power, P o = P r,nak, can be calculated for an input power, P i = P r,prop_loss, as given in (17): ( ) (17) where m Nak is the Nakagami-m shape parameter and the Nakagami distribution is calculated using the gamma distribution (4) with v = 1, α=m Nak, and β=p r,prop_loss /m Nak. Substituting (17) into (6), where P i = P r,prop_loss and P o = P r,nak, leads to (18): ( ) (18) Further substituting (18) and (6) into (8) results in (19-20): ( ) (19) ( ) (20) ( ) Slow fading shadowing effects of obstacles (i.e., L s of (9)) are modeled using the Simple Obstacle model of [42], in which obstacle shadowing path loss, L s,o, is dependent on both per-wall-attenuation and per-meter-attenuation as given generally in (21): ( ) (21) where I s is the shadowing loss indicator function (i.e., 1 = include obstacle fading / shadowing and 0 = exclude), α is the attenuation per wall, in db, n is the number of walls penetrated, β is the attenuation per meter, in db, and m o is the distance, in meters, traveled through obstacles. Prior to the development and contribution to the code base of our obstacle shadowing model, a deterministic obstacle shadowing model did not previously exist in ns-3. In the 30

42 obstacle model, two-dimensional polygons represent obstacle boundaries which have attenuation-per-wall and attenuation-per-meter propagation loss attributes. Using the obstacle model, an obstacle-aware shadowing model leverages the Computational Geometry Algorithms Library (CGAL) [66] to count the number of walls as obstacle intersections, calculate the distance travels through obstacles as interior intersection lengths and implement deterministically the shadowing effects of wireless transmissions for obstacle line of sight (OLOS) pathways. For performance optimizations, obstacle locations are stored in a binary space partition (BSP) and inter-vehicle obstacle obstructions are cached. Fig. 6 gives the pseudo-code for the algorithm that determines for two vehicles (i.e., points) the number of obstacle walls penetrated and total distance traveled through obstacles. Substituting (15), (16), (20), and (21) into (9), the total path loss, L t, is expressed as (22) and (23): (22) ( ) ( ) ( ( )) ( ) ( ) (23) A VANET simulation script previously completed by the author and contributed to ns conducts the overall VANET simulation. For each simulation, the mobility trace files produced by SUMO simulations play back for 200 seconds of network simulation time, during which network characteristics are modeled in ns-3. 31

43 Algorithm GETOBSTRUCTEDDISTANCEBETWEEN (p 1, p 2, B) Input. Locations, p 1 and p 2, of two vehicles and a pre-determined binary search partition (BSP), B, of obstacles. Output. The total obstructed distance, m o, and the number of obstacle edge intersections, n. 1. m o 0; n 0 2. Initialize a maximum range, r, the distance from either point p 1 or p 2 to an obstacle center-point, that is used to filter the set of obstacles to the subset which are sufficiently nearby for calculation purposes (i.e., for optimization, exclude far-away obstacles). 3. Create a bounding box, b, for p 1 and p 2 and extend in all direction by r. 4. Get the set of potential obstacles, O. O the range search of b within B. 5. For every obstacle o O, do: 6. if the distance from p 1 or p2 to the obstacle center is within range, r, then 7. for each edge s o, 8. if s intersects a ray from p 1 to p 2 9. n n save the min. and max. distances from {p 1, p 2 } to the intersection pt. 11. m o m o + distance between min., max. values in step 10 (i.e., obstructed distance) 12. return m o and n Fig. 6 Pseudo-code for Algorithm GETOBSTRUCTEDDISTANCEBETWEEN that determines the number of obstacle wall intersections and obstructed distance between two points. RQ2: How does an obstacle model effect packet delivery ratio as compared to other VANET fading and shadowing models? Developed and implemented in the ns-3 network simulator, the models in (9) assess the impact of fading and shadowing in a VANET, and experiments quantitatively compare them. For example, identical experiments are repeated and results compared using different path loss models for 32

44 i) two-ray ground propagation loss only, ii) iii) two-ray ground propagation loss and Nakagami-m fast fading, and two-ray ground propagation loss and obstacle shadowing, as described in Table 3 where experimental parameters that differ from others are shown in bold: 33

45 Experiment (a) (b) (c) Path Loss Two-ray ground propagation loss only Two-ray ground propagation loss and Nakagami-m fast fading Two-ray ground propagation loss and obstacle shadowing Propagation loss Two-ray ground (deterministic) Two-ray ground (deterministic) Formula ( ) ( ) ( ) ( ) d Distance, in m, between Distance, in m, between two vehicles two vehicles G t G r h b 1.5m 1.5m 1.5m h m 1.5m 1.5m 1.5m Two-ray ground (deterministic) ( ) ( ) Distance, in m, between two vehicles Fading None Nakagami-m None (stochastic) Formula I f m Nak N/A { N/A Ω N/A P r for (two-ray N/A ground) propagation loss at distance, d Shadowing None None Obstacle (deterministic) Formula ( ) ( ) ( ) I s α N/A N/A 9 db per wall β N/A N/A 0.4 db per meter n N/A N/A Determined by number of intersections with building walls m o N/A N/A Determined as cumulative distances through building interiors. Table 3 Experimental parameters for path loss using (a) two-ray ground propagation loss only, (b) two-ray ground propagation loss and Nakagami-m fast fading, and (c) two-ray ground propagation loss and obstacle shadowing. 34

46 RQ3: What is the impact to VANET application safety performance of a deterministic obstacle fading model, as compared to other fading and shadowing models? Simulation trials capture comparative evaluation metrics, such as packet delivery ratio (PDR), after which each experiment computes an analytical safety performance assessment, extending the approach in [3]. Using a target safety performance threshold of 95%, we compute the safety communications range for 5, 7, and 9 out of 10 messages per second, leading to a quantitatively comparative assessment of the highway, residential, and urban scenarios. 35

47 Chapter 4 Experimental Setup and Results 4.1 Experimental Setup To evaluate our methodology, simulation experiments assess safety communications awareness range as vehicular and obstacle densities vary. The open highway, residential neighborhood, and urban downtown environments use the OSM topographic data [64] for the surrounding Raleigh, NC, USA area. These three scenarios provide different topologies and densities of roadway networks, and obstacles such as buildings and foliage. Figures 7, 8, and 9 show for these scenarios representative satellite imagery views from GoogleEarth and corresponding vehicle and building/obstacle-only simulation views in SUMO. The open highway scenario is selected from an interstate area near the Raleigh-Durham International Airport (RDU) that includes a few traffic-light-controlled off-interstate areas and buildings, although the densities of both are much less than the residential and urban downtown scenarios. (a) (b) Fig. 7 GoogleEarth view (a) and SUMO view (b) of the open highway scenario near Raleigh, NC USA. 36

48 (a) (b) Fig. 8 GoogleEarth view (a) and SUMO view (b) of the residential neighborhood scenario near Raleigh, NC USA. (a) (b) Fig. 9 GoogleEarth view (a) and SUMO view (b) of the urban downtown scenario near Raleigh, NC USA. Routes within each environment are randomly selected and vary from 50 to 250 in steps of 50. Each vehicle traverses a minimum 1.2 km route with source and destination edges chosen randomly such that traffic originates with greater probability on the fringe of the road network and follows longer, wider roads. Vehicle-route start times are staggered in 0.01s increments. Vehicles are assumed to enter the scenario from external locations using a reasonable lane and speed and thereafter obey traffic light signals (TLS) and car-following rules according to the Krauss model [36]. Vehicular simulation experiments executed in SUMO 0.19 [34] generate corresponding ns-2 movement files for later playback in the ns-3 37