Chapter XIV Simulation of VANET Applications

Transcription

1 258 Chapter XIV Valentin Cristea University Politehnica of Bucharest, Romania Victor Gradinescu University Politehnica of Bucharest, Romania Cristian Gorgorin University Politehnica of Bucharest, Romania Raluca Diaconescu University Politehnica of Bucharest, Romania Liviu Iftode Rutgers University, USA Abstract This chapter systematically presents actual issues regarding the simulation of VANET applications. Some of them refer to challenges in developing VANET simulators. The chapter discusses simulator architectures, models used for representing the communication among vehicles, vehicles mobility features, and simulation tool implementation methods. A critical analysis of the solutions adopted in some well-known actual simulators is also included. Other issues relate to the use of simulation in the evaluation of applications that aim at improving the traffic safety and control. Representative city and highway application scenarios are discussed, and results that can be obtained by simulation, along with ways these results can be exploited by VANET developers and users are highlighted. Future trends in the development of simulators that produce more accurate results and their use for the evaluation of more sophisticated traffic control solutions are also included. Copyright 2009, IGI Global, distributing in print or electronic forms without written permission of IGI Global is prohibited.

2 Introduction Vehicle-to-vehicle communication is a very actual and challenging topic. Vehicles equipped with devices capable of short-range wireless connectivity can form a particular mobile ad-hoc network, VANET Vehicular Ad-hoc NETwork. The existence of such networks opens the way for a wide range of applications. Two of the most important classes of such applications are those related to route planning and traffic safety. Route planning aims to provide drivers with real-time traffic information, which, in the absence of a VANET, would require an expensive infrastructure. By contrast, the VANET approach is scalable and has low maintenance costs. Moreover, short-range wireless communication technologies (such as ) have no associated cost, other than the communication devices. Safety applications involve disseminating urgent information, which is not present in the driver s field of view, or it is difficult to notice for reasons such as fog or other vehicles obstructing the line of sight. For instance, a lot of accidents occur in foggy conditions because drivers notice too late that some kind of incident has occurred in front of them. Safety at intersections could also be enhanced, since the risk of collisions could be detected in advance and the driver could be warned seconds before what would otherwise be an imminent accident. The evaluation of VANET protocols and applications could be made through real outdoor experiments, which are time-costly and claim for a large number of resources in order to obtain significant results. Instead, simulation is a much cheaper and easier to use method. Obviously, this leads network and application developers to use simulation in order to evaluate different simple or complicated and innovative solutions before implementing them. In turn, this stimulated the interest for the development of simulators that easily integrate the models and respond to the requirements of VANET applications. Simulators have become indispensable tools at least in the initial phases of the VANET application engineering process. This chapter aims to present the features and trends in VANET applications simulation, to explain the sources of its complexity and to describe actual solutions. The main challenges are presented in the next section, which introduces the three models involved in VANET simulations: the network and traffic models are determined by the currently used technology; the third model corresponds to the application itself. Then, general features of VANET simulators are presented. We describe an architecture that integrates the first two models and allows an easy interfacing with the third model. Details about the network / communication model are also given, and several components of the traffic / mobility model are then discussed. Current challenges in developing VANET simulators, and the description of solutions adopted so far are subjects of a separate section, in which issues regarding the performance of different simulation tools are presented. Another section is dedicated to application scenarios, the results that can be obtained by simulation, and the ways these results can be exploited by VANET developers and users to improve the car-to-car communication and the traffic control. The chapter ends with conclusions and future trends. Background Simulating a VANET involves two different aspects. First, there are issues related to the communication among vehicles. Network simulators, like The Network Simulator ns-2 (2008) and Jist/ SWANS (2008) cope with communication issues and focus on network protocol characteristics. The second very important aspect is related to the mobility of the VANET nodes. Traffic simulators take into account the traffic model, not necessarily in conjunction with VANETs. For example, TRANS Traffic Network Simulator 259

3 (Katz, 1963) was used to optimize the timing of traffic lights in the absence of inter-vehicular communication. In the majority of VANET applications, vehicles react to messages. For instance, if a driver receives a message saying that the road ahead is congested, that driver will probably change her/his route. In order to study such reactions, an integrated simulator is needed, one in which the relationship between the vehicle mobility module and the network communication module is bidirectional. Choffnes and Bustamante (2005) showed that the vehicular mobility (traffic) model is very important, and its integration with the wireless network model could produce more significant results. The authors present an integrated simulator that uses an original vehicular traffic model called Street Random Waypoint (STRAW) implemented on top of JiST/SWANS (2008). The authors have used the simulator to show that studying routing protocols for a vehicular network without an accurate vehicular traffic model is a wrong approach. In this respect, they compared their own results with those obtained with the Random Waypoint model (Broch, 1998), which is a very inaccurate representation of a vehicular network. The mobility model implemented in some simulators is not a sufficiently accurate representation of actual vehicle mobility. For example, in the model of Saha and Johnson (2004), each vehicle moves completely independent of other vehicles, with a constant speed randomly chosen. Multi-lane roads or traffic control systems are not taken into consideration. Other authors (Mangharam et al., 2005) make similar simplifying assumptions and do not consider multi-lane roads or car following models. The mobility model of Choffnes and Bustamante (2005) is more complex: the motion of a vehicle is influenced by the preceding vehicle, and traffic control systems are considered. However, multi-lane roads are not taken into consideration. A more accurate model has been used in the development of an integrated vehicular networks simulator, VNSim (Gorgorin, 2006), which includes a complex model for vehicles mobility, a wireless network simulator, and an interface for the emulation of vehicular applications. The simulator can be used to analyze networks of several thousands of nodes in complex city scenarios as well as in highway scenarios. The simulator allows the evaluation of a large range of vehicular computing applications, which cannot be studied by using other simulators, and can be used to improve both car-to-car communication protocols and traffic control applications (Gradinescu, 2007; Diaconescu, 2007). General Features of VANET Simulators VANET simulators can be classified as microscopic or macroscopic. Microscopic traffic simulators emphasize local behavior of individual vehicles by representing the velocity and position of each vehicle at a given moment. This type of simulation is especially helpful for studying localized traffic interactions, but it comes with the price of reduced scalability. Macroscopic simulators capture traffic conditions in a global manner and may use concepts from wave theory. Usually, the traffic is represented in terms of flows (vehicle/ hour), density (vehicles/km) or average speed. Similarly, microscopic network simulators take into account the transmission of each individual packet (or message), while the macroscopic network simulators consider transmission statistical properties. Most authors agree that a macroscopic mobility model is not sufficient to allow the study of a vehicular network, therefore microscopic simulation, although more complex, is required. The evaluation of vehicular applications for online traffic control (that produce real time feedback to the driver), such as safety applications, also requires a model of driver behavior. Models for lane changing or car-following are 260

4 some of the features that a microscopic VANET simulator has to implement to allow the analysis of such applications. Simulator Architecture A VANET simulator has two main components: a network component, capable of simulating the behavior of a wireless network, and a vehicular traffic component, able to provide an accurate mobility model for the nodes of a VANET. Depending on the simulation type, the simulator can contain other components as well. To describe them, we refer to the VNSim presented in (Gorgorin, 2006). Figure 1 shows the general view of VNSim. This is a microscopic discrete-event simulator. Consequently, it includes components for event management: an Event Queue, where the events are registered in the order of their occurrence times, the modules for events processing, and a logical clock for the management of the simulation time. The simulation time is quantified in fixed pieces so that each event can be associated to a specific timestamp with an acceptable resolution. At every moment of the simulation time, the Event Engine pulls the current events from the queue, and handles them in a random order by calling the event processing modules. The vehicular traffic component is represented by the Mobility Module, while the wireless network component is incorporated into the Network Module and Scheduler. The Event Queue can hold three types of events: send, receive or GPS. A send event for a specified node triggers the call to the node s procedure responsible for preparing a message. It also schedules the corresponding receive event(s) for the message receiver(s) that are determined by the simulator according to the wireless network model. The receive event is associated with a node (group of nodes) to which the message is transmitted (broadcasted). Its action is to call the appropriate handler in each of the receiving Figure 1. General VNSim architecture 261

5 nodes. The GPS event is scheduled at a regular time interval, for each node, in order to simulate the way an actual VANET application periodically collects GPS data. The mobility module repeatedly updates the position of each node representing a vehicle, according to the vehicular mobility model. This model takes into account vehicle interactions (passing by, car following patterns, etc.), traffic rules, and the behavior of different drivers. The main advantage of this architecture is the possibility to execute (or emulate) the actual code of a vehicular application without significant changes, by using appropriate simulator interfaces. In other cases, the vehicular mobility and wireless network models can be incorporated in different simulation tools. For example, MOVE MObility model generator for VEhicular networks (2008) uses SUMO Simulation of Urban MObility (2008) for simulating a traffic scenario and generating an output file with vehicular mobility traces. This file is offered as an input to a network simulator, such as ns-2 or Qualnet (2008), to obtain the final simulation results. A similar solution is adopted in (Eichler, 2005), where the authors propose the coupling of the CARISMA Traffic simulator with ns-2 and with the application module, in order to study the vehicle-to-vehicle warning-message distribution. Synchronization data is exchanged periodically between the two simulators. The challenge is to ensure the coupling of simulators, i.e. the consistency of vehicle positions in both simulators, at synchronization points. The simple architecture mentioned above makes use of existing traffic and network simulators. The drawback is that it does not support the simulation of complex VANET applications in which the driver behavior would be influenced by the application itself, as already discussed before. An innovative architecture aiming at solving this problem is currently under evaluation by the authors of TraNS (Piorkowski, 2008). They are trying to develop a general framework for controlling the mobility, without being limited to a specific traffic simulator or to a specific network simulator. This architecture promises to offer new possibilities for simulating VANET applications, but the framework has not yet been released. A slightly different architectural approach is the hybrid simulation (Killat, 2007). The simulator architecture incorporates a previously implemented vehicular traffic simulator VISSIM and two new components: VCOM, used to simulate the inter-vehicle communication on the basis of a statistical model proposed by the authors, and an application module. The simulator has to respond to specific user requirements regarding the functionality (it was aimed to investigate the impact of car-to-x communications on transport efficiency) and the performance (scalability, short response time, and easy of use). The relevant characteristics of a VANET simulator are the simulation efficiency and accuracy for large scale VANETs, which we will address in a separate section. Communication Models Communication models highlight the information flows between vehicles and are of special interest when vehicles move with high relative speeds. VANET applications are affected by wireless networking aspects such as transmission delay, packet loss or network access scheme. However, accurate network simulation introduces additional complexity and makes several large scale VANET applications unsuitable for simulation. In this section, we discuss the network characteristics that are mostly relevant for VANET simulation, as well as the accuracy that can be reached with their implementation. At the physical layer, the reliability of the wireless network, in terms of packet loss and transmission delays, is relevant to simulation. Several models are used for the radio waves propagation, which consider estimates of the power of a received signal. This signal can be received by an antenna only if the Signal-To-Noise (SNR) ratio is over a specific SNR Threshold (SNRT). 262

6 Wave propagation models have been studied extensively for decades in telecommunication and recently have been correlated with vehicular short range communication. The radio signal can be degraded by fading. This is caused by signal propagation over multiple paths (with different values for the attenuation, delay and phase shift), which produces interferences at the receiving antenna. There are several examples of such propagation models that were implemented in wireless simulators. Free Space Model is the simplest one, which considers ideal conditions: a single radio signal that attenuates with distance. Two Ray Ground propagation model considers two rays between the source and the destination: the direct line and the ground reflection. It gives more accurate estimates than the free space model, especially for long distances. VNSim (Gorgorin, 2006) implements this model and includes the signal attenuation with distance in order to test different ways to optimize the network traffic. Rayleigh fading (Sklar, 1997) models multiple indirect paths between the sender and the receiver, while the Ricean fading model considers only one dominant line of sight and multiple indirect signals. Reports from real experiments have shown that vehicular short-range wireless networks can be modeled using the Nakagami fading model (Sklar, 1997), which considers fast fading to characterize the wireless channel. Concerning the protocols, DSRC Dedicated Short-Range Communications Protocol (Taliwal, 2004) is a multi-channel wireless protocol, which is based on the IEEE a Physical Layer and the IEEE MAC Layer. It operates over a 75 MHz licensed spectrum in the 5.9 GHz band allocated by the Federal Communications Commission, and supports low latency vehicle-to-vehicle and vehicle-to-infrastructure communications. The motivation behind the development of DSRC is mainly the need for a more tightly controlled spectrum for maximized reliability. The use of hand-held and hands-free devices on the 2.4 and 5 GHz bands along with the increase of WiFi could cause intolerable and uncontrollable levels of interference, which would significantly decrease the reliability and effectiveness of vehicular safety applications. But, even with a licensed band, some issues arise, such as fair access to all applications, including priority scheduling of traffic between different application classes (safety over nonsafety). Unlike , multi-channel coordination is a fundamental capability of DSRC. Consistent effort has been put into matching simulation results to real outdoor experiments. Yin et al. (2004) observe significant degradation of the network performance under the high node mobility. The study evaluated the DSRC Physical Layer and the MAC Layer, in the context of a traffic safety application, on top of a customized simulator combining the CORSIM (2008) traffic simulator and the Qualnet (2001) network simulator. It showed a long term average throughput of 50-60%. Using the Nakagami m-distribution fading model implemented in ns-2, Killat et al. (2007) estimated the probability of successful reception at a given distance depending on vehicle density and data load. Their idea is to combine discreteevent simulation and mathematical modeling, starting from the observation that the data transmission that is actually interesting and relevant for a VANET application represents only a small percentage of the total data exchanged during a traffic scenario. There is a lot of background traffic, which can be much more efficiently simulated with statistical methods, while keeping the more accurate discrete-event simulation techniques only for interesting data. Their work is based on the concept of hybrid simulation and offers much better performance compared to purely discrete-event approaches. Although actual outdoor experiments have been made, the behavior of the wireless network (the DSRC protocol) under stressed city-like scenarios with high node density is still an open issue. This is why simulation frameworks now focus on large scale VANET simulations. 263

7 A lot of work has been done to study the routing layer protocols in mobile ad-hoc networks, MANETs. They can be grouped in two categories. The first category includes protocols that use the IP addresses to locate nodes, such as: DSDV Destination-Sequence Distance Vector Routing (Perkin, 1994), DSR Dynamic Source Routing (Johnson, 2001), AODV Ad hoc on demand distance vector (rfc3561, 2008), and MPR Mobile Relay Protocol (Nain, 2004). Protocols in the second category are based on location information. Some examples are: DREAM (Basagni, 1998), GPSR Greedy Perimeter Stateless Routing (Karp, 2000), and ABRP Anchor Based Routing Protocol (Huaizhi, 2005). Classical MANET protocols have been proven to perform poorly (Blum, 2004) due to specific characteristics of vehicular ad-hoc networks, namely: rapid topology changes, small network diameter, and limited redundancy. This is why other position-based forwarding solutions have been proposed (Festag, 2004) for vehicular environments, which take into account some elements of the vehicle mobility model such as the distance between communicating partners and the available paths towards destination. This led to a combined position-based and map-based ad hoc routing. Mobility Models Mobility models emphasize local behavior of individual vehicles by representing the velocity and position of each vehicle at a given moment. This type of simulation is especially helpful for studying localized traffic interactions. Vehicular traffic mobility models have been studied for a long time. However, until recently, the purpose of developing and studying accurate mobility models has not been to use them for VANET simulation. In this section, we review vehicular mobility models, from the perspective of a VANET simulator architect. 264 Generally, when choosing a mobility model for a VANET simulator, the complexity and the accuracy are the most important issues need to be considered. The benefit of using a very complex and accurate mobility model, together with the performance penalty, should be carefully evaluated, keeping in mind the purpose of the VANET simulator and the type of applications for which it is designed. In the Random Waypoint mobility model (Broch 1998), each mobile node has a position and a velocity, and moves towards a random destination in a predictable way that can be calculated as a function of time. This model was used originally in ns-2 (2008). It was designed for general-purpose mobile networks and does not accurately represent vehicular motion. Jardosh et al. (2003) define a more complicated model that corresponds to urban settings, in which obstacles are introduced to constrain vehicles mobility and the wireless transmission. Two other models have been defined by Potnis and Mahajan (2006). The Stop Sign Model (SSM) introduces stop signs at every intersection. Every vehicle arriving at the intersection would wait for a fixed period of time before proceeding to its destination. The Traffic Sign Model (TSM) introduces traffic lights at every intersection. A vehicle arriving at an intersection waits for a random period. Both models introduce queues at intersections and interdependencies among vehicles on the same street. A thorough analysis of these models is described in (Mahajan et al., 2007). The authors studied the dependency of the packet delivery ratio and delay on the characteristics of the urban settings such as street layout, traffic rules, etc. In some studies (Baumann et al., 2008), specific models such as the Manhattan model and GIS models, are used to establish the street layouts. In the model of Saha and Johnson (2004), each vehicle starts at some randomly selected point on a road and moves until a destination point on another randomly selected road. Several new elements are added to the model and to its use.

8 First, the path between the starting and destination points is calculated by using the Dijkstra s shortest path algorithm. Second, the vehicle speed is also randomly calculated using a uniform distribution within a specified interval. Third, the authors use the TIGER map database to create a more realistic environment of the simulation study. The mobility model of Choffnes and Bustamante (2005) is more complex: the motion of a vehicle is influenced by the preceding vehicle (car following model), and traffic control systems are considered to introduce some actual traffic conditions such as traffic congestion. In addition, the model includes communication details, such as the percentage of cars actively communicating, and road maps details, such as characteristics of street segments and number of lanes in each direction. The study of safety and traffic control applications requires an accurate model of driver behavior. One example is the model developed by Wiedemann (1974, 1991) and further studied by Fellendorf and Vortisch (2001). This psychophysical model is used in the commercial traffic simulator VISSIM (2008). The basic assumption is that a driver can be in one of four modes, which are classified according to the driver s convenience: free driving (superior convenience), approaching, following, and braking (lowest convenience). In the free driving mode there is no influence from preceding vehicles on the same lane. In this situation, the driver will seek to obtain and maintain a desired speed. The desired speed and the acceleration depend on driver s personality, and on road s characteristics. In the approaching mode, there is a slower preceding vehicle that influences the driver. In this situation, the driver decelerates in order to obtain the same speed as the preceding vehicle. The deceleration is a function of the distance between the two vehicles, their speeds, as well as other parameters. In the following mode, there is a current and a preceding vehicle, but the speeds of the two vehicles are practically equal. In this situation, the driver will seek to keep the speed constant. In the braking mode, there is a slower preceding vehicle, very close in front. In this mode, due to the immediate danger, the driver will apply high deceleration rates. This model was implemented in VNSim, together with a lane-changing model for multilane roads, based on lane-usage rules valid for several European countries. According to these rules, the use of the first lane is required, unless it is occupied. It means that a driver will always try to stay on the lower lanes, except when overtaking another slower vehicle. Overtaking on the right side is not allowed. Whenever a driver is in a different mode than free driving, she/he will always check if the higher lane can provide a superior mode. If this is the case, the driver will switch to a higher lane. Similarly, whenever a driver is in a different mode than braking, she/he will always check if the lower lane provides at least equal conditions. In this case, the driver will switch to a lower lane. Different driver profiles (aggressive, regular, calm) can easily be modeled by using several parameters: distance between vehicles, maximum speed, desired speed, driver reaction time, critical distance for driver reaction, etc. Each driver class is represented by a specific set of values for these parameters. In order to further differentiate the drivers, the model allows small deviations from the specified values, which are randomly computed for each individual driver. Fellendorf and Vortisch (2001) showed that the model is accurate, by comparing simulated traces with actual measurement data taken from a German freeway and from a US freeway. To further calibrate the model, the authors of VNSim performed an extra validation, by measuring real-time traffic in intersections and comparing the results to those obtained by simulation (Gradinescu, 2007). Estimating fuel consumption and pollutant emissions is an increasingly important issue when studying vehicular traffic. The model of Akcelik and Besley (2003) considers the estimation of the relation between fuel consumption and emissions, 265

9 on one side, and the speed and acceleration of the vehicle, on the other. A simplified version of this was used in VNSim to take into account only light vehicles. Statistics and global measures have been easily obtained. Maps A digital map is required in almost any VANET application. Each vehicle that is part of the system should have access to such a digital map. A very popular option for digital maps used by VANET simulators is TIGER Topologically Integrated Geographic Encoding and Referencing (2008), which is freely-available, actual digital maps for the USA. Most of the VANET simulators use TIGER files as the source for their digital maps. TIGER files contain detailed geographical information about the roads in a region, from large highways to small streets. For each road, the TIGER files specify its end points, along with as many intermediate points as needed, depending on the road s shape. Furthermore, for each road, a class information is given (whether it is a small street, a local road, a State Route, or an Interstate Highway). Unfortunately, the TIGER database lacks more detailed traffic-specific information, like the number of lanes, or traffic control systems, like traffic lights, yield or stop signs. In more realistic simulations, such information is required. The authors of VNSim have chosen to add the extra information, based on simple heuristics and on the road class information included in the TIGER files. For example, more lanes are considered for higher class roads, yield or stop signs are introduced for lower class roads, traffic lights are used at intersections between equally important roads, etc. 266 Implementation issues Approaches to Simulator Development The development of VANET simulators follows two different approaches. One is to implement the mobility model from scratch. In this situation, the integration between the mobility model and the network communication model is ensured in the design phase. The integration is implicitly achieved since the developers start with this integration in mind. On the other hand, the complexity of realistic models for communication and mobility makes this approach a very challenging one. There is a huge amount of work required to implement complex models. A typical example of this approach is GrooveSim (Mangharam, 2005), which is a simulator for geographic routing in VANETs. It has several modes of operation that allow the evaluation of network protocols, extraction of mobile wireless propagation models in real time, the generation of test scenarios for simulation, hybrid simulation with real and virtual vehicles, and others. The most important limitation of GrooveSim is the simplistic mobility model, which does not consider car-following or lane-changing models. The authors acknowledge this limitation and mention that the focus of their work is to provide basic models and evaluate their impact on key performance metrics. STRAW STreet RAndom Waypoint (Choffnes, 2005) builds a mobility model that constrains vehicles movement according to the map of a city, different traffic profiles (for example vehicular congestion) and specific control mechanisms. It is implemented as an extension to a wireless network simulator developed by Barr (2004) at Cornell University. The mobility model of STRAW is also simplistic, and does not consider multi-lane roads. Saha and Johnson (2004) based their simulation on building a realistic mobility model compatible with the ns-2 network simulator. Their

10 development makes use of the TIGER database. The mobility model which they implement is very simple, without considering car-following or lane-changing models. However, their simulations are among the first ones which prove that using a different mobility model than the wellknown random waypoint model will generate significantly different results, thus proving that the mobility model is an important component of a VANET simulator The other approach is to use an existing mobility model implementation and integrate it with a network simulator. This approach takes advantage from existing vehicular traffic simulators, which already implement complex, validated mobility models. The challenge of this approach is to achieve an online integration, in which the movement of vehicles is directly influenced by the VANET applications running on some of the vehicles. The easier way for integration is to generate traffic patterns offline, and then use these results as an input to a network simulator. VanetMobiSim (Harri et al., 2006) is a generator of realistic vehicular movement traces for network simulators. It considers macro-mobility features, such as road topology, number of lanes, speed limits, etc., and micro-mobility characteristics, like crossroads regulated by traffic signs or lane change capability for vehicles. Similarly, the Generic Mobility Simulation Framework GMSF (Baumann et al., 2008), which models and simulates node mobility, is designed to work together with a network simulator. GMSF has some important distinctive features: it implements new GIS-based models with detailed road maps, and includes microscopic features such as traffic lights management. An example of using mobility traces is given in (Yang et al., 2007). The authors propose an evaluation framework for VANET applications. It utilizes a hybrid emulation testbed TWINE, which combines real implementation of high level network layers (closer to the application) with the simulation of low level layers (physical, MAC). The framework offers several advantages to its users: it allows the execution of real applications in a simulated network environment; it integrates realistic mobility traces and deployment data; it uses evaluation metrics related to the application. A similar approach is described in (Rehunathan et al., 2007), where a microscopic traffic simulator, MITSIMLab is used to generate vehicle movement traces to be used in ns-2. This approach, however, does not allow for an accurate evaluation of a VANET application or protocol, because the mobility pattern cannot be modified as a reaction to the VANET application. Using an existing traffic simulator, which is based on already validated complex traffic models, has the important advantage of avoiding re-inventing the wheel. Among the most recent attempts to implement a VANET simulator by using this approach are MOVE (Karnadi et al., 2007) and TraNS Traffic and Network Simulator (Piorkowski et al., 2008). Both use the open-source microscopic traffic simulator SUMO (2008), which implements complex validated vehicular traffic mobility models. And both use two steps for the VANET simulation. In the first step, traffic traces are generated offline using the traffic simulator. In the second step, a network simulator processes the traffic traces and evaluates the network communication between the nodes. The authors of both simulators agree that the approach does not allow the simulation of complex VANET applications, in which the driver behavior would be changed by the application itself. However, both developer teams mention that this is a goal for future work that will achieve a real integration between the traffic mobility model and the network communication model. Furthermore, the authors of TraNS claim that their simulator is the first open source project that attempts to realize this highly pursued coupling for application-centric VANET evaluation (Piorkowski et al., 2008). Sommer et al. (2008) describe the simulation framework Veins (Vehicles in Network Simula- 267

11 tion) that is based on the bidirectional coupling of OMNeT++, a network simulator capable of modeling realistic communication patterns of VANET nodes, and of SUMO, the microscopic traffic simulator. The bidirectional coupling is able to support the analysis of more realistic traffic models that reflect the impact of inter vehicle communication on the road traffic. Simulator Performance Accurate simulation of vehicles mobility and wireless transmission is time-costly. Also, for an application designer, the study of many application scenarios requires experiments with large maps and with large numbers of vehicles. This is why the scalability of VANET simulators is a main performance metric for which we present some guiding figures. Experiments performed with VNSim on a 1.6GHz uni-processor showed that it can support around network events per second. This value is lower than the throughput of the widelyused network simulator ns-2, which can simulate over events per second, on a 2GHz uniprocessor. This is explained by the of the network simulator with a complex node mobility model, Figure 2 VNSim performance evolution with the increase of the number of nodes number of seconds necessary to simulate A 300 seconds scenario number of nodes which is responsible for accurately computing the motion of all nodes, in each time quantum. The graph in Figure 2 shows the evolution of the simulator s performance with the increase of the number of nodes. The results are based on simulations performed in a highway traffic scenario, with all the nodes running a neighbordiscovery and update protocol, with a 1 second period for the beacon messages. VNSim is able to simulate motion of 1000 vehicles, in a complex city-scenario in real-time (1 second of simulation in 1 second of real time). Experiments have been done by the authors with a squared region of 1km by 1km in Manhattan downtown, which has a high number of intersections and traffic lights (see Figure 3). VNSim was also used to perform the simulation of more complex scenarios, involving up to vehicles. NCTuns (Wang, 2008), a highly-integrated simulator, cannot be used for large-scale simulations. Thus, for the simulation of a scenario with 100 vehicles, 22 minutes are required. For the same scenario, if the number of vehicles doubles then the amount of time required for the simulation grows approximately 4 times, to 84 minutes. These results, obtained on a 1.8 GHz uniprocessor, show that the simulation of large-scale vehicular networks (with tens of thousands of vehicles) can be achieved. Figure 3 Squared region in Manhattan downtown

12 Some performance results are presented in (Choffnes, 2005) for the STRAW simulator. Here, STRAW s overhead has been evaluated for two models: a simple one in which vehicle movement is determined randomly at each intersection, and a more complicated one in which the movements are determined according to pre-computed shortest paths. The evaluation was done by disabling the wireless communication simulator. The experiments were performed on a Pentium 4 at 2.4GHz computer with HyperThreading enabled. When the number of nodes considered in the experiment increased from 100 to 1600, the runtime increased four times for the simple model and 6.5 times for the more complicated model. The highest overhead of 10% of the simulation time has been registered for the complex model with 1600 vehicles. One solution to improve simulators performance is parallelization. Fujimoto et al (2003) report experiments performed on packet level network simulators. Parallelization is achieved by federating several copies of the same simulator or of different simulators. This approach was used to produce two parallel simulators, PDNS Parallel / Distributed Network Simulator (PDNS) based on ns-2 and GTNetS developed at Georgia Tech. The results showed a very good performance measured in number of transmissions per second, PTS Packet Transmissions that can be simulated per Second. More specific, for the best case of parallel simulations, the number of PTS was 50 times higher than the number of PTS obtained in the best sequential case. MoVES, a Mobile Wireless Vehicular Environment Simulator (Bononi et al., 2006) is also parallel and distributed. Its idea is to distribute subsets of the model entities to different processors that communicate via message passing. The authors find that the parallel approach gives good results when simulating massive wireless vehicular networks. Higher performance can also be achieved by the use of hybrid simulators (Killat, 2007). The authors present the results obtained for to vehicles in a scenario dedicated to the analysis of the impact of the penetration rate of radio-equipped vehicles on the average speed. The speedup factor was 500 as compared to the NS-2 simulation. The result is due to the use of the hybrid simulation concept, which combines the discrete-event simulation with analytical traffic modeling. As a general conclusion, developing a highlyintegrated simulator, with complex models for both the vehicular mobility and the wireless communication has the advantage of higher accuracy but has a severe impact on performance. Integrating existing vehicular traffic simulators and network simulators can lead to better performance but it is not able to support more complex scenarios involved in safety and traffic control applications. Improvements can be achieved by parallelization, which distributes the simulation task to different computing resources, or by selecting a model that reduces the computation load. Simulation of representative city and highway application scenarios This section presents several simulation use cases together with results obtained in realistic simulation experiments performed with the VN- Sim simulator. The way these results could be exploited by VANET developers and users are also highlighted. An important class of applications that can be deployed over VANETs is concerned with traffic operations and maintenance: dynamic route planning, weather conditions publishing and adaptive signal control in intersections. Traffic Coordination in Intersections Traffic coordination in intersections is a very challenging and not enough studied topic. As 269

13 size, the probability of the successful transmissions decreases due to wireless interferences. A solution is the use of aggregation mechanisms. A possible method is the use of a probabilistic scheme in which only some leader vehicle, chosen randomly with a probability depending on the traffic density, transmit the information. The simulation experiments highlighted the great decrease in the network load while preserving the functionality of the solution. The data dissemination algorithm supports the implementation of the traffic light control that establishes a signaling plan well suited for the intersection. Considering the traffic light is a node participating in the dissemination process, it can gather traffic metric values (flows, delays) on all the concurrent roads. Optimum cycle length and green split phases are computed using traffic engineering methods, such as Webster s formula (Gradinescu et. al 2007). One of the simulation scenarios that were used to analyze the efficiency of the algorithm was a 4-way intersection in Bucharest during rush hour. As shown in Figure 5, the intersection recovered sooner from congestion when using the adaptive strategy described above, compared to the pre-timed system deployed in the intersection. The total delay measured for all vehicles was decreased by 28%. An additional component of the VNSim simulator is the module for estimating the pollutant emissions and fuel consumption, which considers the particular behavior of each vehicle speed and acceleration. This feature was used in Figure 4. Traffic lights and vehicles communicate to improve traffic efficiency Figure 5. Average control delay in intersection Average control delay (s) an example of a VANET application, we present an adaptive traffic light system based on wireless communication between vehicles and fixed controller nodes deployed in intersections (Gradinescu et al., 2007). We describe the steps required to deploy such an application in VNSim. Furthermore, we present the results showing significant improve of traffic fluency in intersections, and the advantages over other architectures in terms of cost and performance. The goal of the VNSim project was to reduce congestion and the number of accidents in intersections by enabling wireless communication between traffic lights and vehicles. In this way, the traffic lights can get a real-time image of the traffic, and make an informed decision on how to set signal plans. As a secondary benefit, the traffic lights can also detect dangerous situations and alert drivers about the intersection status to increase their awareness. The following aspects need to be taken into consideration when such an application is simulated. First, there must be a data dissemination system that supports the information flow from vehicle to vehicle and to the traffic light. Based on the TrafficView system (Dashtinezhad et. al 2004), data dissemination was designed so that vehicles forward information about other vehicles they know about (position, speed, direction) to all the other vehicles in their range. The simulation of this propagation scheme in VNSim demonstrated that a lot of redundant data is transmitted and a high data charge is put on the network. Moreover, for large packet simulation time (min) Pre-timed 270 Adaptive - Ideal Adaptive - Real

14 this application to demonstrate the benefits of the adaptive solution due to increased traffic fluency: 6.5 % improvements in fuel consumption and CO 2 emissions were measured. Collaborative Route Computation Finding the best route to destination in congested roads traffic is a major challenge in the attempt to improve traffic conditions. Collaborative route planning is the second VANET application we use to show how ad-hoc communications can improve the driving experience. The simulation of a collaborative, dynamic routing application requires several features from the simulation tool. First, in order to study a congested roads network, a large number of vehicles, moving on a large map, should be supported. Therefore, the simulator s performance and scalability are crucial. Furthermore, this kind of application supposes that the mobility of the vehicles is influenced by the application itself: if drivers find out, by using the VANET route planning application, that a road is less congested, then they might change their initial path in order to take advantage of this information. A simulator which does not integrate the vehicular mobility module with the application module will not be able to represent dynamic route changes. Diaconescu et al. (2007) used VNSim in order to study the dynamic route computation problem. The authors discuss two different approaches. The first is to exclusively use the ad-hoc network for all the information exchanges. The second approach is based on the existence of fixed nodes, placed in intersections, which also participate to, and facilitate the communication. In the first approach, each driver first computes a route towards its destination. This initial computation is based on distance, on the roads types, and on the assumption that drivers tend to prefer the main roads. After this initial computation, the driver starts the journey. Along the way, periodically, queries about the traffic conditions ahead are sent via the ad-hoc network. Depending on the replies (for example, the reception of information about traffic jams ahead), re-calculations of the route occur. The simulation experiments showed that global improvements in travel times occur when only a percentage of the vehicles use the algorithm, not all of them. When all vehicles change their route, the traffic jam moves to another road. Starting from this observation, the authors suggested the use of fixed-nodes placed in intersections, which are able to communicate with vehicles and, at the same time, are inter-connected through a separate highspeed network. Of course, this solution requires a more extensive infrastructure. In large cities, a large part of this infrastructure might already be present. The fixed-nodes gather information about the actual traffic (number of vehicles traveling) on each road and execute a collaborative route planning with the other fixed nodes. The results are sent to the vehicles in proximity informing them about the best routes. Highway Lane Reservation An innovative VANET application was proposed in (Ravi et al., 2007). The authors argue that time guarantees of other means of transportation (like train or plane) can be incorporated into the highway system. The idea is to allow drivers to reserve an entry slot onto dedicated lanes of a highway by paying a premium price. Reservations will be allowed for these lanes up to their carrying capacity, so that the dedicated lanes will never become jammed and the system can guarantee the trip duration between any two highway points, in the absence of accidents. The general architecture of the system is shown in Figure 6. The proposed system is complex and is made up of several subsystems. First, a reservation system is required to allow users make advance reservations of time-slots on the dedicated lanes. Then, an enforcement system is needed to ensure that only drivers with valid reservations travel on 271

15 Figure 6. Lane reservation general architecture the dedicated lanes. This could be achieved by using RFID tags located on vehicles, and RFID readers placed on the highway. The lane entrance must be assisted by a VANET. More specific, since the dedicated lane will often be used at near-to-maximum capacity, it is important that vehicles that enter or leave the lane do not create disruptions in the traffic flow on this lane. In order to allow smooth entering and leaving, a VANET protocol can be designed. For example, the car that wants to enter the lane will send a request to a vehicle already on the lane, asking for permission to enter. In order to evaluate such a large-scale system, a proper simulation tool is needed. It should be scalable enough to be able to simulate a large-scale highway traffic scenario. Furthermore, emulating the application code run by each node would also be required. The authors plan to investigate the possibility of using VNSim in order to study and validate a lane reservation system. Conclusion This chapter discussed the simulation of VANET applications. The focus was on the new trends 272 in communication protocols and traffic models, and on new facilities incorporated in simulation tools. The advances in the VANET technology and protocols support the adoption and use of more complex mobility models and of more flexible and adaptable traffic controls. VANETs have specific characteristics such as the rapid topology changes or the changes in the vehicles mobility as reaction to traffic changes. Obviously, these changes must be captured by the simulation models, which become more or less complicated and include more elements that constrain the vehicle mobility: maps, real traffic conditions (congestion), driver behavior, fuel consumption, pollutant emissions, etc. Recent research referenced in this chapter targeted the validation of these models and the exploitation of their facilities in different applications. Realistic models, such as those used in lane reservation, were proposed and will eventually constitute the subject of further research. In order to more accurately evaluate the traffic control applications, more powerful microscopic simulator tools have been developed. The tools integrate the network communication model and the traffic / mobility model, and facilitate the interfacing with the core application modules. One drawback is the high time consumption required by simulation experiments besides other aspects like usability and flexibility. Since microscopic simulators demonstrated their utility in the VANET applications design process, the increase of tool performance, scalability, and flexibility will be subjects to future research work. The technological progress will stimulate new applications aiming to improve the traffic safety and control strategies. More vehicles will be equipped with GPS receivers, communication devices compatible with different protocols (for example DSRC), radars for identifying neighbors positions, equipment for storing digital maps, etc. These and other similar components will facilitate the development of more complex safety applications (like imminent collision warning, obstacle detection, collision avoidance, etc.) and