Safety Communication for Vehicular Networks: Context-Aware Congestion Control and Cooperative Multi-Hop Forwarding. Le Zhang

Transcription

1 Safety Communication for Vehicular Networks: Context-Aware Congestion Control and Cooperative Multi-Hop Forwarding by Le Zhang A thesis submitted in conformity with the requirements for the degree of Doctor of Philosophy Graduate Department of Electrical and Computer Engineering University of Toronto c Copyright 2015 by Le Zhang

2 Abstract Safety Communication for Vehicular Networks: Context-Aware Congestion Control and Cooperative Multi-Hop Forwarding Le Zhang Doctor of Philosophy Graduate Department of Electrical and Computer Engineering University of Toronto 2015 Vehicular safety applications have the potential to make travel on our roads and highways much more safe. These applications require both reliable and up-to-date knowledge of the local neighbourhood, as well as reliable multi-hop propagation of safety alert messages. Under the IEEE Wireless Access in Vehicular Environments (WAVE) standard, the envisioned platform for vehicular communication, the former is attained through an exchange of single-hop broadcast safety beacons in the control channel. However, congestion of these periodic broadcast safety packets remains an obstacle to the large-scale deployment of vehicular ad hoc networks (VANETs). Excessive offered load on the shared control channel results in a deterioration of network performance and a subsequent reduction in the safety level at the application layer. Existing methods focus on providing fairness in the resources allocated, but fail to account for the different network performance requirements of vehicles in different driving situations. First, we address the problem of beaconing rate adaptation in response to network congestion and the driving context. We define a delay constraint profile for each communication link based on the driving context of the vehicles. We formulate the problem as a load minimization problem subject to a probabilistic constraint on the probability of violation of the delay requirements for each link in the network. We also develop an alternate formulation of the problem as a weighted network utility maximization (NUM) problem. We propose distributed algorithms to solve this problem in a decentralized ii

3 manner and validate their performance through simulations. Some analytical results on their convergence are also provided. Next, we turn to the problem of providing reliable multi-hop forwarding in vehicular networks. Cooperative vehicular multi-hop schemes achieve reliability using broadcast transmissions and multiple forwarding relays at each hop. However, packet duplication must be controlled to circumvent the broadcast storm problem. Existing multi-hop dissemination schemes do not account for the presence of periodic safety beacons on the shared safety channel. We propose a cooperative forwarding protocol for highway vehicular networks, which extends a reliable vehicular broadcast medium access control (MAC) protocol based on positive orthogonal codes (POC). Multiple cooperating relays act as a virtual relay and schedule their transmissions to correspond to a single POC codeword. The proposed method exploits spatial diversity while mitigating the effect of hidden terminals. By allocating separate POC-based schedules for multi-hop packets and the periodic broadcast of safety heartbeat packets, the proposed protocol reduces the interference between the two types of safety transmissions. The performance of the protocol is studied through analysis using a Markov model and validated via simulations. iii

4 Dedication To Mom and Dad. iv

5 Acknowledgements First and foremost, I would like to express my utmost gratitude to my supervisor, Professor Shahrokh Valaee. Your mentorship, guidance, and constructive criticism were instrumental to my research. Without his invaluable contributions, this thesis would not have been possible. I would also like to thank my WIRLAB colleagues for their support and friendship. In particular, I would like to thank Behnam Hassanabadi for his contributions in our collaborative research and for being a sounding board for my ideas. I wish to also thank my parents for all their support over the course of my graduate studies, moral and otherwise. I couldn t have done it without your help. Finally, I d like to thank my wife Melissa for her love and understanding throughout my adventures in academia. v

6 Contents 1 Introduction Motivation Vehicular ad hoc networks Standardization Vehicular applications Use Case Scenarios Contributions Thesis organization Related Work Repetition-based medium access control Congestion control and network utility maximization Congestion control in VANETs Proactive protocols Reactive protocols Rate Adaptation: Safety-based Delay Constraints System Model Safety-based Maximum Delay Constraint Fairness in Constraint Violation Probability Simulation and Performance Centralized Optimization Problem Distributed Subgradient Algorithm Message Passing and Update Steps Local Coordinate-wise Subgradient Algorithm Primal Approximation Via Averaging Convergence Performance Evaluation vi

7 3.7.1 Simulation Results Summary Rate Adaptation: Safety-Weighted Network Utility Maximization Safety-Weighted Utility Maximization Utility Function Centralized Optimization Distributed Algorithm Convergence Analysis Performance Evaluation Simulation Setup Simulation Results: 3-Node Scenario Simulation Results: 100-Node Scenario Summary Cooperative Forwarding Using Repetition-based Vehicular MACs Related Work Proposed Cooperative Forwarding Protocol Distributed Relay Selection Location-based Code Allocation Code Allocation for Virtual Relays Time Slot Assignment Analysis Packet Collisions Due to Interference, Q Packet Loss Due to Channel Erasure, P Performance Evaluation Simulation Setup Protocols Used For Comparison Simulation Results Summary Conclusion Future Work Rate of Convergence Delay Constraint Profile and Weighting Functions Adaptive POC-MAC Optimal Relay Selection vii

8 A Primal Approximation Convergence Proofs 93 A.1 Proof of Lemma A.2 Proof of Lemma A.3 Proof of Proposition A.4 Proof of Proposition Bibliography 97 viii

9 List of Tables 1.1 Some categorized vehicular applications Variable dependency and latency Parameters of the simulation Parameters of the simulation POC assignment example Partitioning of SFR transmission patterns according to their interference pattern Partitioning of POC transmission patterns according to their interference pattern Parameters of the simulation ix

10 List of Figures 1.1 FCC channel allocation Network diagram for notation used in the general network model Linear weight functions Network topology Transmission probability at t = 250s Delay constraint violation probability at t = 250s Network-wide maximum delay constraint violation probability Total network load of GTP vs. LTP for a single cluster Iteration t is performed at the start of transmission frame t Epsilon-prime vs. subgradient size The network topology at various snapshots in time Transmission probability assignments Transmission probability at each node: DSM-LC vs. PULSAR Channel Busy Ratio observed at each node: DSM-LC vs. PULSAR Packets received: DSM-LC vs. PULSAR Broadcast delivery ratio: DSM-LC vs. PULSAR Weighted broadcast delivery ratio: DSM-LC vs. PULSAR Safety weight as function of relative velocity and distance The network topology at various snapshots in time The transmission probabilities assigned by PULSAR and D-NUM over a 2 s sliding window Packets received at Node 1 over a 2 s sliding window Cumulative packets received at Node The transmission probabilities of Node 49 over a 2 s sliding window in 100 node scenario Packets received at Node 49 over a 2 s sliding window in 100 node scenario Cumulative packets received at Node 49 in 100 node scenario x

11 4.9 Cumulative packets received at Node 49 from Node 45 in 100 node scenario Cooperative forwarding using multiple vehicles as virtual relays Time slot assignment in a transmission frame Trellis representation for number of active relays Expanded pairs of state columns for each hop for POC Expanded pairs of state columns for each hop for SFR Comparing multi-hop forwarding schemes Analysis vs. simulation results for end-to-end success probability after three hops Packet reception ratio vs. distance from destination Average hops vs. distance from destination PSM Reception Probability vs. distance of neighbour xi

12 List of Abbreviations ACK Acknowledgement AIMD Additive-Increase Multiplicative-Decrease BSM Basic Safety Message CALM Communication Access for Land Mobiles CCH Control CHannel CPF Cooperative POC-based Forwarding CTS Clear-to-Send GP Geometric Program GPS Global Positioning System HTP Hidden Terminal Problem IEEE Institute of Electrical and Electronics Engineers IP Internet Protocol ITS Intelligent Transportation System MAC Medium Access Control MANET Mobile Ad hob NETwork OBU On-Board Unit POC Positive Orthogonal Code POC-MAC POC-based MAC xii

13 PRR Packet Reception Ratio PSM Periodic Safety Message QoS Quality of Service RSU Road-Side Unit RTS Request-to-Send SAE Society of Automotive Engineers SCH Service CHannel SFR Synchronous Fixed Repetition SPR Synchronous p-persistent Repetition TO Transmission Opportunities UTC Coordinated Universal Time V2I Vehicle-to-infrastructure V2V Vehicle-to-vehicle VANET Vehicular Ad hoc NETwork WAVE Wireless Access in Vehicular Environment WSM WAVE Short Message WSMP WAVE Short Message Protocol xiii

14 Chapter 1 Introduction 1.1 Motivation The United Nations has declared to be the Decade of Action for Road Safety [1]. It is estimated that over 1.3 million people died in 2010 due to traffic accidents, which makes motor vehicle crashes the eighth-leading cause of death, accounting for 2.5% of all deaths worldwide [2]. According to the World Health Organization, if unchecked, the death toll is projected to surpass HIV/AIDS and become the fifth leading cause of death by 2030 [3]. In 2012 alone, there were 123,963 motor vehicle collisions in Canada, resulting in a total of 2,077 fatalities and 165,172 personal injuries [4]. This growing problem is the motivation for researchers in academia and industry to develop innovative technologies to improve the safety of the roads and highways of the future. These applications, known collectively as Intelligent Transportation Systems (ITS), make use of advanced technologies to improve the safety and comfort of automotive travel. Some examples of ITS safety applications are: emergency vehicle notification, collision avoidance systems, lane-change warning, and left-turn assist [5]. Realized on a wide-scale, ITS represent the next frontier in the enhancement of vehicular safety. Since many envisioned ITS applications require vehicles to be able to communicate with one another, their viability depends on the realization of a reliable low-latency vehicular communication network. 1.2 Vehicular ad hoc networks Vehicular ad hoc networks (VANET) are wireless networks composed of radio-equipped vehicles, referred to as on-board units (OBU), and fixed access points called roadside 1

15 Chapter 1. Introduction 2 units (RSU). The communication between these two types of nodes can be categorized into the following two types of wireless communication patterns: Vehicle-to-infrastructure (V2I) communication occurs between the mobile vehicular nodes (OBU) and the stationary roadside access points (RSU). Vehicle-to-vehicle (V2V) communication occurs between vehicular nodes (OBU). Although VANETs can be considered as a special type of Mobile Ad hoc Network (MANET), there are properties of VANETs that are not generally true of MANETs. Some of these distinguishing characteristics are listed as follows: Non-random mobility: The mobility of vehicles are restricted to the topology of the road, and are subject to traffic regulations, human driver behaviour and physical limitations of the vehicle. High-capacity battery: Since the OBUs will use the power reserve of the vehicle s battery, power and energy constraints are not typically a limiting design consideration. When the transmission power is limited, the goal is usually to reduce the interference and channel congestion. Dynamic network topology: Vehicular nodes move with high speeds. In cities, the typical speed limit is kmph; on highways, the speed limit is usually kmph. There are multiple lanes of vehicles, and usually lanes with opposing directions of traffic flow. Therefore, the network topology of VANETs is highly dynamic. Variable network density: The traffic density may be very high on major roads and highways during rush hour, while becoming very low at night or in rural areas. Frequent fragmentation: In areas of low traffic density or while the penetration rate of OBUs in vehicles is low, communicating vehicles may be far apart and frequently move in and out of range of each other. Large scale: When the technology is adopted on a large scale, the communication platform should be robust enough to handle the high node densities associated with rush-hour traffic in large metropolitan areas. Harsh vehicular channel: The high relative speeds of nodes and the presence of large obstructions such as trucks can cause harsh fading and shadowing effects on the wireless channel between VANET nodes. This can mean very unreliable wireless channels in the vehicular environment.

16 Chapter 1. Introduction 3 V2V safety communications SCH 175 SCH 181 SCH 172 SCH 174 SCH 176 CCH 178 SCH 180 SCH 182 SCH GHz GHz GHz GHz GHz GHz GHz GHz GHz Reserved High-power, long-distance public safety communications Figure 1.1: FCC channel allocation 1.3 Standardization The United States Federal Communication Commission (FCC) designated a 75 MHz band at GHz for Dedicated Short-Range Communications (DSRC) in 1999 [6]. This spectrum is reserved for providing short to medium range reliable communication for ITS applications. The allocated spectrum consists of seven non-overlapping channels of 10 MHz each, as illustrated in Figure 1.1. Channel 178 was designated as the control channel (CCH), and the other six channels were designated as service channels (SCH). SCH 172 was designated as high availability, low latency channel for V2V transmission of safety information, such as periodic packets required by cooperative awareness safety applications. SCH 184 was designated for high-powered, long-range public safety communication, such as intersection collision warning. The remaining four SCHs (174, 176, 180, 182) may be used for either safety or non-safety applications. Furthermore, each of the two adjacent pairs of these four SCHs can be combined to form a 20 MHz channel, which are named SCH 175 and SCH 181. The International Organization for Standardization (ISO) is developing a set of international communication standards for ITS called Communication Access for Land Mobiles (CALM) [7]. The CALM umbrella covers a wide array of communication technologies and communication modes, with the goal of developing a future-proof abstraction layer between ITS applications and the protocols of the communication technologies

17 Chapter 1. Introduction 4 they rely upon. Communication methods covered by CALM include satellite, cellular, Bluetooth, and DSRC. The SAE J2735 standard [8] defines a message set dictionary for DSRC, including the data frames and elements of each message type. J2735 contains message formats for a variety of ITS applications, including vehicular safety, emergency vehicle warning, automated tolling, etc. Of particular relevance to this thesis, the Basic Safety Message (BSM) is defined as a short heartbeat message to be broadcast periodically by vehicles to near-by neighbours. The BSM is to include state information of the transmitting vehicle, such as the position, speed, acceleration, The current IEEE standard for vehicular communication in this DSRC frequency band is the Wireless Access in Vehicular Environment (WAVE) [9]. The WAVE standard relies upon the IEEE standard for the specification of the PHY and MAC layer operation in vehicular communication. These specifications were in the IEEE p amendment [10], but have since been incorporated into the latest IEEE standard [11]. The WAVE PHY layer is based on the OFDM PHY specified in IEEE a, which operates at the nearby 5 GHz band. However, some modifications were made to account for harsh multipath of the vehicular environment. WAVE uses 10 MHz channels instead of the 20 MHz channels of a, and has double the OFDM timing parameters and half the data rate of the latter. The transmission power were also adjusted to account for the longer required operating range of vehicular radios. IEEE defines the multichannel operation of the WAVE MAC, using a synchronized scheme based on Coordinated Universal Time (UTC). Time is divided into Sync Intervals of 100 ms, which are then split into two 50 ms sub-intervals: the CCH Interval and the SCH Interval. All WAVE devices must monitor the CCH during the CCH Interval at the beginning of each Sync Interval. Devices are permitted to switch to a SCH during the SCH Interval. The first 4 ms of both types of intervals is a guard interval reserved to allow for channel switch to be performed. The MAC operation within each channel falls under the scope of the IEEE p amendment. It defines a new WAVE BSS (WBSS), designed to allow rapid connection between fast-moving vehicular nodes without the slow association and authentication steps required under vanilla IEEE However, the operation of the IEEE p CSMA/CA scheme, especially in broadcast mode suffers from the hidden terminal problem (HTP), as there is no Request-to-Send (RTS)/ Clear-to-Send (CTS). Moreover, a lack of an acknowledgement (ACK) scheme in broadcast mode leaves no way to ensure reliable packet delivery. These factors motivated the proposal of a family of repetition-

18 Chapter 1. Introduction 5 based MAC protocols for reliable vehicular broadcast of short safety packets. An extensive discussion of repetition-based vehicular MAC protocols can be found in Chapter 2. The IEEE 1609 family of standards complete WAVE in defining the architecture, communications model, protocols located in the upper layers of communication: Security Services Networking Services Multi-channel Operations Over-the-air Electronic Payment Data Exchange Protocol for ITS Identifier Allocations In additional to the standard Internet Protocol (IPv6) stack, the WAVE standard defines a WAVE Short Message Protocol, which allows upper layer applications to control the lower layer parameters of each transmission, such as the channel number and the transmission power. While IP packets are restricted to SCHs, WAVE Short Messages (WSM) using WSMP can be transmitted on any channel. 1.4 Vehicular applications There are many envisioned ITS applications for improving the safety and comfort of automotive travel. They can be categorized into safety and non-safety applications. A cooperative collision avoidance application is an example of the former, while providing point-of-interest-based advertisements would be an example of the latter. These applications can be further divided into those which require on some V2I communication with roadside infrastructure (RSU) and those that rely solely on V2V communication. Table 1.1 presents some proposed ITS applications and their classification in terms of these categories Use Case Scenarios In the first scenario, cooperative awareness information is exchanged among OBUs through BSM messages broadcast using WSMP on a designated safety channel, which may be the CCH 178 or the safety SCH 172. There is a high density of vehicular nodes in the area of interest. We wish to adjust the beaconing rate of the heartbeat BSMs to reduce the congestion on the shared channel. However, vehicles will find themselves in different situations and their safety applications would require different levels of service. For instance,

19 Chapter 1. Introduction 6 Safety Non-safety V2I Curve Speed Warning Left Turn Assistant Intersection Collision Warning Road Condition Warning Internet Access Advertisement Traffic Light Scheduling V2V Cooperative Collision Warning Lane Change Assistant Emergency Vehicle Warning Post-Crash Warning Peer-to-Peer File Transfer Gaming Social Networking Table 1.1: Some categorized vehicular applications. consider a two-lane undivided highway with opposing direction of travel in each lane. Two vehicles travelling toward each other are in a potentially more hazardous situation than two vehicles moving away from each other. The distribution of network resources should reflect the importance of each communication link in terms of the potential hazard of their end-point vehicles. In the second scenario, a vehicle has suffered an accidental collision, which triggers the transmission of an emergency alert message. This alert message should be disseminated to all nodes within a certain geographical area of the accident. This area may be too large to be covered by a single hop, and thus multi-hop forwarding of the alert message is necessary. This would provide oncoming drivers without line-of-sight of the accident with advanced warning, allowing them to respond more quickly and avoiding a potentially chain collision. Also, a multi-hop propagation of the alert message to a nearby RSU would inform the local emergency services (police, paramedics, etc) of the traffic accident. Meanwhile, the designated safety channel remains responsible for the periodic safety BSMs. We would like for both the periodic BSMs and the event-based multi-hop alert messages to have high performance in the event of a traffic accident. 1.5 Contributions This thesis covers two problems in vehicular safety communications, in the context of the repetition-based broadcast MAC. The first is the distributed adaptation of the transmission rate of single-hop broadcasts of BSMs, which is done according to both the dynamic driving context of vehicles and network congestion. The second is the multi-hop propagation of emergency alert messages on the CCH in the presence of other broadcasting vehicular nodes. The results are divided into three chapters and are as follows: Chapter 3 considers the congestion control problem from a novel perspective, addressing the non-homogeneous QoS requirement of different vehicles throughout the

20 Chapter 1. Introduction 7 network. We express the driving context using a maximum delay constraint profile which is dependent on the distance and the relative speed between pairs of vehicles. We then define a set of probabilistic constraints based on these context-based delay constraints. We apply these constraints to modify a well-known additive-increase multiplicative-decrease congestion control algorithm. We show that by using the probability of satisfying the delay constraints as the fairness goal instead of the transmission rate itself, improved performance can be gained in certain circumstances. We then formulate the network load minimization problem subject to these probabilistic delay constraints as a geometric program (GP). Using the method of dual decomposition, we derived a distributed subgradient method algorithm. We propose two variants of this distributed algorithm to attempt its convergence speed: one using local coordinate-wise optimization steps and the other using the method of primal approximation through time averaging over a moving window. We present convergence results for the latter method and all three variants are compared with a well-known congestion control algorithm through NS-2 simulations. This work is published in [12] Chapter 4 takes an alternative approach to the problem in the previous chapter. Using the same system model, the differential driving context of vehicular nodes is expressed using a weight for each communication link, which is based on the distance and relative speed of the end-points. These weights are using to express the congestion control or transmission rate adaptation problem as a weighted network utility maximization (NUM) problem, using the expected delay or inter-packet reception time as the measure of utility. We propose a distributed algorithm, based on successive local optimization to solve this problem and provide an analysis of its convergence. The performance of the proposed algorithm is evaluated through NS- 2 simulations of a car-passing scenario and is shown to have improved performance over a well-known AIMD congestion control algorithm. The work in Chapter 4 is published in [13]. Chapter 5 considers the problem of reliable multi-hop forwarding in vehicular networks, which is required by many safety applications. We extend the repetitionbased POC-MAC protocol to handle multi-hop transmission of alert messages, resulting in the proposed Cooperative POC-based Forwarding (CPF) protocol for highway vehicular networks. We introduce the notion of a virtual relay, which is composed of multiple cooperating relays at each forwarding hop. Members of the virtual relay schedule their transmissions to correspond to a single POC codeword,

21 Chapter 1. Introduction 8 thereby adhering to the POC-MAC. CPF exploits spatial diversity while mitigating the effect of hidden terminals. By allocating separate POC-based schedules for multi-hop packets and the broadcast of periodic safety messages, the CPF protocol reduces the interference between the two types of safety transmissions. The following steps of the distributed protocol are specified: a) Distributed Relay Selection b) Location-based POC codeword Allocation c) POC codeword Allocation for Virtual Relays d) Timeslot Assignment. Although the CPF protocol was first proposed in [14], in this thesis we present the novel analytical study of the protocol s performance using a Markov model. We derive the transition probabilities for three variants of CPF based on different repetition-based MAC schemes (SPR, SFR, POC). This model allows a network designer to evaluate the end-to-end success probability of a multi-hop packet under various network parameters. Through NS-2 simulations, the Markov model is validated and the performance of the CPF is compared with several alternative multi-hop transmission schemes. The work in this chapter is published in [15, 16]. 1.6 Thesis organization The remainder of the thesis is organized as follows. In Chapter 2, we review some proposed vehicular MAC protocols, as well as the major works on the topic of congestion control. In Chapter 3, we study the congestion control problem in VANETs as a load minimization problem subject to a set of safety-based constraints on the maximum delay. These constraints adhere to a delay-profile which depends on the driving context of each pair of vehicles. The centralized problem formulation is presented and a distributed subgradient algorithm is proposed and studied. These delay constraints are then applied to enhance the performance of a recent AIMD congestion control algorithm. In Chapter 4, we take a network utility maximization approach to the rate adaptation problem, expressing the driving context of vehicular nodes as a weight on the utility. In Chapter 5, we extend the repetition-based POC-MAC to a cooperative forwarding scheme. A review of the literature relating to multihop message propagation for VANETs is found within the chapter itself for the convenience of the reader. Finally, the concluding remarks and future avenues of research are discussed in Chapter 6.

22 Chapter 2 Related Work In this chapter, we shall begin with an overview of repetition-based vehicular MACs, which have been developed for reliable broadcast between vehicular network nodes. Next, a review of the recent works in congestion control for VANETs will be provided. Finally, we give a discussion of the related works in multi-hop forwarding for vehicular networks. 2.1 Repetition-based medium access control As previously mentioned, the broadcast mode of the WAVE MAC lacks a way of providing reliable delivery and dealing with the hidden terminal problem (HTP). Unlike in unicast transmissions where a RTS/CTS handshake is used to notify interfering hidden terminals, the broadcast mode does not have any such mechanism. Broadcast mode also lacks an acknowledge/re-transmit scheme to ensure reliability. Since broadcast safety packets are expected to be short, a complicated handshaking scheme would mean significant overhead. Motivated by these issues, a family of repetition-based broadcast MACs have been proposed for the reliabile transmission of short safety messages in VANETs [17]. In these schemes, the time on the channel is divided into time slots that correspond to the transmission duration of a safety packet or BSM. The useful lifetime of the safety packet is called a transmission frame and consists of L time slots. All nodes are assumed to be synchronized in their slot times using UTC from GPS devices or some other synchronization method. Each node will be actively transmitting repetitions of its safety packet in a certain subset of the L total time slots in a transmission frame. The various schemes differ in how these active transmission time slots are selected. The work in [17] introduced the repetition-based vehicular broadcast MAC and proposed random selection of the transmission time slots. In Synchronous p-persistent Retransmission (SPR), a node transmits its packet in each time slot with probability p. In 9

23 Chapter 2. Related Work 10 Synchronous Fixed Retransmission (SFR), a node randomly selects a fixed number of time slots out of L in which to transmit. Reliability is enhanced by exploiting temporal diversity in the channel via the repeated transmissions. Expanding upon this concept, POC-MAC uses structured transmission patterns based on Positive Orthogonal Codes (POC) [18]. A POC is a binary code of fixed length L, where the cross-correlation between any pair of codewords is no more than λ. For example, if x and y are two different codewords in a POC of length L, then L x, y = x i y i λ. i=1 Under this definition, codewords do not necessarily have constant weight, as long as the cross-correlation property holds. Under POC-MAC, each vehicle is permitted to transmit only in the time slots corresponding to the 1 bits of its assigned POC codeword. Thus, a node i with a codeword of weight w i and length L will repeat the transmission w i times in a transmission-frame of L time slots. This scheme was shown to further reduce collisions and thereby improve packet reception ratios [18]. 2.2 Congestion control and network utility maximization In 1988, Transmission Control Protocol (TCP) was proposed by Jacobson in [19] as congestion control mechanism for the Internet, which is implemented at the end-points of the data transmission path. Chiu and Jain analysed a congestion control scheme which used a single-bit feedback mechanism from a fair resource allocation perspective in [20]. Following these seminal works, a vast family of literature has been written on the topic of congestion control for both wired and wireless communication networks. In the late 1990 s, Kelly took this resource allocation view further and presented a network utility maximization (NUM) framework for the congestion control problem in [21, 22]. Different notions of fairness can be achieved using particular utility functions of the resource allocation. Among these, proportional fairness was examined in [21], and minimal potential fairness in [23], whereas the max-min fairness concept was discussed in [24]. These notions were encapsulated in the alpha-fairness framework of [25], which introduced a common utility function parametrized by α. The utility functions for propor-

24 Chapter 2. Related Work 11 tional fairness, minimal potential fairness, and max-min fairness can be obtained using the alpha values of 1,2, and, respectively. A comprehensive treatment of more than three decades of work on this topic would not be feasible within the scope of this thesis. However, excellent overviews can be found in Srikant s survey [26], and in Shakkottai and Srikant s monograph [27], among others. 2.3 Congestion control in VANETs The works on congestion control in VANETs can be divided into two categories: proactive and reactive. Proactive methods seek to minimize each node s offered load, regardless of the current state of the channel. In reactive protocols, the channel is monitored for congestion. When the latter is detected, each node responds by adapting their share of the network resource, for example via the transmission power or the beaconing rate Proactive protocols Reference [28] proposes a framework which removes redundant information in BSMs. In [29], the scheme uses a series of model-based estimators to estimate the tracking error of all neighbours, and vehicle s beaconing rate is minimized subject to a threshold on this tracking error. In [30], the authors of the OPRAM protocol note that, in select safety applications, cooperating vehicles require fewer packets to be exchanged than vehicles acting autonomously. The network load can be reduced further in some cases by introducing awareness of the traffic context. The authors propose a power adaptation scheme to ensure delivery of at least one safety packet with some probability for specific traffic contexts, such as during a lane change. Based on the mobility information of the local nodes and depending on the traffic context, a warning distance is found within which the sending vehicle should notify a particular neighbour via a safety beacon. The protocol estimates the number of packets N T that the vehicle has time to transmit within this warning distance. Based on some desired maximum failure probability, the algorithm finds the required erasure probability of each of the N T transmissions. The algorithm then adapts the transmission power to ensure that the erasure probability is experienced at the target vehicle for each transmitted packet, based on the channel model and the link length.

25 Chapter 2. Related Work Reactive protocols Torrent-Moreno, et al. propose a transmission power control algorithm called distributed fair power adjustment for vehicular environments (D-FPAV) [31]. The authors cite simulation results from a previous work [32], which demonstrate that increasing the beaconing rate results in a lower packet reception probability due to the increased number of collisions. Furthermore, they show that extending the communication range by increasing the transmission power results in a congested channel and packet reception rates for nearby neighbours will suffer. The authors chose to fix the beaconing rate at the minimum level mandated by the standards of safety applications and adapting each vehicle s transmission power to keep the offered load on the channel below a desired threshold. The scheme aims to enforce max-min fairness in terms of the power levels over the network. For a given node i, the carrier-sensing range depends on the transmission power, which in the interval [P min, P max ]. Let CS i denote the set of neighbouring nodes within node i s carrier-sensing range. Let CSi max denote the carrier-sensing range corresponding to the maximum transmission power. The beaconing load BL i is defined as the number of other network nodes j for which i CS j. Lastly, the maximum threshold for the beaconing load is denoted MBL and is common to all nodes. The distributed algorithm is derived from a centralized predecessor (FPAV) which takes a water-filling approach and assumes global knowledge of the network. This centralized algorithm has two phases. First, it incrementally increases the global transmission power of all nodes in the network until the MBL threshold is reached for at least one node. The second phase allows individual nodes to further increase their transmission power incrementally until the beaconing load constraints are tight at all nodes. However, the authors discard this second optimizing step in the design of D-FPAV and reason that the marginal gain is not worth the increased complexity in the distributed algorithm. The D-FPAV is quite simple and requires two-hop message passing. First, it assumes that node i knows the positions of all two-hop neighbours, i.e. those in the set CS max i. Node i then calculates the minimum common power level for all nodes within this set such that BL j MBL, j CSi max. Next, this value P i disseminates over two hops to all nodes in CSi max. After messages are exchanged, node i then sets its power to the minimum over P i and the power values it has received from its neighbours. The authors acknowledge that accurate information of the two two-hop neighbourhood (CS max ) is difficult to obtain in practice in a highly dynamic vehicular environment. Without perfect knowledge, the D-FPAV algorithm cannot guarantee a strictly fair power assignment. The authors argue that the distributed algorithm nevertheless provides an effective approximation to a fair power assignment based on simulations.

26 Chapter 2. Related Work 13 The authors also investigate the trade-off between the accuracy of the two-hop neighbourhood knowledge and the required overhead in additional information added to the safety packets. In order to obtain two-hop neighbour map, each node must transmit the status information of its one-hop neighbours. This large amount of information is piggybacked on beacons and can result in up to a tripling of the packet size. The authors propose sending this large extended beacon every k-th beacon, where k is a parameter in different simulations. For lower values of k, the neighbourhood information is more upto-date, but the large amount of overhead results in a lower overall power level assigned to each node for a given offered load threshold. The rate adaptation scheme in [29] is extended in [33, 34] to adapt the transmission range or power in response to congestion detected through the channel busy rate (CBR). An information dissemination rate (IDR) metric is defined as the number of packets successfully received by a node s neighbours per unit time. The IDR vs. CBR curve is found to be independent of the values of beaconing rate and transmission range, and the maximum IDR can be achieved by adapting either one of the controllable parameters. The locally observed CBR is used as feedback in the closed-loop transmission power adaptation scheme based on gradient descent. An analytic study of this approach is presented in [35], which also includes the spatialtemporary priority notion by adding a distance-dependent weight to the IDR metric (widr). One finding is that the widr-maximizing values of CBR are spread over a larger range of values than for IDR, which makes optimized designs more dependent on the transmission rate [35]. These widr vs. CBR operating curves were derived by sweeping network-wide global parameters of beaconing rate and vehicular density. The authors recognize that an individual node cannot easily discover the appropriate values of these parameters in order to choose the correct operating curve. They resort to a robust design where each node adjusts its transmission range to keep its local CBR measurement within a range of interest. However, the role of the distance weights have been subsumed into a CBR range; the algorithm itself does not distinguish between vehicles with different safety-related needs for channel resources as long as they are experiencing the same local CBR. The authors of [36] show through simulations that the node density affects the optimal beaconing rate but not the optimal transmission range. Thus, their PULSAR (Periodically Updated Load Sensitive Adaptive Rate) algorithm fixes the latter and adapts the rate in an additive increase multiplicative decrease (AIMD) manner based on a comparison of the two-hop CBR information with an optimal CBR value that corresponds to the minimum average inter-packet reception time (IRT). The magnitude of the adaptation is

27 Chapter 2. Related Work 14 modulated by a comparison with the average rate of the local nodes in order to improve the scheme s convergence. The authors also note that the algorithm in [35] does not result in a fair power allocation, as a node can contribute to congestion at a node that it does not know about, and therefore they proposed piggybacking congestion information over two hops. The temporal-spatial priority of BSMs is also described via the various safety benefit vs. transmission rate curves in [36]. These curves were used to find a pair of min/max Tx rates for each vehicle, which forms the interval within which the rate is allowed to vary. However, the algorithm does not seek to provide an allocation of rate that results in a fair distribution of safety benefit to all the nodes in the network. Furthermore, we must consider the worst case scenario when the channel cannot accommodate even the minimum Tx rate of all the vehicles. In [37], a linear scheme adapts the rate proportionally to the difference between the current and desired aggregate rate in a formulated collision domain. While it does not require synchronous operation like [36], it assumes that all nodes in the network experience the same CBR and all nodes desire an equal share of the available channel resources. A recent work in [38] examines vehicular congestion control methods within the context of an intersection assistance application, and studies the implications of a globally fair resource allocation on safety. The authors found that the fair allocations of current approaches cannot provide sufficient safety to the studied intersection assistance application. A hybrid adaptive method was proposed, wherein temporary exceptions are granted to hazardous vehicles to the fairness constraints of the original vehicular congestion control algorithm.

28 Chapter 3 Rate Adaptation Through Safety-based Delay Constraint Profile An important requirement of many safety applications is the reliable exchange of broadcast packets called Basic Safety Messages (BSMs). These periodic transmissions allow each vehicle to obtain an up-to-date knowledge of the state information of its neighbouring vehicles. However, the reliability of such transmissions suffer from the harsh vehicular wireless channel and the lack of ACKs in WAVE s broadcast mode. Furthermore, high vehicular density and high node mobility in some networks may cause channel congestion upon the large-scale deployment of VANETs. This problem motivates the search for an effective distributed congestion control scheme for the periodic broadcast of BSMs. In order to mitigate channel congestion, congestion control schemes must reduce the offered load on the shared control channel. However, this load reduction must be done while ensuring fairness. If network-wide throughput is maximized at the expense of a particular vehicle, the latter would be unable to inform others of its presence, resulting in a potentially dangerous situation. While other works in literature seek to obtain fairness in the allocated resource, we note that not all nodes have the same requirement. Vehicles in different driving contexts pose different amounts of hazard to each other, and subsequently have different quality of service (QoS) constraints on the exchange of BSMs. We use the term safety benefit as an umbrella term including any performance metric for a safety application which may be the subject of such QoS constraints. We propose a congestion control scheme which takes the varying driving context, and the subsequently varying safety benefit constraints, into account. We aim to achieve a fair distribution in the amount of safety benefit we provide to each vehicle. To do 15

29 Chapter 3. Rate Adaptation: Safety-based Delay Profile 16 so, we must account for the variation in local topology, the distance-varying quality of wireless links, and the spatial-temporal nature of BSM priority. The latter refers to the fact that closer neighbours are potentially more dangerous and should be informed with lower delay/higher frequency. Thus depending on the distance of their neighbours, two vehicles may require different portions of the channel resources to reach the same degree of safety. The remainder of this chapter is organized as follows. Section 3.1 presents the system model, including a review of the underlying MAC scheme. Section 3.2 presents the driving context-based constraint on the maximum delay. A safety-aware AIMD scheme called SALSA, which aims to achieve fairness in the probability of delay constraint violation, is proposed in Section 3.3. Section 3.4 details the proposed centralized problem formulation for minimization the network load subject to the delay constraints. Section 3.5 presents the distributed subgradient method algorithm derived through dual decomposition. Section 3.6 gives a variant distributed algorithm which uses a moving average of the primal solutions from each iteration. Finally, the performance of the proposed algorithms are evaluated through simulations in Section System Model The vehicular network is represented by the graph G(Ω, E), where Ω = {1, 2,..., n} denotes the set n of vehicular nodes. The edges in E are represented by an n n binary adjacency matrix and are determined by the transmission range R N. For any pair of distinct nodes i, j Ω, E i,j = 1 if the distance between the nodes x i,j R N, and equal to 0 otherwise. We denote the set of neighbours of node i as N i = {j Ω j i, x i,j R N }. Each edge in the network has an interference-free probability of packet reception p i,j = p(x i,j ), which is a function of the distance between nodes and depends on the channel fading and path loss model. Collisions on the shared channel are modelled by an interference range R I. A collision occurs at receiving node j if more than one node located within R I of node j, including node j itself, transmits concurrently. We denote the set of nodes whose transmissions may collide at node j as M j. For the tractability of the following analysis, we shall assume that the interference range is equal to the transmission range, as illustrated in Figure 3.1. In real-world systems, the interference range is often much greater than the transmission range and obtaining information from all nodes within the latter would require two-hop exchange of information. Our assumption, which pertains to the system model used in both chapters 3 and 4, allows for simpler and more tractable problem formulations.

30 Chapter 3. Rate Adaptation: Safety-based Delay Profile 17 Rn Rn i j Ri Ri Mj Neighbor nodes of interest for node i, of set Ni Figure 3.1: Network diagram for notation used in the general network model. Therefore, with this caveat, we shall proceed with this assumption that R I = R N. However, the performance evaluation sections show that the algorithms developed under this system model and this assumption nevertheless perform well in simulations with realistic network parameters where the interference range is greater than the communication range. A simplification of this multi-hop network is the fully-interfering network model, where the network graph G(Ω, E) is a complete graph. This is a single cluster network where any concurrent transmissions in the network will collide at all nodes, and thus for all i Ω, M i = Ω and N i = Ω \ {i}. The proposed congestion control scheme operates on top of the timeslot-based Synchronous P-persistent Repetition (SPR) MAC, which was proposed for reliable broadcast of periodic BSMs for VANETs in [17, 18]. Time on the control channel is divided into timeslots corresponding to the transmission during of one safety packet. Due to IEEE p s lack of an acknowledgement and retransmission mechanism when operating in broadcast mode, packets are repeated to increase their chance of reception. These repetitions are justified by their relatively small size when compared to the network overhead of broadcast acknowledgement schemes. In SPR, a repetition of the most recently generated BSM is broadcast with probability α in each timeslot. Studies of these repetition-based MACs in [17] and [18] have found that the optimal value of a global α for all nodes, depends on factors such as the network density, packet size, packet range, data rate, etc. In previous studies, all nodes in the network used the same global parameters, and often only for single-cluster fully-interfering networks. In this work, we propose methods for adaptively selecting the transmission probability α i for