Vehicular Ad-Hoc Networks (VANETs) Simulator and Hierarchical Clustering Algorithm

Transcription

1 Vehicular Ad-Hoc Networks (VANETs) Simulator and Hierarchical Clustering Algorithm Thesis submitted in partial fulfillment of the requirement for the degree of master of science in the Faculty of Engineering Submitted by: Mr. Efi Dror Advisers: Dr. Chen Avin, Dr. Zvi Lotker Department of Communication Systems Engineering Faculty of Engineering Ben-Gurion University of the Negev April 22, 2012

2 Abstract Vehicular Ad-Hoc Networks (VANETs) offer communication between vehicles and infrastructure devices located at road sideways and junctions. Warning messages, among other messages, can be used to alert drivers, and thus improve road safety. To adapt to the unique nature of VANETs, which demands the delivery of time sensitive messages to nearby vehicles, fast topology control and scheduling algorithms are required. A clustering approach, which was initially offered for Mobile Ad-Hoc Networks (MANETs), can be adapted to VANETs to solve this problem. In this work, Hierarchical Clustering Algorithm (HCA), a fast randomized clustering and scheduling algorithm is presented. HCA creates hierarchical clusters with a diameter of at most four hops. Additionally, the algorithm handles channel access and schedules transmissions within the cluster to ensure reliable communication. Unlike other clustering algorithms for VANETs, HCA does not rely on localization systems which contributes to its robustness. The running time and message complexity of the algorithm were analytically analyzed. In order to evaluate such unique networks, a designated VANETs simulation tool was developed as well. The Wireless Vehicular Networks Simulator (WVNS) allows a comprehensive study of algorithms and protocols for VANETs over a variety of realistic scenarios. Each scenario consists of several topographic environments, each characterized by different mobility and physical channel attributes. WVNS was used to examine HCA. The algorithm was simulated under several mobility scenarios. The simulation results confirm that the algorithm behaves well under realistic vehicle mobility patterns in terms of cluster stability. Keywords: Vehicular Ad-Hoc Networks, VANETs, Clustering, MAC, Leader Election, Simulation, OMNeT++, Dominating Set 2

3 Acknowledgements I would like to thank Dr. Chen Avin and Dr. Zvi Lotker for giving me the opportunity of working on this research project. I would like to acknowledge Arik Sapojnik, Naaman Sittsamer, Oren Tzfati and Eran Nahum for their contribution to this project. This research was conducted in collaboration with Mobilicom LTD, and was supported by the office of the Chief Scientist of the Ministry of Industry, Trade & Labor, Israel. Grant No

4 Contents 1 Introduction Objectives Contributions Outline Related Works Vehicular Ad Hoc Networks - General Overview Simulation Of Vehicular Ad Hoc Networks Dominating Set Problem Clustering and Channel Access Algorithms Wireless Vehicular Networks Simulator (WVNS) Design Overview Modules Details Channel Modeling Mobility Patterns Algorithm Module Hierarchical Clustering Algorithm for Vehicular Ad-Hoc Networks Problem Definitions

5 4.2 Hierarchical Clustering Algorithm (HCA) ClusterRelays Selection Phase ClusterHead Selection Phase Cluster Formation and Scheduling Phase Cluster Maintenance Phase Time and Message Complexity Correctness Run-time Complexity Message Complexity Simulation Results K-ConID Algorithm Static Scenarios Number of Clusters Convergence Time Messages Dynamic Scenarios Number of Clusters Cluster Stability Simulation Summary Conclusions and Future Work 76 5

6 List of Figures 3.1 Wireless Vehicular Network Simulator (WVNS) General Structure Example of multi-environments configuration Path loss calculation for two environments The API between OMNeT++ and the implemented code of the algorithm HCA state machine of role transitions SYNC message slots scheme Number of ClusterHeads vs. the total number of nodes in different playground sizes. The numbers in the legend denote the square of the simulation playground size in meters Number of ClusterHeads vs. the total number of nodes and different densities The average number of rounds during HCA s third phase The average number of rounds during HCA s first three phase using WVNS. The numbers in the legend denote the square of the simulation playground size in meters The average number of slots during HCA s first three phases using WVNS. The numbers in the legend denote the square of the simulation playground size in meters

7 5.6 The number of rounds during HCA s first third phases vs. the number of clusters using WVNS. The numbers in the legend denote the square of the simulation playground size in meters The average number of messages sent during the first three phases The average number of messages sent and the average number of slots during the first three phases. The upper and lower surfaces denote the number of slots and message respectively A screen-shot of the simulation with the region which is used (a suburb of Tel- Aviv). ClusterHeads, ClusterRelays, and Slaves are denoted by rectangles, triangles, and circles respectively The average number of ClusterHeads vs. the maximal speed for 100 vehicles network The average number of cluster switches per node during the simulation time (300 sec), vs. vehicles maximal speed for 100 vehicles network Average percentage of time in which nodes were not members of any cluster, vs. the maximal speed for 100 vehicles network

8 List of Algorithms 1 ClusterRelay Selection ClusterHead Selection Cluster Formation and Scheduling Cluster Maintenance Phase - Part Cluster Maintenance Phase - Part

9 List of Tables 5.1 Static Simulation Configuration Dynamic Simulation Configuration

10 Chapter 1 Introduction Vehicular Ad-Hoc Networks (VANETs) serve as the basis for intelligent transportation systems (ITS) by utilizing Vehicle to Vehicle (V2V) and Vehicle to Infrastructure (V2I) communication. Dedicated hardware are planned to be incorporated into vehicles and some roadside units, while a wireless medium will be utilized to enable V2V and V2I communication. VANETs can serve several purposes [10]: First, safety and warning messages could be used to alert drivers as to dangerous and unpredicted situations, and thus reduce the number of car accidents or and their severity. In addition, such systems can improve road utilization by managing traffic flows. Traffic updates can be delivered in real time to allow time saving and lower fuel consumption. Finally, commercial and entertainment services can be distributed via these systems. For example, advertisement can be sent to users based on their location and Internet access can be provided for passengers comfort. Therefore, research on VANETs has been receiving increasing interest in the last couple of years, both on the algorithmic aspects e.g., [15, 37, 33, 36, 13] as well as standardization efforts like the IEEE p [19] and IEEE 1609 [18] standards (named WAVE - Wireless Access in Vehicular Environments). The p standard handles the MAC and PHY layers for each individual channel, while IEEE 1609 standards deal with upper layer protocols and 10

11 multi-channel operation [21, 39]. VANETs have some unique characteristics and requirements [4]. On one hand, safety applications require extremely low message delay in order for them to be effective. On the other hand, unlike some MANETs (Mobile-Ad-Hoc Networks), power and computational abilities are sufficient since the devices are carried by vehicles. Another key characteristic of V2V communication is the high relative velocity between the vehicles engaged in such communication, which results in short term sessions. Consequently, solutions based on the current packet radio communication protocols cannot guarantee the required quality of service (QoS). In addition, safety messages are only required to be transmitted in a small radius in order for them to be effective. Combining these two key requirements leads to the basic approach - clustering vehicles into groups [33, 15, 37] to ensure high QoS and fast propagation of messages in a limited area. The clustering approach was initially offered for MANETs [6, 27, 3, 2, 12, 7]. These works include k-hops clusters in which the number of hops between any of the nodes in the cluster is at most k [2, 12, 7]. More recent work adopted the clustering approach to VANETs. Examples include [15, 37, 33, 36, 13], however, they are all based on a localization system such as GPS to improve their performance. Unfortunately, this leads to undesirable dependence upon these systems in some situations. We elaborate more on this in the related work section. 1.1 Objectives Motivated by an industry-based collaboration with Mobilicom LTD [38], our goals were to create 4-hops clusters (where each node in the cluster is at most 2 hops from a cluster head) as fast as possible without the use of GPS or any other localization systems. The clustering algorithm should provide a comprehensive solution for wireless channel access in a mobile environment. In addition, it should guarantee a tight time delay bound for safety messages 11

12 by selecting a set of leader nodes (i.e., cluster heads) that would synchronize and schedule channel access for the entire network. A preliminary step towards the development of such algorithms is the development and implementation of an appropriate simulator to allow the evaluation of such an algorithm. The algorithm was implemented in a way that would allow executing the same code on the developed simulator and on a dedicated vehicular wireless device. 1.2 Contributions Following the unique nature of VANETs expressed in short term sessions, a fast clustering algorithm is required. In this work we offer a fast randomized clustering and scheduling algorithm, HCA, to allow a quick network setup. Our basic motivation was that 4-hops are sufficiently local for safety and other messages types, and that the initial setup time is more critical than cluster quality. The initial setup, in turn, can be followed by a maintenance phase that improves the clusters. Therefore, the suggested algorithm creates hierarchical clusters in which the maximal distance between a ClusterHead (CH) vehicle and any other vehicle in the cluster is two hops. Additionally, unlike other k clustering algorithms such as [2, 12, 7], the algorithm does not assume any lower layer connectivity, and the algorithm handles the channel access method and transmission scheduling within the cluster to avoid message collisions. In order to evaluate the proposed algorithm prior to deployment and to examine its compatibility to vehicular networks, a designated simulator for VANETs was developed. The suggested algorithm was simulated using the designated simulator in several scenarios and was compared to K-ConID algorithm [7]. The studied and compared parameters include the number of the formed clusters and the stability of clusters measured by the number of cluster switches. Simulations show that the algorithm performs better under realistic mobility patterns rather than under a random mobility pattern, which indicates that the 12

13 algorithm suits VANETs. In addition, even though more clusters are formed by the HCA than by K-ConID, it forms much more stable clusters. A partial review of the results can be seen in [11]. Furthermore, the algorithm s cluster formation process was studied analytically and simulated, showing that the algorithm creates clusters within O(m log log m) time slots using O(m) messages with m denoting the maximal cluster size in the network (of size n). 1.3 Outline This work is organized as follows: in chapter 2 we discuss related work regarding VANETs, VANETs designated simulators and simulations tools used to evaluate algorithms and protocols. In addition, some solutions to the the Dominating Set problem and clustering algorithms for MANETs and VANETs are reviewed. Chapter 3 presents the designated simulator for VANETs which we developed, and its unique characteristics. This is followed by a detailed description of the suggested algorithm and its running time analysis. The scenarios that were used to evaluate the algorithm, and the results that were obtained are presented in chapter 5. This work concludes with the conclusions chapter. 13

14 Chapter 2 Related Works In this chapter, previous work will be reviewed according to the logical structure of this work. First, VANETs development, challenges and standardization efforts will be presented, followed by an overview of existing simulation tools, particularly tools that were used in this work. This is followed by a review of the Dominating Set (DS) problem and its solutions. This chapter is concluded with an overview of known clustering algorithms for Mobile Ad Hoc Networks (MANETs) in general and Vehicular Ad Hoc Networks (VANETs) in particular. 2.1 Vehicular Ad Hoc Networks - General Overview 75 MHZ of Dedicated Short Range Communication (DSRC) spectrum were allocated in the US for vehicular communication with its primary goals being public safety and improvements of traffic flows. In [10], the authors present feasible applications for Intelligent Transportation Systems (ITS). ITS require reliable data communication, therefore an appropriate and mature technology must be used. Several wireless technologies are reviewed including infrastructure and infrastructureless methods of connecting vehicles and road side units including cellular networks, WiMAX, WLAN and DSRC/WAVE which will be further explained next. 14

15 Several standardization efforts were made in recent years, mostly led by IEEE organization in order to standardize such communication. The authors of [39] review the Wireless Access in Vehicular Environments (WAVE) protocol stack as defined by the IEEE. The lower layers (Physical and MAC) are covered by the p[19] standard, while the Logical Link Control (LLC) layer and upper layers are covered by the IEEE 1609.x [18] standards. The WAVE architecture supports two protocol stacks above the same Physical and MAC Layers. The first protocol stack is the IPv6 used for traditional communication, while the other is WAVE Short Message Protocol (WSMP) which handles high-priority, time-sensitive communications. The IEEE p architecture is based on IEEE and specifies requirements to allow communication in the rapidly varying vehicular environment. It defines 10 MHZ channels, which include one control channel and six service channels. The 1609 standards define multi channel operation which requires that all stations must monitor the control channel during a common time interval. In addition it defines the WAVE services at the network and transport layers. Additional services, such as resource management and security services, are also defined by this standard. The authors of [21] review additional features of the p standard. One of its main feature allows quick network setup without much of the overhead required for the wireless networks. This is attributed to the fact that nodes using WAVE mode, are allowed to transmit and receive data frames without the need to belong to any Basic Service Set (BSS). Additionally, in order to enable quick network setup, joining a WAVE BSS (WBSS) is done by exchanging only one data frame without the association and authentication processes, as defined by the standard. Furthermore, the PHY layer is based on the a standard with minor changes: each channel bandwidth was minimized; and improved, transmission and reception capabilities were added. The authors of [4] and [26] investigated the differences between VANETs and MANETs. In [4] the differences were presented using a simulation case study. The authors point out that inter vehicle networks exhibit characteristics that are dramatically different from many 15

16 generic MANETs due to mobility constraints, driver behaviour and high mobility. The presented case study demonstrates four unique VANETs characteristics: 1) Rapid topology changes and short link life time; 2) Fragmentation of the network due to possible low deployment ratios; 3) Links disconnection prior to utilization; 4) Limited number of multiple routes. In addition in [26] the authors point out that MANET and VANET share similar characteristics such as short wireless transmissions range, self organization and management and the mobility of nodes. However VANET can be distinguished from Ad Hoc Networks in terms of the following aspects: Frequent disconnections when vehicular density is low Sufficient power and computational power Short term mobility prediction since vehicle move on defined routes Various communication environments - Urban, sub-urban etc. Hard delay constraints in life saving and critical applications Therefore these aspects must be taken into consideration when designing protocols and algorithms for VANETs. 2.2 Simulation Of Vehicular Ad Hoc Networks Simulation is an important tool used for study and evaluation of complex systems. Simulation of networks and protocols enables development and study of the suggested protocols prior to deployment. Well known networks simulators include Qualnet, and the open source simulators NS2 and OMNeT++. One of the broadly used simulation tools in academy is the open source network simulator OMNeT++ [40]. OMNeT++ provides a powerful networking simulation tool, however, it lacks the modelling of wireless communication. The authors 16

17 of [24] present MiXiM - a framework for simulating wireless channels. It provides detailed models of wireless channels, connectivity, mobility and MAC layer protocols for OMNeT++. The MiXiM framework is divided into two logical parts: Base framework and protocol library. The Base framework provides general functionality such as connection management, mobility and channel modelling, while the protocol library provides implementation of some known MAC layer protocols. Mobility of nodes has to be modeled for an accurate simulation results. In [34], the authors present different approaches for simulating the mobility of nodes in VANET simulations. In addition, they outline the main advantages and disadvantages for each approach. A node s mobility can be simulated in several ways: random/mathematical models allow simple implementation, while real world vehicular traces relay on traces recorded from vehicles trajectories. Additional implementation of mobility simulation are based on traffic simulators. These implementations include simulated traces and bidirectional coupled simulators which allows the network simulator to influence vehicles trajectory, which might resemble a realistic scenario. The conducted simulations show that the results obtained in scenarios with a realistic mobility patterns differ from the results achieved in random mobility scenarios, which leads to the conclusion that realistic mobility patterns are favorable, since they resemble practical scenarios. In order to allow the most accurate modeling of vehicular movements mobility of vehicles, the authors in [34] present hybrid simulation framework which is composed of a network simulator OMNeT++ and a road traffic simulator SUMO[25]. SUMO is an open source, highly portable, microscopic road traffic simulation package designed to handle large road networks. The mobility model consists of several sub-mobility models, such as: Car- Following-Model, Lanes Changing, Traffic Lights and more. The combined simulator was constructed to allow modelling of realistic communication patterns of vehicles engaged in VANETs. It simulates the influence of real time communication on road traffic. The two simulators exchange commands using a TCP connection and 17

18 synchronize the movements of vehicles and locations. Therefore, commands sent from the communication model can influence vehicles behaviour (i.e. warning messages exchanged among vehicles) enabling an accurate modeling of traffic and mobility. 2.3 Dominating Set Problem The algorithm presented in this work was highly influenced and is based on the the Dominating Set problem. The Dominating Set problem, requires finding a small subset S of nodes in a graph G = (V, E) such that each node is either in S or adjacent to a node in S [20]. It is a well known fact that the Dominating Set problem is an NP problem, therefore several heuristics and distributed algorithms were suggested. In [20], a distributed randomized algorithm is presented. Each node calculates the number of its uncovered neighbors, and broadcasts that measure. The nodes with the highest degree in their neighborhood are selected as candidates. This is followed by uncovered nodes which count the number of candidates covering them, named as their support. Candidate nodes are than added to the Dominating Set with a probability inverse to the median of the support of nodes they cover. This yields a O(log ) expected size of the optimal Dominating Set, with denoting the maximal degree. The authors of [30] present a dominating set construction algorithm in Unit Disk Graphs (UDG) which takes into account message loss, due to simultaneous transmissions. The distributed algorithm requires O((log n) 2 ) time, in which Minimal Independent Set (MIS) is constructed. The authors state that the number of nodes in the Minimum Dominating set is at most 5 times the maximal number of nodes in any MIS in the case of UDG, therefore they construct a MIS in a similar way to the algorithms presented in [1, 28]. A Maximal Independent Set is a maximal set of vertices in a graph, such that no two vertices in the set are adjacent[28]. The algorithm in [28] is executed by parallel processors, each implementing a vertex. At each round, a vertex which was not selected or adjacent to a node 18

19 which was selected, is added to the MIS with probability inverse to its current degree. This requires O(log n) rounds of O(log n) time each. 2.4 Clustering and Channel Access Algorithms Clustering is a well known method used in Ad Hoc networks. The main advantages using clustering schemes are presented in [6] and include benefits in network management, it allows introducing QoS, enabling efficient resource allocation and it might reduce topology changes overhead. Some clustering algorithms include their own unique channel access method derived from the clustering process. In [6], the authors present a distributed one hop clustering algorithm for Ad-Hoc networks. They firstly introduce a channel access scheme in a slotted system with nodes contending for transmission for a partial time-fraction of the slot. The channel access scheme is then analyzed in order to decrease collision probability. The clustering algorithm is than presented; each node that doesn t belong to any cluster, sets a random period of time and broadcasts a message declaring the node as a cluster head. Neighbouring nodes that have overheard this message join that cluster. Simulation studies were conducted to show average number of cluster heads and the time required for clustering. The results show that the algorithm requires only a small constant number of slots, due to the suggested channel access scheme. In addition setup and maintenance overhead and leader (cluster head) change rate were studied as well. Some clustering algorithms [27, 3] use different metrics (e.g., id, mobility) for the cluster formation process in order to form stable clusters. In [3], MOBIC, an algorithm for cluster formation and cluster head selection, is presented. It uses a mobility metric as the basis for this process. It is stated, that mobility is the main cause for changes in clusters, therefore, the metric is constructed according to the ratio of the received power between two successive messages. An aggregated mobility metric is defined as the variance 19

20 of all set of samples among neighbouring nodes. Nodes with low aggregated mobility metric are preferable since they will suffer from fewer disconnections. The algorithm uses this metric in a similar way to the Lowest id algorithm [27]: All nodes send Hello messages, while two successive measurements for each neighbour are required for calculation of the aggregated metric. This is followed by nodes exchanging their relative mobility values and select the lowest as their cluster head. Both algorithms [27, 3] were compared in terms of cluster head changes rate. MOBIC outperformed the Lowest id algorithm since it used a preferable metric rather than lowest id. However, these two algorithms require an overhead of exchanging messages between all pairs of nodes before a decision can be taken which increases the algorithm s run time. Several works deal with extending the span of the created clusters in order to reduce their number. In [2] the authors prove that the D-Hops Dominating Set problem is NP complete. As a result, they present a heuristic for forming clusters with d as the maximal number of hops between any cluster member to its ClusterHead. The algorithm starts by flooding the maximal node id that nodes encounter. This step lasts d rounds followed by d rounds of flooding the minimal id. Converge cast process is used for cluster formation. The run-time complexity is O(d) rounds, however the analysis lacks the complexity of the shared channel access process. A simulation study was conducted in which the number of cluster heads, cluster head duration, cluster sizes and cluster member duration were studied. It is worth mentioning, that the algorithm does not run continuously, but on constant time intervals, then it sampled. MobDHop is presented in [12]. It is a mobility based clustering algorithm with a varying diameter. The diameter varies according to the group mobility pattern detected by the algorithm as extracted from RSSI measurements. Stability metric is calculated according to the variation in the received power of exchanged messages. Following the creation of one hop clusters, nodes adjacent to two clusters, initiate a merger between clusters with similar mobility patterns. 20

21 The authors of [7] modify the Lowest Id Algorithm [27] and different degree based algorithms to create k-hop clusters named K-ConID, i.e., the maximal number of hops between each cluster member to its ClusterHead is k. Nodes flood their identifiers, which consist of the number of k-degree neighbors and id, to k hops away and select the larger identifier as their ClusterHead. Ties are broken according to the lowest id. A key feature enabled by clustering is the ability to reuse resources. In [17] a slotted channel access-based scheme is presented. Nodes declare themselves as cluster heads and transmit this announcements to their direct neighbors, which join the first cluster that they heard its cluster head. The transmission of such control messages is affected by the outcome of the suggested channel access process: Each slot has several phases and a RTS/CTS mechanism for collision detection. In addition, according to their previous work, the authors mention that scheme that is based on low id is more stable and leads to less cluster head changes, than a scheme based on the degree of nodes. During the last couple of years, several researches suggested using clustering algorithms in VANETs scenarios. Such works use exact location knowledge by using devices such as GPS receivers. In [33] an algorithm that forms clusters with low relative velocity between cluster members is presented. The algorithm utilizes a GPS receiver and knowledge of future location to form stable clusters i.e., clusters that are characterized by long cluster head duration and low rate of cluster changes. The algorithm requires message exchange between each vehicle and its one hop neighbors, to allow a distributed ClusterHead selection. Such a procedure requires precious convergence time, before a sensible decision can be made. Similar approach was used in [36] which presents an algorithm that deals with frequent changes in clustering topology triggered by occasional nearing due to gatherings around traffic lights. The algorithm is beacon based and its main purpose is to extend the cluster life time which is measured by ClusterHead life time. The authors follow the general idea used in [3], however prior to any state change a set of rules must be fulfilled, to avoid frequent changes. Rules define merging of clusters only after a contention period to avoid redundant changes. The 21

22 least suitable cluster head, measured by the variance of relative location to all neighbouring nodes, drops its role to reduce the number of clusters. Another simple one hop clustering algorithm is presented in [13]. The algorithm leverages ClusterHead duration and driving direction information extracted from Hello messages in order to join stable clusters. Initial setup of clusters is made by each node declaring itself as ClusterHead, which causes slow initial clusters setup. In [23] clusters are formed between nodes with similar velocities. The protocol s control messages are piggybacked on data messages in order to avoid massive exchange of control messages such as used in [33, 27, 3], at the initial stage. The algorithm is based on selection of ClusterHead (CH) according to contention. A node that desires to send a message will broadcast the message. Neighboring nodes that aren t assigned a CH role will contend for this role, if they travel with a similar speed. The first node to declare its victory, will announce its new role and all neighboring nodes will change their CH identifier. Inter-cluster interferences are avoided by adopting CDMA that has to be coordinated only locally among neighboring clusters. In addition the suggested algorithm doesn t rely on GPS which is reflected in more robust algorithm, since it s not affected by the reception of a GPS signal. Some of the works that deal with Vehicular Ad-Hoc Networks use the clustering algorithm to improve network stability by introducing new MAC layer algorithms and protocols. In [15] a clustering scheme that enables channel access for VANETS is presented. The authors present a substitution to the MAC layer in VANETs. The algorithm calculates a weighted factor for the node s compatibility to act as a ClusterHead. The weighted factor requires the knowledge of the surrounding nodes speeds and locations. Hence, each control message exchanged between nodes includes these parameters. Messages are sent using a TDMA scheme, with the ClusterHead scheduling all nodes in the cluster. In addition, 10% of the slots are reserved for new nodes to join the cluster. Collisions at border nodes are resolved by ClusterHeads exchanging their local schemes and solving conflicts for neighboring nodes. However, this solution leads to some scaling issues. 22

23 In [37] a clustering scheme that forms clusters and enables intra-cluster communication using TDMA and inter-cluster communication using MAC is presented. The suggested scheme requires the use of two transceivers, with one used for delay sensitive communication within the cluster, while the other is used for inter-cluster data transfer. Cluster joining is done by a triple hand shake mechanism, with a new joiner, declaring itself as cluster head. Two approximate cluster heads under certain distance will be merged to one, in order to reduce the number of clusters. An analysis and study of the required parameters to allow the delivery of safety messages within their time limit is given. A similar medium access technique for clustered VANETs is presented in [31]. The suggested protocol is a hybrid protocol that uses TDMA access technique for intra-cluster communication and contention based access for inter-cluster communication. The authors adopted their protocol to the DSRC protocols and use all 7 channels specified in the DSRC standards. Unlike the authors of [37] that use a similar approach, the authors require that vehicles will be equipped with only one radio set. A delay analysis is provided for the suggested algorithm and simulation results. However the clustering algorithm is not presented but only the channel access technique. An another channel access algorithm for VANETs that does not require formation of clusters is presented in [5]. The algorithms coordinates distributed channel access in order to avoid collisions due to the hidden terminal problem which is common in wireless scenarios such as vehicular networks. The algorithm is based on R-ALOHA protocol [9] for reservations of the shared channel, in which each node reserves a slot for its own use and each frame (group of slots) is followed by an indication of each of the slots being busy or idle (each node views the network state two hops away). This algorithm allows introducing QoS, however, there is a limit on the number of vehicles in the same communication area. Enchantments to the RR-Aloha algorithm [5] are presented in [8] creating RR-Aloha+. The authors modify some of the original ideas presented in RR-Aloha to allow improved coping with hidden terminal problem and blocking of newly joined nodes. In addition the authors 23

24 propose to raise the refresh rate of allocated slots, so slots will not remain allocated in cases where nodes finish their transmissions, or left the original transmission range. The modified algorithm was simulated and compared to CSMA/CA. Main results show that this algorithm is feasible and produces better results than CSMA/CA in terms of message delay. However this is limited to only scenarios with small scale of contending nodes. An additional improvement to former protocols [5, 8] is the Mobile Slotted Aloha For VANETS presented in [32]. Since RR-Aloha+ [8] suffers from scalability issues, due to lack of resources reusing, the authors suggest creating soft clusters and limit the propagation of slot state messages, so slots can be reused in different clusters. Simulations were used to compare this protocol to RR-Aloha+, in which the improved algorithm outperformed RR-Aloha+ in terms of packet delivery rate. 24

25 Chapter 3 Wireless Vehicular Networks Simulator (WVNS) This chapter describes in detail the Wireless Vehicular Networks Simulator (WVNS) that was developed in this work. WVNS allows comprehensive study and evaluation of protocols and algorithms for Vehicular Ad-Hoc Networks prior to deployment on real devices carried by vehicles. The presented simulator, combines both traffic modeling simulator and a communication modeling simulator. Both simulators have mutual impact on each other, since traffic can be affected according to messages that can be exchanged between vehicles. On the other hand, communication is affected according to instantaneous locations of the vehicles engaged in such communication. OMNeT++ [40] simulator is used as the main simulation development environment in WVNS. OMNeT++ has a modular structure, with its cornerstone denoted as csimplemodule. Relations and connections between modules are defined within Network Description (NED) files, while the activities are defined by C++ classes. The NED files allow grouping together several simple modules to a compound cmodule, for simulating more complex be- 25

26 havior. Communication between modules is handled via messages transfer, or C++ pointers to other models. Message transfer is made via gates connecting two hosts or entities. Gates can be created statically at initialization or dynamically during run-time to allow dynamic topologies. Network topology is also defined within the NED files, and modules can be created and destroyed statically at start/finish time as defined in the NED files or dynamically during simulation runtime. All OMNeT++ classes are C++ classes derived from the csimplemodule class. Each deriving class has to overload two functions: virtual void initialize() A method that is invoked during module s startup. This function has to initialize all module s parameters and retrieve pointers to other modules. In addition the initialization can be made in stages to support more complex behaviour (see OMNeT++ user manual for further details). virtual void handlemessage (cmessage* msg) A method that handles all incoming messages - messages from other modules, and self messages used for timing. This function is usually used as an event dispatcher that calls the relevant methods of the algorithm. This chapter describes OMNeT++ modules used in our simulator, their relations and their interfaces. The simulation includes global modules for simulation management and compound modules for simulating moving entities (vehicles). As OMNeT++ basic configuration does not include any modules, the MiXiM [24] framework is used as the building blocks for developing new modules. MiXiM is a framework for simulating wireless networks. It provides detailed models of wireless channels, connectivity, mobility and MAC layer protocols for OMNeT++ simulator. The MiXiM framework is divided into two logical parts: Base framework and protocol library. The Base framework provides general functionality such as connection management, mobility and channel modelling, while the protocol library provides implementation of some known MAC layer protocols. 26

27 WVNS has a generic structure which is not limited to a specific scenario and it can be used for all general VANET simulations regardless of a specific algorithm or protocol. 3.1 Design Overview The simulator provides a tool for Vehicular Ad-Hoc Networks research; hence the main focus is on three key aspects. First, providing an accurate channel modeling according to the studied scenario and the topographic environment. Thus, several channel propagation characteristics were combined into one scenario according to the topographic environment yielding different channel models for each topographic area (urban, suburban, open space). Secondly, we focused on developing various mobility patterns based on [34]. Moreover, the simulator includes the ability to characterize different mobility features (such as different maximal speeds) for each topographic area. Finally, the simulator must include a platform for evaluating algorithms and protocols prior to deployment on their dedicated hardware. We require that the evaluated algorithm s code will not have to be modified to allow execution on a dedicated hardware for VANETs. Therefore, the simulator must comply with the services provided by the hardware. Next the simulator structure will be presented: Figure 3.1 shows the general structure of WVNS, the components and their relations. WVNS consists of three major parts: (1) OMNeT++ simulation framework; (2) Optional external mobility modules; and (3) Algorithm s code designed for real vehicular dedicated communication devices. Other modules in this simulator provide complementary roles and are used for simulation management. 3.2 Modules Details The WVNS defines a Car module which defines a node in a VANET. Each Car module consists of the following modules: 27

28 Figure 3.1: Wireless Vehicular Network Simulator (WVNS) General Structure Channel Modeling Channel modeling in a wireless network plays a major role in simulating interferences and signal propagation. Modeling such phenomenon requires understanding of signal propagation, and the fact that signal propagate differently in different topological environments such as urban, rural or open space. Due to this fact, the channel modeling has to take this into consideration when calculating channel attenuation, and combine several channel propagation characteristics into one scenario according to the topographic environment. This multi-environments feature is integrated into WVSN, to allow accurate modelling. The simulation playground is divided into 3 different topological areas (environments), each with different characteristics. The 3 current implemented environments are: Urban Dense area with high vehicular and building density. Sub-Urban Residential area with moderate vehicular and building density. Open Space Rural area with few buildings and vehicles. 28

29 Figure 3.2: Example of multi-environments configuration Each of these environments is characterized by different channel parameters, and can be configured at simulation initialization. Figure 3.2 displays an example of such configuration, in which a real map is characterized according to the 3 environments, while Figure 3.3 presents a scenario in which a vehicle transmits a message across two environments. The path loss is calculated according to the relative distances D1 and D2 each with different path loss characteristic according to the topographic environment. The channel modeling is divided into two sections presented next: Network Interface (NIC) This MiXiM based module is used to simulate Network Interface Card (NIC), of a VANET dedicated hardware. It doesn t have any C++ class associated with it; it s used as a structure for holding other simple modules: MAC Layer and Physical Layer. The Medium Access Control (MAC) module is used to allow access to the shared channel according to the simulated system. It also provides basic handling of lower layer messages and frame queuing capabilities. In addition, it schedules message sending only at discrete time points i.e., the 29

30 Figure 3.3: Path loss calculation for two environments beginning of time slots - to allow granularity of transmission. The Physical Layer is the core part of the wireless node in WVNS. It is responsible for sending and receiving messages and models the physical channel (attenuation and Bit Error Rate - BER). The Physical Layer module contains three main components as it was derived from the MiXiM frame work: 1. The PhyLayer OMNeT++ module itself, providing interface to upper layer and the channel. 2. AnalogueModel used as attenuation filter of the physical channel. 3. Decider which calculates bit error rate (BER) and packet error rate (PER), makes a decision rather a message was received correctly or has to be discarded. The PhyLayer OMNeT++ module provides an interface to the MAC layer. It is directly connected to the MAC layer via OMNeT++ gates. In addition, the PhyLayer module is capable of sending messages to other physical layers. Two important entities are encapsulated by this module; the AnalogueModel is used to calculate attenuation of the received message according to the traversed path the message has passed, hence it allows calculating path loss and shadowing of the signal by obstacles. The path loss is calculated separately for 30

31 each traversed region according to it s specific parameters and then summed to describe the accurate received signal strength. In a similar manner the shadowing is calculated separately in each region and its maximal value is selected. The topographical data is stored in a sectors database to allow all physical layers modules to access that data when calculating the received signal strength. The Decider is used to determine whether the received signal is a receivable message, a noise or an interference. It calculates BER and PER for the receivable messages according to the used transmission scheme. In addition it holds parameters such as, thermal noise value, sensitivity for receiving messages and information regarding all messages currently transmitted on the channel to calculate interferences to the referenced message. Connection Manager This module is used for establishing and tear down of connections dynamically according to the updated location of nodes. A connection is maintained among two nodes which are within each others maximal interference distance. The maximal interference distance is P calculated according to Friis transmission equation which is given by: r ( P t = G t G λ 2. r 4πR) Where P r and P t are the received and transmitted power respectively, G t and G r are the antenna gains of the transmitting and receiving nodes respectively, λ is the wave length in meters and R is the distance between the nodes measured in meters. This is used to reduce computational time and resources for messages transmission process, since it s not required to connect nodes that their distance exceeds the distance encountered by this equation (given a minimal power value required for message reception). It is worth mentioning that current location of nodes must be tracked to allow this calculation. 31

32 3.2.2 Mobility Patterns Mobility pattern plays a major role in every Mobile Ad-Hoc Network (MANET) in general, or Vehicular Ad-Hoc Network (VANET) in particular. Unlike MANETs simulations in which the mobility patterns can be simulated by a random or mathematical model, VANETs simulations require more realistic mobility patterns for better understanding of their unique characteristics as presented in [34]. Among more realistic mobility patterns one might find traces of vehicular movements derived from real vehicles and coupling of traffic engineering simulators into the communication simulators. As a result the presented simulator is not limited to one type of mobility pattern and has a generic structure to allow it using all the mentioned above mobility patterns. In addition, different vehicles may have different mobility patterns and can be forced to adjust their velocity or other parameters according to the topological region they are driving through. This corresponds to the regions defined in 3.2.1, since mobility parameters can be characterized according to three regions: urban, sub-urban and open space. The WVNS mobility pattern implementation is divided into two parts: Mobility Manger (MM)- This module is responsible for general mobility issues. The simulation holds one instantiation of this module, for the whole simulation. The MM can operate according to two operation approaches: (1). Centralized - Describes a more centralized approach in which hosts are forced to set their next position according to an algorithm executed at the MM model. External components such as mobility traces, or TraCI - interface to traffic simulation tool SUMO [35] are also included in this category. Movement of vehicles in the road traffic simulator SUMO is reflected in movement of nodes in the OMNeT++ simulation. Nodes can then interact with the running road traffic simulation, in order to simulate the influence of Inter Vehicle Communication on road traffic; (2) Distributed - Describes a distributed approach in which the MM receives position updates from mobility client modules, which implement their own 32

33 mobility pattern. Mobility Client - This module is an autonomous component within the Car module, therefore it does not couple with the network layers modules. Its main purpose is to define the vehicle s movement pattern. It can be done independently or by retrieving the next location from the mobility manager. Since OMNeT++ is a discrete time simulator, the location updates can be done in discrete time points. An adjustment in the time interval is available to adjust the trade off between accuracy and computational complexity Algorithm Module The main purpose of WVNS is to study and evaluate the algorithms and protocols developed in this work. The implementation is designed in a way that allows executing the same code written in C language, both on a PC within OMNeT++ simulator and on a real dedicated hardware carried by vehicles. The algorithm module wraps the core functions used for each studied algorithm and provide all the required services for the algorithm as it was provided by the real device, such as: Timers and interrupts. Memory allocation of buffers for Tx and Rx. Message sending and receiving. In addition the module will have to dispatch messages according to the correct protocol and invoke the correct methods. Each algorithm has its own set of methods implemented in C code. Each host has its own identical set of functions and its unique database. In order to allow the code to use the services provided by the OMNeT++ simulator (as a replacement to the services provided by the real device), an API between the simulator and the code is 33

34 1 -cback OmnetAlgorithmModule -cback : ccallback +handlemessage() +sendmessage(out msg) +settimer() +.() +.() +.() * ccallback -parent 1 -parent : OmnetAlgorithmModule -contextptr : ccallback +API functions() * -Setting context pointer -Function Call * -Service API function * -DataBase +API functions() +Core functions() Code Figure 3.4: The API between OMNeT++ and the implemented code of the algorithm defined and implemented. This module is also used to collect algorithm specific statistics about relevant parameters to allow comprehensive study of the examined algorithm. Figure 3.4 presents the mechanism used for this feature. 34

35 Chapter 4 Hierarchical Clustering Algorithm for Vehicular Ad-Hoc Networks A shared channel access protocol is required to allow vehicles to communicate via wireless channel. VANETs specific feature, such as requirements to fulfill certain QoS, must be taken into consideration when developing such a protocol. In addition, the suggested solution must not rely on wireless access-points or on any other infrastructure devices. Since random access cannot guarantee a tight bound to the delivery of warning messages, we selected to approach this problem differently. Using Clustering approach can be beneficial [6] since it allows to divide the network into smaller sections (Clusters). In each of this clusters, a dynamic synchronization entity is required to synchronize and schedule channel access and allocate bandwidth to communicating nodes. Such entity can be referred as a Leader or alternatively a ClusterHead and both terms will be used further in this work. The suggested algorithm is a distributed randomized two hops clustering algorithm (i.e., the maximal number of hops between any node to a ClusterHead is two). Our basic approach was to create a distributed randomized Dominating Set in G 2. The initial creation of the Dominating Set consists of the following 3 phases: 1) Fast randomized selection of 35

36 assisting nodes named ClusterRelays, 2) Fast randomized selection of a relatively central ClusterHead, and 3) Topology learning and transmission scheduling within the cluster. This is followed by a phase aimed at adapting the cluster to topology changes caused by the mobility of nodes. The quick nature of the proposed algorithm is analytically analyzed, and the results are presented in this chapter. 4.1 Problem Definitions First some notions and definitions which will be used throughout this chapter are introduced: G 2 of a graph G = (V, E) is obtained by adding an edge (u, v) E if node u is at most 2 hops away from v in G. A Dominating Set (DS) of a graph G = (V, E) is a subset V V such that each node in V \ V is adjacent to some node in V. A ClusterHead is a node that belongs to the Dominating Set. It manages and handles the channel access method. Cluster - Each node has a single ClusterHead. A cluster is a subset of nodes with the same ClusterHead. ch v and slot v are respectively the ClusterHead of v and the time slot in which v transmits, for v V. Next, we formulate the problem of creating clusters in which each node is at most 2 hops from a ClusterHead and scheduling them (i.e., nodes that belong to the same cluster have different transmission time slots) as Distributed G 2 Dominating set and Scheduling problem named DG 2 DS. Definition: The DG 2 DS Problem: 36

37 Instance: Undirected graph G = (V, E) Solution: 1. A Dominating Set of G 2, i.e., a subset V V such that each vertex is either in V or has a maximal two hops path to at least one of the vertices in V. The nodes within the Dominating Set will be referred as ClusterHeads. 2. Each node u V has a single ClusterHead - ch u V 3. Each node is allocated a unique transmission slot in its cluster i.e., v, u V such that if ch v = ch u than slot v slot u The solution is measured in three aspects: the size of the Dominating Set, the time and the number of exchanged messages required to form clusters and assign a unique time slot for each of the nodes. There is obviously a tradeoff between the time and size metrics. For example a solution can be a Set which is composed of all the nodes in the network, which can be obtained immediately. However it is not an acceptable result and we would like to minimize the size of the Dominating Set. We take a randomized approach similar to [1], which allows us to benefit both from small Dominating Set and fast convergence. Next is presented a solution to the DG 2 DS problem: 4.2 Hierarchical Clustering Algorithm (HCA) Next we present a distributed randomized solution to the DG 2 DS problem: Hierarchical Clustering Algorithm (HCA). The algorithm was influenced by the work of [1], in which an Independent Set is constructed in a randomized manner. In [1], a node is selected and removed with its immediate neighbors from the graph at each phase, until all the nodes are either in the Independent Set or adjacent to any of the nodes in it. 37

38 The HCA forms TDMA-like synchronized clusters. In order to eliminate collisions by simultaneous data transmissions in the same cluster, transmissions are only allowed on unique slots as assigned by the ClusterHead. The algorithm is comprised of 4 phases. The first 3 phases refer to a static scenario in which the clusters are formed, while the fourth phase handles topology changes caused by the mobility of nodes. The nodes are not required to travel according to a particular mobility pattern, and are allowed to move freely. Therefore, the algorithm can be used in VANETs or in any other Mobile Ad-Hoc network that consists of mobile entities (such as networks used for military or search and rescue purposes). It is worth mentioning that the suggested algorithm differs from the algorithms that were presented in chapter 2 in two key aspects. First this algorithm handles the channel access and does not assume any lower layer connectivity. Secondly, it does not require the knowledge of the nodes location. This feature enhances the algorithm s robustness since it does not rely on localization systems (e.g., GPS), which sometimes is preferable. We overcome the lack of information regarding the nodes location by inferring connectivity from received messages. To ease the formation of clusters, in which the maximal distance from a ClusterHead to any other node in its cluster is two, 3 hierarchies are defined. Each hierarchy is characterized by a role carried out by the nodes. The Slave role is executed by regular nodes that are at most two hops away from a ClusterHead. Such nodes are at the lowest hierarchy of the cluster, and it is the initial role of all nodes executing this algorithm. The ClusterRelay (CR) role refers to nodes that relay messages from the ClusterHead to the Slaves, and help extend the reach of nodes within the cluster. ClusterRelay nodes are used to forward the control messages in both directions: from ClusterHeads to Slaves and vice versa. The ClusterHead role refers to nodes that manage and synchronize the shared channel access for all the other nodes in the formed cluster (a ClusterHead is a member of the Dominating Set of G 2 as defined in section 4.1). Such nodes are in the highest hierarchy of the algorithm. The state machine of possible roles and transitions among them is presented in Fig The numbers in this figure denote the relevant Procedure and line associated to the transition. 38

39 Figure 4.1: HCA state machine of role transitions Some control messages are required for HCA s execution - SYNC and ACK. Each message contains the following data: Source MAC address. Destination MAC address. Sender s ClusterRelay MAC address. Sender s ClusterHead MAC address. Message type. Node s role. Maximal distance within the cluster measured in hops. SYNC messages are synchronization messages that are created by ClusterHeads and used to assign slots to cluster member nodes. These messages are sent from ClusterHeads to Slave nodes and are forwarded by ClusterRelays. This type of scheduling is required in order to eliminate the collisions caused by simultaneous transmissions in the same cluster. The 39

40 Figure 4.2: SYNC message slots scheme message contains some unassigned slots for nodes to select randomly and join the cluster. Fig. 4.2 shows a possible assignment of slots as carried by a SYNC message and includes the suggested scheduling scheme. The first slot is reserved for the ClusterHead. Next, slots are allocated for each ClusterRelay, followed by slots which are assigned for Slave nodes. Additional m empty slots are reserved for random access of new members, with m denoting the maximal cluster size. The slots schemes completes with slots which are designated for ACKs sent by ClusterRelays ACK messages are messages required to acknowledge reception of SYNC messages. These messages are sent from Slave nodes to the ClusterHeads. This is done in order to allow topology learning by the ClusterHead. Slaves reply with ACK messages on their assigned slot to avoid collisions with other nodes within the cluster. ClusterRelays collect these ACKs and forward an accumulated ACK to their ClusterHead, with the identifiers of these nodes. Additionally, we define a round as the time period between two successive SYNC messages sent by the same ClusterHead. The execution of the proposed algorithm can be divided into the following four phases, which are executed asynchronously: (1) ClusterRelays Selection (2) ClusterHead Selection (3) Cluster Formation and Scheduling (4) Cluster Maintenance 40

41 The first three phases refer to the initial phase of the algorithm in which the DG 2 DS is created, while the fourth phase handles topology changes and maintenance tasks, required to handle mobility of nodes in such dynamic networks. Each node stores the following parameters: ch v, cr v, role v and N(v) which refers to the node s ClusterHead, ClusterRelay, role and neighbors list respectively. Additionally, each node has a parameter d max which denotes its current maximal distance in hops within the cluster. This measure is used to determine whether cluster merging is advisable without the disconnection of other nodes ClusterRelays Selection Phase In the ClusterRelays Selection phase (described in Procedure 1 below) ClusterRelays that help with the cluster formation are selected. Initially, each node draws a random slot of 2m slots which would end its listening period, with m denoting the maximal number of nodes in a cluster. If a node has not received a message before its listening period elapses, it will announce itself as a ClusterRelay. Otherwise, (i.e., it is in the range of other ClusterRelay) it will become a Slave ClusterHead Selection Phase In the ClusterHead Selection phase (as described in Procedure 2 below), a relatively central ClusterHead is selected. A Slave node which has heard a SYNC message from at least two ClusterRelays will announce itself as a candidate to become a ClusterHead. Such nodes set up a random period of time, drawn uniformly from a parameter set to 2m slots, in which they listen for incoming messages from other ClusterHead candidates. At the end of this period, candidates which did not receive a SYNC message will send a message declaring themselves as ClusterHeads. The remaining candidates, if any, will drop their nomination and remain Slave nodes. In case there is only one ClusterRelay, and neither of the Slaves is in range of two ClusterRelays, the ClusterRelay will become a ClusterHead by itself, hence forming a 41

42 Procedure 1 ClusterRelay Selection 1: role v 2: cr v 3: ch v 4: Select random slot from 2m slots for contention period On reception of SYNC from u such that role u = CR 5: role v Slave 6: cr v u 7: go to ClusterHead Selection On Timeout 8: role v CR 9: cr v v 10: Send SYNC message 11: Set timer to become ClusterHead 12: go to ClusterHead Selection 42

43 one hop cluster. Procedure 2 ClusterHead Selection On Reception of two SYNC messages from two ClusterRelays 1: if role v = Slave then 2: Select random slot of 2m slots for contention period and wait for timeout On Reception of SYNC from ClusterHead 3: Cancel contention period 4: go to Cluster Formation and Scheduling On Timeout 5: role v CH 6: ch v v 7: Send SYNC 8: go to Cluster Formation and Scheduling Cluster Formation and Scheduling Phase The Cluster Formation and Scheduling Phase (Procedure 3) is used for cluster formation and topology discovery by the ClusterHead and other nodes. The elected ClusterHead is not familiar with surrounding nodes, other than the neighboring ClusterRelays. Therefore it broadcasts a SYNC messages to the ClusterRelays such that each ClusterRelay is assigned two unique slots - one for relaying the SYNC and the other for ACK transmission at the end of the round. ClusterRelays rebroadcast these messages to their neighboring Slaves. A Slave node that receives a SYNC message from its ClusterHead, sends an ACK message in the slot which was assigned by the ClusterHead in the SYNC message. Additionally, SYNC messages contain m unassigned slots for new nodes to randomly select a slot and join the cluster. If 43

44 the node has not yet been assigned a slot, meaning it is not a member of the cluster yet, it will select an unassigned slot randomly and send an ACK to inform of its joining. In each round, the ClusterRelays add the identifiers of Slave nodes, as extracted from ACK messages, and stores them in its neighbors list. At the end of each round, an accumulated ACK is sent on behalf of all Slave nodes to the ClusterHead, in order to allow learning of the exact topology and assigning slots for the following round. Since nodes select a random slot for their first ACK, some collisions may occur. In case of such a collision, all colliding nodes will not receive a slot, and therefore they will contend in the next round. An analytical analysis (described in section 4.3) guarantees a bounded limit of rounds for this process. The process continues until all nodes obtain their slot and reply with an ACK Cluster Maintenance Phase Finally, the Cluster Maintenance phase (Procedures 4 and 5) is used to maintain up to date clusters and reduce the number of redundant clusters by merging relatively small clusters. Furthermore, it is required for forming new clusters due to topological changes. This phase is executed in parallel to phase 3 of the HCA algorithm. In order to prevent the propagation of outdated information in the network, each node executes an aging process to eliminate irrelevant entries from the neighbors list. A Slave node that did not receive SYNC message from its ClusterRelay during an aging period will erase its ClusterRelay identifier and look for neighboring ClusterRelays within the same cluster. As a last resort, the Slave node will start its random listening period in the Initial State in order to join a different cluster. All timeout periods were chosen after an extensive simulation study to allow optimal performances. In order to prevent an unlimited drift of ClusterHeads and the creation of redundant clusters, some rules for clusters merging are defined. A rule of thumb for these role changes is to avoid role transitions that might cause disconnection of other nodes. First, a ClusterHead 44

45 Procedure 3 Cluster Formation and Scheduling 1: if role v = CH then 2: Periodically (with addition of some random slots) send SYNC message with unique slot for each of the known nodes with m unassigned slots On reception of SYNC from u 3: if role v = CR and u = CR v then 4: if SYNC contains assigned slot for v then 5: ch v u 6: Broadcast SYNC from u on the assigned slot 7: else 8: Send ACK on a random slot 9: if role v = Slave and role u = CR then 10: if SYNC contains assigned slot for v then 11: ch v ch u 12: Reply an ACK on the assigned slot 13: else 14: Select a random unassigned slot from the SYNC message and send ACK On reception of ACK from u 15: if role u = Slave and role v = CR then 16: Store the ACK and send ACK on behalf of all nodes in the end of the round 17: if role u = CR and role v = CH then 18: Reserve a slot for all the nodes which are present in the ACK On Aging Timeout 19: role v = CH 45

46 Procedure 4 Cluster Maintenance Phase - Part 1 On Reception of SYNC from u 1: if role v = CH then 2: if role u = CH and d max < 2 then 3: role v CR 4: ch v ch u 5: cr v v 6: if role u = CR, ch u v and d max < 1 then 7: role v Slave 8: ch v ch u 9: cr v u 10: if role v = CR then 11: Forward SYNC 12: if role v = Slave then 13: if SYNC contains assigned slot for v then 14: Reply an ACK on the assigned slot 15: ch v ch u 16: else 17: Select a random slot of the unassigned slots and send ACK 46

47 that was abandoned by all of its members, and encounters a ClusterRelay associated with another cluster, will be merged into the other cluster. Second, if a ClusterHead of a cluster with maximal distance within the cluster of one hop encounters another ClusterHead, the first ClusterHead will change its role to ClusterRelay and switch clusters together with all of its nodes. Additionally, a ClusterRelay that did not receive a SYNC message from its Cluster- Head will try to join another cluster. If this did not succeed it will become a ClusterHead by itself. In addition, Slave node will send their ACK messages to the nearest ClusterRelay within the same cluster to allow smooth transition of Slaves to other ClusterRelays in the same cluster, according to topology changes caused by mobility. Procedure 5 Cluster Maintenance Phase - Part 2 On Aging Timeout 1: if role v = CR then 2: if v is neighbor of u s.t role u = CH then 3: ch v u 4: else 5: role v CH 6: ch v v 7: if role v = Slave then 8: cr v 9: if u N(v) s.t role u = CR and ch u = ch v then 10: cr v u 11: else 12: Start over HCA The main feature of this algorithm is its quick clusters setup. This allows quick node synchronization by the ClusterHeads in order to enable a reliable communication. Next we present a run time analysis of HCA s first three phases, and a proof that it solves the DG 2 DS 47

48 problem. In addition, a message complexity analysis is provided as well. 4.3 Time and Message Complexity This section describes the key results of the analysis of HCA s first three phases (ClusterRelay selection; ClusterHead selection; and Cluster Formation and scheduling), in terms of time and messages complexity. These three phases are assumed to take place while the nodes are static. Additionally, we assume that messages that are sent between nodes which are within each others hearing range will be received correctly, unless a simultaneous transmission occurs. Furthermore, we require that each node in the network will be represented by a unique identifier, and that each node in G 2 has a maximal degree of m, i.e., the maximal size of a cluster in G is m. Moreover we assume that phases are executed synchronously i.e., all nodes finish executing the same phase prior to starting the execution of the following phase Correctness The following Theorem proves the correctness of the algorithm: Theorem 4.1 Upon termination of the third phase, the Hierarchical Clustering Algorithm solves the DG 2 DS problem, i.e., 1. The ClusterHeads form a Dominating Set in G 2 2. Each node has one ClusterHead 3. Each node has a unique slot for transmission within the cluster Proof. In order to prove that the ClusterHeads form a Dominating Set in G 2, it will be shown that by the end of the second phase, each node is either a ClusterHead or within two hops from at least one of the ClusterHeads. Upon termination of the first round, 48

49 all the nodes will either transmit a SYNC message (line 10, procedure 1), or will receive a SYNC message from at least one of their neighboring ClusterRelays (line 5, procedure 1). Since we assume that the phases are synchronous, by the beginning of the second phase, all nodes will be either ClusterRelays or adjacent to at least one ClusterRelay. In the second phase, ClusterHeads are selected either by transmitting the first SYNC in their neighborhood after their contention period had elapsed (line 2, procedure 2), or by nodes which didn t receive a message from a ClusterHead (line 5, procedure 2). Therefore, by the end of the second phase, all ClusterRelays will be adjacent to at least one ClusterHead, or have become ClusterHead by themselves. Since that all Slaves are adjacent to ClusterRelays and all ClusterRelays are adjacent to ClusterHeads, all nodes will be within two hops from the ClusterHeads. During the third phase, nodes are static and do not engage in role changing other than to ClusterHead, therefore it can be concluded that by the end of the third phase the ClusterHeads form a Dominating Set in G 2. The second and third parts of this proof will be jointly proved, since nodes select their ClusterHead immediately after the reception of their unique slot (lines 6 and 13, procedure 3). Line 2 in procedure 3 states that the ClusterHead sends a SYNC message with assigned slots to all the known nodes within the cluster, with each node assigned a unique slot. Each node that receives this message (lines 5, 11, procedure 3) stores the assigned slot and the ClusterHead identifier. Additionally, a timer expiration, will trigger role changing of nodes that cannot receive the SYNC message correctly, therefore, such nodes will become ClusterHeads by themselves and will obtain their slot and ClusterHead immediately. This concludes the correctness Theorem Run-time Complexity Since HCA is a randomized algorithm an analysis of its run-time complexity with high probability 1 is provided. 49

50 Theorem 4.2 The first three phases of the Hierarchical Clustering Algorithm, will terminate with high probability 1 in O(m log log m) steps where m denotes the maximal size of a cluster. In order to prove theorem 4.2 we ll analyze the running time of each of the algorithm s phases while executed in a single cluster and their failure probability. The Balls and Bins model, as presented in [29] will be used to calculate the probability of failure in the first two phases. Failure probability is referred as the probability that neither of the transmitting nodes had succeeded in transmitting in a unique slot during the first t slots. The Balls and Bins model referrers to a model in which m balls are thrown with uniform distribution into n different bins. During the algorithm s analyses each slot will be referred as a bin and a transmission would be referred as a ball. In order to simplify calculations, we would like to assume that the number of balls in each bin is an independent variable, which would enable us to find an upper bound for the exact case probability. Using Poisson approximation for relatively rare events can simplify calculation and an upper bound can be derived for the failure probability. Therefore, the number of balls in each bin is Poisson distributed with mean m, such that the expected total number of balls in all the bins is n m. In order to assume so, the authors of [29, Corollary 5.9] found an upper bound to the probability of an event E in the exact scenario (Balls and Bins) denoted as P [E BB ] using Poisson approximation: P [E BB ] e m P [E P oisson ] (4.1) with P [E P oisson ] denoting the probability of event E in the Poisson approximation case. Let x i be a Poisson distributed random variable with parameter λ = m, which denotes n the number of balls in bin i, while the expected number of balls in all the bins is n. The probability that a bin contains exactly one ball is denoted by p 1 : p 1 = P [x i = 1] = e λ λ (4.2) 1 Event E m occurs with high probability if probability P [E m ] is such that lim m P [E m ] = 1 50

51 Since x 1, x 2... x n are independent Poisson variables the probability that neither of the first t bins contain one ball is given by P [t]: P [t] = (1 e λ λ) t (4.3) Claim 4.1 Let z be a random variable which denotes the index of the first slot with exactly one ball, the distribution of z is given according to Geometric distribution: (1 e λ λ) i 1 (e λ λ) i 1 P [z = i] = 0 otherwise Thus the expected number of required slots before a successful slot is given by (4.4) E [z] = 1 p 1 = eλ λ (4.5) Lemma 4.1 The first phase of the Hierarchical Clustering Algorithm (selection of Cluster- Relays) requires O(log 4 m) slots w.h.p. This phase terminates when all nodes in the network 3 are ClusterRelays or within one hop from a ClusterRelay. Proof. To prove this lemma we will bound the probability of failure denoted as P [failure]. Failure occurs when all the contending nodes collide during the first t = c 1 log 4 m slots, i.e., 3 neither of the first t slots contain only one transmission, with t denoting the time required for this phase in slots and c is a constant. Setting n as 2m yields λ = 1/2, and t in equation 4.3 yields: [ ] ( P [failure] = P t = c 1 log 4 m = e) c1 log 43 m ( 3 c1 log 43 m ( 1 ) c1 = (4.6) 4) m Next P [failure1], the probability that the first phase would fail, will be found by finding an upper bound to the exact case (Balls and Bins). Using equation 4.1 yields: P [failure1] e [ ] m P t = c 1 log 4 m = e ( 1 ) c1 ( 1 ) c2 m = e (4.7) 3 m m 51

52 This means that at least one node will successfully select a unique slot among the first c 1 log 4 m slots and declare itself as a ClusterRelay w.h.p. All its direct neighbors will receive 3 its message and terminate the execution of this phase. This implies that the first phase will terminate in c 1 log 4 m slots, while the expected number of required slots is given by equation 3 4.5: E [FirstSlot] = eλ λ = 2 e (4.8) Lemma 4.2 The second phase of the Hierarchical Clustering Algorithm (selection of a ClusterHead) requires O(m) time slots w.h.p to complete. This phase terminates when all the nodes in the cluster are within two hops from a ClusterHead. Proof. In order to prove lemma 4.2, we will bound the probability of failure in this phase - P [failure2]. During the second phase Slave nodes that have heard at least two ClusterRelays (will be referred as candidates) will propose themselves to act as a Cluster- Head. This is done by selecting a slot among the first 2m slots and transmitting during that slot. Since there are only 2m slots to select from, failure will occur only if all the candidates will collide during their proposal, i.e, non of the other nodes in the cluster can receive their proposal. Since the number of candidates k is unknown (it varies from 1 to m 2 nodes) the analyses will be divided into two cases: 1) log m k m 2 and 2) 1 k < log m. For values of k that are larger than log m a similar approach to the one used in lemma 4.1 will be used. The analyses will consider the worst case (highest probability of failure) i.e,. k = log m (since it is the highest number of contending nodes in this case), hence λ = log m m. Using equation 4.3 with t = c 1 m with c (to allow the required failure probability) denoting a constant, (since a failure would occur if the first 2m slots will not contain a single 52

53 nodes transmission) yields: Hence P [t = c 1 m] = ( 1 1 ) c1 m ( = e λ e λ λ 1 λ ) c1 m 1 eλ 1λ e λ 1 λ ( 1 c1 m e e) λ λ (4.9) P [t = c 1 m] ( 1 e) c1 m e log m m log m m ( 1 m) c1 e log m m ( 1 m ) c1 (4.10) Finding an upper bound to the exact case P [failure2] yields: P [failure2] e ( 1 ) c2 m P [t = c 1 m] (4.11) m For small values of k i.e., 1 < k log m, a different technique would be used. We ll look from the nodes perspective in the worst case i.e., the number of contending nodes is log m. We ll define 3 mutually exclusive events that relate to the number of transmissions in each slot and analyze their occurrence probability. e 1 At least one slot with at least 3 transmissions e 2 All slots with transmissions contain exactly 2 transmissions e 3 At least one slot with 1 transmission, given that the maximal load is at most 2 The event of failure F (There is not a single slot with a single transmission) can be expressed by F e 3 = therefore F e 1 e 2 and it s obvious that P [failure2] = P [F ] P [e 1 ]+P [e 2 ]. Therefore P [e 1 ] and P [e 2 ] will be presented. P [e 1 ] is first introduced: The probability that 1 3 nodes select the same slot is given by m. The number of distinct sets that contain 3 m 3 nodes that choose from log m slots is ( ) log m 3. Therefore: ( 1 ) ( ) log m ( log 3 m ) P [e 1 ] O (4.12) m 2 3 m 2 Next P [e 2 ] will be derived. First we ll define a Pair Set; A Pair Set S is a set of unordered 53

54 pairs of numbers. Given S, the probability that the set will be ordered in some order is given by: P [S] = 1 1 ( m 1 m m 1 ) ( m 2 m m 1 ) ( m log m... 1 ) m m m (4.13) This is obtained by the probability that each odd indexed node selects a slot that wasn t already selected and that the following node selects the same slot as the previous one. Next the number of distinct Pair Sets will be found. There are log m! permutations of log m nodes and since the order between the pairs themselves is not important this is divided by log m 2!. The number of permutations among two nodes given log m 2 pairs is 2 log m 2. Therefore the number of distinct Pair Sets is: log m! 2 ( ) (4.14) log m! 2 log m 2 From equations 4.13 (probability per a Pair Set) and 4.14 (number of Pair Sets) an upper bound can be derived: P [e 2 ] = ( 1 ) 2 log m log m! P [s] (m 1)(m 2)... (m log m) m s S 2 log m 2 ( ) log m! ( 1 2 log m P [e 2 ] m log m) m P [e 2 ] ( 1 m) log m log m log m 2 2 log m 2 log m (log m 1)... ( log m 2 + 1) 2 log m 2 ( 1 ) log m ( log m m 2 ) log m 2 2 ( 1 ) log m P [e 2 ] m ( m 2 ) log m 2 1 m (4.15) From equations 4.12 and 4.15: ( ) 1 P [failure2] P [e 1 ] + P [e 2 ] O m (4.16) 54

55 It can be seen that the failure probability, P [failure2] in scenarios that the number of contending nodes, k, is smaller than log m, is relatively small. Hence it can be shown w.h.p that there will be at least one slot with a single transmission. Remark: In case that all nodes are within the hearing range of the first woken node (ClusterRelay), neither of the neighboring nodes will select a slot (hence k = 0). In that case the first node will become a ClusterHead in O(m) time slots. Lemma 4.3 The third phase of the Hierarchical Clustering Algorithm (Cluster Formation and Scheduling Phase) requires less than O(m log log m) slots with high probability. This phase terminates when all the nodes in the cluster were assigned a unique slot by the Cluster- Head. Proof. The third phase analyses is based on work made by Grable and Panconesi [14]. The authors present a simple distributed edge colouring algorithm and its running time analysis. The algorithm starts when each edge is uncolored, and the algorithm terminates when all edges are colored and there are no two adjacent edges colored with the same color. At each stage, each edge selects independently with uniform distribution a color of its current colors palette. Followed by each stage, messages are exchanged by edges, and colors of adjacent edges are removed from the color palette. If conflicts arise, the edges will contend in the next round for another color. The authors state that If every edge e s initial palette has a 0 (e) = Ω(n c/ log log n ) colours then, the algorithm colours the graph within O((1 + c 1 ) log log m) rounds [14]. The authors show that the probability to success by selecting a unique color increases by a double exponential rate as the algorithm advances. Therefore, the algorithm requires O((1+c 1 ) log log m) rounds of communication, while the initial palette size is Ω(m c/ log log m ), with m denoting the number of nodes in a graph and c denoting a constant. The authors show that the algorithm fails with probability o(1). 55

56 The Hierarchical Clustering Algorithm s third phase sets a unique transmission slot for each of the nodes in the cluster. In each round, nodes that were not assigned a slot in previous rounds, contend over a pool of m unassigned slots. Each node that selects a unique slot, gains channel access in this slot, and therefore will not contend in the successive round. However, nodes that collided, will contend again in the following round in order to gain their unique slot. This can be presented as finding an edge colouring setup in a clique with nodes denoting the edges in the graph and slots denoting the colors palette. Therefore, c will be set to log log m (initial palette of Ω(m) colors) yielding a running time of O(log log m) rounds of communication. Each round of HCA requires m free slots, in addition to the already allocated slots which can be up to m slots as well, summing up for a total of O(m log log m) time slots. Proof. (Theorem 4.2) Since each of all three phases of the Hierarchical Clustering Algorithm requires less than O(m log log m) slots, it can be concluded that the algorithms run-time is O(m log log m) with high probability. It is worth mentioning that, since the algorithm does not assume any previous knowledge regarding neighboring nodes such process must be present. Therefore, the process of neighborhood learning requires at least m slots for transmissions Message Complexity Next we will analyze the number of messages sent in a single cluster that contains m nodes, during the algorithm s first three phases. Theorem 4.3 HCA requires O(m) messages, with high probability as 1 1, with m denoting m the number of nodes in the cluster. Proof. First, one can see that during the first and second phases of HCA s execution each node will transmit only once or not at all, therefore only O(m) messages will be sent. Next we analyze the number of transmitted messages during the third phase. The number of 56

57 transmitted messages during this phase consists of one SYNC messages sent at each round, and ACK messages sent by contending nodes in order to gain a slot. The number of SYNC messages is derived from Lemma 4.3: at each round a single SYNC message is sent by the ClusterHead, followed by a small number of rebroadcasts by the ClusterRelays 2. Therefore the number of sent SYNC messages is O(m). The number of sent ACK messages is altered by the number of collisions caused by simultaneous transmissions, i.e., Slave nodes that collided during the previous round contend over m slots at each round. Using a Poisson approximation technique, in a similar way that was used in section (equation 4.1), allows us to find an upper bound to the probability that more than O(m) messages were sent. Let x j be a Poisson random variable which denotes the number of transmissions in slot j. During the first round, m nodes contend over m slots, therefore we set n = m yielding λ = 1 and the probability that a given slot contains one transmission is given by: p 1 = P [x j = 1] = e λ λ = 1 e (4.17) Let I xj be: 1 x j = 1 I xj = 0 else And S be the number of slots occupied with a single transmission at the end of the first round, therefore S = n j=1 I x j. Next we define an event E i as the event of failure in round i, i.e., less than 1/4 of the contending nodes have succeeded to uniquely transmit in a slot, if more than one node has contended in that round. 2 We assume that the number of ClusterRelays is smaller than m log log m 57

58 Using Chernoff bounds as presented in [29, Theorem 4.5], we bound the probability that the sum of Poisson trials S deviate from its mean, E [S] = µ: P [S (1 δ)µ] e µ δ2 /2 (4.18) Next we set µ = n p 1 (according to Binomial distribution), and δ = 1 e 4 shown that at least 1 4 probability in the first round is given by: so it can be of transmissions will succeed in the first round, therefore the failure [ P [E i ] = P S m ] e µ (1 e 4 ) 2 2 = e m (1 4 e )2 2 e = e m c (4.19) 4 With c denoting (1 e 4) 2 2 e. The number of contending nodes will decrease at each consecutive round while the number of slots remains m, hence the probability of successful transmission will only increase. Therefore a union bound to the probability that a failure will occur in any of the first k rounds, with k denoting the last round (i.e., only 1 node contending 3 ) is given by: P [E 1 E 2... E k ] k P [E i ] k E 1 = k e m c log log m e m c (4.20) i=1 And an upper bound to the exact probability is given by P [slots > O(m)] e m log log m e m c 1 m (4.21) with m. The worst case states that 3/4 of the contending nodes in round i will have to transmit again in round i + 1. Therefore the total amount of ACK transmissions caused by contending nodes is bounded by: m k ( 3 i ( 3 ) i m = 4m = O(m) (4.22) 4) 4 i=0 i=0 3 k is derived from lemma

59 Summing all, yields that the total number of messages is O(m). scenarios. Next we present an extensive simulation study of all four phases of HCA in different 59

60 Chapter 5 Simulation Results The hierarchical Clustering Algorithm was evaluated using the Wireless Vehicular Networks Simulator which was presented in chapter 3. In this chapter, the scenarios that were used to evaluate the algorithm s performance and the results that were obtained are presented. Various static and dynamic scenarios were considered for measuring the performance of the algorithm. The static scenario was used to evaluate the convergence process, i.e., Cluster- Head selection, and the cluster formation in aspects of the number of exchanged messages, the time complexity and the number of formed clusters. The dynamic scenarios which were studied, simulated circumstances that are more resembling to real life, in which vehicles engage in different mobility patterns. Additionally, a comparison to a known clustering algorithm for MANETs - K-ConID [7] was provided to show the supremacy of HCA in terms of cluster stability. 5.1 K-ConID Algorithm K-ConID [7] is a clustering algorithm which considers connectivity and lower ID as a primary and secondary criteria respectively, for selecting ClusterHeads. The authors, modified the 60

61 Lowest Id [27] and degree based algorithms, to create k-hop clusters, i.e. the maximal number of hops between each cluster member to its cluster head is k. Participating nodes flood their identifiers, which consist of number of k-degree neighbors, and their id, to k hops away. This is followed by a flooding phase, in which each node selects the largest identifier among its k hop neighbors. The algorithm was simulated and examined in terms of the average number of created clusters, the average ratio of border nodes and the average cluster size. In order to compare the two algorithm in similar configuration we set k to be 2, therefore both algorithms form 2 hops clusters. Since 2-ConID does not define a channel access method like HCA does, we used MiXiM s implementation as a channel access scheme. It is worth mentioning that 2-ConID does not handle scheduling, and therefore only clustering aspects (the number of clusters and cluster stability) were compared. The authors of K-ConID assumed that each node knows its K hops neighbors. This assumption does not comply with realistic scenarios, therefore the algorithm was modified to allow learning of neighboring nodes by exchanging beacons between nodes. 5.2 Static Scenarios This section describes a study that is used to evaluate the execution of HCA in a static scenario in order to estimate the creation of DG 2 DS. A varying number of nodes were placed randomly on different simulation playgrounds, creating different densities of nodes. Each scenario consisted of 30 runs to allow finding average results. Table 5.1 presents some of the key simulation parameters used in the static scenarios. Next we present the studied metrics: Number of clusters, time complexity and message complexity. 61

62 Table 5.1: Static Simulation Configuration Number of Nodes Simulation Playground Size m 2 Path Loss Coefficient 3.5 (Sub-urban area) Carrier Frequency 5.85 GHZ Simulation Time 300 sec Transmission Rate 2.5 M bps Slot Length sec Vehicular Speed 0 m/s Repetitions 30 Communication Range 200 m Number of Clusters One of HCA s main requirements was to find a small Dominating Set of G 2 (as defined in 4.1). Our goal was to avoid the creation of redundant clusters as much as possible without the knowledge of nodes location. Therefore, the first metric that was studied, measured the number of clusters that were created. Figure 5.1, exhibits the number of clusters formed in different playgrounds according to the number of nodes participating in the simulation. Two main conclusions can be derived from this figure. It can be seen that the main cause for the increase in the number of clusters is the playground size. Obviously, a larger playground requires more ClusterHeads for complete coverage. However, this is not the only cause for a larger number of clusters. It can be seen that the number of nodes in the network influences that measure as well. This can be explained by the fact that the probability that nodes will be placed in uncovered areas increases, therefore some additional clusters need to be formed. This observation is strengthened in figure 5.2. When measuring the number of clusters (denoted by the number of ClusterHeads) in different node densities, it can be seen that 62

63 Figure 5.1: Number of ClusterHeads vs. the total number of nodes in different playground sizes. The numbers in the legend denote the square of the simulation playground size in meters only in very low densities the number of clusters significantly increases. Yet, the number of nodes has some influence as well, but mostly on very low nodes densities Convergence Time An additional requirement from the Hierarchical Clustering Algorithm was to enable communication, as fast as possible. This requirement was originated from an adaptation of a channel access protocol to vehicular networks, in which nodes might require a fast network setup (due to the high mobility of vehicles). Hence, initial convergence time has enormous significance upon the algorithm s performance. A comprehensive study of both the number of slots, and the number of rounds required for the execution of HCA was conducted. We compared the analytical results obtained in 63

64 Figure 5.2: Number of ClusterHeads vs. the total number of nodes and different densities Theorem 4.2 to the simulation results. First we studied the number of required rounds for HCA s third phase - the cluster formation and synchronization phase. Wolfram Mathematica [41] software, was used to study the process of random selection of slots in that phase. Mathematica allowed us to study the behaviour of this process in a much larger scale, since the location, the mobility and the transmission capabilities of WVNS were not required in this experiment. Figure 5.3 presents the required number of rounds, for all nodes within a single cluster to obtain their slot. It can be easily seen, that HCA behaves as expected, comparing to the analytical bound. Therefore, the ratio between the simulated number of rounds and the analytical bound aspires to a constant number. When comparing the execution of HCA s first three phases, using WVNS, similar results are obtained, though much smaller networks can be examined. Figures 5.4 and

65 Figure 5.3: The average number of rounds during HCA s third phase present the average number of rounds and slots respectively, before HCA s first three phases terminate, in different playground sizes. It is worth mentioning that due to the fact that the size of the playground varies, several clusters are created, therefore each cluster contains less nodes than the total number of nodes in the network. As expected, the number of rounds and slots is relatively close to the analytical bound, even when combining the first three phases. Additionally, it can be seen in Figure 5.6 that the average number of rounds slightly increases when comparing it to the number of clusters. This can be explained by interferences from adjacent clusters causing message collisions, hence more rounds are required for all the nodes to obtain their unique transmission slot. 65

66 Figure 5.4: The average number of rounds during HCA s first three phase using WVNS. The numbers in the legend denote the square of the simulation playground size in meters Messages A study of the number of exchanged messages during the first three phases of the Hierarchical Clustering Algorithm was conducted. We studied the number of messages required for HCA to converge and create DG 2 DS. We examined the simulation results with different number of nodes and different playground sizes, yielding a different number of clusters. Figure 5.7 demonstrates indeed, that the number of messages is close to be linear in the number of nodes. Additionally, it can be seen that the number of messages is slightly affected by the number of clusters. Since adjacent clusters are not synchronized, simultaneous transmissions in bordering clusters occur. Therefore, an increase in the number of clusters yields intercluster collisions, causing message losses. This leads to retransmission of ACKs by the Slaves during the following rounds, introducing an increase in the number of messages exchanged. 66

67 Figure 5.5: The average number of slots during HCA s first three phases using WVNS. The numbers in the legend denote the square of the simulation playground size in meters When comparing the number of slots to the number of messages, a significant difference is seen (as observed in Fig. 5.8). Both of these measures correspond to the analytical results presented in section Dynamic Scenarios This section describes a simulation study which was used to study and evaluate HCA s suitability to more real scenarios - scenarios in which vehicles travel according to different mobility patterns. A suburb of Tel-Aviv metropolitan area, sized m 2, was extracted from Open-Street-Map project [16] and adapted into the SUMO traffic simulator (see Fig. 5.9 for screen-shot). Different mobility patterns were considered for the purpose of comparing the impact of the mobility pattern on the number of clusters and their stability, as formed 67

68 Figure 5.6: The number of rounds during HCA s first third phases vs. the number of clusters using WVNS. The numbers in the legend denote the square of the simulation playground size in meters by HCA and K-ConID with K set to 2. Two mobility patterns were compared: i) Random Way Point (RWP) [22] in which vehicles select a random location and speed and change their location accordingly; And ii) TraCI - a traffic simulator generated traces model (as described in subsection 3 and adopted from [34]). This model simulates a more realistic scenario in which vehicles are obligated to move along roads according to the extracted map and road regulations. In this scenario four different types of vehicles (differing in their length, acceleration and de-acceleration) selected one of six different source points. Each vehicle selected a random destination from a set of six possible destinations and drove towards that point according to the possible routes and traffic lights as simulated by the traffic simulator (SUMO). All simulations were based on a network consisting of vehicles, executing HCA and 2-ConID for 300 seconds of simulation time, in order to allow significant number 68

69 Figure 5.7: The average number of messages sent during the first three phases Figure 5.8: The average number of messages sent and the average number of slots during the first three phases. The upper and lower surfaces denote the number of slots and message respectively 69

70 Figure 5.9: A screen-shot of the simulation with the region which is used (a suburb of Tel- Aviv). ClusterHeads, ClusterRelays, and Slaves are denoted by rectangles, triangles, and circles respectively of cluster creations and tear downs in highly dynamic scenarios. Table 5.2 summaries key parameters used in this simulation study. Next we present the results that were obtained for the two studied measures (Number of ClusterHeads and stability) Number of Clusters One of the goals of the HCA was to minimize the number of formed clusters. A study was conducted to observe the impact of different speeds under different mobility models on the average number of ClusterHeads. We measured the number of ClusterHeads at each time step and calculated the mean number of clusters for each simulation run. It can be seen in 70

71 Table 5.2: Dynamic Simulation Configuration Playground Size m 2 Path Loss Coefficient 3.5 (Sub-urban area) Carrier Frequency 5.85GHZ Simulation Time 300 sec Transmission Rate 2.5M bps Slot Length sec Repetitions 30 Communication Range 200m Fig that 2-ConID produces fewer clusters than HCA, under both mobility patterns. This can be explained by the fact that the 2-ConID forces clusters to merge when every two ClusterHeads approach each other, while HCA tries to minimize cluster switching as can be seen in the following subsection. It can be observed that for both algorithms, the mobility pattern indeed influences the number of clusters. It is clear that in the RWP scenario more clusters are created rather than in the TraCI scenario. This can be explained by the fact that the TraCI scenario imitates real life traffic, which constrains vehicles to move along roads. Due to the fact that vehicles usually move in groups, and perhaps not all of the simulated area contains roads, fewer clusters are required. An additional observation derived from Fig. 5.10, shows that an increase in the maximal speed, results in a decrease in the number of ClusterHeads in the RWP pattern for both algorithms. A reason for this phenomenon is that as speed increases, less nodes find themselves isolated for long periods of time, hence they can be merged by other clusters. However, a different behaviour is recognized when comparing both algorithm under more realistic mobility pattern (such as TRaCI). In that case, increasing the speed has a 71

72 Figure 5.10: The average number of ClusterHeads vs. the maximal speed for 100 vehicles network minor effect on the number of created clusters under high velocities Cluster Stability Cluster stability was measured in terms of the average number of cluster switches throughout the simulation and the percentage of time in which nodes were not members of any cluster, denoted as disconnection time. We aimed to form stable clusters since cluster switching might incur some penalty or overhead in terms of higher layer communication and connectivity. HCA outperforms 2-ConID in terms of cluster switches as can be seen in Fig. 5.11, which describes the average number of cluster switches per node in both algorithms while vehicle travel according to the two mobility patterns. From it we conclude that a realistic mobility model (as simulated in the TraCI scenario) indeed minimizes the number of cluster 72

73 Figure 5.11: The average number of cluster switches per node during the simulation time (300 sec), vs. vehicles maximal speed for 100 vehicles network switching and extends the membership period, therefore creating more stable clusters. Comparing these results to the 2-ConID algorithm s results, shows HCA s supremacy in terms of cluster stability. This can be explained by the fact that as opposed to the 2-ConID algorithm, HCA does not force clusters merging when it can lead to disconnection of members. Additionally, Fig shows that an increase in the maximal speed, results in a higher number of cluster switching. This can be noticed both for the HCA and 2-ConID algorithm, however it is noticeable that HCA handles the speed increase much better. Therefore HCA is a much more suitable solution for VANETs. An additional dimension of stability was measured in terms of the percentage of time in which nodes were disconnected from a ClusterHead, i.e., not members of any cluster. It is worth mentioning that this parameter was not compared to the corresponding parameter 73

74 Figure 5.12: Average percentage of time in which nodes were not members of any cluster, vs. the maximal speed for 100 vehicles network in the 2-ConID algorithm because nodes that were disconnected from their cluster, in the 2-ConID algorithm will immediately join a neighboring cluster or form a new one, therefore such node can be seen as constantly a cluster member. HCA tries to minimize the number of 1 node clusters, and thus disconnected nodes might extend their contention period before declaring themselves as ClusterHeads which results in disconnection periods. Fig presents the average percentage of time that each node executing HCA was disconnected from a cluster. First, it can be seen that the RWP model causes the clusters to be less stable than TraCI. Secondly, it can be observed that maximal speed increment degrades HCA s performances in a term of disconnection periods. However, a real life scenario allows an acceptable measure of disconnection periods. 74