REDUCTION 2011-2014 Deliverable 1.3 VANET Packet ... · D1.3 [VANET Packet Scheduling/Routing and Information Dissemination] Executive Summary This report describes the network architecture

REDUCTION2011-2014

Deliverable 1.3VANET Packet Scheduling/Routing

and Information Dissemination

30-8-2014

D1.3 [VANET Packet Scheduling/Routing and Information Dissemination]

Public Document

2


Project acronym: REDUCTIONProject full title: Reducing Environmental Footprint based on Multi-Modal Fleet managementSystems for Eco-Routing and Driver Behaviour Adaptation

Work Package: WP1

Document title: VANET Packet Scheduling/Routing and Information DisseminationVersion: 3.0Official delivery date: 31/08/2014Actual publication date: 31/08/2014Type of document: ReportNature: Public

Authors: Dimitrios Katsaros (UTH), Leandros Tassiulas (UTH), Iordanis Koutsopoulos (UTH),Leandros Maglaras (UTH), Donatos Stavropoulos (UTH)

Approved by: REDUCTION consourtium partners

Version Date Sections Affected

1.0 August 17, 2012 Initial version

1.1 August 25, 2012 Review comments processed

1.2 August 26, 2012 Updated to reflect changes in the imple-mentation

1.3 August 27, 2012 Updated to reflect changes in the imple-mentation

1.4 December 11, 2012 Updated to reflect reviewer’s comments

1.5 December 19, 2012 Updated to correct some minor issues2.0 February 12, 2014 Updated to incorporate 2nd review com-

ments3.0 August 22, 2014 Explanation of the term ‘sociological’

3


Executive Summary

This report describes the network architecture of REDUCTION. Each layer is composed ofthe individual vehicles (i.e., V2V communication). The vehicles are organized in platoons (clus-ters), with some (dynamically elected) representative as the leader of each platoon. Protocols forplatoon formation and maintenance are presented while three low overhead high performanceclustering techniques are evaluated. A custom simulator for traffic and network simulationsis created and the performance of Spring clustering compared to other clustering methods isevaluated. The Enhanced Layered backpressure method that models the vehicular network as aConstraint Queueing Network is also presented. Constraint Queueing Networks (CQN) and sub-sequently packet scheduling/routing policies with guaranteed (optimal) throughput and stabilityproperties, which is tolerant to topology changes, is of major importance for VANETs. Thisversion of the backpressure is a centralized one and a formal proof of its troughput optimality ispresented.

4


Contents

1 Introduction 101.1 Project Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101.2 Work Package Objectives and Tasks . . . . . . . . . . . . . . . . . . . . . . . 111.3 Objective of this Deliverable . . . . . . . . . . . . . . . . . . . . . . . . . . . 131.4 Vehicular communications and networking aspects in REDUCTION . . . . . . 131.5 Relevance of VANET networking aspects to CO2 reduction . . . . . . . . . . . 16

2 Related Work and REDUCTION 182.1 Background on Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.1.1 Why clustering? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182.1.2 Clustering techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . 192.1.3 Clustering in VANETs . . . . . . . . . . . . . . . . . . . . . . . . . . 20

2.2 Routing in VANETs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222.2.1 Topology-based routing protocols . . . . . . . . . . . . . . . . . . . . 222.2.2 Position-based routing protocols . . . . . . . . . . . . . . . . . . . . . 232.2.3 Broadcast routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232.2.4 Geocast routing protocols . . . . . . . . . . . . . . . . . . . . . . . . 232.2.5 Cluster-based routing protocols . . . . . . . . . . . . . . . . . . . . . 23

3 Framework and Methodology 253.1 Evaluation of three low-overhead, high performance clustering schemes . . . . 25

3.1.1 Simulation Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273.1.2 Simulation platform . . . . . . . . . . . . . . . . . . . . . . . . . . . 273.1.3 Performance metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . 283.1.4 Simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293.1.5 Tuning of RanMIS . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293.1.6 Comparison of competing clustering protocols . . . . . . . . . . . . . 35

3.2 The SPRING clustering algorithm . . . . . . . . . . . . . . . . . . . . . . . . 403.2.1 Network identification . . . . . . . . . . . . . . . . . . . . . . . . . . 423.2.2 Clustering process and protocol structure . . . . . . . . . . . . . . . . 433.2.3 Special role of vehicles . . . . . . . . . . . . . . . . . . . . . . . . . . 443.2.4 Cluster-head election parameters . . . . . . . . . . . . . . . . . . . . . 453.2.5 The cluster formation algorithm . . . . . . . . . . . . . . . . . . . . . 463.2.6 Cluster maintenance . . . . . . . . . . . . . . . . . . . . . . . . . . . 463.2.7 Demo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473.2.8 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

3.3 Sociological vehicle clustering . . . . . . . . . . . . . . . . . . . . . . . . . . 493.3.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493.3.2 Network model: Definition of the system . . . . . . . . . . . . . . . . 513.3.3 Network model: Road network communities . . . . . . . . . . . . . . 523.3.4 Network model: Collection of personalized data . . . . . . . . . . . . 52

5


3.3.5 Mobility of nodes and semi-Markov Model . . . . . . . . . . . . . . . 533.3.6 Transition probability matrix . . . . . . . . . . . . . . . . . . . . . . . 543.3.7 Equilibrium probabilities . . . . . . . . . . . . . . . . . . . . . . . . . 543.3.8 Communication phase: Vehicle leaving the area of interest . . . . . . . 543.3.9 Communication phase: Vehicle entering the area of interest . . . . . . 553.3.10 Communication phase: Mean sojourn times . . . . . . . . . . . . . . . 563.3.11 Sociological patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . 563.3.12 Social Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 573.3.13 Sociological pattern of vi . . . . . . . . . . . . . . . . . . . . . . . . . 603.3.14 Route stability of vehicle vi . . . . . . . . . . . . . . . . . . . . . . . 613.3.15 Simulation and performance evaluation . . . . . . . . . . . . . . . . . 623.3.16 Sociological Pattern Clustering . . . . . . . . . . . . . . . . . . . . . . 633.3.17 Route Stability Clustering. . . . . . . . . . . . . . . . . . . . . . . . . 66

4 Delay-aware cross-layer packet scheduling 684.1 The quest for delay-aware packet scheduling protocols . . . . . . . . . . . . . 704.2 Distributed backpressure-based packet scheduling . . . . . . . . . . . . . . . . 734.3 The original Backpressure algorithm . . . . . . . . . . . . . . . . . . . . . . . 744.4 Layered backpressure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

4.4.1 The Layered BackPressure protocol . . . . . . . . . . . . . . . . . . . 754.5 The Enhanced Layered Backpressure policy . . . . . . . . . . . . . . . . . . . 784.6 Dynamic networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

4.6.1 Experimental setting . . . . . . . . . . . . . . . . . . . . . . . . . . . 794.6.2 Experimental results . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

5 Risk Assessment 81

6 Conclusion 82

List of Tables

1 Summary of objectives, ITS tasks and methodologies. . . . . . . . . . . . . . . 132 L Parameter Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283 Interpretation of used symbols. . . . . . . . . . . . . . . . . . . . . . . . . . . 284 Speed per lane. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 475 Speed per lane. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 476 Path table of vehicle i for time period j. . . . . . . . . . . . . . . . . . . . . . 557 Road segment travel times table RST . . . . . . . . . . . . . . . . . . . . . . . 568 Minimum sensitivity in receiver antenna according to data rate. . . . . . . . . . 629 Simulation parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6310 Route distributions according to different social patterns of vehicle i. . . . . . . 67

6


List of Figures

1 An example where vehicle clustering and delay-aware packet scheduling cansave fuel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2 The Randomized MIS clustering algorithm. . . . . . . . . . . . . . . . . . . . 263 Impact of D on the number of Rounds required by RanMIS to complete back-

bone construction w.r.t. network size and its density. . . . . . . . . . . . . . . . 304 Impact of D on the number of messages exchanged by RanMIS for backbone

construction w.r.t. network size and its density. . . . . . . . . . . . . . . . . . . 315 Impact of D on the number of clusters produced by RanMIS for backbone con-

struction w.r.t. network size and its density. . . . . . . . . . . . . . . . . . . . 326 Impact of M on the number of rounds required by RanMIS to complete backbone

construction w.r.t. network size and its density. . . . . . . . . . . . . . . . . . . 337 Impact of M on the number of rounds required by RanMIS to complete backbone

construction w.r.t. network size and its density. . . . . . . . . . . . . . . . . . . 338 Impact of M on the % percentage of Max rounds required to complete the back-

bone construction w.r.t. network size and its density. . . . . . . . . . . . . . . . 349 Impact of M on the number of messages exchanged by RanMIS for backbone

construction w.r.t. network size and its density. . . . . . . . . . . . . . . . . . . 3410 Impact of M on the number of messages exchanged by RanMIS for backbone

construction w.r.t. network size and its density. . . . . . . . . . . . . . . . . . . 3511 Impact of M on the number of clusters produced by RanMIS w.r.t. network size

and its density. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3512 Impact of M on the number of clusters produced by RanMIS w.r.t. network size

and its density. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3613 Rounds required by each competitor algorithm to complete backbone construc-

tion w.r.t. network size and its density. . . . . . . . . . . . . . . . . . . . . . . 3714 Messages required by each competitor algorithm to complete backbone con-

struction w.r.t. network size and its density. . . . . . . . . . . . . . . . . . . . 3715 Clusters produced by each competitor algorithm w.r.t. network size and its density. 3816 Distribution of nodes in clusters after backbone construction w.r.t. network size

and a density of D = 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3917 Distribution of nodes in clusters after backbone construction w.r.t. network size



and a density of D = 15. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4020 Network Diameter diversification (in meters) after backbone construction w.r.t.

network size and its density. . . . . . . . . . . . . . . . . . . . . . . . . . . . 4121 Network Diameter diversification (in Hops) after backbone construction w.r.t.

network size and its density. . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

7


22 CH Distance Distribution after backbone construction w.r.t. network size and itsdensity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

23 Relative mobility at node i with respect to node j. . . . . . . . . . . . . . . . . 4524 The correct choice of the clusterhead plays significant role . . . . . . . . . . . 4525 Simulation environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4826 Average cluster change per vehicle for Sp-Cl, LPG and Low-Id methods for

different transmission ranges(125m, 80m). . . . . . . . . . . . . . . . . . . . . 4927 The average total number of formed clusters for Sp-Cl and Low-Id methods for

different transmission ranges(125m, 80m). . . . . . . . . . . . . . . . . . . . . 4928 The average cluster lifetime Sp-Cl and Low-Id methods for different transmission

ranges(125m, 80m). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5029 Semi-Markov model of vehicle vi in a simple network topology with one entry

and one exit. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5330 Vehicle packet design. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5531 Transition table → social pattern. . . . . . . . . . . . . . . . . . . . . . . . . . 5732 Procedure for social patterns extraction. . . . . . . . . . . . . . . . . . . . . . 5733 Relative mobility at node i with respect to node j. . . . . . . . . . . . . . . . . 5934 States of a vehicle. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6135 A typical region of interest. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6436 Lifetime of SPC versus VFVC for a typical urban scenario [2 flows, 70 % prob-

ability of following the social pattern] for different communication ranges. . . . 6537 Number of heads produced by all methods during the simulation. . . . . . . . . 6638 Lifetime and mean cluster changes versus range. . . . . . . . . . . . . . . . . . 6739 Lifetime and mean cluster changes versus speed. . . . . . . . . . . . . . . . . . 6840 Lifetime and mean cluster changes versus number of social patterns. . . . . . . 6941 Lifetime and mean cluster changes of SPC versus probability of following a social

pattern. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7042 Main road and the flows that split traffic. . . . . . . . . . . . . . . . . . . . . . 7143 Lifetime of RSC. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7144 Performance of the S Q−BP policy in a 4x4 grid network (M = 2). . . . . . . 7245 Impact of the number of clusters on the grid network topology. . . . . . . . . . 7346 An example of the L ayBP execution. . . . . . . . . . . . . . . . . . . . . . 7647 Dynamic network with 3 layers. . . . . . . . . . . . . . . . . . . . . . . . . . 7948 Impact of moving nodes on the delay performance of BP , L ayBP and

E L−BP for the dynamic network with the 3 layers. . . . . . . . . . . . . . 80

8


49 Impact of moving nodes on the throughput performance of BP , L ayBP andE L−BP for the dynamic network with the 3 layers). . . . . . . . . . . . . . 80

Glossary

Acronym Expanded term Description

VANET Vehicular Ad Hoc Network Network of vehicles communicating in an ad hoc manner

V2V Vehicle-to-vehicle Communication among a vehicle and another vehicle

RSU RoadSide Unit Fixed infrastructure element, e.g., proxy, base station

V2I Vehicle-to-infrastructure Communication among a vehicle and fixed infrastructure

V2R Vehicle-to-Roadside Communication among a vehicle and a RSU, e.g., a proxy

V2X Vehicle-to-X X could stand for V, R, I

9


1 Introduction

1.1 Project Overview

Reduction of CO2 emissions is a great challenge for the transport sector nowadays. Despiterecent progress in vehicle manufacturing and fuel technology, still a significant fraction of CO2emissions in EU cities is resulting from vehicular transportation. Therefore, additional innova-tive technologies are needed to address the challenge of reducing emissions. The REDUCTIONproject focuses on advanced ICT solutions for managing multi-modal fleets and reducing theirenvironmental footprint. REDUCTION collects historic and real-time data about driving behav-ior, routing information, and emissions measurements, that are processed by advanced predictiveanalytics to enable fleets enhancing their current services as follows:

1. Optimizing driving behavior: supporting effective decision making for the enhancementof drivers’ education and the formation of effective policies about optimal traffic opera-tions (speeding, braking, etc.), based on the analytical results over the data that associatedriving-behavior patterns with CO2 emissions;

2. Eco-routing: suggesting environmental-friendly routes and allowing multi-modal lets toreduce their overall mileage automatically; and

3. Support for multi-modality: offering a transparent way to support multiple transportationmodes and enabling co-modality.

REDUCTION follows an interdisciplinary approach and brings together expertise from sev-eral communities. Its innovative, decentralized architecture allows scalability to large fleets bycombining both V2V and V2I approaches. Its planned commercial exploitation, based on itsproposed cutting edge technology, aims at providing a major breakthrough in the fast growingmarket of services for “green” fleets in EU and worldwide, and present substantial impact to thechallenging environmental goals of EU.

10


1.2 Work Package Objectives and Tasks

WP1 deals with hardware design and development issues and also with basic wireless com-munication infrastructure and networking aspects. Its objective is to develop the on-board tech-nology taking also into account the requirement for supporting multi-modal fleets for passengertransport. WP1 consists of three tasks:

T1.1 Design and architecture.

T1.2 Onboard technology.

T1.3 VANET packet scheduling/routing and information dissemination.

The present deliverable is the report on T1.3. We briefly describe the goals of this task. Thefirst goal of this task is to define a two-layer network architecture. The lower layer will be com-posed by the individual vehicles (i.e., V2V communication). The vehicles will be organized inplatoons (clusters), with some (dynamically elected) representative as the leader of each platoon.The upper layer will be composed by the leaders and by the fixed (infrastructure) servers, andthey will collectively compose an overlay. Protocols for platoon formation and maintenance willbe the objectives of this subtask. These protocols will be evaluated via simulation (in customand established VANET simulators) against competing strategies. The second goal of task T1.3involves the modelling of the vehicular network as a Constraint Queueing Network (CQN) andsubsequently the development of packet scheduling/routing policies with guaranteed (optimal)throughput and stability properties, which will be tolerant to topology changes. The task aimsto develop new theory and algorithms for optimal control of vehicular signals, and their imple-mentation through distributed coordination among the signals using wireless communication. Inparticular, it will develop and use CQN theory to design traffic signalling policies that throughdistributed coordination can optimize metrics like travel time, network stability, and can respondquickly to changing traffic patterns and emergency situation requirements. Moreover, the taskwill investigate the design of data packet scheduling and routing algorithms that can efficientlyimplement the signal control policies through scalable wireless communication/networking. Itwill study of tradeoffs between efficiency of traffic signalling and complexity of control messageexchanges, and design of localized signalling algorithms that are optimal for a given wirelesschannel. Finally, it will build a simulation platform to implement, test and evaluate distributedsignal control policies taking into account the wireless communication necessary for informationexchange and coordination among the signals.

A third goal of this task is the design of cooperative communication protocols for informa-tion dissemination and storage/caching based on solid theories, i.e., social and complex networkanalysis. These traditional theories will be enhanced with temporal aspects to suit the needs ofprotocols for a highly dynamic network. These protocols will concern stable routing, ‘en-route’caching, with the ultimate aim of delivering both time-critical and non-time-critical informationto information seekers, i.e., vehicles at minimum delay. The protocols will strive to optimize thealternative communication paths, i.e., ad hoc and via infrastructure, with an emphasis though onutilizing the vehicle communication capacities. Since alternative paths (V2V and/or V2I) ex-ist for enabling communication between communicating parties, several optimization problems

11


arise that will be addressed. The information dissemination/retrieval and caching protocols willcapitalize on the mobility of the vehicles and on the two-layer architecture; they will exploit con-cepts such as ‘data density and ‘speed of spread’ in distributed settings where control decisions(e.g., cache admission control) will be taken with minimal amount of information exchange,and only between neighbouring vehicles. For all the components of the developed protocols,detailed simulations will take place to investigate their throughput and delay properties, as wellas their appropriateness and robustness in cases of network partition and congestion.

12


1.3 Objective of this Deliverable

The aim of REDUCTION’s first work package (WP1) is to address the challenges involvedwith the hardware and the (communication and networking) protocols needed to realize an en-vironment where information will travel from vehicles to other vehicles or to infrastructure el-ements, and vice versa. The role of the University of Thessaly within this WP is to developall the necessary communication infrastructure and wireless communication/networking pro-tocols. In accordance with DoW, the very first goal involves the modeling of the vehicularnetwork as a Constraint Queueing Network (CQN) and subsequently the development of packetscheduling/routing policies with guaranteed (optimal) throughput and stability properties; thesepolicies should also be delay-aware, since many application-level tasks of REDUCTION aretime-critical. We did significant steps towards this goal, as described in subsection 4.5 of thisreport. A second goal which was pursued in parallel was the design of communication proto-cols for information dissemination based on vehicle clustering, aggregation, caching, and so on.Since this goal had many ‘subtargets’, we strove for the investigation of the fundamental one,which is the development of clustering techniques; later on, the rest of the aims can be followed,since aggregation can be directly done by clusterheads, and cooperative caching protocols canbe implemented with the aid of clustering; for instance (the late) Denko [20] exhibited the feasi-bility of the idea. In summary, Table 1 presents the goals of this deliverable (i.e., task) and howto achieve them.

Objective to achieve ITS task Methodology

Platoon formation Clustering Spring & Sociological clusteringEvaluation via simulation Traffic, Network simulators SUMO, OMNET++Modelling as a CQN Backpressure Enhanced Layered backpressureCooperative communication Air indexes DiSAInprotocols

Table 1: Summary of objectives, ITS tasks and methodologies.

The vast amount of the work described in this deliverable was done during the first year ofthe project. Even the ideas of vehicle clustering based on social factors was conceived duringthe very begining of the project, but the final research article [47] that polished and finalised theideas was prepared later. This may create the wrong impression that the work was funded duringa period that the relevant project task was already over.

1.4 Vehicular communications and networking aspects in REDUCTION

To enhance the safety of drivers, to provide a comfortable driving environment and to con-tribute to fuel economy, messages for different purposes need to be sent to vehicles throughintervehicle communications. Unicast routing is a fundamental operation for vehicle to con-struct a source-to-destination routing in a VANET. Multicast is defined by delivering multicastpackets from a single source vehicle to all multicast members by multi-hop communication.Geocast routing [48] is to deliver a geocast packet to a specific geographic region. Vehicles

13


located in this specific geographic region should receive and forward the geocast packet; other-wise, the packet is dropped. Broadcast protocol is utilized for a source vehicle sends broadcastmessage to all other vehicles in the network. Apart from 1-hop broadcast operations, the othertypes of communications, require some lower-level packet scheduling algorithms and specificnetwork organization. Among the most promising cross-layer scheduling methods is the back-pressure protocol [72]; on the other hand, in order to address the broadcast storm problem [76]in VANETs, some form of vehicle clustering is required.

As already described in Deliverable 1.1 (February 29, 2012), in order to address the scalabil-ity problems [76] arising in any ad hoc network, and in particular in a VANET, we chose a twotier architecture for REDUCTION, which necessitates the creation of clusters. More argumentsin favor of clustering can be found at subsection 2.1.1. Thus, we had to survey and evaluatethe existing MANET and VANET clustering protocols, and depending on the outcome of theevaluation to develop new clustering protocol(s). Indeed, in subsection 3.1 we evaluated threegraph-theoretic clustering protocols [24] that were proposed for MANETs, but seemed promis-ing for VANETs as well, due to their low message overhead (low congestion/delay) and theirfast convergence (ideal for rapidly changing topologies). One of them was a newly proposedrandomized algorithm that creates maximum independent sets with optimal message complex-ity; randomization seems a promising technique for clustering and has recently received atten-tion [22]. Our study showed that there is no single clear winner for all situations; we extendedour efforts and developed a new clustering protocol in 3.2, namely the SPRING clustering pro-tocol, that is proposed as a communication-efficient and low-delay solution. Our investigationis not yet complete, and we plan to extent the current performance analysis of the protocol [42]a) with more simulations and comparisons against the state-of-the-art VANET clustering proto-col(s), and b) with implementation of it in a real small-scale VANET using the NITOS testbedof the University of Thessaly.

Information dissemination is a crucial issue for VANETs [54], but unfortunately there isno single solution to fit all needs. Despite the wealth of routing and information dissemina-tion protocols for VANETs in the literature [37, 39, 71], none of them can offer at the sametime throughput and delay optimality as needed for REDUCTION. We searched for solutionto this problem in the family of the so-called cross-layer solutions, and in particular those pro-tocols which are based on Lyapunov optimization [72]. Unfortunately, even though they arethroughput-optimal, they suffer from long delays in packet delivery. Thus, we worked on a lineof research that combines this packet scheduling/routing cross-layering technique, with cluster-ing; our results [43] – cf. subsection 4.5 – extent our earlier work for static ad hoc networks [41].Still, we work on this problem to convert it into a fully distributed protocol.

(Uni-Multi-Geo-Broad)cast routing can be implemented with the aid of four basic opera-tions that were originally specified by the GEONET project [56]. These operations are thegeo-unicast, geo-broadcast, geo-anycast and topo-broadcast. VANET clustering and packetscheduling are essential protocols for developing these operations, which in their turn are vitalto REDUCTION, and they will be implemented over the hardware provided by DELPHI (theother partner in WP1), namely their DSRC communications box.Geo-unicast.The GeoUnicast forwarding algorithm is executed by a router to relay a packet to the next hop.

14


The ETSI standard defines two GeoUnicast forwarding algorithms:

• Greedy Forwarding (GF) algorithm. With the Greedy Forwarding (GF) algorithm, therouter uses the location information of the destination carried in the GeoUnicast packetheader and selects one of the neighbors as the next hop. The algorithm applies themost forward within radius (MFR) policy, which selects the neighbor with the smallestgeographical distance to the destination, thus providing the greatest progress when theGeoUnicast packet is forwarded.

• Contention-based forwarding (CBF) algorithm. With the Contention-based forwarding(CBF) algorithm, a receiver decides to be a forwarder of a GeoUnicast packet. Thisis in contrary to the sender-based forwarding scheme specified in GF, where the senderdetermines the next hop. The CBF algorithm utilizes timer-based re-broadcasting withoverhearing of duplicates in order to enable an implicit forwarding of a packet by theoptimal node. With CBF, the router broadcasts the GeoUnicast packet. All neighbours,which receive the packet, process it: the router buffers the packet in its CBF packet bufferand starts a timer with a timeout that is inversely proportional to the distance between therouter’s local position and the destination’s positions.

Geo-broadcast – Geographically-Scoped broadcast (GBC).The GeoBroadcast forwarding algorithm is executed by a router to relay a packet to the nexthop. The ETSI standard defines three forwarding algorithms; the two are experimental, so thefollowing one is the dominant:

• Simple GeoBroadcast forwarding algorithm. The algorithm utilizes the function F(x,y)specified in order to determine whether the router is located inside, at the border or outsidethe geographical target area carried in the GeoBroadcast packet header. If the router isinside or at the border of the area, the packet shall be re-broadcasted. If it is outside thearea, the packet shall be forwarded by the GF algorithm.

F(x,y) =

⎧⎪⎪⎪⎪⎨⎪⎪⎪⎪⎩

= 1, for x = 0 and y = 0 (at the center point),

> 0, inside the geographical area,

= 0, at the border of the geographical area,

< 0, outside the geographical area.

Geo-anycast – Geographically-Scoped Anycast (GAC).It is the forwarding mechanism that transports data from a single node to any of the nodes withina geographically area. Compared to geographically-scoped broadcast, with geographically-scoped anycast a packet is not forwarded inside of the geographic area when the packet hasreached the area. Therefore, the operations for GeoAnycast packet handling are similar to thosefor GeoBroadcast packet with minor changes to the source, forwarder and receiver node opera-tion, to support this slightly different functionality.Topo-broadcast – Topologically-scoped broadcast (TSB).TSB offers re-broadcasting of a data packet from a source to all nodes that can be reached incertain number of hops (all nodes in an n-hop neighborhood). Single-hop broadcast is a specialcase of TSB that is used to send heartbeats including application data payload.

15


1.5 Relevance of VANET networking aspects to CO2 reduction

In the previous paragraphs we used a strictly ‘technical networking language’ to present theconcepts and developments in the context of REDUCTION with respect to wireless communi-cations and networking. The astute reader will have already understood the implication of thoseprotocols to fuel economy and subsequently to CO2 reduction. For the sake of completenessthough, we will provide concrete evidence of the positive impact of such kind of protocols todriver safety and to fuel economy with the aid of an example (see Figure 1) that involves anaccident in a highway at a peak time.Event takes place. We suppose that an accident happens in a highway at a point in time wheretraffic is intense, and we also suppose that the vehicles approaching the place of accident areable to ‘detect’ the accident. The accident results in the highway being blocked.Traffic management improvement. The approaching vehicles send warning messages to the im-mediately following vehicles to reduce speed so as to avoid a new car crash. This is a typical,widespread safety application in today’s automotive industry. If the vehicles are organized intoclusters, the clusterhead of each group suppresses the redundant messages; instead of naiveflooding we employ cluster-based routing. Therefore, we have fewer message collisions, wehave shorter MAC access time, and thus faster delivery of the message. So far fuel reductionis not present. These messages are propagated further back to the other following vehicles thatwill quickly be notified about the blocked highway; thus they will have the possibility to followan exit of the highway. Apparently, the faster the message spreads backwards, the more vehicleswill get it before reaching the exit.Congestion avoidance. By having vehicles following these exits, the highway is not severelycongested.Travel time reduction. Those vehicles that followed exits and ‘escaped’ from the accident site –instead of standing still and consuming fuel – are now able to go to their destination by alterna-tive routes in shorter time.Fuel and CO2 reduction. Shorter travel time (usually) translates into saving more fuel.

Figure 1: An example where vehicle clustering and delay-aware packet scheduling can save fuel.

The potential of V2V communication for improving fuel efficiency has also been demon-strated in [80], showing that vehicular communications can assist to reduce average fuel con-

16


sumption especially under high traffic density and long traffic light cycles. Other projects haveinvestigated the impacts on fuel efficiency when using wireless communications between vehi-cles (V2V) or vehicle-to-infrastructure (V2I) employing different algorithms to smoothly slowdown at a red traffic light or to reach it at the next green phase [79, 15].

17


2 Related Work and REDUCTION

2.1 Background on Clustering

For exchanging information about the current driving situation traffic or weather conditions,hazard areas or road conditions vehicles form a spontaneous network, known as a vehicularad hoc network (VANET). Due to the distributed network nature many messages are generateddescribing the same hazard event, hence, these messages can be combined to a single aggregatemessage. Since VANETs have a very limited capacity, it is desired that the number of messagescan be reduced e.g., using aggregation.

The dynamic nature of wireless ad hoc networks though, requires that solutions for multihopnetwork protocols must be distributed. Of the solutions proposed for scaling down networks withlarge numbers of nodes, network clustering is among the most investigated for mobile ad hocnetworks [4, 5, 9, 10, 16, 29, 35, 69, 81, 75, 87], for sensor ad hoc networks [21, 86], and forvehicular ad hoc networks [42, 52, 77].

The basic idea is that of grouping network nodes that are in physical proximity, therebyproviding the flat network a hierarchical organization, which is smaller in size, and simpler tomanage. The subsequent backbone construction uses the induced hierarchy to form a commu-nication infrastructure that is functional in providing desirable properties such as minimizingcommunication overhead, choosing data aggregation points, increasing the probability of aggre-gating redundant data, and so on.

In a clumpy approach, clustering algorithms can be divided in two major families. The firstwith its roots in graph theory, exploits the localized network structure for estimating dynami-cally the clusterheads (CHs), while the second provides mechanisms to ameliorate the fact thatnodes belonging to the backbone are solely responsible for carrying out all communication, thusrunning out of energy very soon. Namely, the latter family addresses the energy consumptionproblem, that is essentially proposes ways to rotate the role of CH among nodes of clusters e.g.,the SPAN [17], the LEACH [30] and the HEED [85] protocols. The proposed methods use theresidual energy of each node in order to direct its decision about whether it will elect itself as aCH or not. However, this family’s methods ignore topological features of the nodes.

The former family of protocols encompasses the most representative and successful solu-tions in the area of network management. This is because while it exploits the wealth of in-formation extracted from the network topology particularities, it can also take into account ap-plication specific requirements, such as QoS. Moreover, algorithms of this family can easily becombined with a round robin CH rotation method as was described in [21], and thus exploit allthe advantages the energy-efficient algorithms provide. Apparently, for the case of VANETs,energy consumption by vehicles is not an issue, since their battery can provide nearly unlimitedpower.

2.1.1 Why clustering?

Exchange of information between vehicles can be either V2V or vehicle-to-roadside (V2R).Forming V2V-based VANETs has some advantages as compared with the V2R-based VANETs.First, the V2V-based VANET is more flexible and independent of the roadside conditions, which

18


is particularly attractive for the most developing countries or remote rural areas where the road-side infrastructures are not necessarily available. Also, V2V-based VANET can avoid the fastfading, short connectivity time, high frequent hand-offs, and so forth caused by the high relative-speed difference between the fast-moving vehicles and the stationary base stations. However, thelink qualities in V2V communications can also be very bad due to multipath fading, shadowing,and Doppler shifts caused by the high mobility of vehicles. V2V communication can be used asthe basic means of communication between vehicles and Roadside units may help in places ofhigh vehicle density.

One of the many challenges for VANETs is the dynamic and dense network topology, result-ing from the high mobility and high node-density of vehicles especially in urban environments.This dynamic topology causes routing difficulties. A clustered structure can make the networkappear smaller and more stable in the view of each vehicle.

To implement a robust clustering algorithm, two problems need to be solved: a clusteringmetric has to be selected and clustering methods need to be designed. A clustering metricshould provide the quantitative definition of closeness which is fundamental to the clusteringproblem. Furthermore, clustering metric should reflect the characteristics of its target networks.In VANETs, the network topology is frequently changing. If a metric can resist against theunnecessary changes while maintaining connectivity, it is robust. A suitable metric can reflectfeatures of nodes and improve performance of the clustering algorithm. Clustering methodsare the approaches to assign nodes into clusters. Strictly speaking, in vehicular networks, theproblem is not only how to form a cluster structure, but also how to maintain the cluster structurein a dynamic network.

2.1.2 Clustering techniques

One of the many challenges for VANETs is the dynamic and dense network topology, re-sulting from the high mobility and high node-density of vehicles [73] especially in urban envi-ronments. This dynamic topology causes routing difficulties. A clustered structure can make thenetwork appear smaller and more stable in the view of each vehicle.

A well-known mobility-based clustering technique is MOBIC [11], which is an extension ofthe Lowest-ID algorithm [27]. In Lowest-ID, each node is assigned a unique ID, and the nodewith the lowest ID in its two-hop neighborhood is elected to be the cluster head. This schemefavors nodes with lower identifiers to become CHs without taking in mind mobility patterns ofthe nodes.

In MOBIC, an aggregate local mobility metric is the basis for cluster formation instead ofnode ID. The node with the smallest variance of relative mobility to its neighbors is elected asthe cluster head. The relative mobility for a certain node is estimated by comparing the receivedpower of two consecutive messages from each neighboring node which is not a so easy task inhigh dense environments. If we consider certain scenarios where attenuation of radio signalsinevitably exists, the power of signals for estimating mobility can be very limiting and mayprovide inaccurate measurements In cluster maintenance also a clusterhead is not guaranteedto bear a low mobility characteristic relative to its members, As time advances the mobilitycriterion between clustrer members is somewhat ignored. If mobile nodes move randomly andchange their speeds from time to time, the performance of MOBIC may be greatly degraded.

19


Many clustering methods have been introduced lately which aim at establishing stable clus-ters, where clusterhead reelection is reduced. In DDVC clustering (DLDC) [59] the Dopplershift of communication signals is used in order to create clusters. Affinity propagation is an al-gorithm for image processing, and APROVE has proved that its distributed case can be utilizedfor VANETs [63]. In Distributed group mobility adaptive clustering (DGMA) [88] group mobil-ity information which contains group physical center’s coordinates, group size, group velocityis used for clustering. Density based clustering is based on a complex clustering metric whichtakes into account the density of the connection graph, the link quality and the road traffic con-ditions [34]. Blum et al. [13] proposed a Clustering for Open IVC Networks (COIN) algorithmwhere cluster-head election is based on vehicular dynamics and driver intentions.

2.1.3 Clustering in VANETs

In cluster-based routing protocols vehicles near to each other form a cluster. Each cluster hasone cluster-head, which is responsible for intra and inter-cluster management functions. Intra-cluster nodes communicate each other using direct links, whereas inter-cluster communicationis performed via clusterheaders. In cluster based routing protocols the formation of clusters andthe selection of the cluster-head is an important issue. In VANET due to high mobility dynamiccluster formation is a towering process. Many clustering techniques for VANETS have beendeveloped lately. The cluster based methods are divided in five major categories. The methodsthat are focused in Urban environments, those that suitable for the VANET environment onhighways , methods that combine V2V and V2I communication and the two-tier architectures.One other category of clustering methods may be those that were initially produced for Manetscan be used in Vanets with some modifications.Urban environments.Affinity propagation is an algorithm for image processing, and APROVE has proved that itsdistributed case can be utilized for VANETs [63]. APROVE distributively elects clusterheads byusing affinity propagation from a communications perspective. The method is tuned for highwayscenarios.

Density-based clustering is based on a complex clustering metric which takes into accountthe density of the connection graph, the link quality and the road traffic conditions [34].

In [2]The proposed algorithm is based on the assumption that each vehicle knows its exactlane on the road via a lane detection system and an in-depth digital street map that includeslane information.The clusterhead is selected based on the flow of the majority of traffic. A laneweight (LW) metric is applied for each traffic flow (LT, RT and NT). The clusterhead selectedis towards the middle of the cluster and continuing in the same direction as the majority of thetraffic flow.Clustering in highways.In [25] during the process of the clusterhead selection, the closest position to the average andthe closest velocity to the average of all proximal vehicles are calculated along with connectivitylevel to determine the most stable clusterhead. The method performs well in highways but notin Urban envirnonments where traffic flow must be taking in mind.

Cluster-based directional routing protocol (CBDRP) [67] is focused on highway scenarios.Clusters are formed from vehicles that move to the same direction. In this method the header of

20


a cluster selects another header according to the moving direction of vehicle to forward packets.In LORA-CBF, there are three types of nodes: headers, gateways and members. The cluster-

head maintains information about its members and gateways.Packet forwarding follows the sameprinciples as the greedy routing. Clusterhead and gateways send out the location request (LREQ)packets when the location of the destination is not available as well as the phase of the LocationReply (LREP) messages. The mobility of the vehicles on a motorway has been used to evaluateaverage route discovery time, end-to-end delay and other performance metrics.

The network in [62] is divided to multiple clusters each having a cluster head and a group ofmembers within the transmission range. Packets are forwarded from source to destination by aprocedure similar to AODV. The main difference is that only cluster heads and gateways forwardpackets.

Cluster head selection in COIN [13] is based on vehicular dynamics and driver intentions.Relative mobility between a cluster head and a member node should be low, so they remain inradio contact for as long as possible.Two Tier architecture.Hierarchical Cluster Based Routing [82] is designed for highly mobility adhoc networks. HCBis two-layer communication architecture. In layer-1 mostly nodes communicate with each othervia multi-hop routing. Among these nodes some also have another interface with long radiocommunication range called super nodes which exist on both layers. Super nodes are able tocommunicate with each other via the base station in layer-2.

In [74] authors propose a two-tier architecture where vehicles are first organized into groupsin VANETs.Traffic information is broadcasted and exchanged among vehicles through IVC.Some vehicles in the groups are selected to form a high-tier P2P overlay through infrastructurewireless communication. These vehicles are called superpeers and serve as a bridge between thehigh-tier and low-tier networks to handle message exchanges and lookups.

In [12], an integration of VANET and 3G networks using mobile gateways is introduced.A VANET-3G integrated network architecture is presented which defines the concept of mobilegateways. A mobile gateway refers to the dual-interfaced vehicle that relays data from othervehicle sources to the UMTS backhaul network.V2I-V2V protocols.In [70] authors propose an RSU-assisted cluster head selection and backup protocol in order tocope with situations where the communication is lost between two adjacent cluster heads.

Frequent disconnected Network is a common feature of VANETS due to high mobility ofvehicles. In order to ensure the service quality of nodes’ communications, [78] make use of thehandoff idea of cellular networks and propose a new protocol which is special for VANET, theproposed scheme is called Traffic Infrastructure Based Cluster Routing Protocol with Handoff.MANET clustering methods.One of the many challenges for VANETs is the dynamic and dense network topology, resultingfrom the high mobility and high node-density of vehicles [73] especially in urban environments.This dynamic topology causes routing difficulties. A clustered structure can make the networkappear smaller and more stable in the view of each vehicle. Many clustering methods initiallycreated for Vanets cannot cope with these characteristics of the VANETS.

A well-known mobility-based clustering technique is MOBIC [11], which is an extension of

21


the Lowest-ID algorithm [27]. In Lowest-ID, each node is assigned a unique ID, and the nodewith the lowest ID in its two-hop neighborhood is elected to be the cluster head. This schemefavors nodes with lower identifiers to become CHs without taking in mind mobility patterns ofthe nodes.

In MOBIC, an aggregate local mobility metric is the basis for cluster formation instead ofnode ID. The node with the smallest variance of relative mobility to its neighbors is elected asthe cluster head. The relative mobility for a certain node is estimated by comparing the receivedpower of two consecutive messages from each neighboring node which is not a so easy task inhigh dense environments. If we consider certain scenarios where attenuation of radio signalsinevitably exists, the power of signals for estimating mobility can be very limiting and mayprovide inaccurate measurements In Cluster maintenance also a clusterhead is not guaranteedto bear a low mobility characteristic relative to its members. As time advances the mobilitycriterion between clustrer members is somewhat ignored. If mobile nodes move randomly andchange their speeds from time to time, the performance of MOBIC may be greatly degraded.

Many clustering methods have been introduced lately which aim at establishing stable clus-ters, where clusterhead reelection is reduced. In DDVC clustering (DLDC) [59] the Dopplershift of communication signals is used in order to create clusters. In Distributed group mobil-ity adaptive clustering (DGMA) [88] group mobility information which contains group physicalcenter’s coordinates, group size, group velocity is used for clustering.

[58] is a Cluster Based Vehicular Adhoc Network Model for Simple Highway Communica-tion.Vehicle within a radio coverage range can communicate directly or though roadside unit. Inthis model a very few Fixed roadside units are assumed. All the cluster heads are synchronizedin a specific time interval.

2.2 Routing in VANETs

In VANET, the routing protocols – despite many detailed categorizations – can in general beclassified into five categories: topology-based routing protocols, position-based routing proto-cols, cluster-based routing protocols, geocast routing protocol and broadcast routing protocol.These protocols are characterized on the basis of area/application where they are most suitable,and in the next paragraphs we provide a brief survey of them.

2.2.1 Topology-based routing protocols

These routing protocols use links information that exists in the network to perform packetforwarding. They are further divided into Proactive, Reactive and Hybrid Protocols.

1. Proactive routing protocolsThe proactive routing means that the routing information, like next forwarding hop is main-

tained in the background irrespective of communication requests. The advantage of proactiverouting protocol is that there is no route discovery since the destination route is stored in thebackground, but the disadvantage of this protocol is that it provides low latency for real timeapplication. The various types of proactive routing protocols are: FSR, DSDV, OLSR, CGSR,WRP, TBRPF.

2. Reactive/Ad hoc based routing

22


Reactive routing opens the route only when it is necessary for a node to communicate witheach other. Reactive routing consists of route discovery phase in which the query packets areflooded into the network for the path search and this phase completes when route is found. Thevarious types of reactive routing protocols are AODV, PGB, DSR,TORA, JARR.

3. Hybrid ProtocolsThe hybrid protocols are introduced to reduce the control overhead of proactive routing

protocols and decrease the initial route discovery delay in reactive routing protocols

2.2.2 Position-based routing protocols

Position based routing consists of class of routing algorithm. They share the property ofusing geographic positioning information in order to select the next forwarding hops. Positionbased routing is broadly divided in two types: Position based greedy V2V protocols, DelayTolerant Protocols.

1. Non DTN2. DTN Position Based Routing Protocols3. Hybrid Position Based Protocols

2.2.3 Broadcast routing

Broadcast routing is frequently used in VANET for sharing, traffic, weather and emergency,road conditions among vehicles and delivering advertisements and announcements. The variousBroadcast routing protocols are BROADCOMM, UMB, V-TRADE, and DV-CAST.

2.2.4 Geocast routing protocols

Geocast routing is basically a location based multicast routing. Its objective is to deliverthe packet from source node to all other nodes within a specified geographical region (Zone ofRelevance ZOR). The various Geo cast routing protocols are IVG, DG-CASTOR and DRG

2.2.5 Cluster-based routing protocols

Cluster based routing is preferred in clusters. A group of nodes identifies themselves to be apart of cluster and a node is designated as cluster head will broadcast the packet to cluster. Goodscalability can be provided for large networks but network delays and overhead are incurredwhen forming clusters in highly mobile VANET. The various Clusters based routing protocolsare COIN, LORA-CBF, TIBCRPH, CBDRP.

1. CBDRP : Cluster-Based Directional Routing Protocol [67]Cluster-based directional routing protocol (CBDRP) is focused on highway scenarios. Clus-

ters are formed from vehicles that move to the same direction. In this method the header of acluster selects another header according to the moving direction of vehicle to forward packets.

2. TIBCRPH: Traffic Infrastructure Based Cluster Routing Protocol with Handoff [78]Frequent disconnected Network is a common feature of VANETS due to high mobility of

vehicles. In order to ensure the service quality of nodes’ communications, it makes use of the

23


handoff idea of cellular networks and propose a new protocol which is special for VANET, theproposed scheme is called Traffic Infrastructure Based Cluster Routing Protocol with Handoff.

3. LORA-CBF: Location Routing Algorithm with Cluster-Based Flooding [61]In LORA-CBF, there are three types of nodes: headers, gateways and members. The cluster-

head maintains information about its members and gateways.Packet forwarding follows the sameprinciples as the greedy routing. Clusterhead and gateways send out the location request (LREQ)packets when the location of the destination is not available as well as the phase of the LocationReply (LREP) messages. The proposed LORA-CBF shows highly heterogeneous performanceresults.

4. COIN: Clustering for Open IVC Network [13]Cluster head selection in COIN is based on vehicular dynamics and driver intentions instead

of ID or relative mobility as in conventional clustering methods. Relative mobility between acluster head and a member node should be low, so they remain in radio contact for as long aspossible.

5. HCB: Hierarchical Cluster-based routing [82]It is a novel based Hierarchical Cluster routing protocol designed for highly mobility adhoc

networks. They attempt to mitigate the impact of rapid topology change by exploiting the nodeheterogeneity and organizational structure of clusters HCB is two-layer communication archi-tecture. In layer-1 mostly nodes communicate with each other via multi-hop routing. Amongthese nodes some also have another interface with long radio communication range called supernodes which exist on both layers. Super nodes are able to communicate with each other via thebase station in layer-2.

6. CBLR: Cluster-based location routing [62]This algorithm assumes all vehicles can gather their positions via GPS. The network is di-

vided to multiple clusters each having a cluster head and a group of members within the trans-mission range. The cluster-head and members are formed as follow: A new vehicle transmits aHello Message. If the vehicle gets a reply from the cluster-head vehicle, the new vehicle wouldbecome a member of the cluster. If not, the new vehicle becomes the clusterhead.

7. CBR: Cluster-based routing [40]CBR (Cluster Based Routing) Protocol is a routing protocol based on position and clusters.

The geographic area is divided into square grids. Each node calculates optimal neighbor clusterheader to forward data to the next hop by using geographic information. This protocol does notconsider velocity and direction which are important parameters in VANET.

24


3 Framework and Methodology

3.1 Evaluation of three low-overhead, high performance clustering schemes

Recently, RanMIS [1] considered the problem of computing a maximal independent set(MIS), a fundamental distributed computing procedure, that seeks to elect a set of local leadersin a network. It generates the maximal set of nodes in such a manner that, no two of themare neighbors, by using an extremely harsh broadcast model that relies only on carrier sensing.Since the set is maximal, every node in the network is either in the MIS or a neighbor of a nodein the MIS.

The model consists of an anonymous broadcast network in which nodes have no knowledgeabout the topology of the network. It has its roots to the solution of a similar problem that evolvesduring the development of the fly’s nervous system, when sensory organ precursor (SOP) cellsare chosen [1].

According to the specified assumptions, nodes receive as input, an upper bound on the num-ber of nodes in the network n, and an upper bound D, on the number of neighbors any node canhave (if no such bound is known, then D is set to n). Furthermore, it is assumed that they allwake up together at the same synchronous round (and start executing the algorithm), also thatthey can detect collisions, and finally that no failures occur.

The algorithm proceeds in log D phases, each consisting of M log n steps, where M is aconstant. Initially, all nodes are active. Each step in each phase i consists of two exchanges.

In the first exchange, each active node broadcasts a message to its neighbors with probabilitypi. Such as in the biological model, the probability pi increases with i. In the second exchange, anode that has broadcasted a message in the first exchange joins the MIS if none of its neighborshad broadcasted at the first exchange. Such node broadcasts again a message to its neighbors,telling them to become inactive, and exits the algorithm.

For the sake of completeness we give in Figure 2 the pseudocode of the RanMIS algo-rithm [1] which is synchronously executed by all nodes.

RanMIS competitors were selected from two distinct protocol categories. The first categoryis oriented on providing the network with a two-layer hierarchical organization comprised bygroups of nodes, i.e., clusters. One ‘special’ node of each cluster (the CH) participates in theso-called backbone; the backbode nodes form a dominating set (DS) over the flat networktopology. This means that, each node that is not in the backbone, has at least one backbone nodeas its neighbor. Backbone nodes are joined via gateway nodes. Apparently, if the flat network isconnected, it is deterministically guaranteed, that the backbone is connected as well. DistributedClustering Algorithm (DCA ) is a high-performance representative of this category.

The second category retains the layered structure of the first category, and it is oriented onfacilitating routing without the need of gateway nodes. The concept behind this category isthat of building a Connected Dominating Set (CDS) directly, without firstly selecting CHsand then joining them. It was initially introduced by Das et al. in [18] when they proposed theconcept of creating a communication spine inside a flat network topology in order to supportunicast, multicast, and fault-tolerant routing within an ad hoc network. WuLi is a practical androbust representative of this category.

Contrary to RanMIS ’s randomized nature, its competitors use iterative clustering tech-

25


Algorithm: RanMIS (n, D) at node u// n: number of participating nodes in network topology// u: node under consideration// D: upper bound on the number of neighbors// B: 1 bit message// M: constant

1 . For(i = 0 : log D)2 . For ( j = 0 : M log n)3 . * exchange 1 *4 . v = 0;5 . With probability 1/2log D - i,

broadcast B to neighbors;set v = 1;

6 . If received message from neighbor, thenv = 0;

7 . * exchange 2 *8 . If v = 1 then9. Broadcast B;

join MIS;exit the algorithm.

10. Else11. If received message B in this exchange, then

mark node u inactive;exit the algorithm.

12. End.13. End.14. End.

Figure 2: The Randomized MIS clustering algorithm.

niques, meaning that a node waits for a specific event to occur or certain nodes to decide abouttheir role before making a decision themselves.

The first class, which enables WuLi [81], exhibit a high degree of localization. In such aclass as soon as a node has collected information about its surrounding topology, it is able todecide whether it will be part of the backbone or not. The only information it needs to waitfor, is the identity of the node in its (h hop) neighborhood. In the second class, the degree oflocalization shown by the algorithms is limited with respect to that of the first. Such protocolsimplement a distributed version of the heuristics for finding an independent set of nodes whichis maximal, and a dominating set which is minimal. As a representative algorithm of this classwe consider the DCA [9]. Being maximum, the independent set produced by DCA is also aminimal dominating set.

WuLi is a very simple distributed procedure consisting of a few local rules, the execution ofwhich creates the desired CDS. Every node v exchanges its neighbor list with all its neighbors.A node set itself has a dominating node if it has at least two unconnected neighbors. In order

26


to reduce the size of a CDS, the original protocol presents two pruning rules. According to thefirst rule, a node deletes itself from the CDS when its close neighbor set, which includes all ofits direct neighbors as well as itself, is completely included in the neighbor set of a neighboringdominating node and it has smaller ID than the neighboring dominating node. According to thesecond rule, a node deletes itself from the CDS when its open neighbor set, which includes allof its direct neighbors, is completely included in the neighbor sets of two connected neighboringdominating nodes and has the smallest ID.

The DCA protocol assumes quasi-stationary nodes with real-valued weights, which are ini-tially UNDECIDED. Node weights induce a total ordering of the nodes without any ties be-cause node weights reflect real numbers. During the execution of the algorithm each node willdecide to be IN or to be OUT, of the independent set. A node decides when all its neighborswith larger weight have decided. When the time comes, a node decides to be OUT if one of itsneighbors is IN. Otherwise it decides to be IN. The IN nodes form the minimal dominating setrequired for clustering. The execution time depends on possible chains of dependency betweenthe nodes, a negative consequence of the reduced localization DCA presents. The node with thesmallest weight has to wait for all other nodes in the chain.

3.1.1 Simulation Model

As in most studies in multihop wireless networks, we used unit disk random graphs to rep-resent the network topologies used in the performance analysis. A unit disk graph is determinedby node positions and a fixed communication range R, for all nodes. The produced topolo-gies which more precisely are described as Constrained - Connected Random Unit Graphs(C-CRUG), because of the constrains the participating nodes are obliged to adhere, were con-structed by taking in to consideration two crucial independent variables, concerning the finalnetwork connectivity, that is the network density, which is perfectly expressed by the degree dof each node, and the node population N.

3.1.2 Simulation platform

The graphs were normally generated by placing nodes randomly and independently fromeach other with the help of a modified version of Minimum degree proximity algorithm (MIN-DPA) [6], a C-CRUG generator, which aims to distribute node degrees more uniformly whilemaintaining connectivity.

The idea is to place each new node in the vicinity of the node that has the smallest numberof neighbors, while preserving a minimum distance to increase the probability of achieving aconnected graph. The modified version of it though, involves the notion that the average degreeof the topology and the transmission range of a node, were predefined.

The simulations refer to scenarios in which N static wireless nodes with maximum transmis-sion range R are randomly and uniformly scattered in a geographic area of side L. We make theassumption that two nodes are neighbors if and only if their Euclidean distance is less than R.

We emphasize that nodes start the protocol execution at the same time and run the clusteringand backbone formation algorithms to form a hierarchical multi-hop ad hoc network.

27


In our simulations R has been set to 50m, the number of nodes N has been assigned thevalues 100, 500, 1000 and 5000 while L has been set accordingly (see Table 2) to influence anaverage node degree of 4, 7, 10 and 15. This allowed us to test the protocols on increasinglydense networks, from (moderately) sparse to dense networks.

Table 2: L Parameter ValuesNodes Degree#4 Degree#7 Degree#10 Degree#15

100 500 400 250 200500 2000 1500 750 500

1000 5000 4000 1000 7005000 7000 5000 2000 1500

In order to diminish any statistical errors all tests were repeated for 1000 times and the dataused in our graphs are all average values of the results. A short description of the simulationparameters is given in Table 3.

Table 3: Interpretation of used symbols.

Symbol Usage Default valuen Number of Nodes in NetworkD Intended Number of Neighbors 24d Average Node DegreeM Constant 32R Node Transmission rangeL Axis Length (in meters) of AOR

(Area Of Responsibility)

3.1.3 Performance metrics

The assessment of the competing protocols were done along three dimensions: the firstdimension portrays the cost (in delay and energy consumption) incurred by the protocols, thesecond concerns the characteristics of the resulting backbone, and the third one deals with theresilience of the resulting spanner to node failures.Metrics for protocol cost. This family of measures, includes the protocol duration which isa direct measure of the delay incurred until the network spanner is established, and the totalnumber of messages exchanged among nodes which is a measure of the network congestion.

Protocol duration. It is the number of rounds required by the protocol to complete theprocedure of backbone formation.

Total number of messages transmitted. The situation concerns the backbone constructionoperation, i.e., the selection of network nodes which will take on the role of a CH, and thedeclaration of ‘attachment’ of non-CH nodes to a CH. To uncouple our measurements fromMAC particularities and focus only to the pure clustering protocol performance, we assume anideal MAC layer with no provision for retransmissions in case of collisions exist.

28


Metrics for backbone description. This family of measures assess the backbone’s ‘layout’.Backbone size. It is the number of the network nodes that comprise the backbone. A smaller

backbone is desirable for minimizing the routing overhead. For instance, in sensor networkswith nodes that can turn off their radio interface (energy saving sleep mode), the backbone sizeis a measure of how many nodes need to stay awake for data transport. Usually, the nodesin the backbone stay up or have a higher duty cycle for guaranteeing routes to the sink, andhence the smaller the backbone the more nodes can be in energy conserving mode. On the otherhand, a very small backbone could result in situations where CH-to-member communicationwould require either excessive amount of energy to reach the node and vice versa, or multi-hopcommunication within the cluster.

Cluster cardinality distribution. It is the distribution of the number of members for eachgenerated cluster.

Route length. If we consider the subgraph Gb of the network induced by the backboneconstruction (Gb is the graph where there are no links between ordinary nodes), then this metricgives a measure of how well a routing protocol can perform over the backbone.

Inter-clusterhead distance distribution. It records the distances among neighboring CHs.The larger this distance is the better for the performance of the clustering protocol, since, forinstance, a three-hop distance among clusterheads would avoid hidden terminal problems andguarantees the ‘independence’ of clusterhead transmission.

3.1.4 Simulation results

In order to evaluate the performance of RanMIS , we selected two graph-theoretic ad hoc net-works clustering algorithms, namely the WuLi and DCA , both of which create dominating sets,the former produces a connected dominating set and the latter produces a maximum indepen-dent set. These algorithms are simple, practical, extremely popular in the clustering community1 and comprise the base for the development of many similar clustering algorithms [24]. We per-formed a series of experiments to compare the performance of RanMIS against these algorithms.The first set of experiments concerns the tuning of RanMIS with respect to its parameters, andthen it follows its comparison against its competitors.

3.1.5 Tuning of RanMIS

The original article which introduced RanMIS [1], described two parameters that controlits operation and affect its overall performance. The first of these parameters, D, concerns theestimated size of neighborhood of each node, while the second, M, is a constant related to thenumber of rounds required for the algorithm to complete the backbone construction. Theseparameters are independent and thus we have to make an initial choice for one of them (e.g., forD) and investigate the other (e.g., M), and vice versa.

1They both have an excessive number of citations. Seehttp://scholar.google.com/scholar?cites=7672109277762563412&as sdt=2005&sciodt=0,5&hl=enandhttp://scholar.google.com/scholar?cites=13992908027776879246&as sdt=2005&sciodt=0,5&hl=en.

29


Impact of parameter D. To evaluate the impact of parameter D on RanMIS ’s performance,we conducted a series of experiments for various values of D, 6, 12, 24 36, and 48 and forvarious performance measures. D controls the maximum number of rounds RanMIS has inorder to complete the backbone construction and it is consistent with the size of each node’sneighborhood. That is because in a large neighborhood the intuition is that the likelihood tohave a simultaneous transmission and a resultant collision between two adjacent transmittingnodes is increased, and consequently more protocol rounds will be required by the affectednodes in order to decide about their status.

In Figure 3, we see the number of rounds required by RanMIS to complete backbone con-struction for various values of D w.r.t. network size and its density.

Figure 3: Impact of D on the number of Rounds required by RanMIS to complete backboneconstruction w.r.t. network size and its density.

The first observation is that, in accordance with our intuition, when the value of D increases,the number of rounds required by RanMIS to complete backbone construction increases forevery network size. This consequence derives from the fact that there is a direct relation betweenthe probability a node has in order to be chosen as a CH candidate and the setting of parameter D(cf. Figure 2). The higher the setting of D the lesser the probability for a node to become a CHand if so, the probability for a neighboring node to become a CH is also small (thus, we avoid anyunwanted message retransmissions due to collisions). The second observation is that while thenumber of rounds is almost irrelevant to network density for a given network size, the numberof rounds increases linearly to the network size. This is due to the fact that more clusters will bedeveloped and therefore the competition for becoming a clusterhead increases. It is obvious thatthe less the required rounds to complete the backbone is, the best for the protocol is (namely inFigure 3 this observation favors D setting of 6 and 12 to the other). Though, to conclude whichis the best choice for parameter D, we must examine also the number of transmitted messages

30


and the number of clusters created for the various choices of D (cf. Figure 4 and Figure 5). Thus,we do not conclude at this point that the best value for D is 6.

In Figure 4, we see the number of messages exchanged by RanMIS for backbone construc-tion for various values of D w.r.t. network size and its density.

Figure 4: Impact of D on the number of messages exchanged by RanMIS for backbone con-struction w.r.t. network size and its density.

The situation here is reversed with respect to what we saw in Figure 3. In general, the num-ber of messages required by RanMIS to complete the backbone construction process decreasesfor increasing values of D. Therefore, so far the best choice for D is the one that achieves a bal-ance between the number of rounds and the messages transmitted, i.e., D = 24 (we will revisitthis issue when commenting the next figure). Additionally, examining the figure, we observe adecreasing number of transmitted messages for increasing network density (for every networksize). This is because when a node declares itself as a CH, then all of its neighbors (and thereare too many nodes in dense networks) immediately attach themselves to this node. The fi-nal observation is that the relative ordering of the performance curves for various values of Dis maintained across different densities (with some statistically insignificant variation for verysmall networks).

In Figure 5, we see clusters produced by RanMIS for backbone construction w.r.t. networksize and its density for various values of D.

The main observation is that the number of clusters produced is not affected by parameter D.However, protocol execution drives to decreasing number of produced clusters w.r.t. increasingnetwork density for every network size, and the explanation for that is similar to the one for theprevious figure.

Overall, a value of D = 24 is the most appropriate choice since it achieves a good tradeoffbetween network backbone construction delay (due to many rounds) and the transmitted mes-

31


Figure 5: Impact of D on the number of clusters produced by RanMIS for backbone constructionw.r.t. network size and its density.

sages.Impact of parameter M. To evaluate the impact of constant M on RanMIS ’s performance,we conducted a series of experiments for various values of M = 2x (x = 0 - 7), and for variousperformance measures. For these sets of experiments, the value of D was set equal to 24 (cf.subsection Impact of parameter D.)

In Figures 6, 7 we see the impact of M on the number of rounds required by RanMIS tocomplete the backbone construction. In Figure 6, it is clear that an M = 8 setting supersedes anyother setting for practically all network instances (except for 100 nodes), since a small M meansless work in the inner loop of RanMIS , that is less residency in a low probability state for anode to be chosen as a CH. Practically we bias the system to converge faster with small settingsof M. The same observation holds in Figure 7, where for smaller values of M (namely 1, 2) therequired number of rounds for convergence decreases.

On the other hand, a small M setting might create severe competition for the role of CHamong neighboring nodes. Since M value affects the maximum number of rounds that are avail-able for backbone construction, it is possible for a small value of M the protocol not to converge.In Figure 8, we see the impact of M on the % percentage of Max rounds used by RanMIS inorder to complete the backbone construction. For small settings of M (e.g 1, 2) more than 80%of the available rounds were used by the protocol in order to converge.

In Figures 9, 10, we see the impact of M on the number of messages exchanged by RanMIS tocomplete the backbone. The observation is that for practically all network instances the numberof transmitted messages is unrelated to M.

In Figure 11, 12, we see the impact of M on the number of clusters produced by RanMIS .Similarly, the number of clusters produced is not affected by M.

32


Figure 6: Impact of M on the number of rounds required by RanMIS to complete backboneconstruction w.r.t. network size and its density.

Figure 7: Impact of M on the number of rounds required by RanMIS to complete backboneconstruction w.r.t. network size and its density.

Overall, it seems that the authors’ suggestion of M = 34 is not the best choice; it is onlymarginally good for small networks (100 nodes) when we want to minimize the number oftransmitted messages – the difference though is not really significant. We concluded though,that we should be conservative with the setting of M, especially when we have to deal with

33


Figure 8: Impact of M on the % percentage of Max rounds required to complete the backboneconstruction w.r.t. network size and its density.

Figure 9: Impact of M on the number of messages exchanged by RanMIS for backbone con-struction w.r.t. network size and its density.

real life network topologies. Nevertheless, in the rest of this article, we follow the RanMIS ’screators suggestion and set the value of M equal to 32.

34


Figure 10: Impact of M on the number of messages exchanged by RanMIS for backbone con-struction w.r.t. network size and its density.

Figure 11: Impact of M on the number of clusters produced by RanMIS w.r.t. network size andits density.

3.1.6 Comparison of competing clustering protocols

Results concerning the protocol cost. In this series of experiments, we measured the number of

35


Figure 12: Impact of M on the number of clusters produced by RanMIS w.r.t. network size andits density.

rounds required by each protocol to complete and also the total number of messages transmitted.In Figures 13 and 14, we see the relative comparison of the algorithms.

As far as the number of protocol rounds is concerned, we observe that DCA is the clearwinner for all network instances. The performance of the algorithms is consistent with the whatthe theory predicts for them; thus we do not comment further on this issue.

As far as the number of transmitted messages for backbone construction is concerned, we ob-serve that RanMIS is now the clear winner confirming its theoretical behavior. Its performancegap from the other competitors is stable across all network topologies. DCA sends around 40%more messages than RanMIS does, and WuLi sends around 60% more messages than RanMIS .Therefore, even though it takes more time to RanMIS to construct the backbone due to its prob-abilistic nature, it does this with far less messages than its competitors.

Results concerning the backbone description. In Figures 15–22 we present the performanceof the protocols for the metrics belonging to the category of “backbone description”.

The number of generated clusters by each protocol is presented in Figure 15. The firstobservation is that RanMIS and DCA produce almost the same number of clusters; even thoughtheir difference is indistinguishable in the figure, their difference is less than 1% which can beattributed to statistical variation. The reason for this resemblance is that both algorithms arebuilding maximum independent sets using either a weight for each node (assigned dynamicallyor statically) in the case of DCA or using randomization in choice in the case of RanMIS . Onthe other hand, WuLi produces too many clusters, since it is mandatory to create connecteddominating sets, whereas the former two algorithms are building plain dominating sets. Finally,we observe that the number of generated clusters reduces with increasing density, since in densenetworks more nodes are able to find a CH (i.e., dominator) in their 1-hop neighborhood.

36


Figure 13: Rounds required by each competitor algorithm to complete backbone constructionw.r.t. network size and its density.

Figure 14: Messages required by each competitor algorithm to complete backbone constructionw.r.t. network size and its density.

Next, we evaluated the cluster cardinality distribution for various densities of the network.The results are illustrated in Figures 16–19. At this point we should emphasize that this dis-tribution should in general be ‘bell-shaped’ or look like any distribution with probability mass

37


Figure 15: Clusters produced by each competitor algorithm w.r.t. network size and its density.

concentrated around the average node degree of the topology. In Figures 16 and 17, we seethat this property is achieved by RanMIS and DCA for sparse networks, but it is not achievedfor dense networks (in Figures 18 and 19). In these last two figures, it seems that a significantpercentage of clusters – around 22% of them – contain very few nodes, i.e., 1 or 2, including theclusterhead. On the other hand, WuLi exhibits an undesirable pattern across all network topolo-gies, since more than 70% of its clusters contain 1 or 2 nodes, which is also a consequence ofthe large numbers of clusters produced (see Figure 15).

The diameter2 length of the produced backbone by each protocol is presented in Figure 20 (inmeters) and in Figure 21 (in hops). The common observation is that WuLi is the best performingalgorithm which is due to the fact that it produces connected dominating sets and thus there areno gateway nodes intervening among CHs so as to increase the diameter. The second observationis that the diameter in general decreases with increasing network density, since in dense networksit easier to find short routes for any pair of nodes. In some cases where the topology is peculiar,the diameter increases in denser networks (e.g., network with 500 nodes with density equal to 7).

In Figure 22 we illustrate the percentage of adjacent CHs that reside at a distance of 3-hopsfrom each other. Apparently, WuLi produces clusterheads at 1-hop distance from each other,and therefore we do not include it in the plot. For the other two algorithms, which producemaximum independent sets, it holds by the definition of the MIS, that any two adjacent CHs willbe either at a distance of 2 hops or at a distance of 3 hops, with the latter being the preferredone (see paragraph on ‘Inter clusterhead distance distribution’ definition, Section 3.1.3). Wesee that both algorithms achieve approximately the same performance – the gap among them isstatistically insignificant. In particular, we see that the majority (more than 50%) of adjacentCHs in almost all network topologies, are 3-hops away. The only exception appears for very

2The largest of the shortest paths for any pair of nodes.

38


Figure 16: Distribution of nodes in clusters after backbone construction w.r.t. network size anda density of D = 4.


small and very sparse networks (i.e., 100 nodes with degree 4), where this percentage dropsto 48%; this is due to the fact that there are not too many links among the nodes to establish3-hop distance among neighboring CHs.

39




3.2 The SPRING clustering algorithm

The idea is based on force-directed algorithms. The force-directed assign forces amongthe set of edges and the set of nodes in a network. The most straightforward method is to

40


Figure 20: Network Diameter diversification (in meters) after backbone construction w.r.t. net-work size and its density.

Figure 21: Network Diameter diversification (in Hops) after backbone construction w.r.t. net-work size and its density.

assign forces as if the edges were springs and the nodes were electrically charged particles.The entire graph is then simulated as if it were a physical system. The forces are applied tothe nodes, pulling them closer together or pushing them further apart. Every node apply to

41


Figure 22: CH Distance Distribution after backbone construction w.r.t. network size and itsdensity.

its neighbors a force Frel according to their distance and their velocities. Vehicles that moveto the same direction or towards each other apply positive forces while vehicles moving awayapply negative forces. Components of the vector Frel along the east-west Fx and north-southFy axes are calculated. In order to form stable clusters only vehicles that move to the samedirection or towards each other are considered as candidate cluster members. For a specificvehicle that the total magnitude of forces applied to it is negative no clustering procedure istriggered since all the surrounding nodes tend to move away from it. Calculating total forceF helps to avoid re-clustering in many situations - for example, when groups of vehicles moveaway from each other. All nodes are equipped with GPS receivers and On Board Units (OBU).Location information of all vehicles/nodes, needed for clustering algorithm is collected with thehelp of GPS receivers. The only communications paths available are via the ad-hoc network andthere is no other communication infrastructure. The Maximum Transmission Range (R) of eachnode in the vehicular network environment is 250 meters.

3.2.1 Network identification

Neighborhood identification is the process whereby a vehicle/node identifies its currentneighbours within its transmission range. For a particular vehicle, any other vehicle that iswithin its radio transmission range is called a neighbour. All vehicles consist of neighbour setwhich holds details of its neighbour vehicles. The neighbours set is always changing since allnodes are moving. The neighbor set Ni of vehicle i is dynamic and is updated frequently. Everymoving node keeps track of all current neighbors (their id’s) the current and the past distance.

Generally,neighbour node identification is realized by using periodic beacon messages. The

42


beacon message consists of node Identifier (ID), node location, speed vector in terms of relativemotion across the axes of x and y (dx,dy) total force F , state and timestamp. Node location isused in order to calculate the distance between the nodes. Each node informs other nodes of itsexistence by sending out beacon messages periodically.

All nodes within the transmission range of source/packet carrier node will intimate theirpresence by sending beacon messages frequently. After the reception of a beacon, each nodewill update its neighbour set table. For a neighbor that already exists in its neighborhood onlythe current and the past distance are updated. If a node position is changed, then it will updateits position to all neighbours by sending a beacon signal. If a known neighbour, times out, it willbe removed from the neighbour set table. The total number of neighbors of a given vehicle iscalled the ”active neighborhood set” Ni of the vehicle.

3.2.2 Clustering process and protocol structure

As described in the previous section the beaconing thread is responsible for exchanginginformation between neighboring nodes. Another task of this thread is processing and properuse of the messages received from other nodes. Each node constantly updates knowledge aboutneighboring nodes. Each node i using the information of the beacon messages calculates thepairwise relative force Freli j for every neighbor applied to every axes j using the Coulomb law.

Freli jx = ki jxqiq j

r2i j

, Freli jy = ki jyqiq j

r2i j

(1)

where ri j is the current distance among the nodes ki jx (ki jy) is a parameter indicating whether theforce among the nodes is positive or negative depending on whether the vehicles are approachingor moving away along the corresponding axis and qi and qj may represent a special role of anode (e.g. best candidate for Cluster head due to being close to an RSU, or due to following apredefined route (bus)).

In coulombs law a positive force implies it is repulsive, while a negative force implies it isattractive. In our implementation a positive force symbolizes the fact that the specific pair ofnodes is approaching or is moving towards the same direction while a negative force is appliedto nodes that move to different directions. Every node computes the accumulated relative forceapplied to it along the axes x and y and the total magnitude of force F . According to the currentstate of the node and the relation of its F to neighbor’s F , every node takes decisions aboutclustering formation, cluster maintenance and role assignment.

A node may become a clusterhead if it is found to be the most stable node among its neigh-borhood. Otherwise, it is an ordinary member of at most one cluster. When all nodes firstenter the network, they are in non-clustered state. A node that is able to listen to transmissionsfrom another node which is in different cluster can become a gateway. We formally define thefollowing term: (1) relative mobility parameters ki jx and ki jy.

DefinitionRelative mobility parameters ki jx and ki jy between nodes i and j, indicate whether they are

moving away from each other, moving closer to each other or maintain the same distance fromeach other. To calculate relative mobility, we compute the difference of the distance at time, tand the possible distance at time, t +dt for every axis.

43


Relative mobility at node i with respect to node j is calculated as follows:We calculate the distance at every axes between the nodes at time t and the possible distance

at time t +dt according to,

Dcxi j = xi − x j, Df xi j = xi +dxi − x j −dxj (2)

Dcyi j = yi − y j, Df yi j = yi +dyi − y j −dyj (3)

The relative movement dx and dy of every vehicle along the axes x and y are calculated bythe vehicles OBU according to previous data received from the GPS with respect to the trafficahead. According to mobility in every axis relative mobility ki jx and ki jy are calculated accordingto:

i f Dcxi j ≤ Df xi j then ki jx =−axdt. (4)

i f Dcxi j ≥ Df xi j then ki jx = axdt. (5)

where ax and ay are given by

i f Dcxi j ≤ Df xi j then ax = Df xi j −Dcxi j (6)

i f Dcxi j ≥ Df xi j then ax =1

Dcxi j −Df xi j(7)

Parameters ax and ay indicate the significance of the force applied between the vehiclesby reflecting the ratio of divergence or convergence among moving nodes. In equation 6 ax isproportional to the divergence among nodes, since the faster it takes place the more negative theforce must be. In equation 7 ax is proportional to the reverse difference of the distance amongthe nodes, since nodes that approach each other in a fast pace wont probably stay in contact fora sufficient amount of time in order to form cluster and exchange information.

3.2.3 Special role of vehicles

The pairwise relative force Freli j for every pair of nodes depends on the relative mobilityKreli j, the current distance and parameters qi and qj which indicate a special role for the vehi-cles. In the cluster creation procedure, nodes that driver intentions can be predicted like truckdrivers that keep an almost constant velocity must be favored to become clusterheads. Also inurban areas vehicles that follow the same routes constantly may act as clusterheads in such adynamic environment. In a street with many lanes, cars that follow the non turn lane are alsobest candidates for clusterhead since they are expected to stay longer on the street (Figure 24).

44


Current positions of vehicles Future positions of vehicles

Uj = 50 Km/h

Node j

Node iUi = 100 km/h

Node j

Node i

Figure 23: Relative mobility at node i with respect to node j.

Clusterhead

Node turns

New clusterhead

Node turns

ReclusteringNo Reclustering

Figure 24: The correct choice of the clusterhead plays significant role

Historical data about driver behavior may be used in order to select the appropriate cluster-heads. In Sp-Cl parameter qj is used to favour vehicles to become clusterheads. Using Equa-tion 1 to compute the relative force between two nodes, parameter qj is used as follows:

i f ki jx ≥ 0 then qi = 2, i f ki jx ≤ 0 then qi = 1/2 (8)

Positive forces applied to these nodes are strengthened while negative are weakened, inorder to facilitate this node to become a clusterhead. The correct choice of the clusterhead isvery important for the stability of the method, the cluster lifetime and the overhead involved informing and maintaining these clusters.

3.2.4 Cluster-head election parameters

Vehicles use beacon messages in order to broadcast information to neighboring nodes suchas Identifier (ID), node location, speed vector in terms of relative motion across the axes ofx and y (dx, dy) total force F, state and timestamp. Using this information as stated abovenodes calculate the forces applied to each other according to position and relative mobility. Themobility information of the neighbors is needed for the vehicle to initiate the cluster formationrequest, while cluster-head election information for any node is limited to the nodes that arewithin range. After receiving information of all neighboring vehicles, node i calculates:

Fx = ∑j∈Ni

Freli jx and Fy = ∑j∈Ni

Freli jx (9)

45


which is the total force along axes x and y applied to it. The total value of forces (norms) iscalculated for every node according to:

F = |Fx|+ |Fy| (10)

Total force F is used to determine the suitability of a vehicle to become clusterhead accordingto the following criteria:

• The suitability value of the vehicle is calculated by considering the mobility informationof its neighbors (parameters ki jx and ki jy)

• Nodes having higher number of positive neighbors (Freli jx ≥ 0 Freli jy ≥ 0), maintainingcloser distances to their neighbors, should have be qualified to be elected as cluster-heads.

3.2.5 The cluster formation algorithm

In order to execute the algorithm, each vehicle is assumed to maintain and update the Ni

set. At any time each vehicle i recalculates total F and according to total non-clustered memberswithin range try to form a cluster and become the clusterhead.

If the node has the biggest positive force applied to it and there exist at least one free nodein its neighborhood, it declares itself to be a clusterhead. In the opposite situation, where thereexist a free node j with biggest total F in range the vehicle becomes a cluster member of j. Thisalgorithm leads to the formation of clusters which are at most two hops in diameter.

3.2.6 Cluster maintenance

The cluster maintenance procedure follows the following rules:

• For every free node.

When a standalone (non-clustered) vehicle comes within R distance from a nearby cluster-head, the cluster-head and the vehicle compare the total force f applied to them. If therelative force F of the clusterhead is bigger than that of the free node then the cluster-headwill accept the vehicle and will add it to the cluster members list. If the clusterhead hassmaller F then no action is triggered.

If there are other nearby standalone vehicles it compares the values of F in order to forma new cluster.

• For every member node.

If a member node at a certain time finds itself to have bigger F than any of the surroundingclusterheads then it becomes a free node and tries to form its own cluster.

When a cluster member moves out of the clusterhead’s transmission range. it is removedfrom the cluster members list maintained by the cluster-head and it becomes a free nodeagain.

46


• For every clusterhead.

when two cluster heads come within each other’s transmission ranges and they stay con-nected over a time period CCI the cluster merging process takes place. The clusterheadwith the lower F gives up its cluster-head role and becomes a cluster-member in the newcluster.

3.2.7 Demo

An extensive simulation study was conducted to evaluate the performance of our protocol.A custom simulator was used to evaluate our method. In our simulation, we consider differentroad traffic and different network data parameters. The simulation environment is a one direction5-lane highway with a turn in order to evaluate the performance of the scheme.

Scenario The total length of the highway is 2 Km. The stationary LPGs created in eachscenario are of size comparable to communication range of nodes, i.e., if the communicationrange of the vehicles is 80 meters each LPG is of 160 meters long as if the RSU has range of 80meters.

The arrival rate of the vehicles follows the Poison process with parameter λ . The speedassigned to the vehicles is according to the lane it chooses to follow according to table 4.

Lane Speed km/h

1 802 1003 1204 1405 160

Table 4: Speed per lane.

The density of the vehicles depends on parameter λ . The number of vehicles per lane isbetween (2 -15 v/km/Lane) depending on the speed being used and the value of parameter λaccording to table 5.

parameter λ ν/km/Lane

3 8-155 5-97 3-69 2-5

Table 5: Speed per lane.

47


Figure 25: Simulation environment

3.2.8 Evaluation

Cluster stability.In order to evaluate the stability of the clustering algorithm, we measure the stability of thecluster configuration against vehicle mobility. In a high dynamic VANET, nodes keep joiningand leaving clusters along their travel route. Good clustering algorithms should be designedto minimize the number of cluster changes of the vehicle by minimizing reclustering. Thistransitions among clusters are measured in order to evaluate the performance of the algorithm.The basic transition events the vehicle encounters during its lifetime:

• A vehicle leaves its cluster and forms a new one (becomes a clusterhead).

• A vehicle leaves its cluster and joins a nearby cluster or becomes free.

• A cluster-head merges with a nearby cluster.

We compare the average transition events of the vehicles for the Sp-Cl, Low-Id and LPGmethods when different speeds and different transmission ranges are used. From Figure 26, wecan see that the average transitions produced by our Sp-Cl technique is smaller compared to thatproduced by the Low-ID and LPG methods. This means our technique causes less number ofcluster transitions for all different density topologies. Similar figures were produces for othertransmission ranges.

The figures show that the average transitions of the vehicle decreases as the transmissionrange increases. This is because increasing the transmission range, increases the probability thata vehicle stay connected with its cluster-head.Number of clusters.Due to high dynamics of the VANET, clusters are created (new clusters added to the system)and vanished over time. The total number of clusters created over a period of time is a metricof the stability of the clustering method used. Good clustering algorithms should be designedto reduce the rate at which clusters are created and added to the system due to the mobility ofthe nodes. The ability of the clustering method to maintain cluster structure despite vehicle’smobility defines its performance. In this simulation we counted the new clusters which areadded to the system.

48


2

3

4

5

6

7

8

9

3 5 7 9

Ave

rage

num

ber

of c

lust

ers

chan

ged

per

vehi

cle

Density

Sp-ClLow-Id

LPG

2

4

6

8

10

12

3 5 7 9

Ave

rage

num

ber

of c

lust

ers

chan

ged

per

vehi

cle

Density

Sp-ClLow-Id

LPG

Figure 26: Average cluster change per vehicle for Sp-Cl, LPG and Low-Id methods for different trans-mission ranges(125m, 80m).

35

40

45

50

55

60

65

70

75

80

3 5 7 9

Num

ber

of c

lust

ers

Density

Sp-ClLow-Id

55

60

65

70

75

80

85

90

95

100

105

3 5 7 9

Num

ber

of c

lust

ers

Density

Sp-ClLow-Id

Figure 27: The average total number of formed clusters for Sp-Cl and Low-Id methods for differenttransmission ranges(125m, 80m).

The figure shows that the total number of clusters created by Sp-Cl is always smaller com-pared to that produced by the Lo-Id method and this number decreases as the transmission rangeincreases.This is because the Sp-Cl method uses the accumulated forces among as a parameterto create the clusters. Thus, the clusters are more stable and have longer lifetime.Cluster lifetime.The average cluster lifetime is an important metric that shows the performance of the clusteringalgorithm. The cluster lifetime is directly related to the lifetime of its cluster-head. The cluster-head lifetime is defined as the time period from the moment when a vehicle becomes a cluster-head to the time when it is merged with a nearby cluster.

The average cluster lifetime produced by the Sp-CL and the Low-Id methods is evaluated invarious topologies with different transmission ranges.

3.3 Sociological vehicle clustering

3.3.1 Motivation

The technique of clustering is widely investigated in the context of mobile ad hoc net-works [87], and in sensor networks [86]. For both types of networks, and for any kind of ad

49


30

35

40

45

50

55

60

65

70

75

3 5 7 9

Ave

rage

num

ber

of c

lust

ers

chan

ged

per

vehi

cle

Density

Sp-ClLow-Id

35

40

45

50

55

60

65

3 5 7 9

Ave

rage

Clu

ster

life

time

Density

Sp-ClLow-Id

Figure 28: The average cluster lifetime Sp-Cl and Low-Id methods for different transmissionranges(125m, 80m).

hoc network, it brings significant benefits that can be summarized as follows: a) alleviates thebroadcast storm problem [76] which results in reduced congestion, and packet losses, b) de-creases packet delivery latency, c) provides better spectrum utilization in time and space, d)allows for data aggregation, and e) increases network longevity. A search for clustering pro-tocols for these two types of networks will reveal the existence of several hundreds of articles.Thus arises the question of whether VANETs need new clustering techniques. The answer isaffirmative because VANETs are characterized by unlimited power, high but constrained (due tothe road network) mobility, and human sociological factors.

A quick inspection of the literature on MANET and sensor networks clustering will showthat the great majority of these clustering protocols has as primary aim to reduce energy con-sumption in order to increase the network lifetime. Therefore, all these algorithms are not appro-priate for the VANET environment. Some proposals, such as the GESC protocol [21], exploit thetopological relations of nodes in order to detect those nodes that are significant in carrying outcommunication tasks, but these algorithms are not appropriate for highly mobile nodes whichare encountered in vehicular environments.

A significant body of work on MANET clustering is based on the unique IDs of nodes inorder to build connected dominating sets or independent sets and subsequently clusters e.g.,[9, 27]. These ideas have been transferred in the VANETs environment as well resulting in clus-tering protocols such as the MOBIC [11]. The main disadvantage of this category of algorithmsis that they are almost completely road network’s topology-agnostic, exploiting only vehicle’sIDs. Some other vehicle clustering protocols are based on complex data mining-inspired proce-dures, e.g., [34, 63] which are hard to be deployed in any realistic VANET. Some protocols suchas those reported in [52, 60, 68, 88] incorporate the mobility of the nodes into the clusteringprocedure but they do so at a very constrained sense, taking into account only the road networktopology; they ignore the true “intentions” of the vehicles (actually, of their drivers) which is theprimary reason of their mobility. Finally, some protocols are only appropriate for highways e.g.,[60], others only appropriate for urban environments e.g., [44].

Collectively, the existing proposals for vehicle clustering suffer from one or more of thefollowing shortcomings: a) they are not generic enough to be used for both urban and high-way scenarios, b) they are based on sophisticated and unpractical data mining procedures with

50


many hard-to-set administrative parameters, c) they do not exploit the road network’s and/orthe VANET’s topology at all, and d) they exploit at a very localized manner the “intention” ofthe mobility (i.e., speed, direction) which may present significant variations thus confusing theclustering protocols and making suboptimal clustering decisions which harm both the clusterstability and effectiveness.

Here we propose a novel vehicle clustering protocol that avoids the aforementioned short-comings and tries to incorporate the best features of the major vehicle clustering families. Atthe heart of the protocol are the social aspects of vehicles moving in a city or in highway; theirtendency to follow the same routes because their drivers have some final destination in mind.Therefore, our method is based on ‘mining’ the relationships (in terms of movements) of groupsor individual vehicles. The study of the relationships between groups in human societies aretraditionally conducted in the field of sociology. Thus, we will use the term ’sociological’ todescribe the analysis and mutual relation of vehicle (and of their human drivers) trajectories.Such sociological aspects have reported in several studies [23, 32, 53]. The implementation ofthis idea is based on the simple, solid mathematical theories.

In particular, we develop two clustering policies, namely the Sociological Pattern Clustering(SPC) and its specialization, namely the Route Stability Clustering (RSC). Statistical informa-tion gathered by Road-Side Units (RSUs) located in the boundaries of the area of interest areused in order to build the sociological profile of every vehicle. This profile is used in order tocreate clusters with neighbors that will (high probability) have similar behaviors. Therefore, thisapproach makes the following contributions:

• It exploits the “macroscopic” social behavior of vehicles, for the first time in the clusteringliterature.

• It combines this macroscopic behavior with “microscopic” behavior which is based onan earlier proposal under the concept of virtual forces [45] in order to create stable andbalanced clusters.

• Based on this two-level behaviors, it develops the Sociological Pattern Clustering (SPC),and the Route Stability Clustering (RSC) clustering protocols.

• It evaluates the performance of the proposed clustering techniques against the most popu-lar clustering methods for VANETs. The evaluation is done for a large range of parametersand values:

– for both urban and highways scenarios,

– for different transmission ranges and vehicles speeds,

– for varying social behaviors, and so on.

3.3.2 Network model: Definition of the system

Definition 1 Let S = {S1,S2, ...,SM} represent the set of road segments in a given geographi-cal area or map. A road segment Si is represented by a unidirectional edge between two consec-utive junctions.

51


Definition 2 Let V = {V1,V2, ...,VN} be the set of vehicles that are traveling in the givengeographical area during a certain time period.

Definition 3 Let TP = {TP1,TP2, ...,T PK} be the set of Time Periods that the investigatedsystem is segmented.

3.3.3 Network model: Road network communities

In the past few years, complex networks [32, 21, 23] have been studied across many fieldsof science. A number of network features have been discovered, among which the properties ofhierarchical topology and community structure have attracted a great deal of interests recently.Communities are groups of vertices within which connections are dense but between which theyare sparser. Networks often show a hierarchical structure of communities nested within eachother. Identification of these communities can provide insight into better understanding andvisualizing the structure of networks, and applications have ranged from technological networksto biological networks and social networks. In a road network where streets are mapped asedges and intersections as vertices, the intersections that locate close in a small region are morelikely to form a community. The network is then decomposed, with adjacent subnetworks beingloosely connected by the intergroup edges. Many approaches focus more on how to partitionand manage a network such that the number of boundary – border nodes for each subnetworkis uniform and minimized, the subnetworks are approximately of equal size, and so on. We usepartitioning based on [66] where each subnetwork forms an isolated part, and different parts areconnected together via intergroup arcs (arcs that are incident to/from boundary nodes and do notbelong to any subnetwork). The city is then partitioned in small areas that can be investigated inisolation.

3.3.4 Network model: Collection of personalized data

RSUs are assumed to be located at fixed locations at the borders of the region of interest.As vehicles move through the VANET, they record every road segment Sj they traverse in theorder of arrival to those segments. Every second, each RSU will broadcast a short message(DENM) to all vehicles in its vicinity, which will query each vehicle to send their collected setof segments. Upon receipt, the vehicles will create a packet containing the partial path collectedas well as other attributes. Each vehicle has a unique identifier Vi.

Privacy preservation is critical for vehicles. In the vehicular context, privacy is achievedwhen two related goals are satisfied: untraceability and unlinkability [28]. First property statesthat vehicle’s actions should not be traced and second that it must be impossible for an unautho-rized entity to link a vehicle’s identity with that of its driver/owner. On the other hand no trafficregulation or congestion avoidance can be achieved if this privacy protection is not removed.Data concerning owner’s identity of a given vehicle and the path followed along a period oftime are crucial for building that vehicle’s social profile. Security mechanisms should preventunauthorized disclosures of information, but applications should have enough data to work prop-erly [19].

52


3.3.5 Mobility of nodes and semi-Markov Model

We model the mobility of vehicle i with a time homogeneous semi-Markov, with discretetime. The states are represented by the road segments.

The reason for using semi-Markov processes (rather than continuous-time Markov chains)for modeling user mobility is because the sojourn time during which a user is traveling alonga road segment does not always follow the exponential distribution. A semi-Markov processallows for arbitrary distributed sojourn times and can be viewed as a process with an embeddedMarkov chain, where the embedded points are the time instants when a user travels along a roadsegment. A node that moves between two road segments, transitions in the Markov processbetween the corresponding states. We assume the transition probabilities between states havethe Markov memoryless property, meaning that the probability of a node i transiting from stateV i

j to state Vij+1 is independent of state Vi

j−1

A =

⎛⎜⎜⎜⎜⎝

0 a12 0 a14 00 0 a23 a24 00 0 0 0 a35

0 0 a43 0 a45

1 0 0 0 0

⎞⎟⎟⎟⎟⎠

As a vehicle enters a state (road segment) j, it stays there for a time called state holding timeTj,TPk , and then leaves to the next state j′. Note that this sojourn time does not include the timewhen the nodes are in transit between road segments. In order to avoid having absorbing stateswe perform wrap around, connecting each exit state with every entry state with connections thathave equal probabilities.

Exit node

T1

T2

s1

s3

s5s4

T5T4

s2 T3Entry node

Figure 29: Semi-Markov model of vehicle vi in a simple network topology with one entry andone exit.

The state holding time effectively depends on the vehicle speed and road condition (e.g.congestion) that we assume is constant in our application for every time period TPk.

53


3.3.6 Transition probability matrix

A is the transition probability matrix of the embedded Markov chain for vehicle Vi for TimePeriod TPk. Figure 29 shows an example transition probability matrix for vehicle Vi that movesaround.At any of the road segments move to another road segment according to its preferredprobability.

For example, if the node is at s2, it can then

• move to the s4 with probability a24

• or stay in state S2 for time T2,TPk

• or go to the s3 with probability a23

Those mobility probabilities constitute the transition probability matrix A. Note that each nodehas its own transition probability matrix that reflects its trajectory preference for the investigatedtime period TPk.

3.3.7 Equilibrium probabilities

The equilibrium probabilities of the embedded Markov process can be calculated accordingto Equations 11 and 12

π j =M

∑i=0

πiai j, j ≥ 0 (11)

M

∑j=0

π j = 1 (12)

If Tj,TPk is the mean sojourn time at state j for investigated time period k then the equilibriumprobability of the semi-Markov process at that state is calculated using the probabilities of theembedded process using Equation 13

pj =π jTj,TPk

∑i πiTi,TPk

(13)

3.3.8 Communication phase: Vehicle leaving the area of interest

As vehicle Vi leaves the region of interest, goes into the control range of the intersection,RSU device near the boundary of the control range impels this vehicle to send the collected setof segments, also known as partial path PPi and travel time Tj for each road segment j traversedalso known as partial time path PTi. Upon receipt, the vehicles will create a packet containingthe partial time path collected as well as other attributes as shown in Figure 30.

54


pcktype pckTimePeriod

Vehicle Id

PartialPathSize PTi

PPi

Figure 30: Vehicle packet design.

Each RSU has a database that contains all the partial paths that the vehicles traversed in theinvestigated area. The database consists of a separate table PPTik for every vehicle i and forevery Time Period k. That means that for a single vehicle there may be several tables one foreach Time Period according to the segmentation of time.

At the RSU side, received collected partial paths will be added to the RSUs’ databasesaccording to the nature of the received packet. If the received packet contains a vehicle ID thatdoes not exist in the RSU database, the RSU will create a new table for this vehicle ID. Howeverif the vehicle ID in the packet already exists, the partial path is appended to the existing vehicletable PPTik. Every β seconds, the RSUs can provide each other with the information collected.

The investigation time is segmented to time periods TP. It has been observed [65] that inpractice, weekdays and weekends usually exhibit significantly different traffic conditions whileit is shown that all weekdays (weekends) have similar congested and congestion-free trafficpatterns; therefore we group the days and treat different groups separately. The time periodsare pre-dawn: (up until 8AM), morning rush-hour (8-10AM), late morning (10-noon), earlyafternoon (noon-4), evening rush-hour (4-7PM), and night time (after 7PM). The Paths tablePPTi j, shown in Table 6, represents the set of vehicles movement patterns during their trips inthat monitored area. As shown below, each vehicle Vi will have a table in this database describingits movement paths for each Time Period. RSU uses this data to compute the mean holding timeTSj ,TPk for each time road segment Sj and for each time period TPk (see subsection 3.3.10).

VID Partial Paths

V1 [S1,S2,S3,S5]

V1 [S1,S2,S4,S4]

V1 [S1,S4,S3,S5]

Table 6: Path table of vehicle i for time period j.

According to these paths the transition probabilities of table A are updated for vehicle Vi forthe Time Period TP that the vehicle entered the region of interest. That way every vehicle has aunique semi-Markov model for every time period called SMVi,TPk .

3.3.9 Communication phase: Vehicle entering the area of interest

As vehicle Vi enters the region of interest, goes into the control range of the intersection,RSU device near the boundary of the control range sends to the vehicle data (DENM message)

55


that is extracted from its semi-Markov model for the specific time period SMVi,TPk . Since all thecomputations take place off the vehicle and on the RSU the information that the vehicle getsfrom the RSU is limited. This information is either the social number SN - flow of the vehicle orRoute stability metric RN according to the clustering method that the vehicle is going to follow.If the vehicle enters the area of interest for the first time then RSU assigns to the vehicle a socialnumber SN or a Route stability RN metric which is the mean value for the specific Time Period.This data is just one single floating number that is embedded in a simple beacon message.

3.3.10 Communication phase: Mean sojourn times

Each vehicle packet that is transmitted to the RSU contains the time that the vehicle spent oneach road segment traveled for the specific time period that it was in the area of interest. Thesepartial time paths (PTi) are used in order to calculate mean travel time for each road segment jfor each time period k. For each new message new values are added to the table RST . One newrow is added for each distinct road segment, time period and travel time as shown in Table 7.

Road Segment j Travel Time Time Period k

S j T1 TPk

S′j T2 TP′k

S′′j T3 TP′′k

Table 7: Road segment travel times table RST .

Mean travel time for a specific road segment j for a time period k is the sum of all values ofthe second column from Table 7 according to Equation 14.

TSj ,TPk =∑i RST (2, i)

L, RST (1, i) = s j and RST (3, i) = TPk (14)

, where L is the number of rows of Table 7 that satisfy the above constraints.

3.3.11 Sociological patterns

As it was described earlier as a vehicle Vi leaves the area of interest informs the RSU inrange, and through it the central database, about the path it followed during its stay in the area.This path in inserted in the table PPTik that contains all the paths of the vehicle for the specifictime period.

Using these paths, the transition table of Vehicle is created and from it the social patternsof the vehicle are extracted (Figure 31). As shown in Figure 31 the Social Pattern in created bystarting at each entry segment in the network and by following the most possible next transitionof the transition table of vehicle Vi for the specific time period.

Once the social patterns are extracted, a table SPk than contains all the social patterns for thespecific time period in updated:

• If Vi has previous Social patterns then these values are deleted from SPk.

56


0

0

0

0

0 0

0

0

0

0 0

0 0

0

0 0

a

a

a

a

a

a

a

a

12 14

23 24

35

0 4543

53

SP:S1,S2,S4,S5

Figure 31: Transition table → social pattern.

• The Social Pattern is compared to all the social patterns that exist in the central databaseand in the table for the specific Time Period. If the Social Pattern already exists in thedatabase then the vehicle is assigned the number of it and in the Table SPk the Vehicle IDis appended in the corresponding line. If this social pattern is new, the social Pattern isappended to the table and a new number is assigned to it.

This procedure is represented in Figure 32.After the end of this procedure Vehicle Vi is assigned several Social Pattern values, one for

each Time Period and for each entry road segment to the region of interest. According to thesevalues, the time and the road segment that the vehicle uses the next time to enter the region ofinterest a unique Social Number SN is assigned to it. This number is used in order to performclustering and create groups with vehicles that share common habits and behaviors.

Transition Probabilitiesare updated

Received Partial Path

RSU

Social Pattern Vehicle IDs

Social Patterns of Viare extracted

Social Pattern Table (SPk)

Update SPk

Partial Path Table (Vi, PPTik)

Figure 32: Procedure for social patterns extraction.

3.3.12 Social Clustering

In order to create clusters the basic mechanism of Virtual Forces Vehicular Clustering (VFVC)[42, 45, 44] is used. The basic idea lies in modeling vehicles as electrically charged particles.

57


Every node applies to its neighbors a force Frel according to their distance and their relativevelocities. Vehicles that move to the same direction or towards each other apply positive forceswhile vehicles moving away apply negative forces. In order to perform clustering nodes periodi-cally broadcast beacon (CAM) messages. Each beacon message consists of node Identifier (Vid),node location, speed vector, total force F , state and time stamp.

Each node i using the information of the beacon messages calculates the pairwise relativeforce Freli j for every neighbor applied to every axes j using the Coulomb law.

Freli jx = ki jxqiq j

r2i j

, Freli jy = ki jyqiq j

r2i j

(15)

where ri j is the current distance among the nodes ki jx (ki jy) is a parameter indicating weather theforce among the nodes is positive or negative depending on whether the vehicles are approachingor moving away along the corresponding axis and qi and qj may represent a special role of anode (e.g. best candidate for Cluster head due to being close to an RSU, or due to following apredefined route (bus).

In Coulomb’s law a positive force implies it is repulsive, while a negative force implies itis attractive. In our implementation a positive force symbolizes the fact that the specific pair ofnodes is approaching or is moving towards the same direction while a negative force is appliedto nodes that move to different directions. Every node computes the accumulated relative forceapplied to it along the axes x and y and the total magnitude of force F . According to the currentstate of the node and the relation of its F to neighbor’s F , every node takes decisions aboutclustering formation, cluster maintenance and role assignment.

A node may become a clusterhead if it is found to be the most stable node among its neigh-borhood. Otherwise, it is an ordinary member of at most one cluster. When all nodes first enterthe network, they are in non-clustered state. A node that is able to listen to transmissions fromanother node which is in different cluster can become a gateway. We formally define the follow-ing term: relative mobility parameters ki jx and ki jy. Relative mobility parameters ki jx and ki jy

between nodes i and j, indicate whether they are moving away from each other, moving closerto each other or maintain the same distance from each other. To calculate relative mobility, wecompute the difference of the distance at time, t and the possible distance at time, t+dt for everyaxis.

Relative mobility at node i with respect to node j is calculated as follows:We calculate the distance at every axes between the nodes at time t and the possible distance attime t +dt according to,

Dcxi j = xi − x j, Df xi j = xi +dxi − x j −dxj (16)

Dcyi j = yi − y j, Df yi j = yi +dyi − y j −dyj (17)

The relative movement dx and dy of every vehicle along the axes x and y are calculated by thevehicles OBU according to previous data received from the GPS with respect to the traffic ahead(Figure 23). According to mobility in every axis relative mobility ki jx and ki jy are calculatedaccording to:

i f Dcxi j ≤ Df xi j then ki jx =−axdt. (18)

58


i f Dcxi j ≥ Df xi j then ki jx = axdt. (19)

where ax and ay are given by

i f Dcxi j ≤ Df xi j then ax = Df xi j −Dcxi j (20)

i f Dcxi j ≥ Df xi j then ax =1

Dcxi j −Df xi j(21)

Parameters ax and ay indicate the significance of the force applied between the vehicles byreflecting the ratio of divergence or convergence among moving nodes. In Equation 6 ax isproportional to the divergence among nodes, since the faster it takes place the more negative theforce must be. In Equation 7 ax is proportional to the reverse difference of the distance amongthe nodes, since nodes that approach each other in a fast pace wont probably stay in contact fora sufficient amount of time in order to form cluster and exchange information.

Current positions of vehicles Future positions of vehicles

Uj = 50 Km/h

Node j

Node iUi = 100 km/h

Node j

Node i

Figure 33: Relative mobility at node i with respect to node j.

After receiving information of all neighboring vehicles, node i calculates

Fx = ∑j∈Ni

Freli jx and Fy = ∑j∈Ni

Freli jx, (22)

which is the total force along axes x and y applied to it. The total value of forces (norms) iscalculated for every node according to:

F = |Fx|+ |Fy| (23)

Total force F is used to determine the suitability of a vehicle to become clusterhead accord-ing to the following criteria:

59


• The suitability value of the vehicle is calculated by considering the mobility informationof its neighbors (parameters ki jx and ki jy)

• Nodes having higher number of positive neighbors (Freli jx ≥ 0 Freli jy ≥ 0), maintainingcloser distances to their neighbors, should have be qualified to be elected as cluster-heads.

A node may become a clusterhead if it is found to be the most stable node among its neighbor-hood. Otherwise, it is an ordinary member of at most one cluster.

When all nodes first enter the network, they are in non-clustered state. Nodes having highernumber of positive neighbors in terms of relative force Freli j, maintaining closer distances totheir neighbors, are qualified to be elected as cluster-heads. In the initial method [44] a lanedetection algorithm is used to determine the lane the vehicle moves on. According to the lanebeing a turning or a non turning one the method favor vehicles that follow a non turning lane tobecome clusterheads. This method produces stable clusters when focusing on what happens ona central road, where cars enter and leave it all the time. In a more realistic scenario, a large areaof a city, the long lifetime of clusters should not be limited on main roads but all clusters createdin the city must be as stable as possible.

In order to create stable clusters we use the social behavior of vehicles based on historicaldata collected from RSU’s that scattered along the borders of the area of interest. In order toincorporates the social behavior of the vehicles when move in Urban environments we incorpo-rate in every beacon message one additional byte of information about the Social Pattern - flow(SN) the vehicle has. When for specific applications we are interested to create stable clustersalong a central road we introduce a new metric called Route stability metric (RN) as describedin section 3.3.14.

3.3.13 Sociological pattern of vi

The first step in creating a cluster for a every vehicle is to identify its neighbors. Neighbor-hood identification is the process whereby a vehicle/node identifies its current neighbors withinits transmission range. For a particular vehicle, any other vehicle that is within its radio trans-mission range is called a neighbor. The neighbor set is always changing since all nodes aremoving. Every moving node keeps track of all current neighbors (their id’s) the current and thepast distance. In order to perform clustering using social criteria, SPC maintains two differentsets of neighbors. Set Ni is the set of all neighbors in range of vehicle Vi and set NSi is the set ofall neighbors that share a common social pattern.

The clustering procedure consists of two stages.First stage of clustering

In the first stage each vehicle try to create cluster with nodes that have the same SN accordingto these rules:

• At any time each vehicle i recalculates total F and according to total non-clustered mem-bers with same SN within range, try to form a cluster and become the clusterhead.

• If the node has the biggest positive force applied to it and there exist at least one free nodein its neighborhood NSi, it declares itself to be a clusterhead.

60


• In the opposite situation, where there exist a free node j with biggest total F in range thevehicle becomes a cluster member of j.

This algorithm leads to the formation of clusters which are at most two hops in diameter andhave same SN.Second stage of clustering

After the initial clustering phase, some clusters will have been formed. However, there willalso be nodes which could not join any clusters during the initial clustering phase mainly becausethey are surrounded by vehicles having different social patterns. In this occasion clustering isperformed again using total Force applied to each vehicle. At this stage the set Ni of all neighborsis used and clusters of vehicles having different Social Patterns are created.

Figure 34 describes the different states that each vehicles passes while entering, movingaround and leaving the area of interest when sociological pattern clustering (SPC) is used. Whenthe vehicle first enters the area of interest or leaves it, its state is undefined UN since every regionis studied in isolation.

CHCM /

Second stage of clustering

Free

UN

UNNSi Null

Leaving the area

Send packet to RSU

Entering the area First stage of Clustering

NSi not Null

Send ID to RSU

Receive SN from RSU

Figure 34: States of a vehicle.

3.3.14 Route stability of vehicle vi

In order to perform clustering in main streets the vehicles that meet each other exchangebeacon messages. Each beacon message consists of node Identifier (Vid), node location, speedvector in terms of relative motion across the axes of x and y (dx,dy), Route Stability R, state andtime stamp.

Route stability (Reliability) of Vehicle Vi that moves on road segment Si the time period TPk

is calculated be Equation 24:

RVi = ∑j

p j, j ∈ PPi (24)

where PPi is the set of the road segments that belong to the same street and pj are the long termprobabilities of vehicle Vi that have been received by the RSU the specific time period.

In a multi-lane street vehicles that have bigger route stability are better candidates to becomeclustrheads since they are going to stay longer on the street. If a vehicle that is going to leavethe street soon, is elected as a clusterhead then major re-clustering is going to take place whenit turns to another road segment, since it leaves all of its members orphans. In case a member

61


node Vi leaves the street in order to follow another edge of the network, only this vehicle tries tofind a nearby cluster to enter.

3.3.15 Simulation and performance evaluation

This section evaluates the performance of SPC and RSC. The evaluation is based on simu-lated traffic. Unfortunately, there are no large-scale vehicle mobility pattern available to use, andseveral earlier [53] and new studies [50] use simulated data. The traffic simulation is conductedwith SUMO [33] and the trace files are injected to our custom simulator in order to perform clus-tering. In the simulation, we use the road network of city of Erlangen. Using the hierarchicalcommunities method we are able to divide the city in isolated regions and study the mobility ofvehicles in the one marked on Figure 35. The only communications paths available are via thead-hoc network and there is no other communication infrastructure. The power of the antenna isPtx = 18dBm and the communication frequency f is 5.9Ghz.

Data Rate (Mb/sec) Minimum Sensitivity(dBm)

12 -7718 -7024 -6927 -67

Table 8: Minimum sensitivity in receiver antenna according to data rate.

The reliable communication range of the vehicles is calculated according to Table 8. Thereliable communication range is calculated for every pair of nodes at every instance based on thediffraction caused by obstructing vehicles [45]. In our simulations, we use a minimum sensitivity(Pth) of −69dBm to −85db which gives a transmission range of 130 to 300 meters.

According to [8] an acceptable communication range for VSC applications, which use thesame broadcast messages to our clustering methods, is about 300m. This range that can beachieved by low transmission power, like the one we use in our simulations, is enough forcorrect dissemination of a message in a neighborhood while it improves spatial reuse in heavytraffic. In rural environments, in scenarios with low data rate (3MBPs) authors in [8] showedthat Packet Delivery Ratio (PDR) of 60 % can be achieved for such medium distances.

All the simulation parameters with their default values are represented in Table 9.All nodes are equipped with GPS receivers and On-Board Units (OBU). Location infor-

mation of all vehicles/nodes, needed for the clustering algorithm is collected with the help ofGPS receivers. By default, 80 to 160 vehicles move in the network, and their movement patternis determined by Krauss following model. The vehicles have maximum velocities from 40 to50km/h, large speed deviation (60 to 140 % of legal speed limits) with 2 to 4 different flows –social profiles.

While one would like to have deterministic social profiles for every driver - vehicle , this isnot possible due to the nature of driving. Even if a driver follows a standard route every day,it is still possible for the driver to deviate from it once in a while. Circumstances like a doctor

62


appointment, road construction or alternative route due to congestion may cause driver to changethe route he is predicted to follow according to his social profile. All of this points to the fact thatthe prediction of driver intent must be probabilistic. For these reasons vehicles are injected tothe map at random sequence and follow their path according to their social profile with defaultprobability of 67 % and range from 67 % to 97 % (see Figure 41).

Independent parameter Range of values Default value

Velocity (m\sec) 40, 50 42Number o f Vehicles 80,120,160 120

Probability o f f ollowing the social pattern (%) 67,97 67Number o f Sociological patterns 2,3,4 2

Communication Range (m) 130 - 300 130Number o f RSUs 5 5

Table 9: Simulation parameters.

To show the performance of our proposed Social clustering (SPC, RSC) methods, we com-pare them with the Lowest-id (Low-id) and Dynamic Doppler Value Clustering (DDVC) pro-posed in [27] and [60] respectively. The Lowest-ID algorithm forms of clusters which are atmost two hops in diameter.

The basic concepts of Lowest id are the following. Each node is given a distinct ID and itperiodically broadcasts the list of its neighbors (including itself). A node which only hears nodeswith ID higher than itself is a “clusterhead” (CH). The Lowest-ID node that a node hears is itsclusterhead, unless the Lowest-ID specifically gives up its role as a clusterhead (deferring to ayet lower ID node). A node which can hear two or more clusterheads is a “gateway”. Otherwise,a node is a free node.

In DDVC a cost metric derived from the Doppler shift property, the Doppler Value, is used inorder to create clusters. The Doppler Value (DV ) is related to the relative velocity. We simulateDDVC with parameter nmin having value 1.

3.3.16 Sociological Pattern Clustering

In order to evaluate the stability of the algorithm, we measure the stability of the cluster con-figuration against vehicle mobility. In a high dynamic VANET, nodes keep joining and leavingclusters along their travel route. Good clustering algorithms should be designed to minimize thenumber of cluster changes of the vehicle by minimizing reclustering. This transitions amongclusters are measured in order to evaluate the performance of the algorithm. The basic transitionevents the vehicle encounters during its lifetime:

• A vehicle leaves its cluster and forms a new one (becomes a clusterhead).

• A vehicle leaves its cluster (due to communication range) and joins a nearby cluster orbecomes free.

• A cluster-head merges with a nearby more stable cluster.

63


Figure 35: A typical region of interest.

The average cluster lifetime is another important metric that shows the performance of theclustering algorithm. The cluster lifetime is directly related to the lifetime of its cluster-head.The cluster-head lifetime is defined as the time period from the moment when a vehicle becomesa cluster-head to the time when it is merged with a nearby cluster.

SPC versus VFVC

Initial Spring Clustering and the enhanced VFVC method behave well when investigatinghighways where vehicles don’t change direction so often as in a real urban environment (theformer), or when we are focused on main streets when we care about the stability of the clusteron the street and not at the whole area (the latter). In addition VFVC gives good outcome whenthe road lanes clearly clarify the possible direction of the car that is traveling on the road. In amore realistic scenario when small road segments of a city consist of one lane, VFVC degradesto the initial Spring Clustering. As shown in Figure 36 the performance of VFVC compared toSPC in a city region when most of the roads consist of one lane is much lower but still betterthan the performance of DDVC since except from relative speed that DDVC uses to performclustering, VFVC uses virtual forces among nodes that are affected by relative mobility, currentand future distance in both x and y axes. In urban environments, where the mobility of nodescompared to a highway is more dynamic, SPC has clear impact on cluster formation and stability(Figure 36).

SPC versus Low-Id and DDVC

We compare the performance of SPC, Low− Id and DDVC methods when different trans-mission ranges (Figure 38) and different speeds (Figure 39) are used. We also investigate theperformance of SPC against different number of Social Patterns that the vehicles have (Fig-ure 40) and against different probabilities in following the correct pattern (Figure 41).Number of clusters over time. The number of clusters created by a clustering algorithm is asignificant parameter of the clustering procedure; too many, and thus small, clusters implies thatthe benefits reaped due to clustering will diminish, since the broadcast storm is not really cured,and too much communication has to take place to forward messages (too many clusterheads and

64


0

20

40

60

80

100

120

140

160

180

130 150 200

Mea

n lif

etim

e of

Clu

ster

s

Mean Communication Range

SPCVFVCDDVC

Figure 36: Lifetime of SPC versus VFVC for a typical urban scenario [2 flows, 70 % probabilityof following the social pattern] for different communication ranges.

too many gateways participate in the forwarding process). On the other hand, the existence ofonly a few, and thus quite large, clusters is also not desirable since the channel is shared amongtoo many members of the same cluster and hence the communication latency increases. Wepresent an experiment – with the default values of the parameters of Table 9, and the results areillustrated in Figure 37. It depicts the total number of clusters created by the competing methodsover the simulation time of the experiment. We see that SPC creates a moderate number ofclusters, less than that created by Low− Id, but more than those created by DDVC for mostof the time. Analogous observations were made for other values of the parameters. Therefore,SPC can achieve the best of two worlds; relatively small transmission latency, and relatively fewrebroadcast messages.

In order to investigate how stable the clusters created by each method are, we measure clus-ter lifetime and mean transitions that each vehicle encounters during the simulation. We tunedifferent parameter each time and we can see, from the sections that follow, that the averagetransitions produced by our SPC technique is smaller compared to that produced by the Low-IDand gives comparative numbers to DDVC.Cluster stability versus communication range. Figure 38 shows that the average transitionsof the vehicle decreases and mean cluster lifetime increases as the transmission range increaseswhen SPC is used. This is because increasing the transmission range, increases the probabilitythat a vehicle stay connected with its cluster-head. Communication range doesn’t have any im-pact on Low− Id’s performance and though it slightly improves DDVC it has a major impact onSPC stability. Since SPC creates clusters of nodes sharing common social profiles, as communi-cation range increases the probability that such nodes stay interconnected for a longer time alsoincreases. In Low− Id only vehicle’s ID is used in order to elect clusterheads and that way al-

65


Figure 37: Number of heads produced by all methods during the simulation.

though increased communication range may have a positive impact on nodes connectivity it alsoaffect them in a negative way since nodes more often find a neighbor with lowest id and performreclustering. In DDVC increase of communication range does not have so big positive impactsince in an Urban environment vehicles always change directions, accelerate and decelerate inorder to follow different road segments thus causing the method to create new clusters often.Cluster stability versus speed. In Figure 39 we observe that the impact of different vehiclespeeds in a urban environment is not so clear. This is due to the fact that in urban areas the maxvelocity cannot be easily reached by vehicles since they always have to stop at intersections orchange speed due to turns and congestion. For the maximum speeds investigated SPC has muchbetter performance compared to both DDVC and Low− Id.Cluster stability versus social patterns. As social patterns increase, meaning that vehiclesfollow less common routes, the performance of SPC decreases (see Figure 40). From this figure,it could be seen that SPC follows the theoretical model closely, yet the actual cluster lifetime isa always better than the other methods give. DDVC also degrades as the mobility of vehiclesbecomes more chaotic.Cluster stability versus pattern following probability. The mean lifetime that our method pro-duces, even when the probabilities that a car follows its Social Pattern drops to 67 % (Figure 41),is always better than those that the other two methods give.

3.3.17 Route Stability Clustering.

In order to evaluate the performance of RSC we are interested in a main street of the area ofErlangen. The street is shown in Figure 42. The area consists of many intersections. On the mapthree main flows of vehicles are shown. These flows split the traffic of the main road of interest.

We focused only on one traffic direction. Vehicles follow three different route distributionsaccording to Table 10. These distributions are used in order to represent the social patterns ofthe vehicles and are based on their historical data.

66


0

20

40

60

80

100

120

140

160

180

130 150 200

Mea

n lif

etim

e of

Clu

ster

s


SPCDDVC

Lowest-Id

0

10

20

30

40

50

60

70

130 150 200

Ave

rage

Num

ber

of C

lust

er c

hang

es /

vehi

cle


SPCDDVC

Lowest-Id

Figure 38: Lifetime and mean cluster changes versus range.

We follow vehicles until they leave the section of the road turning left or right. In that waywe are focusing on what happens on a central road, where cars enter and leave it all the time ifwe favor cars that follow the non turning lane to become clusterhead. Using Equation 24 andthe data from Table 10 we calculate the route stability Ri of each vehicle regarding the street ofinterest. Route stability is then incorporated in Equation 22 as parameter qi, in order to favorvehicles that are more possible to stay on the street for the longer time become clusterheads.

Route Probability to Stability Parameterfollow the route RVi qi

1 90% Low 0.52 90% High 23 90% Medium 1

Table 10: Route distributions according to different social patterns of vehicle i.

67


0

20

40

60

80

100

120

140

40 50

Mea

n lif

etim

e of

Clu

ster

s

Mean Velocity

SPCDDVC

Lowest-Id

0

10

20

30

40

50

60

70

40 50

Ave

rage

Num

ber

of C

lust

er c

hang

es /

vehi

cle

Mean Velocity

SPCDDVC

Lowest-Id

Figure 39: Lifetime and mean cluster changes versus speed.

For the scenario of RSC we used the values of Table 9 in terms of velocity and communica-tion range. We compare the performance of RSC, SPC, Low− Id and DDVC methods and theresults are presented in Figure 43.

The results of simulations conducted show that RSC algorithm outperforms the other inves-tigated methods, in terms of average cluster lifetime (higher) which is translated in increasedcluster stability, lower percentage of orphan nodes and larger cluster sizes. The other parametersthat determine the stability of a clustering method like cluster-head changes, total number ofclusters and null nodes also give better values for RSC compared to the other methods.

4 Delay-aware cross-layer packet scheduling

The development of a throughput-optimal routing algorithm for packet radio networks which isalso robust to topology changes was first presented in [72]; it is based on the Lyapunov drift the-ory, and it is known as the backpressure (BP ) packet scheduling algorithm. Subsequently, theoriginal concept spawn several lines of research in the topic. The performance of backpressure

68


0

20

40

60

80

100

120

140

2 4

Mea

n lif

etim

e of

Clu

ster

s

Social Patterns

SPCDDVC

Lowest-Id

0

10

20

30

40

50

60

70

2 4

Ave

rage

Num

ber

of C

lust

er c

hang

es /

vehi

cle

Social Patterns

SPCDDVC

Lowest-Id

Figure 40: Lifetime and mean cluster changes versus number of social patterns.

deteriorates in conditions of low, and even of moderate, load in the network, since the packets“circulate” in the network, i.e., the backpressure algorithm stabilizes the system using all possi-ble paths throughout the network. The net effect of this mechanism is to increase delay and alsoto increase the energy consumption of the nodes that play the role of relays. Delay and energyare interconnected; the minimization of the average time that the packets stay in the system,implies a reduction in the average number of hops that the packets travel until they reach theirdestination, which in turn implies a reduction in the total energy consumption. Delay and energyconsumption problems are particularly significant for underwater acoustic sensor nets which arethe target of our work.

To circumvent the delay problems of backpressure, the shadow queue algorithm [14] forcesthe links to stay inactive in order to lead the network to work in a burst mode, since for periodswhere the load of the network is low or moderate, link activation is prevented by a parameter M,leading to a delay increase. On the other hand, the relatively high computational cost incurredat each node by backpressure (maintenance of a queue for each possible destination, and updateof these queues at each new arrival) inspired approaches based on node grouping in order to

69


0

20

40

60

80

100

120

140

160

180

67 90 97

Mea

n lif

etim

e of

Clu

ster

s

Probability %

0

2

4

6

8

10

67 90 97

Ave

rage

Num

ber

of C

lust

er c

hang

es /

vehi

cle

Probability %

Figure 41: Lifetime and mean cluster changes of SPC versus probability of following a social pattern.

reduce the number of these queues [84] and thus the computational overhead, and as a side-effect, reduce the delay. To alleviate the delay problems of backpressure, scheduling based onthe combination of backpressure and shortest-paths have been proposed [83], which thoughdemands an excessive number of queues and repeated calculation of all-node-pairs distances incase of topology changes. In summary, all these approaches either impose unpractical and nonscalable assumptions, or are not very efficient. Finally, some recent ideas [31, 49] could beincorporated in an orthogonal way to improve analogously the delay performance of all policiesat the expense of throughput optimality, but this is beyond the scope of this paper. Here, wetake an holistic approach in designing an efficient delay-aware backpressure policy, that is alsopractical – it has low computational overhead and it is robust to topology changes.

4.1 The quest for delay-aware packet scheduling protocols

In the shadow queues (S Q−BP ) methods [7, 14], backpressure is used with a parame-ter M that forces the links to stay inactive as long as their differential backlog doesn’t exceed

70


Figure 42: Main road and the flows that split traffic.

0

20

40

60

80

100

120

140

160

130 150 200

Mea

n lif

etim

e of

Clu

ster

s

Social Patterns

RSCSPC

DDVCLowest-Id

Figure 43: Lifetime of RSC.

the value of M, i.e., links with backpressure less than M can not be scheduled. This means thatpackets stay in queues longer, which can lead to higher delays. Indeed, we tested this intuitiveresult by evaluating the delay performance of these methods (see a sample performance in Fig-ure 44) and confirmed this behavior. Therefore, since the S Q−BP methods are not delaycompetitive, we do not consider them as viable competitors any more.

The authors of [84] identified the scalability problems of the backpressure policy in wire-line networks with millions of routers (e.g., Internet) due to the maintenance of several queuesper node (one queue per destination, as it is mandated by the original BP ), and proposed

71


0

200

400

600

800

1000

1200

1400

1600

1800

2000

0.01 0.02 0.05 0.07 0.1 0.12 0.15 0.17 0.2 0.3

Mea

n D

elay

Arrival rate

BPSQ-BP

Figure 44: Performance of the S Q−BP policy in a 4x4 grid network (M = 2).

the creation of clusters3 of nodes so as to reduce the number of queues per node, which as a“side-effect” has the additional benefit of delay reduction. Their policy, namely cluster-basedbackpressure (C B−BP ), requires maintaining one queue per gateway at each relay nodewhich leads to an excessive number of gateways, which is turn alleviates any performance gains(i.e., increases delays) when the number of clusters becomes, say, more than ten. Even worse,all contemporary clustering algorithms [3, 10] (such as the Distributed Clustering Algorithm,Max-min d-hop clustering algorithm) for wireless ad hoc networks produce quite a large num-ber of clusters and thus several dozens of gateways even for relatively small networks. In thesecases, the delay of C B−BP approaches asymptotically the delays of the original BP (cf.Figure 45c). Moreover, the strong dependence of the policy on the identity of the gatewaysmakes it problematic in cases of gateways breakdowns. Finally, considering the technicalitiesof the C B−BP method, it is evident that even when a packet has already reached the destina-tion cluster (i.e., the cluster where the destination nodes resides), the method is agnostic on thisinformation and it may happen to send it again out of the cluster seeking alternative paths to thedestination.

The Shortest-paths backpressure (S P−BP ) method [83] assumes the precomputationof all pairwise-node distances and then application of the backpressure notion on the shortestpath(s) between source and destination. Apparently, this method achieves the lower bound of thedelay, it does so at the expense of a very complex initialization phase (all-node-pairs distancesmust be computed). Moreover, it must maintain a quadratic number of queues at each node(n×(n−1)), whereas the original backpressure maintains only a linear number of queues (n−1)at each node. Also, during the running phase of the algorithm, the processing of such a hugenumber of queues is time-consuming, which in turn would increase delays. Apart from thesecomputational-type problems, frequent topology changes would lead the method to break down,since many shortest paths would not exist anymore. Finally, in topologies where there is only

3The authors did not propose any specific clustering algorithm.

72


0

100

200

300

400

500

600

2 4 8 16

Mea

n D

elay

Arrival rate

LBPBP

CBP

Figure 45: Impact of the number of clusters on the grid network topology.

one shortest-path per node pair (which is the most common case), S P−BP would rapidlyconsume the energy of that path, shortening the longevity of the network. This problem becomesworse in topologies where a lot of shortest paths traverse the same set of nodes (i.e., nodes withhigh betweenness centrality [32]); in such topologies, these nodes deplete their energy so fast,that the network gets partitioned very rapidly.

4.2 Distributed backpressure-based packet scheduling

Max-weight scheduling, also known as backpressure routing, is a cross-layer control algo-rithm that is well-known to be throughput optimal since it provides queue stability within thenetwork for all traffic injection rates within the network’s capacity. Max-weight scheduling re-lies on global knowledge of full system state information of the network which is not realisticin ad-hoc networks. Implementing backpressure in a distributed way is a challenge. In orderto implement the backpressure in ad-hoc networks, nodes need to disseminate some kind ofinformation in order to make the optimal cross-layer control decisions.

The centralized scheduling algorithm selects the node with the highest differential queuebacklog first and thus gives higher priority to a congested node for transmission. In distributedand random access based fashion, a node with higher differential queue backlog should be al-lowed to access the medium with higher probability.

Authors in [38] developed a distributed scheduling algorithm for multi-channel ad hoc wire-less networks that automatically track the changes in the network topology and offered load.They introduce a two stage queueing mechanism for packet scheduling. The method has a sig-nificant lower capacity region than Greedy Maximal Scheduling.

In [57] authors explore the idea of disseminating only a limited amount of state feedback forthis algorithm and evaluate the subsequent impact on performance. Queue backlogs are repre-sented by K bits. The values are calculated using threshold values uniformly It is shown when

73


using uniformly distributed threshold values, a tradeoff exists between the overhead requiredin information dissemination and performance gains from more effective scheduling The exactnumber of bits that is necessary for the correct behavior of the system is a major drawback ofthe proposed method.

In [64] when a packet is transmitted to any destination, the transmitting node inserts its queuelength for that destination onto the packet. All one-hop neighbors overhear this transmissionand update their queue length values for the node-destination pair. The algorithms are fullydistributed and requires only one-hop information.

Lately backpressure has been used in traffic signal control. In [36] a distributed algorithm ispresented where the signal at each junction is locally controlled independently from other junc-tions. Their approach relies on constructing a set of local controllers, each of which is associatedwith a junction. These local controllers are constructed and implemented independently of oneanother. The method seems to overcome SCATS, an adaptive traffic signal control systems thatis being used in many cities.

4.3 The original Backpressure algorithm

Backpressure [72] is a joint scheduling and routing policy which favors traffic with highbacklog differentials. The backpressure algorithm performs the following actions for routingand scheduling decisions at every time slot t.

• Resource allocationFor each link (n,m) assign a temporary weight according to the differential backlog ofevery commodity (destination) in the network:

wtnmd(t) = max(Qdn −Qd

m,0).

Then, define the maximum difference of queue backlogs according to:

wnm(t) = maxdεD

wtnm(t).

Let d∗mn[t] be the commodity with maximum backpressure for link (n,m) at time slot t.

• SchedulingThe network controller chooses the control action that solves the following optimizationproblem:

μ∗(t) = argmax∑n,m

μnmwnm(t),

subject to the one hop interference model. In our model, where the capacity of every linkμnm equals to one, the chosen schedule maximizes the sum of weights. Ties are brokenarbitrarily.

• RoutingAt time slot t, each link (n,m) that belongs to the selected scheduling policy forwards onepacket of the commodity d∗mn[t] from node n to node m. The routes are determined on thebasis of differential queue backlog providing adaptivity of the method to congestion.

74


Backpressure algorithm is throughput-optimal and discourages transmitting to congestednodes, utilizing all possible paths between source and destination. This property, leads to un-necessary end-to-end delay when the traffic load is light. Moreover, using longer paths in casesof light or moderate traffic wastes network resources (node energy).

4.4 Layered backpressure

This section describes a delay-efficient backpressure algorithm based on the creation of lay-ers4 in the network. The main idea is to split the network into layers according to the connectivityamong them, which also (usually) implies geographic proximity, as well. These layers dividethe initial graph to k smaller networks. Then the algorithm forwards packets according to thedestination layer ID, thus effectively reducing the long paths. In some sense, these layers playthe role of attractors, that attract the packets destined for them and then “disallow” the packetsto leave the layer. Apparently, if we have only one layer then the policy reduces to the originalbackpressure algorithm.

For the implementation of the L ayBP scheduling policy, every node n keeps a separatequeue for each destination going through it and also holds the layer Layer(n) that the destinationbelongs to. The backpressure scheduling is executed using only the IDs of these layers. This is asignificant difference from the work of [84], because L ayBP does not require the knowledgeof gateways’ queue lengths. The “correct” partitioning of nodes into layers and a proximity-based naming of layers is of paramount importance in order to improve the mean end-to-enddelay. Apparently, the concept of “correctness” is algorithmic, and we only need to set a parti-tioning criterion; we proposed to determine the number of layers based on the network topology,i.e., connectivity. For the description of layer-creation and layer-naming algorithms, the reader isdirected to [41], whereas subsection 4.4.1 presents the Layered Backpressure packet schedulingalgorithm.

4.4.1 The Layered BackPressure protocol

After the completion of the grouping and the assignment of IDs to the layers, the actualpacket scheduling is performed as follows. Each node n maintains a separate queue of packetsfor each destination. The length of such queue is denoted as Qd

n[t]. For every queue Qdn [t] the

node computes the parameter Dleveldn which represents the absolute difference between currentand destination node’s layer number: Dleveldn = |Layer(n)−Layer(d)|. At each time slot t, thenetwork conrtoller observes the queue backlog matrix Q(t) = (Qd

n(t)) and performs the follow-ing actions for routing:

Layered BackPressure at time slot t:

• Each link (n,m) is assigned a temporary weight according to the differential backlog

4We use the term layer to describe node partitioning, since the target of this research is underwater sensor networksfor surveillance applications, where sensors are placed in layers beneath the sea surface.

75


wtnmd(t) = (Qdn −Qd

m) and parameter Anmd according to:

Anmd =

⎧⎨⎩

2, if Dleveldn > Dleveldm1/2, if Dleveldn < Dleveldm1 otherwise.

• Each link is assigned a final weight according to wnm and parameter Amnd :

wnm(t) = maxdεD(wtnmd(t)∗Anmd)

• The network conrtoler chooses the control action that solves the following optimization:

μ∗(t) = argmax ∑n,m μnmwnm(t)

subject to the interference model mentioned above where adjacent links are not allowedto be active simultaneously.

A simple example is illustrated in Figure 46. The network in Figure 46 consists of 7 nodeswhere all nodes are senders and only nodes with id’s 1,2,3 are destination nodes. In Figure 46 weobserve the status of the queue backlogs at a specific time slot t. The network is divided in twolayers. Layer one includes nodes in the left side of the dot line. Running the initial backpressurealgorithm the commodity that achieves maximum differential backlog for link (n,m) is 2 whilefor the proposed layer backpressure is 3. If we assume that link (n,m) is scheduled at that timeslot then for the initial backpressure algorithm a packet destined for node 2 will be forwardedto node m pushing it further away and forcing it to follow longer path in order to reach itsdestination (node 2). With the layer backpressure policy at the same time moment the link (n,m)if scheduled forwards a packet of commodity 3 from node n to node m which is closer to thedestination of the packet. Furthermore node m belongs to the same cluster as node m, whichaccording to the proposed layer backpressure policy will prevent the packet from returning tonode n.

23

6

2

Dlevel

4

−1

3

w{nmd}wt{nmd} Anmd

−2

2

1/2

1/210

0

1 0

1

mnQnd

1

2

3

BP : Optimal commodity for link (n,m) on slot t is 2. w(n,m) = 6 LBP : Optimal commodity for link (n,m) on slot t is 3. w(n,m) = 4

Layer 2Layer 1

21

3

3

1

1

2

mn

n m

Figure 46: An example of the L ayBP execution.

Theorem [Throughput optimality].The L ayBP algorithm is throughput optimal.

76


Proof.The original backpressure algorithm is throughput optimal. In order to prove that L ayBP isalso throughput optimal, the Lyapunov stability criterion is used. The idea behind the Lyapunovdrift technique is to define a non-negative function, called the Lyapunov function, which repre-sents the aggregate congestion of all queues (Qd

n) of the network. The drift of the function attwo successive time slots is then taken, and in order for the policy to be throughput optimal, thisdrift must be negative, when the sum of queue backlogs is sufficiently large. For both strategies,we use:

L(Q) = ∑nd

θdn (Q

dn)

2

as the Lyapunov function.Recall that each link is assigned a final weight according to wnm and parameter Amnd :

wnm(t) = maxdεD

(wtnmd(t)∗Anmd).

This equation can be rewritten in the following form:

wnm(t) = maxdεD

(Anmd ∗Qdn −Anmd ∗Qd

m).

which is equivalent to :wnm(t) = max

dεD(θd

n ∗Qdn −θd

m ∗Qdm),

where weights θdi are used to offer priority service similar to [51].

Queue dynamics at each time slot satisfy :

Qdn(t +1)≤ max[Qd

n(t)−∑b μdnb(t),0] + Ad

n(t)−∑a μdan(t),

where μdnm(t) are routing control variables, representing the amount of commodity d data deliv-

ered over link (n,m) during slot t and Adn(t) represent the process of exogenous commodity d

data arriving to source node n.According to [26], we have:

(Qdn(t+1))2 ≤ ([Qd

n(t))2+(∑b μd

nb(t))2+(Ad

n(t)+∑a μdan(t))

2−2[Qdn(t)(∑b μd

nb(t)−Adn(t)−

μdan(t))

Multiplying both sides with θdn , summing over all valid entries (n,d) and using the fact that

the sum of squares of non-negative variables is less than or equal to the square of the sum wetake:

∑nd θdn (Q

dn(t +1))2 ≤ ∑nd θd

n (Qdn(t))

2 +∑nd θdn (∑b μd

nb(t))2

+∑nd θdn (A

dn(t)+∑a μd

an(t))2 −2∑nd θd

n Qdn(t)(∑b μd

nb(t)−Adn(t)−μd

an(t))

It is not difficult to show that:

ΔL(Q)≤ 2BN −2∑nd θdn εQd

n(t),where

B � 12N ∑

nεNθmax[(μout

max,n)2 +(Amax

n +μ inmax,n)

2].

77


Using the above we can rewrite drift inequality:

ΔL(Q)≤ 2B′Nθmax −2∑nd θdn εQd

n(t),where

B′ � 12N2 ∑

nεN

[(μoutmax,n)

2 +(Amaxn +μ in

max,n)2].

This drift inequality is in the exact form for application of the Lyapunov drift lemma, provingthe stability of the algorithm.

The sum of all queues is:

limsupt→∞1t ∑t

τ=0 E{∑n,d θdn Qd

n(τ)} ≤ NB′θmaxεmax

�

4.5 The Enhanced Layered Backpressure policy

Routing protocols must be dynamic in order to cope with mobility of nodes in modern wire-less networks. Widely varying mobility characteristics are expected to have a significant im-pact on the performance of routing protocols that are based on node grouping (like C B−BP ,L ayBP ) in order to route packets even if links among nodes are updated. In case of grouping-based routing protocols, high mobility of nodes which lead them to change groups, degrades theperformance of the methods since this ‘wrong’ information is used in the routing procedure.Although L ayBP does not use gateways, it still suffers from this behavior if the layer thatthe moving nodes belong to, are not updated. The differential backlog of each link is com-puted according to the difference between current and destination’s node layer. It is clear thatL ayBP behavior can be affected of ‘misplaced’ nodes. In this case packet may be forwardedto layers different than the desired making the method inappropiate.

In order to cope with node mobility we incorporate in L ayBP algorithm another step inwhich moving nodes recalculate their cluster according to their neighborhood. In the initiation,phase every node has a counter C0nl for every layer ID, indicating how many neighbors belongto it and a variable Layern indicating the layer that node n belongs to. Every time slot t thefollowing actions are performed:

• Calculate Cnl , the total number of neighbors that belong to every layer detected.

• if Cnl >=C0nl for l = Layern then moving node remains in the same layer.

• else calculate the layer with the most neighbors Mnl = maxCnl . if Mnl >Cnl then movingnode belongs to layer Mnl .

• assign for every layer l, C0nl =Cnl as new initial values for next time slot.

This algorithm can be executed every k time slots according to how fast we want the systemto adapt to node mobility. Only moving nodes update the information on their counters in orderto find the appropriate layer. The algorithm does not perform reclustering, but only ‘helps’ layersincorporate moving nodes.

78


4.6 Dynamic networks

4.6.1 Experimental setting

We evaluated the performance of the cluster-based algorithms in networks with mobile nodesin order to measure the adaptivity of the E L−BP method. It is important here to mention thatE L−BP copes with topology changes in terms of mobility nodes, which is different than thecase evaluated earlier, where a link breakdown is used in order to show the lack of adaptivityof S P−BP method to topology changes. For the evaluations we conducted, we assume thatmoving nodes update their link information, according to some handshaking mechanism, but areunaware of layer changes.Network topologies. As network topologies, we considered that illustrated in Figure 47. Thetopology is a network of 13 nodes where node 13 moves from layer 1 (time slot 0) to layer 2 (timeslot 1000) and finally to layer 3 (time slot 2000). BP , C B−BP and L ayBP are unawareof layer change while all the active links of the moving node are updated. In E L−BP method,update algorithm runs at every time slot and moving nodes join the appropriate layer accordingto neighbor’s information.

13 13 13

Layer 2

3

2 9 10

12118

1

4 6

5 7Layer 1

T_0: (time slot 0) T_1: (time slot 1000)

Layer 3

T_2: (time slot 2000)

Figure 47: Dynamic network with 3 layers.

Packet generation. In order to have a clear view of the performance of the cluster-based meth-ods we generated at each experiment only one flow from node 1 to the moving node. Theexogenous arrival processes are independent Bernoulli processes with rate λ .

4.6.2 Experimental results

Delay and throughput performance. It is clear that C B−BP brakes down in situationswhere nodes change cluster. Based on the fact that packets only after entering the correct clus-ter can be directed to the correspondent queue, makes it impossible for ’misplaced”’ nodes toreceive packets. Because of this behavior CB−BP is not simulated in these experiments.The plots in Figure 48 show the delay performance of the L ayBP , E L−BP and the origi-nal backpressure. It is clear L ayBP behaves worse than the original BP in cases of movingnodes. The E L−BP gives the best outcome since the moving node updates the layer informa-tion according to the neighborhood. It is clear that in situations where nodes move along many

79


layers (second dynamic topology) the behavior of L ayBP degrades, while E L−BP retainsits behavior.

10

20

30

40

50

60

70

0.1 0.2 0.5 0.7

Mea

n D

elay

Arrival rate

BPLay-BPEL-BP

Figure 48: Impact of moving nodes on the delay performance of BP , L ayBP andE L−BP for the dynamic network with the 3 layers.

The plots in Figure 49 depict the throughput performance of the competing methods5

0

500

1000

1500

2000

2500

0.1 0.2 0.5 0.7

Tot

al th

roug

hput

Arrival rate

BPLay-BPEL-BP

Figure 49: Impact of moving nodes on the throughput performance of BP , L ayBP andE L−BP for the dynamic network with the 3 layers).

5An extensive performance analysis of E L−BP is described in [43].

80


5 Risk Assessment

The failure to have OBU’s installed on vehicles on the field trials, in order to implement theREDUCTION vehicle routing algorithms is at low risk. Field trials are used in order to evaluatethe accuracy of our methods but do not affect the designing of them. In case where no fieldtrial can be performed in real time, network simulation frameworks (such as OMNET++) incombination with road traffic simulation packages (such as SUMO) can be used to evaluate theperformance of the new communication methods. Off line GPS data can also be used in orderto simulate real traffic scenario’s and test the performance of our communication protocols. TheREDUCTION project finally has access to approximately 370 million measurements from 120vehicles with a total 500,000 trips.

81


6 Conclusion

Our work on the wireless communications and vehicular networking aspects of REDUCTIONrevealed the benefits of the two layer architecture in combination with a packet scheduling/routingmechanism that guarantees throughput optimality and at the same time strives for low delay inpacket dissemination.

In the clustering context, UTH team has published extensively these results in internationalfora (in conferences [42, 44, 45] and journals [47, 55]), in order to confirm that these are orig-inal ideas that advance the current of the art. The results confirmed that vehicle clustering canindeed address the broadcast storm problem that any ad hoc network faces. In particular, theproposed methods proved superior to existing work in terms of number, size and stability ofclusters produced, and therefore comprise viable solutions for the REDUCTION networkinginfrastructure.

Similarly, the developed backpressure controller [46], which is a cross-layer solution forpacket scheduling and routing for ad hoc networks, meets the goals set by the REDUCTIONneeds, i.e., to be able to exploit the clustered nature of the underlying ad hoc network, to berobust to frequent topology changes due to the high vehicle mobility, and finally to achieve lowlatencies which is a prerequisite for real-time applications.

Overall, this report provides viable solutions that advance the state-of-the-art to the prob-lems/goals set by task T1.3 and can be used both for REDUCTION and future EU projects.

82


References

[1] Y. Afek, N. Alon, O. Barad, E. Hornstein, N. Barkai, and Z. Bar-Joseph. A biologicalsolution to a fundamental distributed computing problem. Science, 331(6014):183–185,2011.

[2] M. S. Almalag and M. C. Weigle. Using traffic flow for cluster formation in vehicular ad-hoc networks. In Proceedings of the IEEE Conference on Local Computer Networks(LCN), pages 631 –636, 2010.

[3] A. D. Amis and R. Prakash. Clusterhead selection in wireless ad hoc networks. US Patent6,829,222 B2, 2004.

[4] A.D. Amis, R. Prakash, T.H.P. Vuong, and Huynh D.T. Max-min d-cluster formationin wireless ad hoc networks. In Proceedings of the IEEE Conference on ComputerCommunications (INFOCOM), pages 32–41, 2000.

[5] V.S. Anitha and M. P. Sebastian. (k,r)-dominating set-based, weighted and adaptive clus-tering algorithms for mobile ad hoc networks. IET Communications, 5(13):1836–1853,2011.

[6] F. Atay, I. Stojmenovic, and H. Yanikomeroglu. Generating random graphs for the sim-ulation of wireless ad hoc, actuator, sensor, and internet networks. In Proceedings ofthe IEEE Symposium on a World of Wireless, Mobile and Multimedia Networks(WoWMoM), pages 597–615, 2008.

[7] E. Athanasopoulou, L. Bui, T. Ji, R. Srikant, and A. Stoylar. Backpressure-basedpacket-by-packet adaptive routing in communication networks, 2010. Available at:http://arxiv.org/abs/1005.4984.

[8] F. Bai, D. D. Stancil, and H. Krishnan. Toward understanding characteristics of DedicatedShort Range Communications (DSRC) from a perspective of vehicular network engineers.In Proceedings of the IEEE/ACM International Conference on Mobile Computingand Networking (MOBICOM), pages 329–340, 2010.

[9] S. Basagni. Distributed clustering for ad hoc networks. In Proceedings of the Interna-tional Symposium on Parallel Architectures, Algorithms and Networks (I-SPAN),pages 310–315, 1999.

[10] S. Basagni, M. Mastrogiovanni, A. Panconesi, and C. Petrioli. Localized protocols for adhoc clustering and backbone formation: A performance comparison. IEEE Transactionson Parallel and Distributed Systems, 17(4):292–306, 2006.

[11] P. Basu, N. Khan, and T. D. C. Little. A mobility based metric for clustering in mobile adhoc networks. In Proceedings of the International Workshop on Wireless Networksand Mobile Computing (IWNMC), pages 413–418, 2001.

83


[12] A. Benslimane, T. Taleb, and R. Sivaraj. Dynamic clustering-based adaptive mobile gate-way management in integrated VANET-3G heterogeneous wireless networks. IEEE Jour-nal on Selected Areas in Communications, 29(3):559 –570, 2011.

[13] J. Blum, A. Eskandarian, and L. Hoffman. Mobility management in IVC networks. InProceedings of the IEEE Intelligent Vehicles Symposium (IV), pages 150–155, 2003.

[14] L. Bui, R. Srikant, and A. Stolyar. Novel architectures and algorithms for delay reductionin back-pressure scheduling and routing. In Proceedings of the IEEE InternationalConference on Computer Communications (INFOCOM) Mini, 2009.

[15] Mei Chao-Qun, Huang Hai-Jun, and Tang Tie-Qiao. Improving urban traffic by velocityguidance. In Intelligent Computation Technology and Automation (ICICTA), 2008International Conference on, volume 2, pages 383 –387, oct. 2008.

[16] M. Chatterjee, S. K. Das, and D. Turgut. WCA: A weighted clustering algorithm for mobilead-hoc networks. Journal of Cluster Computing, 5:193–204, 2002.

[17] B. Chen, K. Jamieson, H. Balakrishnan, and R. Morris. Span: An energy-efficient coordi-nation algorithm for topology maintenance in ad hoc wireless networks. ACM WirelessNetworks, 8(5):481–494, 2002.

[18] B. Das and V. Bharghavan. Routing in ad hoc networks using minimum connected domi-nating sets. In Proceedings of the IEEE International Conference on Communications(ICC’97), pages 376–380, 1997.

[19] J. M. De Fuentes, A. I. Gonzalez-Tablas, and A. Ribagorda. Overview of security issuesin vehicular ad-hoc networks. In Handbook of Research on Mobility and Computing:Evolving Technologies and Ubiquitous Impacts, 2011.

[20] M. K. Denko, J. Tian, T. K. R. Nkwe, and M. S. Obaidat. Cluster-based cross-layer designfor cooperative caching in mobile ad hoc networks. IEEE Systems Journal, 3(4):499–508, 2009.

[21] N. Dimokas, D. Katsaros, and Y. Manolopoulos. Energy-efficient distributed clustering inwireless sensor networks. Journal of Parallel and Distributed Computing, 70(4):371–383, 2010.

[22] E. Dror, C. Avin, and Z. Lotker. Fast randomized algorithm for 2-hops clustering in vehic-ular ad-hoc networks. Ad Hoc Networks, 2012. To appear.

[23] O. V. Drugan, T. Plagemann, and E. Munthe-Kaas. Detecting communities in sparseMANETs. IEEE/ACM Transactions on Networking, 19(5):1434–1447, 2011.

[24] K. Erciyes, O. Dagdeviren, D. Cokuslu, and D. Ozsoyeller. Graph-theoretic clusteringalgorithms in mobile ad hoc networks and wireless sensor networks - Survey. Appliedand Computational Mathematics, 6(2):162–180, 2007.

84


[25] P. Fan, J. Haran, J. Dillenburg, and P. Nelson. Cluster-based framework in vehicular ad-hocnetworks. In Ad-Hoc, Mobile, and Wireless Networks, volume 3738 of Lecture Notesin Computer Science, pages 32–42. 2005.

[26] L. Georgiadis, M. Neely, and L. Tassiulas. Resource allocation and cross-layer control inwireless networks. Foundations and Trends in Networking, 1(1), 2006.

[27] M. Gerla and J.T.-C. Tsai. Multicluster, mobile, multimedia radio network. ACMWirelessNetworks, 1(3):255–265, 1995.

[28] M. Gerlach. VaNeSe – An approach to VANET security. In Proceedings of the Interna-tional Workshop on Vehicle-to-Vehicle Communications (V2VCOM), 2005.

[29] J. Ghaderi, L.-L. Xie, and X. Shen. Hierarchical cooperation in ad hoc networks: Opti-mal clustering and achievable throughput. IEEE Transactions on Information Theory,55(8):3425–3436, 2009.

[30] W. B. Heinzelman, A. P. Chandrakasan, and H. Balakrishnan. An application-specificprotocol architecture for wireless microsensor networks. IEEE Transactions on WirelessCommunications, 1(4):660–670, 2002.

[31] L. Huang, S. Moeller, M. J. Neely, and B. Krishnamachari. LIFO-Backpressure achievesnear optimal utility-delay tradeoff, 2010. Available at: http://arxiv.org/abs/1005.4984.

[32] D. Katsaros, N. Dimokas, and L. Tassiulas. Social network analysis concepts in the designof wireless ad hoc network protocols. IEEE Network magazine, 24(6):2–9, 2010.

[33] D. Krajzewicz, G. Hertkorn, C. Rossel, and P. Wagner. SUMO (simulation of urban mo-bility): An open-source traffic simulation. Proceedings of the Middle East Symposiumon Simulation and Modelling (MESM), pages 183–187, 2002.

[34] S. Kuklinski and G. Wolny. Density based clustering algorithm for vehicular ad hoc net-works. International Journal of Internet Protocol Technology, 4(3):149–157, 2009.

[35] S. Leng, Y. Zhang, H.-W. Chen, L. Zhang, and K. Liu. A novel k-hop compound metricbased clustering scheme for ad hoc wireless networks. IEEE Transactions on WirelessCommunications, 8(1):367–375, 2009.

[36] I. Leontiadis, G. Marfiay, D. Mackz, C. Mascolo, G. Pauy, and M. Gerla. On the ef-fectiveness of an opportunistic traffic management system for vehicular networks. IEEETransactions on Intelligent Transportation Systems, 12(4):1537–1548, 2012.

[37] F. Li and Y. Wang. Routing in vehicular ad hoc networks: A survey. IEEE VehicularTechnology magazine, pages 12–22, 2007.

[38] X. Lin and S. Rasool. A distributed joint channel-assignment, scheduling and routing al-gorithm for multi-channel ad-hoc wireless networks. In Proceedings of the IEEE Inter-national Conference on Computer Communications (INFOCOM), pages 1118–1126,2007.

85


[39] Y.-W. Lin, Y.-S. Chen, and S.-L. Lee. Routing protocols in vehicular ad hoc networks:A survey and future perspective. Journal of Information Science and Engineering,26:913–932, 2010.

[40] Y. Luo, W. Zhang, and Y. Hu. A new cluster based routing protocol for VANET. InProceedings of the International Conference on Networks Security Wireless Com-munications and Trusted Computing (NSWCTC), pages 176–180, 2010.

[41] L. A. Maglaras and D. Katsaros. Layered backpressure scheduling for delay reduction inad hoc networks. In Proceedings of the IEEE International Symposium on a Worldof Wireless, Mobile and Multimedia Networks (WoWMoM), 2011.

[42] L. A. Maglaras and D. Katsaros. Distributed clustering in vehicular networks. In Proceed-ings of the International Workshop on Vehicular Communications and Networking(VECON), 2012.

[43] L. A. Maglaras and D. Katsaros. Layered backpressure routing for delay reduction in adhoc networks, 2012. Submitted.

[44] L. A. Maglaras and D. Katsaros. Clustering in urban environments: Virtual forces appliedto vehicles. In Proceedings of the IEEE Workshop on Emerging Vehicular Networks:V2V/V2I and Railroad Communications, 2013.

[45] L. A. Maglaras and D. Katsaros. Enhanced spring clustering in VANETs with obstructionconsiderations. In Proceedin of the IEEE Vehicular Technology Conference - Spring(VTC-Spring), 2013.

[46] L. A. Maglaras and D. Katsaros. Delay-efficient opportunistic routing in wireless ad hocnetworks, 2014. Submitted for journal publication.

[47] L. A. Maglaras and D. Katsaros. Social clustering of vehicles based on semi-Markovprocesses. IEEE Transactions on Vehicular Technology, August 2014. Under thirdround review.

[48] C. Maihofer. A survey of geocast routing protocols. IEEE Communications Surveys &Tutorials, 6(2):32–42, 2004.

[49] S. Moeller, A. Sridharan, B. Krishnamachari, and O. Gnawali. Routing without routes:The backpressure collection protocol. In Proceedings of the IEEE/ACM InternationalConference on Information Processing in Sensor Networks (IPSN), 2010.

[50] D. Naboulsi and M. Fiore. On the instantaneous topology of a large-scale urban vehicularnetwork: The Cologne case. In Proceedings of the ACM Interational Symposium onMobile Ad Hoc Networking and Computing (MOBIHOC), pages 167–176, 2013.

[51] M. J. Neely, E. Modiano, and C. E. Rohrs. Dynamic power allocation and routing fortime varying wireless networks. IEEE Journal on Selected Areas on Communications,23(1):89–103, 2005.

86


[52] M. Ni, Z. Zhong, and D. Zhao. MPBC: A mobility prediction-based clustering schemefor ad hoc networks. IEEE Transactions on Vehicular Technology, 60(9):4549–4559,2011.

[53] G. Pallis, D. Katsaros, M. D. Dikaiakos, N. Loulloudes, and L. Tassiulas. On the structureand evolution of vehicular networks. In Proceedings of the IEEE/ACM InternationalSymposium on Modelling, Analysis and Simulation of Computer and Telecommu-nication Systems (MASCOTS), pages 502–511, 2009.

[54] S. Panichpapiboon and W. Pattara-atikom. A review of information dissemination proto-cols for vehicular ad hoc networks. IEEE Communications Surveys & Tutorials, 2012.To appear.

[55] D. Papakostas and D. Katsaros. A simulation-based performance evaluation of a random-ized MIS-based clustering algorithm for ad hoc networks. Simulation Modelling: Prac-tice And Theory, 48, 2014.

[56] GeoNet EU FP7 project. http://www.geonet-project.eu/.

[57] S. T. Rager, E. N. Ciftcioglu, A. Yener, T. F. La Porta, and M. J. Neely. Distributedbackpressure protocols with limited state feedback. In Proceedings of the IEEE MilitaryCommunications Conference (MILCOM), pages 693–698, 2011.

[58] B. Ramakrishnan, R. S. Rajesh, and R. S. Shaji. CBVANET: A cluster based vehicularadhoc network model for simple highway communication. International Journal of Ad-vanced Networking and Applications, 2(4):755–761, 2011.

[59] E. Sakhaee and A. Jamalipour. A new stable clustering scheme for pseudo-linear highlymobile ad hoc networks. In Proceedings of the IEEE Global Communications Con-ference (GLOBECOM), pages 1169–1173, 2007.

[60] E. Sakhaeee and A. Jamalipour. Stable clustering and communications in pseudolin-ear highly mobile ad hoc networks. IEEE Transactions on Vehicular Technology,57(6):3769–3777, 2008.

[61] R. A. Santos, O. Alvarez, and A. Edwards. Performance evaluation of two location-basedrouting protocols in vehicular ad-hoc networks. In Proceedings of the IEEE VehicularTechnology Conference - Fall (VTC-Fall), volume 4, pages 2287–2291, 2005.

[62] R. A. Santos, R. M. Edwards, and N. L. Seed. Using the cluster-based location rout-ing (cblr) algorithm for exchanging information on a motorway. In Proceedings of theInternational Workshop on Mobile and Wireless Communications Network, pages212–216, 2002.

[63] C. Shea, B. Hassanabadi, and S. Valaee. Mobility-based clustering in VANETs usingaffinity propagation. In Proceedings of the IEEE Global Communications Conference(GLOBECOM), 2009.

87


[64] U. K. Shukla. Backpressure policies for wireless ad hoc networks, 2010. Technical Report,Virginia Polytechnic Institute and State University.

[65] R. Simmons, B. Browning, Z. Yilu, and V. Sadekar. Learning to predict driver route anddestination intent. In Proceedings of the IEEE Intelligent Transportation SystemsConference (ITSC), 2006.

[66] Q. Song and X. Wang. Efficient routing on large road networks using hierarchical com-munities. IEEE Transactions on Intelligent Transportation Systems, 12(1):132–140,2011.

[67] T. Song, W. Xia, T. Song, and L. Shen. A cluster-based directional routing protocol inVANET. In Proceedings of the IEEE International Conference on CommunicationTechnology (ICCT), pages 1172–1175, 2010.

[68] E. Souza, I. Nikolaidis, and P. Gburzynski. A new Aggregate Local Mobility (ALM) clus-tering algorithm for VANETs. In Proceedings of the IEEE International Conferenceon Communications (ICC), 2010.

[69] I. Stojmenovic, M. Seddigh, and J. Zunic. Dominating sets and neighbor elimination-based broadcasting algorithms in wireless networks. IEEE Transactions on Paralleland Distributed Systems, 13(1):14–25, 2002.

[70] L. Sun, Y. Wu, J. Xu, and J. Xu. An RSU-assisted cluster head selection and backup proto-col. In Proceedings of the International IEEE Conference on Advanced InformationNetworking and Applications Workshops (AINAW), pages 581 –587, 2012.

[71] C. Suthaputchakun and Z. Sun. Routing protocol in intervehicle communication systems:A survey. IEEE Communications magazine, 49(12):150–156, 2011.

[72] L. Tassiulas and A. Ephremides. Stability properties of constrained queueing systems andscheduling policies for maximum throughput in multihop radio networks. IEEE Transac-tions on Automatic Control, 37(12):1936–1948, 1992.

[73] M. Torrent-Moreno, M. Killat, and H. Hartenstein. The challenges of robust inter-vehiclecommunications. In Proceedings of the IEEE Vehicular Technology Conference -Fall (VTC-Fall), volume 1, pages 319–323, 2005.

[74] S.-L. Tsao and C.-M. Cheng. Design and evaluation of a two-tier peer-to-peer traffic infor-mation system. IEEE Communications magazine, 49(5):165–172, 2011.

[75] C. Tselikis, S. Mitropoulos, N. Komninos, and C. Douligeris. Degree-based clusteringalgorithms for wireless ad hoc networks under attack. IEEE Communications Letters,16(5):619–621, 2012.

[76] Y.-C. Tseng, S.-Y. Ni, Y.-S. Chen, and J.-P. Sheu. The broadcast storm problem in a mobilead hoc network. ACM Wireless Networks, 8(2–3):153–167, 2002.

88


[77] S. Vodopivec, J. Bester, and A. Kos. A survey on clustering algorithms for vehicular ad-hocnetworks. In Proceedings of the International Conference on Telecommunicationsand Signal Processing (TSP), pages 52–56, 2012.

[78] T. Wang and G. Wang. TIBCRPH: Traffic infrastructure based cluster routing protocolwith handoff in VANET. In Proceedings of the Wireless and Optical CommunicationsConference (WOCC), 2010.

[79] A. Wegener, H. Hellbruck, C. Wewetzer, and A. Lubke. Vanet simulation environment withfeedback loop and its application to traffic light assistance. In GLOBECOM Workshops,2008 IEEE, pages 1 –7, 30 2008-dec. 4 2008.

[80] S. Widodo, T. Hasegawa, and S. Tsugawa. Vehicle fuel consumption and emission esti-mation in environment-adaptive driving with or without inter-vehicle communications. InIntelligent Vehicles Symposium, 2000. IV 2000. Proceedings of the IEEE, pages 382–386, 2000.

[81] J. Wu and H. Li. A dominating-set-based routing scheme in ad-hoc wireless networks.Telecommunication Systems, 18:13–36, 2001.

[82] Y. Xia, C. K. Yeo, and B. S. Lee. Hierarchical cluster based routing for highly mobileheterogeneous MANET. In Proceedings of the International Conference on Networkand Service Security (N2S), 2009.

[83] L. Ying, S. Shakkotai, and A. Reddy. On combining shortest-path and back-pressure rout-ing over multihop wireless networks. In Proceedings of the IEEE International Con-ference on Computer Communications (INFOCOM), 2009.

[84] L. Ying, R. Srikant, and D. F. Towsley. Cluster-based back-pressure routing algorithm. InProceedings of the IEEE International Conference on Computer Communications(INFOCOM), pages 484–492, 2008.

[85] O. Younis and S. Fahmy. HEED: A hybrid, energy-efficient, distributed clustering approachfor ad-hoc sensor networks. IEEE Transactions on Mobile Computing, 3(4):366–379,2004.

[86] O. Younis, M. Krunz, and S. Ramasubramanian. Node clustering in wireless sensor net-works: Recent developments and deployment challenges. IEEE Network magazine,20(3):20–25, 2006.

[87] J. Y. Yu and P. H. J. Chong. A survey of clustering schemes for mobile ad hoc networks.IEEE Communications Surveys and Tutorials, 7:32–48, 2005.

[88] Y. Zhang and J.M. Ng. A distributed group mobility adaptive clustering algorithm formobile ad hoc networks. In Proceedings of the IEEE International Conference onCommunications (ICC), pages 3161–3165, 2008.

89

Documents

REDUCTION 2011-2014 Deliverable 1.3 VANET Packet ... · D1.3 [VANET Packet Scheduling/Routing and Information Dissemination] Executive Summary This report describes the network architecture