Impacts of Self-organized Mechanisms in Wireless Sensor ...theses.insa-lyon.fr/publication/2008ISAL0026/these.pdf · No: 2008-ISAL-0026 Year 2008 Thesis Impacts of Self-organized

No: 2008-ISAL-0026 Year 2008

Thesis

Impacts of Self-organized Mechanisms in

Wireless Sensor Networks

defend at

l’Institut National des Sciences Appliquées de Lyon

for the degree of

Doctor of Philosophy

Ecole Doctorale Informatique et Information pour la Societe

by

JiaLiang LU

dissertation May 6

Committee in charge: AL AGHA Khaldoun (Reviewer)

Professor, University Paris Sud

BARTHEL Dominique (Examiner)

Senior researcher, France Telecom R&D, Grenoble

FLEURY Eric (Co-supervisor)

Professor, ENS Lyon

Noel Thomas (Examiner)

Professor, Louis Pasteur University Strasbourg

Simplot-Ryl David (Reviewer)

Professor, University Lille 1

Ubeda Stephane (Examiner)

Professor, INSA Lyon

VALOIS Fabrice (Co-supervisor)

Associated professor, INSA Lyon

ii

“To observe without the observer”

Jiddu Krishnamurti

iii

Remerciements v

Remerciements

Cette thèse a été effetuée au sein du Centre d’Innovation en Télécommunications et Inté-

gration de Services (CITI), à l’INSA de Lyon, et dans le projet ARES de l’INRIA Rhône-

Alpes, en collaboration avec le Centre de Recherche et Développement de FranceTélé-

com Grenoble et Beijing.

Je tiens tout d’abort à remercier Fabrice Valois sans, qui cette thèse n’aurait, je

pense, jamais eu lieu. Je le remercie de m’avoir encadré non seulement durant la thèse

mais également pendant mon mastère de recherche, en me donnant les outils et surtout

l’esprit de recherche pour réussir dans ce domaine. Je le remercie de m’avoir me toujour

soutenu mais aussi crétiqué pour réussir cette thèse. Je remercie également Eric Fleury

qui m’a co-encadré durant cette thèse et a su partager son expertise scientifique.

Je remercie tout particuliérement Monsieur Khaldoun Al Agha, Professeur à l’Université

de Paris-Sud, et Monsieur David Simplot-Ryl, Professeur à l’Université de Lille qui ont

accepté de juger mes traveaux de thèse et d’en être les rapporteurs.

Je remercie également Monsieur Dominique Barthel, Chercheur Sénior à FranceTélé-

com R&D de Grenoble, Monsieur Thomas Noel, Professeur à l’Université de Strasbourg

et Monsieur Stéphane Ubeda, Professur à l’INSA de Lyon d’avoir accepté de faire partie

de ce jury de thèse et de s’être intéressés à mon travail.

Je remercie également Monsieur Mischa Dohler avec qui j’ai passé ces moments de

concentration riche, tant pour son apport scientifique que personnel.

Je tiens à remercier Yvan Royon et Noha Ibrahim pour les discussions scientifiques,

sociales ou personnelles que nous avons eu. Je remercie également Fabrice Theo-

leyre, Nathalie Mitton, Tahiry Razafindralambo, Karel Heurtefeux, Elyès Ben Hamida,

Thomas Watteyne, Katia Jaffrès-Runser avec qui les nombreuse exchanges scientifiques

ont eu lieu. Je remercie Haila Wang, Song Wang et Yu Zhang qui m’ont aidé à dévelop-

per les activité de recherche à FranceTélécom Beijing. Je remercie également tous les

chercheurs du laboratoire CITI (pêle-mêle Jean-Marie Gorce, Guillaume Chélius, Is-

abelle Augé-Blum, et tant d’autres que je ne nommerai pas par manque de place).

Enfin, je tiens à remercier Yi, mon epouse, qui a su m’accompagner (physiquement et

moralement) dans cette grande expérience scientifique mais surtout personnelle qu’est

une thèse.

Résumé

Un réseau de capteurs est un réseau radio multi-sauts formé par une quantité impor-

tante de capteurs identiques. Contrairement aux noeuds d’un réseau ad hoc, les cap-

teurs ont des nombreuses contraintes, notamment en termes de consommation d’énergie,

de capacité de calcul et de stockage. L’absence d’infrastructure et de contrôle centralisé

implique une collaboration efficace et pertinente des noeuds du réseau de capteurs.

Ces travaux de thèse s’inscrivent dans la problématique d’auto-organisation des

réseaux de capteurs. De notre point de vue, l’auto-organisation est un problème fonda-

mental des réseaux de capteurs conduisant à construire une vue logique de la topolo-

gie du réseau physique et de fournir des protocoles de communication basés sur cette

vue logique. Nous avons donc proposé une architecture d’auto-organisation (nommée

FISCO) permettant d’organiser le réseau - via des interactions locales uniquement-

sous le forme soit d’un arbre soit d’un treillis. Les propriétés structurelles et les perfor-

mances de FISCO ont été évaluées puis comparées aux principaux mécanismes d’auto-

organisation de la littérature. Cette architecture exhibe de très bonnes performances

en termes de stabilité, persistance, robustesse et surtout gestion de l’énergie.

Ensuite, nous avons revisité les principaux défis de la communication des réseaux

de capteurs en se basant sur l’auto-organisation préalablement introduite. Nous nous

sommes donc intéressé aux problèmes clefs suivant : allocation dynamique d’adresses,

protocole d’inondation, dissémination et agrégation de données dans les réseaux de

capteurs. L’impact de l’auto-organisation est clairement constaté en comparant avec

les approches classiques. Notons que la structure d’auto-organisation offre également

une solution très efficace et très flexible pour la gestion de puits mobiles et multiples

dans un réseau de capteurs. Le travail réalisé dans cette partie repose sur l’utilisation

d’outils issus de la théorie des graphes, de l’algorithmique distribuée mais aussi de

l’évaluation de performances.

En parallèle, nous avons cherché à justifier l’utilisation d’approche auto-organisée

(i.e. structurée) par rapport aux approches à plat. Pour ce faire, nous avons introduit

une nouvelle métrique d’évaluation hérité de la notion d’entropie statistique. Cette

métrique est la première métrique quantitative qui mesure l’ordre de l’organisation

dans un réseau sans fil multi-sauts. Cette métrique ouvre de nombreuses pistes de

recherches dans le domaine de l’auto-organisation des réseaux.

Nous avons aussi proposé une plate-forme d’expérimentation composant des 30 cap-

teurs. Cette plate-forme expérimentale est utilisée pour valider les solutions proposées.

Ainsi, les contributions des travaux de la thèse couvrent à la fois le domaine théorique,

mais aussi la simulation et l’expérimentation.

Contents vii

Contents

1 Introduction 1

1.1 Popularity of Wireless Sensor Networks . . . . . . . . . . . . . . . . . . 2

1.1.1 Advanced sensor nodes . . . . . . . . . . . . . . . . . . . . . . . 2

1.1.2 Ease of deployment . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.1.3 Versatile applications . . . . . . . . . . . . . . . . . . . . . . . . 5

1.2 Description of WSN in Our Consideration . . . . . . . . . . . . . . . . . 6

1.3 Motivations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

1.3.1 Setting up and organizing . . . . . . . . . . . . . . . . . . . . . . 7

1.3.2 Managing and maintaining . . . . . . . . . . . . . . . . . . . . . 8

1.3.3 Support of applications and services . . . . . . . . . . . . . . . . 9

1.3.4 Energy efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

1.4 Organization of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2 State of the Art 13

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2.2 Self-configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.2.1 Autoconfiguration in IP networks . . . . . . . . . . . . . . . . . . 15

2.2.2 Address conflict detection based solutions . . . . . . . . . . . . . 17

2.2.3 Distributed DHCP solutions . . . . . . . . . . . . . . . . . . . . . 18

2.2.4 Synthesis on self-configuration . . . . . . . . . . . . . . . . . . . 20

2.3 Self-organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2.3.1 Network model formalism . . . . . . . . . . . . . . . . . . . . . . 22

2.3.2 Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

2.3.3 Virtual backbone . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

2.3.3.1 Connected dominating set . . . . . . . . . . . . . . . . 25

2.3.3.2 Maximal independent set . . . . . . . . . . . . . . . . . 26

2.3.3.3 Relative neighborhood graph . . . . . . . . . . . . . . . 27

2.3.3.4 Local minimum spanning tree . . . . . . . . . . . . . . 28

2.3.4 Source dependent . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

2.3.5 Synthesis on self-organization . . . . . . . . . . . . . . . . . . . . 29

2.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3 FISCO: An Autonomous Architecture for WSN 33

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

3.2 FISCO Highlights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

3.2.1 Problem statement . . . . . . . . . . . . . . . . . . . . . . . . . . 34

Contents

3.2.2 Hierarchical structure . . . . . . . . . . . . . . . . . . . . . . . . 353.2.3 Two-level address allocation . . . . . . . . . . . . . . . . . . . . . 36

3.2.4 FISCO messages . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

3.3 Join of Nodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383.3.1 One-hop address allocation . . . . . . . . . . . . . . . . . . . . . 39

3.3.2 Two-hop address allocation . . . . . . . . . . . . . . . . . . . . . 40

3.3.3 Creation of a new partition . . . . . . . . . . . . . . . . . . . . . 403.4 Departure of Nodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

3.4.1 Departure of a member node . . . . . . . . . . . . . . . . . . . . 41

3.4.2 Departure of a gateway node . . . . . . . . . . . . . . . . . . . . 42

3.4.3 Departure of a leader node . . . . . . . . . . . . . . . . . . . . . 433.5 Partition Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

3.5.1 Partition splitting . . . . . . . . . . . . . . . . . . . . . . . . . . 44

3.5.2 Partition detection . . . . . . . . . . . . . . . . . . . . . . . . . . 453.5.3 Partition merge . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

3.6 Local Re-organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

3.7 Mesh Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 483.8 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

3.8.1 Analysis on FISCO backbone . . . . . . . . . . . . . . . . . . . . 51

3.8.2 Message complexity analysis . . . . . . . . . . . . . . . . . . . . 55

3.9 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . 553.9.1 Self-configuration related properties . . . . . . . . . . . . . . . . 56

3.9.1.1 Configuration message overhead . . . . . . . . . . . . . 57

3.9.1.2 Configuration latency . . . . . . . . . . . . . . . . . . . 583.9.1.3 Evolution of partitions during configuration . . . . . . . 59

3.9.1.4 Energy consumption for configuration . . . . . . . . . . 60

3.9.2 Self-organization related properties . . . . . . . . . . . . . . . . . 603.9.2.1 FISCO backbone . . . . . . . . . . . . . . . . . . . . . . 61

3.9.2.2 Local structure characteristics . . . . . . . . . . . . . . 62

3.9.2.3 Long term message overhead . . . . . . . . . . . . . . . 633.9.2.4 Energy consumption for organization . . . . . . . . . . 63

3.9.3 Impact of re-organization on lifetime . . . . . . . . . . . . . . . . 65

3.9.4 Energy consumption of FISCO mesh organization . . . . . . . . 65

3.10 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

4 Data Dissemination and Data Aggregation 69

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

4.2 Data Dissemination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 724.2.1 Overview of data dissemination schemes . . . . . . . . . . . . . . 72

4.2.2 Backbone based data dissemination . . . . . . . . . . . . . . . . 74

4.2.2.1 Directed query forwarding . . . . . . . . . . . . . . . . 754.2.2.2 Data notification and data forwarding . . . . . . . . . . 76

4.2.3 Management of multiple mobile sinks over BBDD . . . . . . . . 77

4.2.4 Analysis on data dissemination . . . . . . . . . . . . . . . . . . . 78

viii

Contents

4.3 Data Aggregation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

4.3.1 Where to aggregate . . . . . . . . . . . . . . . . . . . . . . . . . 82

4.3.2 When to aggregate . . . . . . . . . . . . . . . . . . . . . . . . . . 83

4.3.3 Adaptive-ARMA model for data aggregation . . . . . . . . . . . 85

4.3.3.1 ARMA model . . . . . . . . . . . . . . . . . . . . . . . 85

4.3.3.2 Local A-ARMA computation . . . . . . . . . . . . . . . 86

4.3.4 Analysis on A-ARMA . . . . . . . . . . . . . . . . . . . . . . . . 88

4.3.4.1 Accuracy and efficiency . . . . . . . . . . . . . . . . . . 88

4.3.4.2 Under erroneous measurements . . . . . . . . . . . . . . 90

4.3.5 Highlights of A-ARMA technique . . . . . . . . . . . . . . . . . . 93

4.4 SODA Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

4.4.1 Spatial packet merge on leader nodes . . . . . . . . . . . . . . . . 95

4.4.2 Performance evaluation on data collection . . . . . . . . . . . . . 95

4.4.2.1 Message cost during data collection . . . . . . . . . . . 97

4.4.2.2 Active time during data collection . . . . . . . . . . . . 97

4.4.2.3 Energy consumption during data collection . . . . . . . 100

4.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

5 Entropy of Organization 105

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

5.2 Original Definitions of Entropy . . . . . . . . . . . . . . . . . . . . . . . 106

5.3 Extended Definition of Entropy . . . . . . . . . . . . . . . . . . . . . . . 107

5.3.1 Formulation of entropy . . . . . . . . . . . . . . . . . . . . . . . 107

5.3.2 Interpretations of this formulation . . . . . . . . . . . . . . . . . 109

5.3.3 Application of entropy on a simple network . . . . . . . . . . . . 109

5.3.3.1 Flat organization scheme . . . . . . . . . . . . . . . . . 110

5.3.3.2 LMST self-organization scheme . . . . . . . . . . . . . . 110

5.4 Entropy Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

5.5 Entropy Variation Evaluation . . . . . . . . . . . . . . . . . . . . . . . . 115

5.5.1 When single node disappears . . . . . . . . . . . . . . . . . . . . 115

5.5.2 When several nodes disappear . . . . . . . . . . . . . . . . . . . 117

5.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120

6 Test-bed 123

6.1 Description of Test-bed . . . . . . . . . . . . . . . . . . . . . . . . . . . 123

6.1.1 Imote2 hardware platform . . . . . . . . . . . . . . . . . . . . . . 124

6.1.2 Basic communication architecture . . . . . . . . . . . . . . . . . 125

6.1.3 MAC layer based on CC2420 driver . . . . . . . . . . . . . . . . 126

6.1.4 Detailed architecture of network layer . . . . . . . . . . . . . . . 126

6.2 Implementation of FISCO . . . . . . . . . . . . . . . . . . . . . . . . . . 128

6.2.1 FISCO FSM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128

6.2.2 FISCO message format . . . . . . . . . . . . . . . . . . . . . . . 129

6.3 Methodology of Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

6.3.1 Funtional test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

ix

Contents

6.3.2 Performance test . . . . . . . . . . . . . . . . . . . . . . . . . . . 1316.4 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

6.4.1 First node configuration . . . . . . . . . . . . . . . . . . . . . . . 1326.4.2 Configure to member . . . . . . . . . . . . . . . . . . . . . . . . . 1336.4.3 Configure to leader . . . . . . . . . . . . . . . . . . . . . . . . . . 133

6.5 Conclusions and Discussions . . . . . . . . . . . . . . . . . . . . . . . . . 134

7 Conclusions and Persperctives 137

7.1 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1377.1.1 Autonomous architecture . . . . . . . . . . . . . . . . . . . . . . 1377.1.2 Self-organized mechanisms . . . . . . . . . . . . . . . . . . . . . . 1387.1.3 Framework solution . . . . . . . . . . . . . . . . . . . . . . . . . 1397.1.4 Entropy for quantifying organization . . . . . . . . . . . . . . . . 1397.1.5 Experimental Imote2 test-bed . . . . . . . . . . . . . . . . . . . . 139

7.2 Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1407.2.1 Enrich the functionalities of the architecture . . . . . . . . . . . 1407.2.2 Security aspects of the solutions . . . . . . . . . . . . . . . . . . 1417.2.3 Scaling laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1417.2.4 Application tunable parameters . . . . . . . . . . . . . . . . . . . 142

x

List of Figures xi

List of Figures

1.1 General architecture of a sensor node . . . . . . . . . . . . . . . . . . . . . 21.2 Imote2 node . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2.1 DHCP server-client configuration . . . . . . . . . . . . . . . . . . . . . . . . 152.2 IPv6 stateless address autoconfiguration . . . . . . . . . . . . . . . . . . . . 162.3 Segmented address pools in Buddy [1] . . . . . . . . . . . . . . . . . . . . . 182.4 MANETConf [2] configuration process . . . . . . . . . . . . . . . . . . . . . 192.5 Prophet [3] address allocation . . . . . . . . . . . . . . . . . . . . . . . . . 202.6 A Unit Disk Graph of 120 nodes and radius=0.16 . . . . . . . . . . . . . . . 222.7 2.5-hop and 3-hop coverage set . . . . . . . . . . . . . . . . . . . . . . . . . 242.8 Example of CDS construction in [4] and [5] . . . . . . . . . . . . . . . . . . 262.9 RNG: pruning the longest edge . . . . . . . . . . . . . . . . . . . . . . . . . 282.10 Snapshots of self-organization structures: a network with 120 nodes, radius=0.16 30

3.1 FISCO structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353.2 LDBR message format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363.3 Address space management in FISCO . . . . . . . . . . . . . . . . . . . . . 373.4 New node configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393.5 One-hop address allocation . . . . . . . . . . . . . . . . . . . . . . . . . . . 403.6 Two-hop address allocation . . . . . . . . . . . . . . . . . . . . . . . . . . . 413.7 The departure of a gateway . . . . . . . . . . . . . . . . . . . . . . . . . . 423.8 The departure of a leader . . . . . . . . . . . . . . . . . . . . . . . . . . . 443.9 Local re-organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 483.10 Mesh backbone with different values of p . . . . . . . . . . . . . . . . . . . . 493.11 A topology with multiple gateways when p = 1 . . . . . . . . . . . . . . . . 503.12 FISCO mesh analysis with radius=0.18 . . . . . . . . . . . . . . . . . . . . 513.13 The number of nodes in S within node u’s neighborhood . . . . . . . . . . . 523.14 Motif: hexagon, octagon, dodecagon . . . . . . . . . . . . . . . . . . . . . . 533.15 Paving the space with hexagons . . . . . . . . . . . . . . . . . . . . . . . . 543.16 Configuration message overhead, radius=0.20 . . . . . . . . . . . . . . . . . 583.17 Configuration latency, radius=0.20 . . . . . . . . . . . . . . . . . . . . . . . 593.18 Evolution of the number of partitions from 1 to 400 nodes . . . . . . . . . . . 603.19 Energy saving in configuration, radius=0.20 . . . . . . . . . . . . . . . . . . 613.20 Cardinality of dominating set, radius=0.20 . . . . . . . . . . . . . . . . . . . 613.21 Properties of FISCO local structures, radius=0.14 . . . . . . . . . . . . . . . 623.22 FISCO sent and received control message cost, radius=0.14 . . . . . . . . . . 633.23 CDS and FISCO control message cost, radius=0.20 . . . . . . . . . . . . . . 64

List of Figures

3.24 Energy saving in organization, radius=0.20 . . . . . . . . . . . . . . . . . . 643.25 Impact of re-organization in FISCO on network lifetime, radius=0.20 . . . . . 653.26 Impact of mesh organization on energy saving, radius=0.20 . . . . . . . . . . 66

4.1 Communication in WSN . . . . . . . . . . . . . . . . . . . . . . . . . . . . 704.2 Data dissemination: data centric routing and rendezvous systems . . . . . . . 744.3 Directed query forwarding structure on FISCO mesh backbone . . . . . . . . 754.4 Data forwarding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 764.5 Support for sink mobility . . . . . . . . . . . . . . . . . . . . . . . . . . . . 784.6 Analysis of communication overhead (average number of queries per sink q=50,

average number of events per source node e=500, N=10000, r=0.1, α=0.3 for

TTDD) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 804.7 Analysis of communication overhead (m=5 mobile sinks, n=10 source nodes,

N=10000, r=0.1, α=0.3 for TTDD) . . . . . . . . . . . . . . . . . . . . . . 814.8 Block diagram of A-ARMA. . . . . . . . . . . . . . . . . . . . . . . . . . . 874.9 Applying ARMA and A-ARMA models on indoor temperatures. . . . . . . . 894.10 Accuracy and efficiency of the A-ARMA(2,2) on 720 samples. . . . . . . . . . 904.11 Accuracy and efficiency under independent errors . . . . . . . . . . . . . . . 924.12 Accuracy and efficiency under consecutive errors . . . . . . . . . . . . . . . . 934.13 Number of messages in data collection . . . . . . . . . . . . . . . . . . . . . 974.14 Average active time of sensor nodes in data collection (ideal MAC scheduling) 984.15 Average active time of sensor nodes in data collection (BMAC) . . . . . . . . 994.16 Factor of active time between BMAC and ideal MAC scheduling . . . . . . . 1004.17 Distribution of active modes among total active time . . . . . . . . . . . . . 1014.18 Average energy consumption on nodes . . . . . . . . . . . . . . . . . . . . . 102

5.1 Flat organization. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1105.2 LMST organization. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1115.3 Entropy value of a topology of 4 nodes . . . . . . . . . . . . . . . . . . . . . 1115.4 Average entropy value on networks of 200 nodes, radius=0.16 . . . . . . . . . 1135.5 Average broadcast cost on networks of 200 nodes, radius=0.16 . . . . . . . . 1145.6 Probability Density Function of entropy variation for simple node disappearance 1165.7 Entropy variation of self-organization schemes for random geometric network,

radius=0.16 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1175.8 Entropy variation of self-organization schemes under different densities . . . . 119

6.1 Basic communication architecture . . . . . . . . . . . . . . . . . . . . . . . 1256.2 Modules description in network layer . . . . . . . . . . . . . . . . . . . . . . 1276.3 Finite state machine of FISCO configuration . . . . . . . . . . . . . . . . . . 1286.4 Message header of FISCO . . . . . . . . . . . . . . . . . . . . . . . . . . . 1296.5 Setup for 30 Imotes test-bed platform . . . . . . . . . . . . . . . . . . . . . 131

xii

Introduction 1

Introduction 1In the past decade, wireless technologies have become key technologies, offering mo-

bile and flexible communications for industries, enterprises and individuals. The GSM

and CDMA wireless mobile networks are reaching the leadership as daily voice com-

munications media, surpassing the wired telephone network. They are evolving from

2nd generation to 3rd generation, with larger bandwidth, higher data transmission rate

and better support of data and video transmissions. On the other side, the widely

deployed wireless LANs with WiFi technology have also been great successes. The

standard IEEE 802.11a/b/g [6] provides a bandwidth up to 54Mbps. This technology

enables Internet connections everywhere as long as one has a laptop or a PDA. Around

each person, Wireless Personal Area Networks (Wireless PANs) based on Bluetooth

[7] technology, begin to appear which offer the possibility to connect different devices

such as wireless handsets, PDAs and game controllers together to enrich the interac-

tions between humans and environment. Another wireless technology Radio Frequency

Identification (RFID) [8], standing for clear and contact-less identification of objects,

enables rapid and automatic data acquisition via radio waves. It is changing the way

of organization in retail, transportation and logistic.

Beside all these wireless technologies, the Wireless Sensor Networks (WSNs) [9, 10]

became popular in the past five years. More and more sensing applications take advan-

tages of WSNs based on the collaborative efforts of a large number of sensor nodes. In

this work, we place our focus on this new type of wireless networks. Particularly, we

try to find autonomous solutions for networking and data communication in WSNs.

In this chapter, we first discuss the popularity of wireless sensor networks from three

principal angles: the designs of advanced sensor nodes, the ease of deployment and

the versatile WSN applications. Secondly, the motivations of this work are stated as

providing an autonomous architecture for WSN. This architecture’s objectives are to set

up, organize, operate and manage the WSN as well as to support various applications

and services.

Introduction

Sensor Broad

RAM Flash

Infrared sensor

Humidity sensor

Temperature sensor

Radio

Micro−processor

Power

Supply

Con−

nector

A/D Converter

Sensor

Con−nector

Antenna

Communication Broad

Power Supply Broad

Figure 1.1: General architecture of a sensor node

1.1 Popularity of Wireless Sensor Networks

Wireless Sensor Network becomes one of the hottest topics in both research and com-

mercial fields. How can we explain its popularity?

1.1.1 Advanced sensor nodes

The individual processing, storage and communication capabilities of sensor nodes are

relatively limited when comparing to personal computers or devices. However each of

them uses low cost components and targets for low power operation. The size of a

sensor node, as an individual communication unit, has also been dramatically reduced

even comparing to the most up-to-date personal devices. Small size, low cost and low

power operation make wireless sensor nodes significantly different from other wireless

devices.

We distinguish a sensor node from a sensor. A sensor node has capability to sense

(taking measure from physical environment), to communicate (to other sensor nodes)

and it has a power supply module (see Fig. 1.1). The sensing part is generally called

a sensor (or a sensor board) which is a transduction device that measures some phys-

ical quantities and converts them to electrical quantities. The sensor is connected to

a communication board, so called the main board, via analog or digital Iuput/Output

(I/O) interfaces. With the increase of versatile wireless sensor network applications,

2

Popularity of Wireless Sensor Networks

many advanced small powerful main boards such as Mica2/MicaZ [11], µAMPS [12]

and Imote2 (Fig. 1.2(a) and 1.2(b)), have been developed to support in-network data

processing and reporting of measurements. Table 1.1 gives the detailed hardware in-

formation of these advanced sensor nodes. Besides, various modules have also been

designed to supply energy to sensor nodes’ communication boards and sensor boards.

Traditional batteries, solar chargers [13] and energy extracted from vibration [14] have

been used as the possible ways of power supply.

Table 1.1: Detailed specifications of MICA2/MicaZ, µAMPS and Imote2

Components Mica2/MicaZ µAMPS Imote2

Processor ATmega128L SA-1110 Intel PXA271Processor speed 7.3728 MHz 206 MHz 13-416 MHzSRAM 4 KBytes 16 KBytes 256 KBytesProgram Memory 128 KBytes 128 KBytes 32 MBytesData storage 512 KBytes 512 KBytes 32 MBytesSerial Interface UART GPIO UART&GPIO&mini-UOther Interfaces DIO, I2C, SPI No SPI I2C, I2SBattery 2x AA 4x AAA 3x AAAExternal Power 2.7 - 3.3 V 3.6 V 3.2 - 4.5 VExpansion Connector 51 pin 256 pin JTAGRadio CC1000[15]/CC2420 Bluetooth[7] CC2420 [16]Software TinyOs µOS TinyOS, Linux[17]

It is worth noting that the research interests are also multiplied with the develop-

ment of event-driven embedded operating system such as TinyOS [18]. TinyOS is an

event-based operating environment/framework designed for using with embedded net-

worked sensors, to support intensive concurrent operations needed by sensor network

applications. It features a component-based architecture which enables rapid innova-

tion and implementation while minimizing code size as required by the severe memory

constraints inherent in sensor networks.

The design of advanced sensor nodes can be summarized as to target the following

three purposes:

1. Low-energy operations. In the most cases, sensor nodes are battery supplied.

Regarding to the variety of deployment terrains, it may be unfeasible to renew

their energy resources. Therefore low power processing and communication are

required to prolong the lifetime of WSNs. In the earlier design of sensor nodes,

low-frequency micro-controller is used (in Mica2/MicaZ series) to ensure low con-

sumption during the processing. In order to increase the processing capability,

3

Introduction

(a) Imote2 hardware components (b) Imote2 architecture

Figure 1.2: Imote2 node

more powerful CPU have been used lately in Imote2. Thanks to the low-frequency

mode of these CPUs, moderate power operation is also possible. However there

is always a trade-off between the processing capability and energy saving. Radio

communication is another source of energy dissipation, even much more costly

than processing operations. A comparison between computation and communi-

cation cost [19] reveals that 3000 instructions can be executed for the same cost

as the transmission of one bit over 100m. Using low power radio transceivers such

as IEEE 802.15.4 [20] compatible radio module becomes the trend.

2. Size and cost constraints. In order to benefit from collaborative operations of

sensor nodes, a wireless sensor network is formed by hundreds of sensor nodes.

Therefore sensor nodes should have small size, light weight and low cost to support

large scale deployments. Mica2/MicaZ, µAMPS and Imote2 have shown the pos-

sibility of embedding processing, communication and sensing on small platforms.

However the cost of sensor nodes is still too high nowadays.

3. Support of operating system. By running an operating system on small sensor

nodes, both the communication protocols and applications can be developed with

more flexibilities.

1.1.2 Ease of deployment

The needs for large scale sensor network and for intensive collaborative operations

and computations change the way of deploying sensors. Sensor nodes are required

to form multi-hop networks to communicate among themselves. It is a significant

4

Popularity of Wireless Sensor Networks

improvement over traditional sensors in the way that nodes can be randomly deployed

in inaccessible terrains in case of disaster relief operation. In addition to the advances

in sensor nodes hardware architectures, the power of wireless sensor networks also lies

on the ability to deploy large number of nodes which configure and organize themselves.

Through advanced distributed networking protocols, the sensor nodes form a virtual

environment that extends the reach of cyberspace out into the physical world. Although

the capability of any single sensor node is minimal, the composition of hundreds and

thousands of nodes offers radical new technological possibilities.

Unlike traditional wireless devices such as cell phones and PDAs, wireless sensor

nodes do not need to communicate directly with the nearest base station, but only

with nodes within its local area. Instead of relying on a pre-deployed infrastructure,

each individual sensor node becomes a part of the overall infrastructure. Therefore

a wireless sensor network should be able to configure sensor nodes under any possible

node placement upon the deployment. Nevertheless real systems must place constraints

on actual node placement; for example the physical topology should be connected to

ensure the connectivity on network level. The wireless sensor network must be capable

of providing feedback when these constraints are violated.

In addition to an initial configuration phase, a WSN also has the capability to adapt

to changing environmental conditions. The envisioned flexible self-organizing network

architectures dynamically adapt themselves to support arrival of new nodes or to ex-

pand to cover a larger geographic region. Furthermore, the network can automatically

adapt to compensate for node failures. Throughout the lifetime of a deployment, nodes

may also be relocated or other objects may appear in the radio environment which

might interfere with the communication in the WSN. The network should be able to

automatically reconfigure in order to tolerate these occurrences.

1.1.3 Versatile applications

The raise of its versatile applications is one of the propelling forces for researches in

WSNs. Wireless sensor network applications cover a large range of applications from

our daily life acme military usage. Agriculture, environment monitoring, health care,

structure monitoring, intrusion detection, traffic monitoring and industries have made

WSNs increasingly popular [21, 22, 23]. We envision that, in the near future, wireless

sensor networks will be an integral part of our lives, more than the present-day personal

computers.

The common characteristics of wireless sensor applications are: large number of

nodes, long lifetime and fault tolerance.

5

Introduction

1. The use of large number of sensor nodes increases the accuracy of the application

and reduce the cost of prior study of the node placements. It also requires that

networking protocols efficiently reduce the communication overhead and provide

a virtual coordination for data report.

2. Running for a long lifetime is required to minimize the human interventions from

replacing sensor nodes, while ensuring the profitability of such deployed sensor

network services and applications. Long lifetime obviously implies a low energy

consumption of sensor nodes during the operation of the network.

3. As the network ages, it is expected that nodes will fail over time. A large number

of sensor nodes are deployed with redundancy, hence the failure of a fraction of the

sensor nodes should not Hamper the operation of the application. Nevertheless

the network should be capable of reconfiguring its nodes to handle node/link

failure or to redistribute network traffic load.

1.2 Description of WSN in Our Consideration

Now that we have established the sensor nodes’ capabilities, constraints and the set of

applications, we eventually consider a WSN as a network having following characteris-

tics:

1. A WSN is formed by small identical sensor nodes. Each sensor node has the

same limited processing, storage and communication capabilities. A sensor node

is supplied by limited energy resource on board.

2. A WSN requires an easy deployment of sensor nodes in the field. Sensor nodes

must configure themselves. An untrained person should be able to place sensor

nodes throughout the environment and have the system simply work.

3. No predefined communication infrastructure should be given when the network

is set up. After the deployment, the sensor nodes should keep organizing among

themselves to deal with the changes, so as to provide a flexible communication

structure.

4. Long lifetime is critical to many sensor network applications. In some applica-

tions (such as environment monitoring, military and tracking applications), the

goal is to have nodes placed out in the field, unattended, for months or years.

Long lifetime also leads to much more convenient usage of WSNs in home and

6

Motivations

health application. The use of low-power electronic components is one way to

provide energy saving. Moreover, the networking protocol used in WSNs should

be energy efficient by keeping the nodes’ activity to a minimum (communication

and computation).

1.3 Motivations

The popularity of wireless sensor networks also promotes a range of research topics,

particularly around providing energy-optimized communication. A wireless sensor net-

work is deployed to collect and report data while reducing human intervention and

labor. The following scenario is an example of desired implementation and operation

of a WSN.

One may buy a box of sensor nodes from a store. These sensor nodes have been

mass-produced and sold at low price. When the user gets them, they are identical

but may be connected to various sensor boards or controllers. Once at home, the user

places nodes around his house: on the ceiling for monitoring the lights, at every corner

to monitoring the temperature or connected to air-conditioning and light switches for

controlling. Once the nodes are deployed, he simply wants the network work! He

wants to get a comfortable living space without configuring nodes one by one to set up

addresses, data paths and communication structures!

This work aims at providing an autonomous network architecture for WSNs to

minimize the user’s tasks. The architecture makes the WSN run as an autonomous

system where sensor nodes organize themselves without external configurations and

interventions.

This architecture is built and maintained by a set of mechanisms executed by all

sensor nodes. It is the unique architecture upon the deployment of a wireless sensor

network. It takes into account the constraints of sensor nodes as well as the requirement

of sensor applications to set up, organize, manage and maintain the autonomous archi-

tecture in WSNs. This work also addresses how to simplify the multiple development

of networking protocols within the architecture.

The motivations for such architecture are stated as follows along with the operation

of WSNs.

1.3.1 Setting up and organizing

Each sensor node should be configured to set up a network, instead of working indepen-

dently. Such configuration should be achieved by sensor nodes themselves, in order to

7

Introduction

save on human interventions. The corresponding mechanism is called self-configuration.

Self-configuration allows sensor nodes to set up their addresses and parameters for data

communications in WSNs. Although it has been addressed as Autoconfiguration prob-

lem [1, 3, 24] in the field of Mobile Ad hoc Network (MANET) [25], the context of

WSNs is different. Unlike a wireless LAN card containing a unique MAC address, sen-

sor nodes are not necessarily manufactured with unique identifiers, because they are

produced in mass quantities at low cost. Self-configuration scheme is the only way

that a sensor node may get a unique address without any pre-defined identifier, man-

ually configuration or centralized servers. A sensor node can communicate with other

nodes only if it can be identified. Therefore self-configuration is one of the preliminary

schemes for setting up communications in a wireless sensor network.

After the self-configuration, the communication between two nodes which share a

wireless channel can be established. However, in our point of view, multi-hop commu-

nication is still impossible, unless a communication structure is built. Self-organization

is the mechanism which aims at providing such a structure, more precisely a logic struc-

ture on the top of physical network topology. All sensor nodes naturally participate in

the overall structure. They collect information via the structure and they make local

decisions to change it at the same time. Self-organization is an autonomous process, in

which decisions of organization is not guided or managed by centralized elements. The

structure of the self-organization is formed by all local structures around each node. It

is a result of information exchanges and local decisions taken by sensor nodes.

1.3.2 Managing and maintaining

Throughout the life of a wireless sensor network, sensor nodes may fail or be relocated.

It is also possible that new sensor nodes are introduced to cover a larger service re-

gion. The autonomous architecture also has the role of compensating the influences

of these events in the network. A new node should still be configured through the

self-configuration mechanism and be integrated to the communication structure by

self-organization. Besides, the failures of sensor nodes also bring about changes in

the communication structure. Hence, the initial deployment and configuration is only

the first step in the network lifetime. In the long term, the total cost (in terms of

message, computation and energy) may have more to do with the maintenance cost.

Self-configuration and self-organization that are used in the network set up, are con-

tinuously running as self-maintenance mechanisms in order to deal with spontaneous

changes in the network.

As the network ages, the residual energy on board the sensor nodes decrease as well.

8

Motivations

Optimizing the energy consumption during the WSN’s life is another goal of managing

and maintaining the network. In practice, some nodes may spend more energy than

others, because of their particular roles in the network. The network management

should also spread the energy cost over all sensor nodes.

1.3.3 Support of applications and services

One purpose of this autonomous architecture is to provide a general platform for easy

implementation of applications and services. The autonomous architecture makes the

physical changes transparent to the upper applications and services. Several proper-

ties are conserved along with the network operation. One may take advantages of

these properties to improve the effectiveness and efficiency of running applications and

services.

Data dissemination and data aggregation are considered in this work as two examples

of basic services in wireless sensor networks. How to provide efficient data dissemination

and data aggregation are the core problems of many applications such as environment

monitoring and tracking applications. Both of them are supported by the autonomous

architecture that we proposed for WSNs. The additional cost of running data dissemi-

nation and data aggregation is dramatically reduced, and the overall message cost and

energy consumption are lower as well. The results (see in chapter 4) confirm the positive

impacts on the applications and services of using an autonomous network architecture.

1.3.4 Energy efficiency

It is worth noting that such an autonomous architecture should achieve energy efficiency

during its generation, maintenance as well as lifelong operation. As explained in section

1.1.1, the wireless sensor nodes are extremely energy constrained due to their small

size and low cost. Whilst the majority of WSNs applications targets long lifetime, the

utilization of the autonomous architecture should optimize the communications to meet

high energy efficiency. Furthermore, the additional energy consumption generated from

the executions of the autonomous architecture should also be minimized.

Unfortunately, a number of networking mechanisms proposed for WSNs still adopt

energy-hungry techniques. The utilization of HELLO message is the most popular tech-

nique for topology control and organization in WSNs, while it generates a significant

control overhead and is very energy consuming.

The autonomous architecture that we propose in this work as well as all of the asso-

ciated mechanisms aim at providing energy savings during the entire life of WSNs. The

9

Introduction

energy efficiency are particularly considered and analyzed as one of the most important

performance metrics.

To summarize, the motivations of this work is to provide efficient and low-cost mech-

anisms from the deployment to the effective execution of services and application in

WSNs.

1.4 Organization of the Thesis

This thesis is organized in 7 chapters. Chapter 2 presents some prior contributions to

the networking of wireless sensor network. We point out that self-configuration and

self-organization are two key mechanisms for WSNs. The aim of self-configuration is to

let nodes set up addresses and parameters without any manual intervention (minimize

user’s tasks), so that nodes may identify each other in a communication. There exist

two class of self-configuration schemes: Duplicated Address Detection (DAD) based

and Distributed Dynamic Host Configuration Protocol (DDHCP) based. Several works

are reviewed. It is shown through discussions that few of them conform with the energy

constraints of WSN. Self-organization aims at structuring the network (i.e. provide a

communication structure) through merely local information and decisions. Clustering,

virtual backbone based and source dependent self-organization schemes are discussed.

Once again the energy efficiency is the most critical problem to these schemes. By

analyzing and identifying the shortcomings in the existing self-configuration and self-

organization solutions, we point out the necessity of our autonomous architecture.

We propose, in chapter 3, an autonomous architecture which combines self-configuration

and self-organization, named FISCO (Fully Integrated Scheme of self-Configuration and

self-Organization). This autonomous architecture is motivated by improving efficiency

of the network formation and organization. Our point of view is that self-configuration

and self-organization are two relevant schemes for wireless sensor networks. A unique

structure is generated and maintained with the autonomous architecture for both

conflict-free address allocation and data communication. The autonomous architecture

is built with event-driven procedures, while the use of periodical actions are optimized

in order to avoid control overhead and provide energy saving during its operation. Our

architecture integrating two mechanisms exhibits significant improvements comparing

to existing solutions, not only in the energy efficiency but also in other properties such

as the cardinality of backbone, etc. The performance results also indicate the reduction

on total energy consumption and the extension on network lifespan.

In chapter 4, we discuss the usage of FISCO architecture, by exploring data dis-

10

Organization of the Thesis

semination and data aggregation. One of the goals of running FISCO structure is to

provide further improvement in intensive data communication, such as easing path set

up between source nodes and sink nodes or providing energy saving. We show in this

chapter that the use of FISCO architecture may provide an flexible and low-cost data

dissemination structure. It also supports data aggregation techniques to provide sig-

nificant energy saving during data reporting. Through a set of simulations, the impact

of FISCO are clearly highlighted, by comparing the cost and the result of running

data dissemination and data aggregation schemes with and without the autonomous

architecture. We also propose a localized temporal data aggregation technique based

on Adaptive AutoRegression Moving Average (A-ARMA) technique. It is shown that

using an adaptive technique such as A-ARMA is better than non-adaptive methods,

because it achieves low complexity, high accuracy and good efficiency at the same time.

Chapter 5 addresses a fundamental question in wireless sensor networks, more gener-

ally wireless multi-hop network: how to define a good organization? Although metrics

such as complexity or self-stability are commonly used for evaluation, to the best of

our knowledge, none of them quantifies the efficiency of building and maintaining an

organization (order) under connection changes. In this chapter, the notion of entropy

is adopted as a metric for evaluating the organization of a network where different

self-organization schemes are used. This approach provides several quantitative and

qualitative insights into the behavior and design of self-organization protocols. For

example, any of the chosen protocols yields a higher organizational state than a flat

topology.

Chapter 6 presents our work on building a test-bed based on 30 Imote2 sensor nodes in

FranceTelecom R&D center in Beijing. We designed and implemented a protocol stack

based on an embedded Linux OS for wireless sensor nodes. A partial implementation

of FISCO scheme is available on this test-bed. Through a set of functional tests, the

correctness of the FISCO implementation is validated. The Imote2 test-bed shows that

it is possible, with commercial sensor nodes, to validate the autonomous architecture

proposed in this thesis. It is also a step from the design to the practice. This test-bed

will be used in FranceTelecom R&D center as an extensible hardware platform with

stable development environment for further demonstrations of WSN.

Chapter 7 summarizes the thesis and gives a prediction of future technological trends.

The major contribution of this thesis is that we designed a complete solution for net-

working in WSNs which covers deployment, configuration, organization and data com-

munications. This work also leads to several perspectives which are not limited to the

functionalities of the architecture, but also extended to security and applications. This

11

Introduction

thesis was financed and supported by France Telecom R&D under CRE No 46130157.

It is realized at CITI Laboratory of INSA Lyon, in the ARES Team of INRIA Rhône-

Alpes, under the direction of Doctor Fabrice Valois and Professor Eric Fleury.

12

State of the Art 13

State of the Art 22.1 Introduction

As addressed in the introduction (section 1.3), the quality of service provided by a

wireless sensor network relies on an autonomous network architecture. It is not a single

scheme, but the combination of mechanisms which support the operation of a WSN

from its deployment onward, including self-configuration, self-organization, self-healing

and self-management.

Self-configuration [26, 27] is the first action to take when sensor nodes are deployed.

The aim of self-configuration is to let nodes set up addresses and parameters without

any manual intervention. Address allocation is known as the core problem of self-

configuration. The main issue of address allocation in multi-hop wireless network with-

out a pre-defined infrastructure (including both ad hoc networks and wireless sensor

networks) is how to ensure the uniqueness of allocated addresses in the entire network.

Solutions using centralized address servers are not optimized to work for large scale

sensor networks. The use of server-client approach even becomes a bottle-neck of con-

figuration performance. Indeed, the self-configuration should achieve a distributed and

dynamic address allocation in WSNs. It should deal with random deployment of spo-

radic node arrival by an immediate join procedure. Furthermore, it should handle the

problem related to address conflicts during partition splitting and partition merging in

the network.

Self-organization [28, 29] in our point of view is defined as a process aiming at struc-

turing the network through purely local information and decisions. The logic structure

of the network is represented by relations between neighboring nodes (as logic links)

and node roles in the network. Using self-organization instead of a pre-defined structure

makes the communication structure scalable, flexible and adaptive. According to local

information and localized algorithm, nodes form local structures. The self-organization

structure is built based on these local structures. It reflects the emerging behavior in

State of the Art

the network from pure local decisions.

Self-healing [30, 31] aims at adapting the network to keep certain properties that

have been established. It is highly related to the notion of stability, because it runs for

maintaining the structure built by self-organization facing spontaneous changes such

as node failures, relocations, etc. Self-healing also adjusts the structure of organization

partially or entirely on-the-fly to optimize the real time performance such as energy

consumption. Hence, we consider self-healing as a part of self-organization.

Self-management [32, 33] deals with the control of the services based on the au-

tonomous architecture. It benefits from the autonomous architecture to improve the

quality of service in the network. Different from other mechanisms, it also involves

application configurations and parameters. As we target an autonomous network ar-

chitecture, the self-management is beyond the scope of this work. Nevertheless it is

discussed as one of our perspectives in section 7.2.

The objective of this chapter is to review the existing works before proposing an

autonomous network architecture, while self-configuration and self-organization are two

indispensable mechanisms for this end. Therefore this chapter is organized as follows:

Section 2.2 reviews prior contributions to self-configuration. The majority of these

works is in the context of ad hoc networks, where self-configuration aims at providing

dynamic address allocation solutions in a network. Section 2.3 reviews the major works

in self-organization in ad hoc and wireless sensor networks. Clustering based, virtual

backbone based and source dependent self-organizations are considered. Based on the

resulting structure, virtual backbone based self-organizations are further classified into

Maximal Independent Set (MIS), Connected Dominating Set (CDS), Local Minimum

Spanning Tree (LMST), Relative Neighboring Graph (RNG). It is after identifying the

concepts and weaknesses in these solutions that we propose an autonomous architecture

in the next chapter.

2.2 Self-configuration

It is worth noting that some WSN applications do not require nodes to be identified by

logic addresses such as IP addresses. In such applications, coordinates [34] or contents

[35] are used for data reporting. However the routes can not be recorded or reused if

nodes can not be identified. Therefore assigning addresses to sensor nodes facilitates the

data communication in WSN. Moreover, for most of WSN applications, the addressing

is indispensable.

Although the majority of self-configuration schemes reviewed in this section were

14

Self-configuration

Client

DHCP_ACK

DHCP_REQUEST

DHCP_OFFER

DHCP_DISCOVER

Server

Figure 2.1: DHCP server-client configuration

proposed for ad hoc networks, it is worth noting that self-configuration is not a new

notion for IP networks, especially known as autoconfiguration in IPv6 networks. Before

embarking on to the self-configuration schemes in ad hoc network, let us first take a

glance at the mechanisms used in IP networks in order to demonstrate that they can

not be directly applied in our context.

2.2.1 Autoconfiguration in IP networks

Already available for IPv4, Dynamic Host Configuration Protocol (DHCP) [36] is the

most used autoconfiguration protocol for address allocation in wired networks and AP-

based wireless networks. It is built on a server-client model, where designated DHCP

servers allocate IP addresses and deliver configuration parameters to hosts. Client-

server exchanges in DHCP involve four messages (as shown in Fig. 2.1). DHCP is also

known as a stateful autoconfiguration protocol because all the address allocation states

are stored (in DHCP servers) during the configuration.

IPv6 [37], Internet Protocol version 6, is designed as an evolutionary step from IPv4

with some main changes such as: expanded addressing capability, improved support

for extensions and options, extensions for authentication and privacy, etc. As one

of the improvements, IPv6 integrates a stateless autoconfiguration [38] for traditional

hierarchical networks to deal with those networks without DHCP servers.

In a IPv6 stateless autoconfiguration, the following steps are performed by a host

once it joins the network on a physical link (see Fig. 2.2):

1. The host generates locally a link-local address that is based on the interface

identifier (IEEE 64-bits Extended Universal Identifier) and a pre-defined link-

local prefix (FF02::1/64). The state of this address is set to Provisioning.

2. A Duplicated Address Detection (DAD) procedure is launched by the host to

have the uniqueness of its provisioning address verified. A Neighbor Solicitation

15

State of the Art

Provising link−local @

Yes RA

No RARA

RS

No NA

NA Yes NA

NS

Working with link−local @Obtain prefix

Link−local @

Figure 2.2: IPv6 stateless address autoconfiguration

(NS) message is sent out with the provisioning address as the target address. The

source address is the all-zero address; the destination IP address is the solicited-

node multicast address. If the address is already in use by another host in the

network, then it replies with a Neighbor Advertisement (NA) message. An address

conflict is henceforth recognized. Either a manual configuration or a regeneration

of link-local address is required. If there is no answer to the NS within the timeout,

the address is assigned to the interface and the state of the address changes to

Preferred.

3. The host sends a Router Solicitation (RS) in order to determine the global prefix.

The RS message is sent to the all-routers multicast group of FF02::2.

4. If there is a router on the link, then it replies with a Router Advertisement (RA).

Because the uniqueness of EUI-64 has already been verified in DAD with link-

local interface, there is no need to repeat it for the global address. The host

simply combines the prefix with the interface identifier. If no RA is received in

timeout, the host keeps communicating with its link-local address.

Although DHCP and IPv6 stateless autoconfiguration are widely used as the standard

address configuration techniques in the networks which have pre-defined infrastructures,

neither of them is applicable in ad hoc and wireless sensor networks. The major problem

comes from the scalability. In ad hoc or wireless sensor networks, nodes do not reside

on the same physical links. The discoveries of DHCP servers as well as duplicated

addresses are not time bounded. They are achieved through a flooding because there

is no pre-defined structure in the network. It is not only message costly but also time

16

Self-configuration

consuming. To reduce the configuration cost and duration, self-configuration has been

investigated as a particularly important problem in ad hoc and wireless sensor networks

where two approaches have been proposed.

2.2.2 Address conflict detection based solutions

The first approach relies on DAD [24] mechanism. If the network diameter is known in

advance, then the duplicate address can be detected within a timeout. This is known

as Strong DAD. However when the network diameter is unknown, which is actually the

case in ad hoc and wireless sensor networks, the DAD can not be achieved in a given

timeout. Hence Weak [39] and Passive DAD [40] mechanisms are proposed to resolve

this issue of Strong DAD. Contrary to Strong DAD, they only prevent a packet from

being routed to a wrong destination.

Weak DAD [39] tolerates duplicate addresses in the network after the local generation

of addresses as long as all packets are delivered correctly. A unique per-node key is

assumed to be included in the routing control packets and the routing table entries,

but is not embedded in IP address. If two nodes happen to have chosen the same

IP address, then they can still be identified by their keys. It is worth noting that a

new implementation of Weak DAD is required every time a different MANET routing

protocol is used (eg. its implementation over link state routing is different from that

over dynamic source routing). This increases the development cost of Weak DAD.

Passive DAD [40] is similar to Weak DAD. Address conflicts are detected passively

by nodes using continuous monitoring on routing control traffic. It was initially built

for link state routing protocols, and then extended to work with reactive routing pro-

tocols as well. A node analyzes incoming protocol packets, including date packets and

control packets, to derive hints about address conflicts. The basic idea is to apply

various Passive DAD algorithms to exploit protocol events that 1) never occur in case

of a unique address, but always in case of duplicate addresses or; 2) rarely occur in

case of a unique address, but often in case of duplicate addresses. In the first case, de-

tection is certain while in the second case a conflict is probably occurring. In order to

have a certain confidence on the conflict detection, long-term monitoring is necessary.

Although this approach does not incur extra overhead in the network, it heavily relies

on the underlying routing protocol. Furthermore, the correctness and effectiveness are

strongly affected by the particular parameter settings of the routing protocol.

17

State of the Art

A

DB

B

80 − 12764 − 79

0 − 127

64 − 1270 − 63A

Address

Figure 2.3: Segmented address pools in Buddy [1]

2.2.3 Distributed DHCP solutions

The second approach of self-configuration is to build a distributed version of DHCP

in ad hoc and wireless sensor networks. The idea is to assign the DHCP functionality

to every node. Each node has capability to allocate addresses to other nodes on the

same wireless link. The allocation is hence limited to one-hop. In order to avoid address

duplication in the entire network, nodes should exchange address allocation information

and maintain this knowledge synchronized among them.

[1] proposes to use Buddy system (a well known method for memory management) in

address allocation. When a node join the network, it asks one of its one-hop neighbor

nodes for an address allocation after a discovery procedure. The latter node acting as

a DHCP server, named Buddy node in the proposal, responses back by giving the half

of its current address pool. The new node now gets its own DHCP address pool and

assigns itself the first address of its address pool. The address space is hence divided

into a binary tree with the arrival of nodes in the network (see Fig. 2.3). Nodes should

synchronize from time to time to keep the records of IP address assignment in the entire

network and detect any IP address leak for recover.

However the binary division of address space has several inconveniences. First, it

causes a non-uniform distribution of addresses. A massive arrival of nodes may cause

some region of the network to run out of addresses, while many addresses are not used

in another region. Secondly, the binary address tree (Fig. 2.3) reaches very quickly

its leaf level (no address is available on leaf nodes) within an exponential speed 2n.

When a node goes out of addresses, it has to go through its buddy nodes (parent nodes

in the binary address tree) until one buddy node replies to the request with available

addresses. An unpredicted amount of message overhead is generated in this address

pool searching. Furthermore, if several partitions are generated during the network

deployment, then each partition has its own binary address tree. It is difficult to merge

18

Self-configuration

nor Allocate Pending tables

Neighbor discovery

Informing all MANET

Selecting an @ which

Select node A as Initiator

A B

MANET node New arrival node

Neighbor_Reply

Neighbor_Query

Requester_Request

@ + Allocated Table +Allocate Pending Table

nodes of this configuration

is neither in Allocated

Figure 2.4: MANETConf [2] configuration process

binary address trees when these partitions meet, because the address pools used in one

partition should be completely transformed to a branch of the other. It increases not

only the message overhead but also the complexity of computation. Hence it does not

meet the requirements of wireless sensor networks.

MANETconf [2] is another distributed DHCP solution for ad hoc networks. In order

to maintain the address information, every node holds an Allocated and an Allocate

Pending tables. The first one contains the set of all IP addresses in use and the second

one notes all addresses used in the procedure of allocation. When a node joins the

network, it sends a one-hop Neighbor Query message to find a configured node (see

Fig. 2.4). The first replying neighbor is considered the DHCP server for the new node.

It allocates an address which is neither in Allocated or Allocate Pending tables. This

node also informs all other nodes of this allocation via flooding. At the same time, the

Allocated and Allocate Pending tables are duplicated in the new node.

In order to deal with partition splitting, merge, initiator crash and concurrent alloca-

tion, additional mechanisms are proposed. However the flooding used for each address

allocation limits the scalability of the solution and generates a significant message over-

head.

Prophet Addressing [3] uses a function f(n) to assign no-conflict addresses to nodes.

The scheme is simple and entails low overhead as long as f(n) has good properties.

Address allocation is considered as an assignment of different numbers from an integer

range to different nodes. The first node A in the network chooses a random number

as its address and uses a random state value as the seed for its f(n) (Fig. 2.5). The

second node, say node B, joins the network as a neighbor of node A and asks node

A for an address allocation. Node A uses f(n) to compute a new address from its

address and its state value. A new state value is also generated for the next round of

19

State of the Art

@_1, state_1

A

A

A

D

B

BC

f(@_2, state_2) = (@4, state_4)f(@_1, state_2) = (@3, state_3)

f(@_1, state_1) = (@2, state_2)

@_4, state_4@_2, state_4@_3, state_3@_1, state_3

@_2, state_2@_1, state_2

Figure 2.5: Prophet [3] address allocation

address generation. The new address and the new state value are sent back to B, from

which the node is able to generate addresses without conflict. The f(n) also creates an

tree-like address hierarchy in the network (Fig. 2.5).

However, the tree-like address hierarchy causes problems in partition merge. First,

nodes need send periodical HELLO message to detect address conflicts in two partitions.

Upon the address conflict detection, all nodes in one partition have to give up their

address and re-configure themselves using the state value of the other partition. This

mechanism generates significant message overhead and increases computation cost dur-

ing address re-configuration.

2.2.4 Synthesis on self-configuration

It is much more difficult to achieve self-configuration in a wireless multi-hop network,

especially in a wireless sensor network, than in traditional IP networks. Nevertheless

it is a key mechanism to WSN deployment for the following reasons:

1. Different from the production way of wired/wireless LAN card, wireless sensor

nodes are required to be manufactured in large quantities at low cost. In order to

reduce the production cost, wireless sensor nodes are note expected to have any

unique hardware addresses. Therefore it is necessary to develop address allocation

mechanisms for WSNs.

2. Because WSNs are multi-hop networks usually deployed in large scale, it is infea-

sible to implement centralized servers. Self-configuration provides dynamic and

distributed address allocations in WSNs.

3. Remember wireless sensor nodes have very limited energy resources. Hence, low

energy consumption and low message overhead are critical for self-configuration

in WSNs.

Although the existing approaches tried to adapt the techniques used in traditional

IP networks to ad hoc network and wireless sensor networks, they do not meet all the

20

Self-organization

requirements, especially because of their significant control message overhead.

1. The technique of Strong DAD is applicable in an ad hoc or wireless sensor network

only if its network diameter is known in advance. Otherwise, a node is not sure

to get a duplicate detection response within a timeout. Furthermore, the DAD

messages injected during the flooding in the network is a harmful overhead for

network deployment.

2. Weak/Passive DAD is an alternative to Strong DAD, in which address conflicts are

detected by examining incoming and outgoing routing messages. This approach

does not generate additional control message only because routing messages are

used to this end. The efficiency of this technique also highly relies on the under-

lying routing protocol, and its implementations vary from one routing protocol

to another. This causes a significant development cost.

3. In distributed DHCP approach, the idea is to make every node a DHCP server in

the network. A node asks one of its neighbor nodes for an address allocation. The

core problem in this approach is how to maintain the address information updated

on all nodes in the network to keep a coherent view for next address allocation

without conflict. To this end, either periodical one-hop HELLO messages or flooding

are used to synchronize addresses in the network. However, this synchronization

on all nodes leads to significant message overhead.

2.3 Self-organization

It is worth distinguishing self-organization from the notion of self-organized. Self-

organized is a properties related to mechanisms. For instance, unicast routing in ad

hoc and wireless sensor networks should be self-organized. However, the notion of self-

organization in our consideration is a basic concept to create order and to provide a logic

structure in the network. Certainly, the mechanisms used in self-organization should

be self-organized. Nevertheless, comparing to individual self-organized protocols, self-

organization has a more general purpose: organizing and structuring the network. It is

also around this idea that we develop our autonomous network architecture. Hence, the

prior works revised in this section all aim at forming a structure in the network. They

are classified into three categories: clustering based, virtual backbone based and source

dependent structure. The virtual backbone based solutions can be further divided into

CDS, MIS, LMST and RNG according to the properties of the resulting structures.

21

State of the Art

Figure 2.6: A Unit Disk Graph of 120 nodes and radius=0.16

The algorithms that we reviewed in this section have also been used to deal with

topology control problems in wireless ad hoc and sensor networks. Topology control

[41] aims at controlling the topology of the graph representing the communication links

between nodes, while reducing energy consumption and/or interference that are strictly

related to the nodes’ transmitting range. Both clustering and virtual backbone based

algorithm can be applied to this end. However self-organization that we focus on in

this section has a more general goal. It not only deals with individual transmission

assignment, but also takes care of an emerging behavior in the entire network.

Before embarking onto different algorithm, we shall subsequently define the formalism

of network model based on which the properties of structures are formulated.

2.3.1 Network model formalism

A wireless network is generally modeled as a graph, where nodes are represented as

vertexes and radio links as edges. If all links are symmetric, then the graph is a non-

oriented graph. If all nodes use the same radio transceiver with identical transmission

powers, then the radio vicinity of a node is a disk of ray R (normalized between 0 tp

1). In this case, the network is modeled as a Unit Disk Graph (UDG). Fig. 2.6 gives

an example of UDG with 120 nodes and radius at 0.16.

We introduce here the notations we used in this thesis:

• G = (V,E): the graph representing a wireless network, where V is the vertex setand E ⊆ V 2 the edge set.

• (u, v): the edge exists between two vertex u and v if and only if v can receive

22

Self-organization

correctly the message sent by node u.

• |X|: the cardinality of a vertex set X.

• Nk(u): the neighbor set of vertex u within k-hop distance to u. By simplification,N(u) = N1(u), where N1(u) = {v|(u, v) ∈ E}).

• ∆k(u): the number of k-hop neighbors (∆k(u) = |Nk(u)|). ∆1(u) is also definedas the degree of node u.

• dist(u, v): the hop distance between u and v.

It is possible to introduce the arrival and departure of nodes as well as node mobility

in this graph model. The arrivals and departures can be modeled as random processes.

For example, we can consider the arrival as a Poisson process, while the lifetime of a

host in the network is exponentially distributed. There are also some mobility models

in the literature [42].

Although a real radio environment does not correspond exactly to the UDG model,

it shows a number of properties which simplify the analysis of geometric structure in

the network. The use of UDG model gives an idea on the bound properties of target

structures. Certainly, it is not considered as the tool which gives detailed performance

results for a solution.

2.3.2 Clustering

Cluster is a basic structure used in network organization. Each cluster is a set of nodes

which is grouped within a geographical area. A clusterhead is in charge of a set of

specific functionalities within its cluster. Organizing a wireless network into clusters

necessitates a distributed clusterhead election algorithm.

[43] is one of the fundamental works in clustering. The election of clusterhead is based

on the lowest identifier: node u is elected as a clusterhead if its identifier is the lowest

among its neighbors N(u). To this end, nodes need communicate their identifiers to

their neighbors at the beginning of the election. Non-clusterhead nodes join the cluster

whose clusterhead has the lowest identifier and become cluster members. Some of them

become gateways, if they have neighbors belonging to other clusters. The network is

partitioned into clusters where each cluster has one clusterhead and several gateways

and members. In such structure, there is no adjacent clusterheads and the clusterheads

are spaced at most by 2-hop. This solution needs synchronization for launching an

election phase. And it takes a number of iterations to finalize the election. We also

note that high message overhead is generated.

23

State of the Art

3−hop coverage

w

vu Member

Gateway

Clusterhead

2.5−hop coverage

Figure 2.7: 2.5-hop and 3-hop coverage set

[44] also uses lowest identifier based clusterhead election, but extends with a localized

backbone construction to connect clusters in the network. To this end, 2.5-hop cov-

erage set of a clusterhead u is defined as a set which includes all the clusterheads in its

2-hop neighbor set N2(u) and the clusterheads having members in N2(u). It is different

from u’s 3-hop coverage set, which includes all the clusterheads in N3(u). In Fig. 2.7,

clusterhead v and w are in 3-hop coverage set of u. However, w is not in 2.5-hop cover-

age set of u, because none of w’s neighbor is in N2(u). Clusterheads collect 1-hop and

2-hop information from their neighbors to construct a 2.5-hop coverage set once they

are selected. By using a greedy algorithm, each clusterhead selects a set of gateways to

interconnect clusterheads in its 2.5-hop coverage set. A node should broadcast several

rounds of messages before it gets a stable role in the network (clusterhead, gateway or

member). This solution does not overcome the high message overhead neither.

[45] proposes to use a metric composed of residual energy and node’s degree to adjust

the backoff time before broadcasting a clusterhead decision. In this way, nodes with

more residual energy have a bigger chance to be elected as clusterheads. Low power

transmissions are used within each cluster, while clusterheads use higher transmission

power to communicate among themselves. However, the energy consumption for the

clusterhead election remains high because the maximum transmission power is used for

sending election messages. Furthermore, the election should be renewed periodically in

order to take into account the energy dissipations on nodes.

2.3.3 Virtual backbone

Here, the algorithms aim at generating a virtual backbone which covers all nodes in the

networks. Some virtual backbones represent a dominating set of the network (in case of

24

Self-organization

CDS and MIS), while the others contain all nodes but with a subset of communication

links (in case of LMST and RNG). Nevertheless, the goal stays the same: simplifying

the physical structure and providing a logic structure based on localized decisions.

2.3.3.1 Connected dominating set

A Connected Dominating Set (CDS) S is defined as a subset of the vertex set V , which

satisfies two characteristics:

• Every node of V is either in S or a neighbor of a node in S.

• S is connected.

The formulation of CDS is:

∀u ∈ V, ∃v ∈ S|v ∈ N(u) (2.1)∀(u, v) ∈ S2, ∃p = Pathu→v|∀w ∈ p,w ∈ S (2.2)

By extending the formula (2.1) to k-hop neighborhood, we obtain the definition of

k-CDS:

∀u ∈ V, ∃v ∈ Sk|v ∈ Nku (2.3)

CDS based self-organizations aim at computing a CDS with small cardinality and

low computation complexity using distributed algorithms. Computation of a Mini-

mum CDS (MCDS) is known as a NP-hard problem. Hence, many distributed CDS

construction algorithms are designed to compute an approximate MCDS.

One localized algorithm is proposed in [4] by Wu and Li. They assume that each

node in the network owns a unique identifier. Nodes send periodic HELLO messages

to exchange neighbor lists with their one-hop neighbors. Each node hence has the

knowledge of 2-hop neighborhood. A node marks itself as dominating node if any pair

of its neighbors is not directly connected. After this marking procedure, all dominating

nodes form a CDS, but the cardinality of this CDS is very big. In order to reduce the

size of CDS, two self-pruning rules are adopted.

1. If node u is a dominating node in current CDS and there exists one neighbor of

u, say v, whose one-hop neighbor set N(v) includes the N(u), and the ID(u) <

ID(v), then node u becomes a dominated node.

25

State of the Art

1

8

5

4

3

2

6

7

9

(a) Marking process

1

8

5

4

3

2

6

7

9

(b) Rule 2 [4]

5

Dominated node

Dominating node

6

2

1

7

9

8 4

3

(c) Rule k [5]

Figure 2.8: Example of CDS construction in [4] and [5]

2. If node u is a dominating node in current CDS and there exist two neighbors of u,

say v and w, whose one-hop neighbor set union N(v)∪N(w) includes N(u), andID(u) = min(ID(u), ID(v), ID(w), then node u becomes a dominated node.

The self-pruning rules reduce significantly the cardinality of CDS. In [5], an enhanced

rule, Rule k, is proposed. If the neighbor set of a dominating node is covered by a set

of k neighbor nodes, and it has the lowest identifier, then it becomes a dominated node.

Rule k is more efficient in reducing dominating set size than the combination of Rules 1

and 2, and has the same communication complexity and less computation complexity.

Fig. 2.8(a)-2.8(c) illustrates the CDS construction algorithm of Wu and Li. In Fig.

2.8(a), nodes {3, 4, 5, 6, 8, 9} are marked as dominating nodes because each of them hasat least a pair of non-connected neighbors. After applying Rule 2 in [4], node 3 finds

its neighbor set is completely covered by two nodes with bigger identifiers: 5 and 6.

It changes itself to a dominated node. When Rule k [5] is applied, node 4 decides to

become a dominated node because its neighbor set is fully covered by nodes {5, 6, 9}.There are other methods of CDS construction such as [46] and [47]. However, interme-

diate structures are formed (MIS and MPR respectively) during the CDS construction.

That is why these solutions are classified in section 2.3.3.2 and 2.3.4.

CDS is used as logic backbone in a network. Although Wu&Li’s CDS algorithm is

completely localized, every node still need two-hop information. To this end, the use

of HELLO message is indispensable. Hence, the control message cost rises significantly.

2.3.3.2 Maximal independent set

An Independent Set (IS) is a subset of vertexes which do not contain adjacent nodes

in a graph. A formal definition of IS is the following:

IS = {u ∈ IS|(not∃v ∈ IS|u ∈ N(v))}

26

Self-organization

The Maximal Independent Set (MIS) is the Independent Set containing maximum num-

ber of vertexes. The particular interest in MIS problem stems on practical importance

of distributed computation in ad hoc and wireless sensor networks. A MIS defines a set

of nodes which can operate in scheduling or parallel processing without interference.

Furthermore, it is possible to obtain approximate Minimum CDS structure based on a

MIS structure.

[46] proposes to construct a MIS based on a rooted tree. It is a distributed algorithm

but not a localized algorithm, because one node should initiate the construction. A

spanning tree is first generated in the network by using a flooding. During the tree

construction, each node is also assigned with a rank. They move to a color-marking

phase to determinate the MIS nodes. All nodes in the network are initially marked

WHITE. The root node marks itself BLACK and broadcasts a BLACK message. A

WHITE node which receives a BLACK message becomes GRAY node and sends a

GRAY message. A WHITE node which receives GRAY messages from all its neighbor

closer to root node will marks itself BLACK and broadcasts a BLACK message. By

propagating BLACK and GRAY messages in the network, all nodes are marked either

BLACK or GRAY. The set of BLACK nodes is the MIS. The algorithm can be extended

to construct a CDS by adding a phase to connect all BLACK nodes.

Although the final CDS has a lower cardinality, the message overhead and time

complexity of the algorithm are higher than those in [4]. Hence, it does not meet the

energy constraint of WSNs. The resulting structure depends on the root node, which

does not leave the solution enough flexibility and scalability.

2.3.3.3 Relative neighborhood graph

Relative Neighborhood Graph (RNG) is a geometric concept proposed by Toussaint

[48] in 1980. The idea is to prune the longest edge in each triangle in the graph. It is

formulated as:

RNG = {(u, v) ∈ RNG|(not∃w ∈ N(u) ∩ N(v)|d(u, v) > d(u,w)d&d(u, v) > d(v,w))}

where d(u, v) is a weight assigned to edge (u, v). The length of an edge is often used as

its weight. As shown in Fig. 2.9, the edge (uv) is not in the RNG subgraph, because

it is the longest edge in triangle (u, v,w).

RNG subgraph is not a loop-free graph, although each local structure around a node

is a tree structure. It is shown that the

Impacts of Self-organized Mechanisms in Wireless Sensor ...theses.insa-lyon.fr/publication/2008ISAL0026/these.pdf · No: 2008-ISAL-0026 Year 2008 Thesis Impacts of Self-organized

Documents