impacts of self-organized mechanisms in wireless sensor...
TRANSCRIPT
No: 2008-ISAL-0026 Year 2008
Thesis
Impacts of Self-organized Mechanisms in
Wireless Sensor Networks
defend at
l’Institut National des Sciences Appliquees 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 these a ete effetuee au sein du Centre d’Innovation en Telecommunications et Inte-
gration de Services (CITI), a l’INSA de Lyon, et dans le projet ARES de l’INRIA Rhone-
Alpes, en collaboration avec le Centre de Recherche et Developpement de FranceTele-
com Grenoble et Beijing.
Je tiens tout d’abort a remercier Fabrice Valois sans, qui cette these n’aurait, je
pense, jamais eu lieu. Je le remercie de m’avoir encadre non seulement durant la these
mais egalement pendant mon mastere de recherche, en me donnant les outils et surtout
l’esprit de recherche pour reussir dans ce domaine. Je le remercie de m’avoir me toujour
soutenu mais aussi cretique pour reussir cette these. Je remercie egalement Eric Fleury
qui m’a co-encadre durant cette these et a su partager son expertise scientifique.
Je remercie tout particulierement Monsieur Khaldoun Al Agha, Professeur a l’Universite
de Paris-Sud, et Monsieur David Simplot-Ryl, Professeur a l’Universite de Lille qui ont
accepte de juger mes traveaux de these et d’en etre les rapporteurs.
Je remercie egalement Monsieur Dominique Barthel, Chercheur Senior a FranceTele-
com R&D de Grenoble, Monsieur Thomas Noel, Professeur a l’Universite de Strasbourg
et Monsieur Stephane Ubeda, Professur a l’INSA de Lyon d’avoir accepte de faire partie
de ce jury de these et de s’etre interesses a mon travail.
Je remercie egalement Monsieur Mischa Dohler avec qui j’ai passe ces moments de
concentration riche, tant pour son apport scientifique que personnel.
Je tiens a remercier Yvan Royon et Noha Ibrahim pour les discussions scientifiques,
sociales ou personnelles que nous avons eu. Je remercie egalement Fabrice Theo-
leyre, Nathalie Mitton, Tahiry Razafindralambo, Karel Heurtefeux, Elyes Ben Hamida,
Thomas Watteyne, Katia Jaffres-Runser avec qui les nombreuse exchanges scientifiques
ont eu lieu. Je remercie Haila Wang, Song Wang et Yu Zhang qui m’ont aide a develop-
per les activite de recherche a FranceTelecom Beijing. Je remercie egalement tous les
chercheurs du laboratoire CITI (pele-mele Jean-Marie Gorce, Guillaume Chelius, Is-
abelle Auge-Blum, et tant d’autres que je ne nommerai pas par manque de place).
Enfin, je tiens a remercier Yi, mon epouse, qui a su m’accompagner (physiquement et
moralement) dans cette grande experience scientifique mais surtout personnelle qu’est
une these.
Resume
Un reseau de capteurs est un reseau radio multi-sauts forme par une quantite impor-
tante de capteurs identiques. Contrairement aux noeuds d’un reseau ad hoc, les cap-
teurs ont des nombreuses contraintes, notamment en termes de consommation d’energie,
de capacite de calcul et de stockage. L’absence d’infrastructure et de controle centralise
implique une collaboration efficace et pertinente des noeuds du reseau de capteurs.
Ces travaux de these s’inscrivent dans la problematique d’auto-organisation des
reseaux de capteurs. De notre point de vue, l’auto-organisation est un probleme fonda-
mental des reseaux de capteurs conduisant a construire une vue logique de la topolo-
gie du reseau physique et de fournir des protocoles de communication bases sur cette
vue logique. Nous avons donc propose une architecture d’auto-organisation (nommee
FISCO) permettant d’organiser le reseau - via des interactions locales uniquement-
sous le forme soit d’un arbre soit d’un treillis. Les proprietes structurelles et les perfor-
mances de FISCO ont ete evaluees puis comparees aux principaux mecanismes d’auto-
organisation de la litterature. Cette architecture exhibe de tres bonnes performances
en termes de stabilite, persistance, robustesse et surtout gestion de l’energie.
Ensuite, nous avons revisite les principaux defis de la communication des reseaux
de capteurs en se basant sur l’auto-organisation prealablement introduite. Nous nous
sommes donc interesse aux problemes clefs suivant : allocation dynamique d’adresses,
protocole d’inondation, dissemination et agregation de donnees dans les reseaux de
capteurs. L’impact de l’auto-organisation est clairement constate en comparant avec
les approches classiques. Notons que la structure d’auto-organisation offre egalement
une solution tres efficace et tres flexible pour la gestion de puits mobiles et multiples
dans un reseau de capteurs. Le travail realise dans cette partie repose sur l’utilisation
d’outils issus de la theorie des graphes, de l’algorithmique distribuee mais aussi de
l’evaluation de performances.
En parallele, nous avons cherche a justifier l’utilisation d’approche auto-organisee
(i.e. structuree) par rapport aux approches a plat. Pour ce faire, nous avons introduit
une nouvelle metrique d’evaluation herite de la notion d’entropie statistique. Cette
metrique est la premiere metrique quantitative qui mesure l’ordre de l’organisation
dans un reseau sans fil multi-sauts. Cette metrique ouvre de nombreuses pistes de
recherches dans le domaine de l’auto-organisation des reseaux.
Nous avons aussi propose une plate-forme d’experimentation composant des 30 cap-
teurs. Cette plate-forme experimentale est utilisee pour valider les solutions proposees.
Ainsi, les contributions des travaux de la these couvrent a la fois le domaine theorique,
mais aussi la simulation et l’experimentation.
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 Rhone-
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 set
and 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 defined
as 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), and
ID(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 has
at 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 degree of RNG is stable no matter the degree of
the initial graph. In [49], each node collects two-hop distance information to compute
local RNG structure. The advantage of using a RNG based self-organization is that
27
State of the Art
u v
w
Figure 2.9: RNG: pruning the longest edge
each node only needs to communicate with a subset of its neighbors instead of the entire
set. Hence the computation and communication complexity are reduced when running
other schemes on this organization such as coverage and broadcast schemes. It is worth
noting that the assignment of edges’ weights requires nodes to collect more information
from the neighborhood, such as distance to neighbors or Radio Signal Strength (RSS).
Hence, the resulting RNG graph depends on the precisions of these measurements.
2.3.3.4 Local minimum spanning tree
Minimum Spanning Tree (MST) algorithms have been widely used for network broad-
casting and routing. However, the computation of MST necessitates complete topology
knowledge of the network, which is infeasible in ad hoc and wireless sensor networks.
[49] proposes a local MST (LMST) version, in which each node computes locally a MST
on its two-hop neighborhood. To this end, greedy algorithm can be used and nodes
should exchange HELLO messages containing one’s one-hop neighbor list. If two nodes
are LMST neighbors of each other, then the edge between them is added in LMST sub-
graph of the network. It is worth noting that the resulting LMST is not a tree structure.
There are loops in the LMST subgraph because of the localized computation and use
of local information. LMST is also efficient to reduce the redundant connections in the
network and keeps a sparser structure than RNG does. It is demonstrated in [49] that
LMST is a subgraph of RNG.
LMST algorithm can be easily used in self-organization schemes to generate organi-
zation which contains only a subset of links in the network. Although all nodes have
the same role in the network (different from clusters, CDS or MIS organization), the
logic structure are represented by LMST links. Nevertheless every node needs send
28
Self-organization
periodic HELLO message in order to compute local structure of LMST. LMST is still of
a perceivable control overhead.
2.3.4 Source dependent
The previous structures are known as source-independent structures which do not de-
pend on the origin of data traffic in the network. On the other hand, Multi-Point Relay
(MPR) enables the creation of a structure depending on the source of data traffic
(source-dependent). It is initially used to achieve broadcast in ad hoc proactive routing
protocol OLSR [50]. A node computes a subset of its one-hop neighbor set which covers
its entire two-hop neighbor set, when receiving a broadcast message. These neighbors
are selected as its multi-point relays, which are informed to forward the message upon
the reception of broadcast message. The MPR structure is generated from the source
of the broadcast to the entire network.
It is also possible to generate a source-independent structure based on MPR. [47]
proposes using two rules to generate a CDS organization based on MPR (MRP-CDS).
A node is in the MPR-CDS if and only if:
• the node has smaller identifier than all its neighbors (rule 1),
• it is a multi-point relay of the neighbor with smallest identifier (rule 2).
2.3.5 Synthesis on self-organization
In our point of view, self-organization is the fundamental mechanism in WSNs, which
provides a structure exhibiting certain properties. To generate and keep these proper-
ties along the network operation, localized algorithms are needed. The local decisions
made according to the corresponding metrics, reflect the order in the network. Figures
2.10(b)-2.10(f) give the snapshots of resulting structures obtained from a flat network
(UDG) (Fig. 2.10(a)) using different self-organization schemes.
Table 2.1 compares virtual backbone based self-organizations through type of infor-
mation, type of message, computation complexity and message complexity. It is shown
that only MIS [46] is not localized, while all algorithms let nodes make their decisions
based on locally collected information. All the solutions adopt periodic HELLO message
to exchange information locally. However the use of HELLO message creates continuous
control traffic in the network. More importantly, the transmissions and receptions of
HELLO messages significantly increase energy consumption on all nodes. Hence HELLO
messages are not a free technique, especially harmful for energy saving. Some algo-
rithms such as LMST and RNG need further position information.
29
State of the Art
(a) UDG (b) CDS WU & LI [4] (c) MPR-CDS [47]
(d) MIS [46] (e) LMST [49] (f) RNG [49]
Figure 2.10: Snapshots of self-organization structures: a network with 120 nodes, radius=0.16
Table 2.1: Synthesis of self-organizations
Algorithm net info msg info computation msgcomplexity complexity
CDS[4] localized HELLO/1-hop O(∆2) O(n)MPR-CDS[47] localized HELLO/1-hop O(∆2) O(n)LMST[49] localized HELLO/1-hop+pos. O(∆) O(n)RNG[49] localized HELLO/1-hop+pos. O(∆2) O(n)MIS[46] distributed HELLO+control/1-hop O(log(∆) · log(n)) O(n)
We also note that a node may benefit from the existing logic structure around it
before taking its own decision. This aspect is not well explored in existing solutions.
In the majority of the existing solutions, nodes only collect the information relevant to
physical topology. A gain might be obtained if nodes take advantages of the existing
structure to operate in the organization.
30
Conclusions
2.4 Conclusions
We have put forward in this chapter the importance of self-configuration and self-
organization for WSNs. The self-configuration allows nodes to set up identifiers for
communication, while self-organization allows nodes to cooperatively generate a com-
munication structure in the network. Both of them are the fundamental mechanisms
for setting up an autonomous network architecture in WSNs.
We have described in this chapter the existing proposals in self-configuration and self-
organization classified into several approaches. Yet, they do not meet all the require-
ments that we address in previous chapter. Motivated by improving efficiency of the
network formation and organization, reducing total control message overhead, we pro-
pose in the next chapter the Fully Integrated self-Configuration and self-Organization
Scheme (FISCO) as the autonomous network architecture for WSNs.
31
State of the Art
32
FISCO: An Autonomous Architecture for WSN 33
FISCO: An Autonomous Archi-tecture for WSN 3
3.1 Introduction
As already addressed in previous chapters, WSNs are required to be quickly and easily
deployed. Due to the large scale of the network, a pre-defined network infrastructure
is not applicable because it can not provide flexibility and adaptability to the network.
Besides, nodes’ addresses should be allocated without human intervention once the
network is deployed. Hence, self-configuration and self-organization stand as two key
schemes during the construction of the autonomous architecture in WSNs.
Most of prior works believe that: self-configuration and self-organization are two
independent problems in WSNs. They have never been linked before our proposal
[IP1]. We believe that self-configuration and self-organization are closely related each
other for three reasons.
1. Nodes can not participate in the organization unless they get validated addresses
through self-configuration. Moreover, the address is often used as a parameter
in self-organization schemes (such as [4, 49, 51]). Self-organization needs self-
configuration.
2. Self-organization could facilitate many network functionalities [52], including ad-
dress allocation. A node may take advantage of the existing organization to get a
validated address more efficiently. Self-organization improves self-configuration.
3. On the one hand, self-organization builds a structure in the network. On the
other hand, self-configuration has to maintain the address space in the network
through some logic structure to avoid address conflicts. It is a waste to generate
one address structure and one organization structure with redundant computation
and communication overheads.
Motivated by improving the efficiency of network formation and organization (i.e.
reducing total message cost and providing long lifetime), we propose Fully Integrated
FISCO: An Autonomous Architecture for WSN
Scheme of self-Configuration and self-Organization (FISCO) [IC2]. This concept not
only ensures the quick participation of a newly arrived sensor node in the network activ-
ities, but also deals with the spontaneous environment changes along with the lifetime
of the network. Different from existing works, we try to particularly avoid the usage
of HELLO messages because it leads to unnecessary energy consumption. The FISCO
not only achieves a good performance in both self-configuration and self-organization
aspects, it also acts has a flexible basis for other mechanisms. Later, in chapter 4, we
show that FISCO efficiently copes with multiple and mobile sinks in data dissemination,
and it supports data aggregation to improve the energy efficiency of data collection in
WSNs.
The chapter is organized as follows. Section 3.2 gives the fundamental concepts of
FISCO. FISCO employs event-driven mechanisms to generate and maintain a hierar-
chical structure and a two-level address allocation. Section 3.3 - 3.5 provide detailed
descriptions of the mechanisms involved when nodes join or depart and when partitions
merge. Section 3.6 introduces a local re-organization technique which extends the life
time of WSNs. Section 3.7 discuss how to provide a mesh structure based on FISCO.
In section 3.8, we evaluate the correctness of the solution through theoretical analysis.
Section 3.9 gives performance evaluation results in order to show the gain of FISCO
over existing self-configuration and self-organization solutions. Section 3.10 lists the
benefits of this autonomous network architecture to WSNs.
3.2 FISCO Highlights
3.2.1 Problem statement
FISCO is an autonomous architecture targeted at WSN sas described in section 1.2.
The WSN are formed with identical wireless sensor nodes which have low processing
and storage capabilities as well as limited energy resources. Neither address nor commu-
nication structure is predefined before the deployment. Nodes are randomly deployed
in a service area. Once they are deployed, they do not change their placements. Sensor
nodes are location unaware. Sporadic arrival and departure of nodes may occur during
the WSN operations. No synchronizations are set up between nodes.
FISCO aims at generating a logic structure through local information and local de-
cisions upon nodes’ deployment. It should also maintain this logic structure through
different network events such as the arrival and the departure of nodes.
34
FISCO Highlights
FISCO snapshot
LM
M
L
G
LM
G
M
M
L
M
FISCO structure
Figure 3.1: FISCO structure
3.2.2 Hierarchical structure
In order to represent FISCO structure, nodes are in one of the three states: leader,
gateway and member. A wireless sensor network running FISCO has a two-level hi-
erarchy: on the high level, a backbone is formed by leaders and gateways; on the low
level, each member node is linked to one leader (see Fig. 3.1).
A leader is in charge of all communications as well as address allocation in its 1-hop
neighborhood. The network state information such as address allocation information
are stored only at the leaders’. A node running under leader states has more respon-
sibilities and activities (e.g. during data dissemination and data aggregation discussed
in chapter 4) than other nodes. It also runs more calculations, spends more time to
communicate, and thus consumes more energy. Therefore the number of leaders in the
network should be limited and that is why leaders are not directly connected to one
another in the network. Leaders are spaced exactly at 2 hop apart. A gateway is in
charge of interconnecting the leaders. Leaders and gateways form FISCO’s high level
hierarchy: FISCO backbone. The cardinality of FISCO backbone is bounded (section
3.8.1) no matter the number of sensor nodes in the entire network.
According to the detection of leaders in their neighborhood, members are attached to
one leader. Once a member is configured, its incoming and out-going communications
are controlled by its leader. A leader-local organization is therefore generated. Although
this looks like a classical cluster formation around each leader, no clusterhead election
algorithm is used. Here, nodes do not need to periodically collect local information and
to elect clusterhead. Instead of collecting physical topology information, nodes take
existing local FISCO structures into account to make decisions.
35
FISCO: An Autonomous Architecture for WSN
16 Bytes for IPv6
Type of LDBR6
1 Byte
61 2 3 5
1
Leader Broadcast Message Format
4
Source IP address Partition ID 32 Old IP
4 Number of Leaders in the Partition
5 Number of Address Pool used in the Partition
4 Byte for IPv41 Byte 1 Byte 1 Byte
4 Bytes for IPv4
16 Bytes for IPv6
Figure 3.2: LDBR message format
Local FISCO structures are indicated by leader broadcast (LDBR) messages, given
in Fig. 3.2. PartitionID, number of leaders and number of address pool are the basic
local information used for new node joining, address update and merge detection events.
The OldIP is reserved for partition merge (in section 3.5). The type of LDBR is a field
reserved for local re-organization event (in section 3.6). Once a node enters into leader
state, it periodically sends LDBR message to its one-hop neighborhood. Different from
many other solutions where every node sends periodic HELLO messages, only FISCO
leaders generate periodic control messages. It optimizes the control traffic by pruning
redundant information.
3.2.3 Two-level address allocation
Based on the FISCO structure, we design a two-level hierarchical address structure (Fig.
3.3) to ensure the uniqueness of assigned addresses. On the low level, a leader node
assigns addresses to its newly arrived neighbor nodes using a stateful procedure based
on an address function f(n). The use of the stateful procedure ensures that no address
conflict is possible within the 1-hop neighborhood of each leader. On the high level,
the address space reserved for the WSN (known in advance) is managed by all leaders.
The full address space is divided into address pools. The division of address space is
possible if the given service area and node’s communication radius are known, because
the maximal number of leaders can be computed in advance (discussed in section 3.8.1).
Furthermore, it is not a problem if the number of address pools is greater than the
number of leaders. Because the allocation of each address pool is dynamic, each leader
may allocate several address pools depending on its local structure.
36
FISCO Highlights
The first address pool of a leader is assigned when it decides to enter into leader
state according to the join algorithm (section 3.3). Furthermore, when it runs out of
addresses in its current address pool, it may allocate a new address pool. Compared to
other types of address structure ssuch as binary addresses tree used in [1] and [3], this
horizontal address space division represents less processing and communication cost in
case of a partition merge (see details in section 3.5.3).
In each low level address allocation, a leader uses f(n) to select an address from
its local address pool. The simplest example f(n) is a denumerable function such as
f(n) = f(n− 1)+1. When the last address in its address pool is assigned out, a leader
should allocate the next address pool which is known by all leaders in its partition. It
should send an update message to other leaders through FISCO backbone (see section
3.3.2). Nevertheless, more sophisticated f(n) can also be implemented easily. For
example, a full period Linear-Congruential Random Number Generator (LCRNG) [53]
may take the role of f(n). The proprieties of a full-period LCRNG can guarantee that
each number is generated only once in its period. Besides, the technique of sequence
splitting can be used for dividing the LCRNG sequence into a set of subsequences with
reasonable computation time.
The design of FISCO combines both high level and low level address allocation to
ensure the uniqueness of addresses in a partition. It is also worth noting that this
address allocation technique takes advantage of the existing self-organization structure.
No additional structure needs to be maintained.
Address Space
............
......
6
L5
5
L4
3
L3
2AL
AL+1
2
L24
32
1
1
L1
......
6AL
5AL+3
5AL+2
5AL+1
5AL
4AL+3
4AL+2
4AL+13AL+3
3AL+2
3AL+1
4AL
AL+3
AL+2
3AL
2AL+3
2AL+2
2AL+1
AL
654
3
2
1Address Pools
Figure 3.3: Address space management in FISCO
37
FISCO: An Autonomous Architecture for WSN
Table 3.1: List of FISCO messages
Message Sender Type Use case
LDBR leader Broadcast/periodic local informationMBNS new node Broadcast/event discovery of neighborhoodMBNA member/gateway Unicast/event indicating node’s presenceMBAR new node Unicast/event requiring for addressLDAA leader Unicast/event address assignmentMBAA member Unicast/event address acknowledgmentMGNT member Unicast/event merge notificationLDAU leader Unicast/event address pool updateMGBA leader Unicast/event merge border leader acknowledgmentLDSC leader Unicast/event local re-organizationGWBR gateway Broadcast/periodic for mesh connecting
3.2.4 FISCO messages
FISCO is mainly built on event-driven procedures. Different from many existing self-
organization schemes such as CDS Wu&Li [4] or LMST [49], FISCO does not adopt
periodic HELLO messages. Nodes in member state do not need to periodically check
neighboring information or to perform calculation on local topology. Only leader nodes
send periodic LDBR messages and they are assigned with some basic computations
such as local re-organization (section 3.6).
Nevertheless, a set of FISCO messages are defined to specify different events in the
network. Table 3.1 gives a list of FISCO messages.
3.3 Join of Nodes
The ease of WSN deployment highly depends on the mechanisms which allow the con-
figuration and integration of new nodes. Hence, we first detail the sequence of events
when nodes join. Fig. 3.4 describes its procedure. Remember that a leader periodi-
cally broadcasts in his neighborhood an LDBR message including the partition identifier
PartitionID, the number of leaders and number of address pools in its partition. A
newly arrived node should listen to the medium during an interval of time, denoted as
LDBR_TIMOUT. From here, three cases are possible:
1. If it detects LDBR messages, then it enters into One-hop Address Allocation (OAA)
as detailed in section 3.3.1.
2. If it does not detect any LDBR, then it considers that there is no leader in its neigh-
38
Join of Nodes
State = "New"
@ allocation
Enter in one−hop
Decide to be the First node
Generate a random partition ID
Allocate an address pool
Using f(n) to generate an address
State = "First"
Finalize configuration
Send periodical LDBR message
whether received
No
Yes
LDBR_TIMEOUT
MBNA message
MBNA_TIMEOUT
No
YesLDBR Messagewhether received
Entre in two−hop @ allocation
Sensing for a MBNA Msg
Switch on
Temporary @
Sensing for LDBR message
Send MBNS message
Figure 3.4: New node configuration
borhood. The new node tries to actively discover its neighborhood by sending a
Member Neighbor Solicitation (MBNS) message and waits for a Member Neighbor
Advertisement (MBNA) message during a MBNA_TIMEOUT. If it detects any MBNA, then
it enters into Two-hop Address Allocation (TAA) procedure as detailed in section
3.3.2.
3. If the node does not detect any MBNA messages, it decides locally to create a new
partition in the network and assigns itself an address and an address pool in the
address space. The first leader also randomly generates a PartitionID to identify
its partition. The detailed procedures are given in section 3.3.3.
After the join procedure, a node may become a leader or a member, depending only
on the local FISCO structures. Intuitively, a node tries to get a snapshot of the current
organization around it, before making its own decision. More precisely, it relies on the
presence of leader nodes in one’s one-hop neighborhood.
3.3.1 One-hop address allocation
After detecting the presence of leader(s) in its one-hop neighborhood, a newly arrived
node extracts the information from LDBR message. If multiple partitions exist around
it (existence of leaders with different PartitionID), then it enters into partition merge
procedure (discussed later in section 3.5). Otherwise it selects one leader for OAA by
sending a Member Address Request (MBAR) message. It acknowledges the address allo-
cation (MBAA message) upon the reception of Leader Address Allocation (LDAA) message.
39
FISCO: An Autonomous Architecture for WSN
���������
���������
N
M
M L
M
M
M
LM
M
LDAA
MBAR
MBAA
Figure 3.5: One-hop address allocation
It is through this two hand-shake procedure (Fig. 3.5) that the new node obtains an
address from the address pool of its leader and acts as a member node.
3.3.2 Two-hop address allocation
A node entering into two-hop address allocation is aware that no leader covers its
current one-hop neighborhood. Based on this information of existing organization, it
decides to act as a leader and sends a two-hop MBAR message to one of the nodes which
have replied with a MBNA message. This node, either a member node or a gateway
node, relays the MBAR message to its own leader in the FISCO structure (Fig. 3.6).
The leader allocates a new address pool from the address space. It also sends Leader
Address Update (LDAU) message to all leaders through FISCO backbone, in order to
maintain a coherent view at high level address structure (address pool level) on all
leaders. The allocated address pool is assigned to the new node via the intermediate
member or gateway. An acknowledgment is returned to the leader. At the same time,
the relay node becomes a gateway if it was not one yet. It is a two hand-shake procedure
extended to two hops, in which an address pool is allocated. If the two-hop address
allocation fails, the address pool is kept by the existing leader for subsequent re-use.
3.3.3 Creation of a new partition
A node which does not find any configured node around itself considers itself the first
node of a partition. The node uses a random number generator to select a PartitionID
and allocates for itself the first address pool in the address space. It selects its address
according to the local function f(n) from its address pool. In order to indicate its
presence, it begins sending periodic one-hop LDBR messages.
40
Departure of Nodes
M
���������
���������
MBAA
MBAA
LDAA
LDAA
MBAR
MBAR
G
L
M
L
N
M
M L
M
Figure 3.6: Two-hop address allocation
To summarize, FISCO has a new node first discover the existing local structure in
its neighborhood, including the presence of one-hop leader and one-hop members or
gateways. Leader nodes play the most important roles in the join mechanisms because
they not only provide the FISCO structure information, but also update the address
allocation information via the FISCO backbone when new leaders are configured. Based
on the knowledge of existing local structure, each new node decides its role in the
organization and selects the most suitable configured node to get a validated address.
3.4 Departure of Nodes
Departure of nodes is also a basic event in WSN. We distinguish the departure of
an individual node and the departure of a set of linked nodes. The latter case is
considered as partition splitting event in section 3.5.1. A departure event may occur
when a sensor node is removed from the service area or when it runs out of energy. The
key for handling the departure of nodes is how to conserve the properties of FISCO
structure and the non-conflict address nature after such event.
3.4.1 Departure of a member node
A member node about to leave the WSN sends a Member Departure Notification
(MBDN) to its leader. The address allocated to this member is picked up by the
leader for re-use. A member has very few responsibilities in FISCO structure. It there-
fore has few impact on the others, except that all messages sent to it are reported as
unreachable. In case of an abrupt departure, a member node does not have time to
notify its leader. In order to recover this address, a leader should activate enhanced
mechanisms such as periodic scan or passive monitoring of data packets to identify
41
FISCO: An Autonomous Architecture for WSN
to provide connectivity
Qualified members reply
SS
@: .1.0M
L
MML
@: .3.0
M GM
L
@: .1.0
@: .2.0
GG
L
MM
@: .3.0
L
L
M
CC
@: .2.0
L
MM
@: .3.0
L
L
M@: .1.0
S
SS
L
@: .3.0
L
L
@: .1.0
@: .2.0@: .2.0
members to enter into sensing modeGateway leaves, leaders ask theirBefore gateway departure
to become Gateways
Leader select the suitable gateway(s)
Figure 3.7: The departure of a gateway
the departure. These mechanisms certainly increase the quantity of the control traffic
in the network. A trade-off, consistent with the application requirements, is always
possible.
3.4.2 Departure of a gateway node
If a gateway node is aware of its departure, then it sends a message to every leader
connected to itself. To deal with abrupt departure of a gateway, we add a leader-
gateway scan mechanism by forcing each gateway to send a unicast reply message to
its leader once it receives the LDBR message. In absence of gateway reply, a leader sends
a solicitation to find the gateway. It eventually finds that the gateway is gone, if no
advertisement is received. Obviously, this active scan mechanism introduces commu-
nication overhead. Nevertheless, the overhead is limited to FISCO backbone which
represents only a small subset of the sensor nodes.
After the detection or the notification of a gateway departure, a leader asks all its
members to enter into a sensing state. A member receiving any LDBR message from
another leader with the lowest address will reply to become a gateway. (see Fig. 3.7).
Once a member node is selected as the gateway by leader, it gets a new address from the
42
Departure of Nodes
leader with the smallest address. It is worth noting that the departure of one gateway
may bring several new gateways in the FISCO structure depending on the distribution
of the sensor nodes. Nevertheless, we show in section 3.8.1 that the cardinality of
FISCO backbone is bounded by its definition. Furthermore, the local re-organization
described in section 3.6 helps to improve the FISCO structure by reducing the number
of gateways around a leader.
3.4.3 Departure of a leader node
If a leader node abruptly leaves the network, then all of its members and gateways do
not receive its periodic LDBR message any more. This local information indicates the
absence of the leader node, thus the lack of local FISCO structure. After k rounds
without an LDBR message, its gateways and members conclude that the leader has gone.
k is a tunable parameter of FISCO. If k is small, then a departure will be quickly
detected. But some false detection might be reported due to bad reception. If k is
big, the false detection occurs less frequently. However, the latency of the detection
increases. Therefore k should be decided based on the requirement of the applications
running over the network.
After detecting their leader departure, gateways propagate a leader departure event
on the FISCO backbone for address pools reuse. After that, these gateways become
member nodes. At the same time, all members enter into a reconfiguration phase.
They give up their addresses and finalize their configuration by either finding another
leader or becoming new leaders (see Fig. 3.8). In fact, all nodes in the locality take this
opportunity to adapt their roles in the network organization according to the remaining
part of FISCO structure. The procedure is somehow close to a local re-organization of
the network (section 3.6).
It is worth noting that the departure of a backbone node creates a significant local
communication traffic, either for a gateway or for a leader. It is related to the hierarchy
in the FISCO structure where gateways and leaders are higher ranked. Because the
information related to the FISCO structure is merely provided by leaders, more control
overhead is generated by nodes, when discovering the existing FISCO organizations
in case of leader absence. However, it prevents a lot of energy wasting from periodic
control message (HELLO message) during network operating phase. According to the
WSN application, the network spends most of its lifetime under stabilized operating
phase. Hence, the gain brought by FISCO is much bigger than the additional cost in
departure events.
43
FISCO: An Autonomous Architecture for WSN
L
M
MM
M
GL
L
M
M
GL
One leader is disappearedBefore leader departure
Local re−organization finishes enter in a re−configuration phase
The members of this leader
M
MM
M
GL
M
M
GL
LL
R
R
R
R M
M R
R
M
M
L
M
L
M
G
L
G
Figure 3.8: The departure of a leader
3.5 Partition Management
A partition is defined as an isolated group of nodes. A network consists of several
disjoin partitions. The communication is possible only within the partition. When the
network is connected, there is only one partition. Many existing works consider only
one partition and do not discuss the network issues related to multiple partitions. Due
to the sporadic arrival and departure of nodes that we assume in WSN, there may exist
several partitions in the same service areas. The interactions between these partitions
should be taken care as well. Three mechanisms are designed in FISCO to manage
partitions in WSNs: partition splitting, partition detection and partition merge.
3.5.1 Partition splitting
One partition splits into two or several partitions if all connections between one part
and the others are completely lost. The objective is to maintain in each partition a
FISCO structure with the same properties. Keeping the existing FISCO structure in
new partitions avoids introducing new control cost. A member takes no action as long
as it is still covered by its leader. Otherwise, it considers itself as a new node and
performs the same mechanisms as described in join of nodes (section 3.3). Locally,
44
Partition Management
a member node does not make any differences between departure of it leader and a
partition splitting. It makes purely local decisions.
The detection of a partition splitting is reported by backbone nodes. When a leader
loses its gateway nodes in the FISCO structure, it first checks whether the link to
its 2-hop distanced leader could be repaired by assigning a new gateway. It is the
mechanism proposed for departure of gateway (section 3.4.2). If this fails, the leader
raises a partition splitting event by generating a new PartitionID and propagates it
through FISCO backbone in its new partition. Although it generates a message load in
partition splitting, the use of new PartitionID ensures the identification of partitions
in the network for merge issues.
3.5.2 Partition detection
We consider a gradual deployment, in which nodes are not scheduled to appear at the
same time. Nodes are randomly deployed and switched on within the region of service.
Therefore it is possible that several disconnected partitions are generated in the network
during the deployment. However, some partitions growing along with the deployment
of sensor nodes may eventually meet each other. In this case, partitions should detect
one another and get ready to merge.
PartitionID is used for detection. It is known that PartitionID is generated by the
first leader using a random number generator. The number of partitions in a region
of service is much smaller than the number of sensor nodes deployed in the network.
Hence, a small number of bits is enough to ensure that the probability to have a conflict
on PartitionID in a WSN is so low to be negligible. This assumption is also admitted
in some works [3]. Two parameters in the LDBR message, number of leaders and number
of address pools used in a partition, also help to distinguish two partitions in case they
would happen to have the same PartitionID.
One partition may detect the presence of another one if their communication vicinities
overlap. As we explained, the multiple partitions commonly occur during a random
deployment of WSN. Although leaders and gateways may detect another partition, we
introduce a new approach for partition detection, new arrival detection, in which the
detection is performed by newly arrived nodes. A newly arrived node takes full use of
the discovery stage in its join procedure to identify different partitions by PartitionID
in the LDBR messages it receives.
45
FISCO: An Autonomous Architecture for WSN
3.5.3 Partition merge
A new node detecting several partitions informs a leader per partition within its neigh-
borhood. It integrates itself into the partition containing the most leaders according to
the information in LDBR message, thereafter called the big partition. The information
on other partitions are also sent to the big partition, so that the new parameters such
as PartitionID, number of leaders and number of address pools are computed by the
leader with which the new node performs its two-hop address allocation. The Parti-
tionID can be re-generated by random number generator, the number of leaders and the
number of address pools can be obtained by summing up all corresponding parameters
in all partitions detected by the new node. After the two-hop address allocation, the
new node sends a Merge Notification (MGNT) message to the other partitions denoted as
small partitions. It then acts as a gateway that interconnects different partitions.
The MGNT message is propagated through the FISCO backbone in small partitions.
Thanks to the design of two-level address structure based on FISCO structure, each
node only needs to do a simple operation to avoid address conflicts in the new partition.
When receiving a MGNT message, a leader of a small partition performs a merge address
translation by skipping all address pools allocated in the big partition. No address
conflict can arise in the new partition because each partition always begins the address
pool allocation with the first addressed pool in the address space. Each leader also
updates the information contained in its LDBR message, so that its member nodes may
perform the same merge address translation as well.
3.6 Local Re-organization
FISCO is an event-driven scheme, hence its structure generated during the deployment
is kept unless new events occur in the network. In FISCO structure, leaders and
gateways are assigned with more tasks than other nodes are. Because they consume
more energy for computations and communications, their residual energy decreases
much more quickly than member nodes. In order to achieve a long lifetime of the
network and to balance the energy consumption, local re-organization mechanisms are
proposed for nodes to change their states during the entire lifetime of WSN. Leaders and
gateways use a local metric to judge if it should let some members replace themselves in
the FISCO structure. Our local re-organization is based on activity scheduling which
has first been applied to dominating structures in [54]. It balances the activities as well
as the energy consumptions on all nodes. The communication structure also benefits
from the local re-organization while some improvements might be achieved.
46
Local Re-organization
We design a metric based on the local residual energy and the properties of one’s
neighborhood. For a leader, the number of its members and gateways reflects its neigh-
borhood. For a gateway, only the number of leaders is taken into account. The expres-
sion of the metric M for a backbone node is the following:
• M(u) = Eresidual
Emax· Nmember+5
Nmember· Ngateway+1
Ngatewayfor leaders.
• M(u) = Eresidual
Emax· Nleader
Nleader+1 for gateways.
A small value of M(u) results to a high probability that a leader/gateway node
changes to a member node. Eresidual stands for nodes residual energy and Emax is a
reference maximum energy (corresponding to the quantity of energy supplied by a fully
charged battery). When the ratio of residual energy becomes low, the value of M(u)
decreases as well. It means that either a leader or a gateway tends to change its state
to member in case its residual energy is low. By giving up its leader/gateway sate, a
node consumes less energy and it keeps itself alive longer.
A leader is able to record the number of its members as well as its gateways because it
has knowledge of its one-hop neighborhood. These two parameters have strong impacts
on the decision of a leader/gateway node in local re-organization. A leader which has
many member nodes can be much more easily replaced by one of its members than a
leader with few member nodes. Therefore we adopt the ratioNmember+5
Nmemberto bring this
impact in the metric M(u). A leader connecting to many gateways is encouraged to
be re-organized, in order to avoid the presence of critical point in the network. This
impact is indicated by the termNgateway+1
Ngateway.
We adopt the same idea when designing M(u) for a gateway. Because a gateway
only has information on the leaders it connects to, the number of leaders is taken into
account. If a gateway is connected to many leaders, then it is preferable to keep its
role in order not to produce too many gateways after its re-organization as indicated
by the term Nleader
Nleader+1 .
After deciding to become a member, a leader sends a LDBR of type SENSE (Fig.
3.9), informing its members to sense for other leaders and to prepare for a local re-
organization. In order to bring minimum changes to the existing FISCO backbone,
members which do not detect any other leaders will reply. This can guarantee that
the new leader is at two-hop distance of other existing leaders. The current leader
selects a new leader from the candidate list, and sends it a leader Successor (LDSC)
message. It informs all its members and gateways of this re-organization in a LDBR
message of type CHANGE. The selected new leader begins its periodic LDBR to update
the local structure information. Any member detecting the LDBR switches its connection
47
FISCO: An Autonomous Architecture for WSN
L
Members not detecting any other leaders reply
Local re−organization finishes
LDBR
G
G
G
G
M
MM
L
L
broadcasts LDBR (type CHANGE)Leader selects succesor and
LDSC
LDBR (CHANGE)
M
M
MM
ML
L
Leader initiates re−organization
LDBRLDBR
LDBR (SENSE)
C
C
LG
M
M
LL
G
M
M M M
M
GL M
GL
LL
M
L
Figure 3.9: Local re-organization
to the new leader. It also checks its neighboring table generated while sensing to decide
whether or not it is required to become a gateway. The members which are not covered
by the new leader will renew their configuration considering themselves as new nodes.
The local re-organization helps to balance the energy consumption on sensor nodes
in the network. It takes three rounds of LDBR messages to form a new local structure
in the region around the old leader. Although this mechanism increases the energy
consumption in the network, it postpones the first node death. The WSN keeps all its
sensor nodes alive for a longer time.
3.7 Mesh Organization
The FISCO backbone is a tree-like structure, where no loop is allowed. The low car-
dinality property of FISCO structure (shown through the simulation results in section
3.9) is highly related to its tree form. However, no redundant paths are provided by
this backbone. In some cases, the usability of the FISCO structure may be restrained
because of robustness or data path length issues.
In order to create redundant paths and to reduce the length of the FISCO backbone,
some member nodes may be added to the backbone. This leads to adding loops on
48
Mesh Organization
(a) FISCO without mesh connecting (b) FISCO 1 -mesh
(c) FISCO 2 -mesh (d) FISCO 3 -mesh
Figure 3.10: Mesh backbone with different values of p
FISCO backbone which becomes a mesh. On the other hand, having more backbone
nodes signifies that more information should be maintained in the network and more
nodes run with higher energy consumption. Hence, there is a trade-off to find between
the maintenance cost and the path redundancy.
We introduce an additional mechanism to transform the basic FISCO tree backbone
into a mesh structure: Mesh connecting. Because leaders are the most energy consum-
ing nodes with the most activities, the mesh connecting aims at adding only gateways
to provide shorter and redundant paths. To this end, gateway nodes should explicitly
indicate their presences the same way leader nodes do, i.e. by sending periodic Gate-
way Broadcast (GWBR) message. A member node decides whether to become a gateway
based on two parameters: the number of leaders and the number of gateways it de-
tects. In a FISCO p-mesh, a member node switches itself to gateway if the following
two conditions are validate at the same time.
49
FISCO: An Autonomous Architecture for WSN
r
L2
L1
G
N
d > r
Figure 3.11: A topology with multiple gateways when p = 1
• Condition 1: more than two leaders are detected.
• Condition 2: less than p gateways are detected.
p is a tunable parameter which represents the desired gateway density around each
leader. Once a node enters into gateway state, it periodically sends GWBR message. All
leaders which detect a new gateway add it in their gateway table.
The parameter p has direct impact on the number of backbone nodes. Figure 3.10(a)
- 3.10(d) compare several FISCO p-mesh structures to FISCO tree structure. When
p = 2, 3, the FISCO backbone appears in a mesh form. Even when p = 1, mesh
connecting may bring new gateways to FISCO backbone. Fig 3.11 illustrates that if
the distance between two leaders is shorter than√
3r (r is the transmission range of
a sensor node), then it is possible that a member node N can not detect the existing
gateway G. Because the distance between them is bigger than r. According to rule 2,
N decides to enter into gateway state.
We apply two rules on a set of random UDGs to evaluate the impact of parameter p
on the number of gateways and the maximal path length in the network. The results
shown in Figure 3.12(a) and 3.12(b) are obtained through numerical analysis. As we
expected, increasing p leads to the generation of more gateways in the network. The
number of additional gateways become significant, when p is set to 2 and 3. On the other
hand, the maximal path length decreases when p turns bigger. It is worth noting that
mesh connecting requires gateway nodes to send periodic one-hop broadcast messages
as well. Hence, a trade-off p value should be set according to application requirements.
We further discuss its impact on energy consumption through simulations in section
50
Analysis
150 200 250 300 350 400 4500
10
20
30
40
50
60
70
Number of nodes
Num
ber
of g
atew
ays
Impact of mesh connecting on number of gateways
FISCOFISCO Mesh p=1FISCO Mesh p=2FISCO Mesh p=3
(a) Number of gateway
150 200 250 300 350 400 45017
18
19
20
21
22
23
24
25
26
27
28
Number of nodes
The
max
imal
pat
h le
ngth
Impact of mesh connecting on path length
FISCOFISCO Mesh p=1FISCO Mesh p=2FISCO Mesh p=3
(b) Maximal path length
Figure 3.12: FISCO mesh analysis with radius=0.18
3.9.4.
3.8 Analysis
In this part, we first prove, through geometric analysis, that the cardinality of a FISCO
backbone is bounded. It is not only an important property of the FISCO structure,
but also important to the address conflict-free nature of FISCO. The complexity of the
algorithm and the message cost are then given in order to confirm that FISCO meets
both the hardware constraints and the WSN application requirements. We note that
we assume circulair radio vicinity and symmetric link models in our proofs.
3.8.1 Analysis on FISCO backbone
We model all nodes as vertices of communication range R in a graph G. We prove
through Lemma 1 to Lemma 5 that the cardinality of the FISCO backbone is bounded,
whatever the number of nodes in the network. Furthermore, this bound only depends
on the size of the service area and the communication range R. First, the number of
leaders generated in a local arbitrary motif is demonstrated to be bounded. Then we
pave a given area with identical motifs. We prove the product of motif number by the
number of leaders in a motif has a maximal value. Therefore the number of leaders in
a given area is limited. The cardinality of FISCO backbone is therefore bounded.
Lemma 1 The leaders generated in FISCO form an Independent Set of graph G.
51
FISCO: An Autonomous Architecture for WSN
non covered
D R R
Da/2
a
Figure 3.13: The number of nodes in S within node u’s neighborhood
Proof : When a new node arrives in the network, it becomes a leaders if and only if
there is no leader in its one-hop communication vicinity. Leaders are at least separated
with two hops from one another. Therefore the leaders form an Independent Set.
Lemma 2 Let S be any Independent Set of graph G, u be an arbitrary node in S and
D be an euclidean distance between R and 2R. We define local(u) as the set of nodes
in S whose distances to node u are between R and D. The number local(u) denoted as
Nlocal
1. is at most[
πarcsin(R/2D)
]
for all D ≤ R2 sinπ/12 .
2. is at most 18 for D > R2 sinπ/12 .
Proof : The nodes in S are at least two-hop distance from one another, therefore
they are also separated by a distance superior to R. The node distribution of maximum
nodes in the annulus centered at u of radii R and D is obtained when all these nodes
are placed just on the circle of radius D. Illustrated by Fig. 3.13, the angle α between
two adjacent nodes on u is
α > 2 arcsin(R/2D) (3.1)
Therefore the max number of nodes in S that we can place on the circle at radius D
from node u is
Nlocal =
[
2π
α
]
(3.2)
For D ≤ R2 sin π/12 , the distance from an intersection of two adjacent circles to arbitrary
node u is shorter than R. Hence from Eq. 3.1 and 3.2, the first part of Lemma 2 is
proven.
52
Analysis
HexagonDodecagonOctagon
Figure 3.14: Motif: hexagon, octagon, dodecagon
For all distances D larger than R2 sinπ/12 , the angle α is bigger than π/6. Fig. 3.13
shows that the intersections between two adjacent nodes are out of the communication
range of u. When Rsinπ/12 < D < 2R, the maximal number of nodes that can be placed
on the circle with radius D is 12. And it is still possible to put a node in each non
covered zone. However, the distance between each non covered zone is smaller than
R. Therefore, at most 6 nodes can be placed among 12 non covered zone (Pigeonhole
Principle [55]). Finally, the annulus contains at most 18 nodes of S. The proof is
complete.
Although second part of Lemma 2 is close to the analysis in [56], the latter can not
be directly applied to prove that the number of independent nodes in a given area (not
an arbitrary area) is bounded. Because we may either use many small motif (with short
D) containing few leader nodes, or use few big motifs (with long D) containing many
leader nodes to pave the same area. We can not judge which paving uses more leader
nodes. It is through Lemma 3 that this result is proven based on the analysis on the
product of the motif numbers and the number of independent nodes in a motif.
Lemma 3 The cardinality of an Independent Set which covers the whole service zone
is bounded.
Proof : Lemma 2 proves the number of independent nodes in a local zone is bounded.
However, for different distance D, the independent nodes set forming certain motif
(Fig. 3.14) covers the communication zone with different sizes. We use a paving to put
maximal number of motifs to cover a given zone of a×a, meeting the condition that the
distance between any two adjacent leaders from different motifs is bigger than R. Fig.
3.15 shows an example of paving with hexagon (with D = R = 16a). Giving a paving
graph, it is possible to calculate the number of motifs needed to cover the given zone.
For any given tuple (a,D,R), we can compute the maximal number of motif Nmotif to
53
FISCO: An Autonomous Architecture for WSN
aR
Figure 3.15: Paving the space with hexagons
cover the network. The product of Nmotif and Nlocal, denoted as NIS , is the number
of of independent nodes which cover the entire service area.
The final expression of NIS is:
=[
a+2R(2D+R·cos α) · a+2R
2D+R sin (α/2)
]
·[
2πα + 1
]
, R < D ≤ R2 sin π
12
=[
a+2R(2D+R·cos α) · a+2R
2DR sin (α/2)
]
· (18 + 1), R2 sin π
12
< D < 2R(3.3)
with α = 2arcsin(R/2D) and function f(x) = [x] returning the closest integer to
variable x.
The derivative of the first part of this function is negative, therefore the maximal
value of N(D) occurs when D = R. The derivative of the second part of N(D) is
also negative, therefore the maximal value of this part occurs when D = R2 sin π
12
. The
cardinality of an Independent Set which covers the whole service zone is bounded by
the Max(N(D = R), N(D = R2 sin π
12
)). Finally, N(D = R) is the maximal number of
N(D) which is 7 ·[
a+2R(2+
√3/2)R
· a+2R2.5R
]
.
Lemma 4 The number of leaders generated in a given service zone is bounded.
Proof : From Lemma 1 and 3, the proof of Lemma 4 is trivial.
Lemma 5 The cardinality of the FISCO backbone generated is bounded for a given
service zone.
Proof : In FISCO (without the mesh connecting mechanism), there is at most one
gateway between two adjacent leaders. As indicated by Lemma 4, the number of
54
Performance Evaluation
leaders is bounded. Therefore the number of gateways in the structure is bounded as
well. The FISCO backbone is formed by leaders and gateways. Hence the cardinality
of FISCO backbone is bounded.
Lemma 5 has consequences not only on FISCO itself but also on address allocation.
Because the cardinality of the FISCO backbone is bounded, the latency of a message
carried by FISCO backbone is bounded no matter the number of nodes in the network.
According to Lemma 4, we know in advance the maximal number of leaders if the
service zone and the communication range are known before deployment of the network.
We can therefore divide the full address space into address pools. This ensures that
FISCO’s two-level address allocation is applicable.
3.8.2 Message complexity analysis
During the execution of FISCO, the message exchange cost for a node in joining proce-
dure and re-organization procedure is O(1) because each node send constant number of
messages per procedure. In FISCO, we especially avoid periodic HELLO messages. Only
leaders send periodic one-hop LDBR to announce their presence. As we have proven that
the number of leaders is bounded, broadcast message complexity is O(1). The message
complexity in the network is thus O(n).
It is worth noting that the message cost of two-hop address allocation is more sig-
nificant than one-hop address allocation when the cardinality of FISCO backbone is
big. Nevertheless, it occurs less frequently than one-hop address allocation. Thus, the
average message cost per node is still low.
Table 3.2 gives a synthesis of FISCO comparing to other self-organization schemes:
Connected Dominating Set (Wu & Li [4] and MPR based [47]), Local Minimum Span-
ning Tree [49], Relative Neighboring Graph [49] and Maximal Independent Set [46].
Hereafter, n stands for the number of nodes in the network, ∆ stands for the max
degree of nodes.
As shown in Table 2.1, FISCO achieves both lowest computation and lowest message
complexity. In addition, nodes do not need position information. FISCO consumes
the minimum information while providing a low complexity and low overhead solution.
These properties are further discussed in the next section through simulation results.
3.9 Performance Evaluation
We use Scalable Wireless Ad hoc Network Simulator (SWANS) [57] as our network
simulation tool. The JAVA based SWANS simulator is able to simulate large sensor
55
FISCO: An Autonomous Architecture for WSN
Table 3.2: FISCO versus other self-organizations: collected information and complexity
Algorithm net info msg info computation msgcomplexity complexity
CDS[4] local HELLO/1-hop O(∆2) O(n)MPR-CDS[47] local HELLO/1-hop O(∆2) O(n)LMST[49] local HELLO/1-hop+pos. O(∆) O(n)RNG[49] local HELLO/1-hop+pos. O(∆2) O(n)MIS[46] local HELLO+control/1-hop O(log(∆) · log(n)) O(n)FISCO local LDBR/local O(∆) O(n)(< n)
networks with a reasonable simulation time and memory occupation.
The positions of the nodes are randomly generated within a square area. Nodes
are also randomly switched on during the deployment phase. Each node employs a
802.15.4 PHY/MAC interface. The node’s radio characteristics are specified according
to Freescale MC13192 transceiver data sheet [58].
As we discussed in section 1.1, the radio transceiver dominates the energy consump-
tion in a sensor node. It is known that if the radio transceiver is in active state, it
consumes a similar amount of energy no matter whether the node is transmitting or
receiving. In order to link the energy consumption to the active time of transceivers,
we adopt BMAC [59] which employs a preambling technique to achieve MAC layer
scheduling.
At network layer, the existing IP stack with static address assignments is replaced by
FISCO. In order to compare FISCO with existing self-configuration, Strong Duplicate
Address Detection [24]and Prophet address [3] allocation scheme are implemented on
network layer. CDS (Wu&Li) and LMST self-organization scheme are also evaluated
by SWANS. The detailed simulation parameters are given in Table 3.3.
Because FISCO is the first integrated scheme of self-configuration and self-organization,
we divide the performance evaluation into two parts: one is dedicated to self-configuration
related properties; and the other one is dedicated to self-organization related properties.
We note that saving energy is key to WSNs. Hence it is evaluated in self-configuration,
self-organization, local re-organization and mesh organization mechanisms respectively.
3.9.1 Self-configuration related properties
We consider a progressive deployment phase where a number of nodes randomly join the
network, during which we compare the performances of FISCO, Strong DAD [24] (sec-
tion 2.2.2) and Prophet [3] (section 2.2.3) schemes. Each scheme represents a different
56
Performance Evaluation
Table 3.3: Simulation parameters
Parameters Descriptions
Radio Channel 802.15.4, 2.4GHzBandwidth 250kbpsTransmission power 0 dBmReception Sensitivity -92 dBmMac layer BMACTransmission energy consumption 100 mWReception energy consumption 87 mwSleep/Signal checking energy consumption 1.4 mWFISCO LDBR time interval. 1 sHELLO time interval in CDS/LMST/Prophet 1 sRadio channel Independent noise modelInter-arrival time between two nodes random value in [673, 1000]
self-configuration approach:
• Strong DAD scheme actively floods the network to find duplicate addresses after
each local address assignment. The address uniqueness is reactively maintained.
• Prophet scheme is based on Distributed DHCP approach. Each node uses a
stateful function to assign conflict-free address to newly arrived neighbor nodes.
The address uniqueness is proactively maintained.
• FISCO provides a two-level address allocation. On the low level, each address is
assigned by a leader from its address pool using a stateful procedure. On the high
level, address pools are allocated by leaders dynamically. The FISCO backbone
carries the address pool update messages for each two-hop address allocation.
Weak/Passive DAD approach is not evaluated, because it does not immediately re-
solve address conflict during a deployment phase. It is not fair to compare it with other
schemes.
3.9.1.1 Configuration message overhead
Fig. 3.16 shows the average message overhead per node in FISCO, Strong DAD and
Prophet. It is observed that FISCO outperforms Strong DAD and Prophet with a
very low average message cost. The main source of message overhead in Strong DAD
comes from the flooding DAD messages, while the use of periodic HELLO messages makes
Prophet scheme costly. Thanks to the two-level address allocation approach, FISCO
57
FISCO: An Autonomous Architecture for WSN
300 400 5000
100
200
300
400
500
600Send msg Fisco
Total nodes
Ave
rage
mes
sage
cos
t per
nod
e
300 400 5000
100
200
300
400
500
600Send msg Strong DAD
Total nodes300 400 500
0
100
200
300
400
500
600Send msg Prophet
Total nodes
300 400 5000
2000
4000
6000
8000Receive msg Fisco
Total nodes
Ave
rage
mes
sage
cos
t per
nod
e
300 400 5000
2000
4000
6000
8000Receive msg Strong DAD
Total nodes300 400 500
0
2000
4000
6000
8000Receive msg Peophet
Total nodes
BroadcastUnicastTotal
Broadcast
Unicast
Total
Figure 3.16: Configuration message overhead, radius=0.20
does not implement either technique. As a result, both sent and received messages
numbers are significantly reduced.
Furthermore, the self-configuration part takes advantage of existing local FISCO
structures in each partition to limit one-hop allocation and a part of two-hop allocation
in local message exchanges. Although two-hop address allocations and partition merges
require nodes to send messages through the network, the propagation of the message is
guided and carried by FISCO backbone. It is much more efficient than flooding based
forwarding. The low control message nature of FISCO demonstrates that distributing
the address server role only to a subset of nodes is very promising for dynamic address
allocation. It is more efficient than the Distributed DHCP approach.
3.9.1.2 Configuration latency
We define the configuration latency as the time it takes for a node to get a validated
address. It also indicates how quickly a node can integrate itself into the network. This
latency is not evaluated for Strong DAD, because it is set by the DAD timeout: each
node considers its pending address is validated after waiting a DAD timeout. For a
given DAD timeout is given, the latency is determined. The latency in Prophet is very
short, because each allocation occurs between two neighboring nodes. In FISCO, the
configuration latency for one-hop address allocation (section 3.3.1) is shorter than two-
hop address allocation (section 3.3.2), because in the second case a node should wait
58
Performance Evaluation
200 400 600 800 10001100
1150
1200
1250
1300
1350
1400FISCO configuration time
Total nodes
Ave
rage
con
figur
atio
n tim
e (m
s) FISCO
Figure 3.17: Configuration latency, radius=0.20
for LDBR and MBNA consecutively. It is shown in Fig. 3.17 that the average configuration
latency in FISCO decreases when the number of nodes increases in the network. This is
related to the fact that most lately arrived nodes perform a one-hop address allocation
because they are already covered by a leader node.
The configuration latency of FISCO is limited by the discovery procedures based
on LDBR and MBNA. Therefore, it is longer than Prophet address allocation in which no
sensing is needed. However, the passive discovery of LDBR avoids the generation of mul-
tiple responses. This technique reduces the message overhead during the configuration.
Hence, the configuration latency is somehow sacrificed in FISCO for energy saving.
3.9.1.3 Evolution of partitions during configuration
Rather than assuming all nodes are deployed before self-configuration, we consider a
random deployment of sensor nodes with sporadic arrivals. Therefore several discon-
nected partitions may appear during the deployment of WSN. Fig. 3.18 shows the
evolution of the number of partitions along with the deployment of sensor nodes. Each
early deployed sensor nodes (upon to 9) creates an isolated partition. The number of
partitions stays constant until the number of nodes reaches 30. Then the number of
partitions decreases, because some partitions begin to meet others with the growth of
their coverage areas. Thanks to partition merge procedure in FISCO (section 3.5.3), a
unique partition covers the entire service zone when the network becomes dense (from
300 nodes onward).
59
FISCO: An Autonomous Architecture for WSN
0 50 100 150 200 250 300 350 4000
1
2
3
4
5
6
7
8
9
Number of nodes
Num
ber
of p
artit
ions
The evolution of partitions during the deployment
Number of partitions
Figure 3.18: Evolution of the number of partitions from 1 to 400 nodes
3.9.1.4 Energy consumption for configuration
The low-energy consumption is the most important property that we aim at achieving
with FISCO. In Fig.3.19(a), it is clearly shown that FISCO has a much lower active
time ratio comparing to Strong DAD and Prophet. It is also shown in Fig.3.19(b) that
FISCO outperforms the other two schemes in terms of energy saving. Both results are
related to the low control message traffic generated in FISCO during deployment phase
which allows sensor nodes to sleep more. Due to the low message overhead, especially
low broadcast message overhead, the nodes’ reception time is significantly reduced.
Furthermore, FISCO provides a stable energy consumptions regarding to the number
of nodes in the network. On the contrary, energy consumptions of Strong DAD and
Prophet scheme increase in a linear fashion when the density of the network grows.
This illustrates that the energy saving provided by FISCO is density invariant.
3.9.2 Self-organization related properties
In this part of performance evaluation, we first look at the properties of FISCO struc-
ture. As a result of self-organization, FISCO generates two levels of structure: the
higher lever is represented by FISCO backbone which can be classified as a CDS struc-
ture; the lower lever is represented by the local structures around leader nodes in the
network. We then evaluate the message overhead and energy consumption of FISCO
compared to other self-organization schemes.
60
Performance Evaluation
250 300 350 400 450 500 5500
2
4
6
8
10
12
14
16
18
20Active time using BMAC scheduling
Total nodes
Per
cent
age
of a
ctiv
e tim
e af
ter
rand
om d
eplo
ymen
t
FISCO
Strong DAD
Prophet
(a) Active time
250 300 350 400 450 500 5500
1000
2000
3000
4000
5000
6000
7000
8000Energy consumption per node
Total nodes
Ave
rage
ene
rgy
cons
umpt
ion
(mJ) FISCO
Strong DADProphet
(b) Energy consumption
Figure 3.19: Energy saving in configuration, radius=0.20
250 300 350 400 450 500 550 600 6500
10
20
30
40
50
60
70
80
90
100Number of dominating nodes in self−organizations
Total nodes
Num
ber
of d
omin
atin
g no
des
FISCO leadersFISCO gatewaysFISCO leaders+gatewaysCDS dominating nodes
(a) Number of dominating nodes
250 300 350 400 450 500 550 600 6500
0.05
0.1
0.15
0.2
0.25Ratio of dominating nodes in self−organizations
Total nodes
Rat
io o
f dom
inat
ing
node
s
FISCO leadersFISCO gatewaysFISCO leaders+gatewaysCDS dominating nodes
(b) Ratio of dominating nodes
Figure 3.20: Cardinality of dominating set, radius=0.20
3.9.2.1 FISCO backbone
We first compare the cardinality of FISCO backbone to CDS [4] (section 2.3.3.1). CDS
organization runs based on the assumption that each node is assigned with a unique
address in the network. In order to avoid address conflict, we apply a static address
allocation in CDS. A node is pre-configured with a unique address before it launches self-
organization. On the other side, FISCO integrates an address allocation mechanism,
therefore no additional mechanism is needed in its simulation.
Fig. 3.20(a) shows the average number of dominating nodes generated according
to each scheme. In FISCO, both leaders and gateways are considered dominating
nodes. Although the number of leaders and gateways in FISCO increases, they do not
61
FISCO: An Autonomous Architecture for WSN
200 400 600 800 10000
10
20
30
40
50
60Properties of FISCO local structures
Total nodes
Num
ber
of n
odes
co
nnec
ted
to a
lead
er
Max number of membersAvg number of membersMin number of membersMax number of gateways
Figure 3.21: Properties of FISCO local structures, radius=0.14
grow as quickly as the number of CDS does. As a result (in Fig.3.20(b)), the ratio of
FISCO backbone nodes decreases with the increase in node number in the network.
Furthermore, this ratio is much lower than it is in CDS.
In FISCO, the newly arrived nodes take into account the existing structure in their
neighborhood. A node arriving in a zone already covered by a leader node, becomes a
member node. Fewer and fewer changes are brought by the later nodes on the existing
FISCO backbone. This is the reason why the cardinality of FISCO backbone is much
smaller when comparing to coverage based backbone’s (i.e. CDS). The low cardinality
property of FISCO backbone also results in low message overhead during address pool
allocation procedure (already shown in Fig. 3.16).
3.9.2.2 Local structure characteristics
We are also interested in local structures around leaders. Fig. 3.21 illustrates the
characteristics of the population connected to leaders. Because the number of leaders
in the network grows very slowly, we observe that both the maximal number and the
average number of members per leader increase in O(n) fashion. On the other hand,
since FISCO does not privilege any nodes to become a gateway node in two-hop address
allocation (section 3.3.2), the maximal number of gateways per leader is stable.
62
Performance Evaluation
200 400 600 800 10000
5
10
15
20
25
30
35
40Statistics on unicast messages (sent)
Total nodes
Ave
rage
num
ber
of u
nica
st m
essa
ges
per leaderper gatewayper memberper node
(a) Number of sent packets
200 400 600 800 10000
10
20
30
40
50
60
70
80Statistics on unicast messages (received)
Total nodes
Ave
rage
num
ber
of u
nica
st m
essa
ges
per leaderper gatewayper memberper node
(b) Number of received packets
Figure 3.22: FISCO sent and received control message cost, radius=0.14
3.9.2.3 Long term message overhead
Control message cost is one of the parameters which measures the efficiency of self-
organization scheme. Fig. 3.22(a) and Fig. 3.22(b) shows the average unicast control
message cost per node in FISCO, classified by roles (leader, gateway and member).
The per node control message cost in FISCO decreases and tends to stabilize with
the increase of the sensor nodes number. It coincides with our analysis on message
complexity of FISCO: message complexity per node is O(1). The same observation
is carried on the average number of message per member. The per leader control
message cost increases when the number of sensor nodes increases. Because leaders
are in charge of communicating with newly arrived nodes for address allocation and
organization, they send and receive many more messages than members do.
Comparing to CDS or LMST organization, FISCO does not employ periodic HELLO
message mechanism. In CDS or LMST, HELLO messages are used as the control messages
which allow nodes to discover the physical topology of the network. In FISCO, only
leaders provide information related to FISCO structure. This is the reason why FISCO
generates fewer control messages comparing to CDS and LMST in Fig. 3.23(a) and
Fig. 3.23(b).
3.9.2.4 Energy consumption for organization
In order to evaluate the energy consumption of self-organization schemes, each simula-
tion is composed of a deployment phase and an operation phase. The operation phase
is several times longer than the deployment phase. Because both CDS and LMST use
HELLO message as their unique control messages, their energy consumption are the same.
63
FISCO: An Autonomous Architecture for WSN
150 200 250 300 350 400 450 500 5500
0.1
0.2
0.3
0.4
0.5
0.6
0.7Statistics on message cost for network formation (sent)
Total nodes
Num
ber
of m
essa
ges
sent
per
nod
e pe
r se
cond
FISCOCDS
(a) Number of sent packets
150 200 250 300 350 400 450 500 5500
2
4
6
8
10
12
14
16
18
20Statistics on message cost for network formation (received)
Total nodes
Num
ber
of m
essa
ges
rece
ived
per
nod
e pe
r se
cond
FISCOCDS
(b) Number of received packets
Figure 3.23: CDS and FISCO control message cost, radius=0.20
200 300 400 500 6000
10
20
30
40
50
60Active time using BMAC scheduling
Total nodes
Per
cent
age
of a
ctiv
e tim
e af
ter
rand
om d
eplo
ymen
t FISCO
LMST
(a) Active time
100 200 300 400 500 600 7000
10
20
30
40
50
Energy saving of self−organizations
Total nodes
Per
cent
age
of e
nerg
y co
nsum
ptio
n co
mpa
ire to
alw
ays
on
FISCOLMST
(b) Energy consumption
Figure 3.24: Energy saving in organization, radius=0.20
Only LMST’s simulation results are represented to compare with FISCO’s.
Because simulations of different numbers of sensor nodes run for different durations,
the active time is normalized by the total simulation time. For the same reason, the
energy consumptions are normalized by the total energy quantity required for keeping
all nodes in transmission state (denoted as always-on).
Fig.3.24(a) clearly shows that FISCO has a lower active time ratio comparing to
LMST. Its energy consumption is therefore much lower (Fig.3.24(b)). The avoidance of
periodic HELLO message in FISCO efficiently reduces the idle sensing time of nodes, thus
energy consumption. Furthermore, FISCO provides a stable ratio of energy savings
(spends only about 5% of the energy of an always-on network). On the contrary,
energy consumption increases in a linear fashion when the density of the network grows
64
Performance Evaluation
50 60 70 80 90 100 110100
200
300
400
500
600
700
800
900Re−organization effect (300 nodes)
Initial energy resource (s)
Life
tim
e (s
)
Fisco with re−organizationFisco without re−organization
(a) 300 nodes
50 60 70 80 90 100 1100
100
200
300
400
500
600
700
800Impact of re−organization (400 nodes)
Initial energy resource (s)
Life
tim
e (s
)
Fisco with re−organizationFisco without re−organization
(b) 400 nodes
Figure 3.25: Impact of re-organization in FISCO on network lifetime, radius=0.20
in LMST. This illustrates that the percentage of energy savings of FISCO is density
invariant.
3.9.3 Impact of re-organization on lifetime
One issue of dominating based organization is that the dominating nodes (leaders and
gateways in FISCO) spend more energy than dominated nodes (members in FISCO).
FISCO adopts local re-organization mechanisms (section 3.6) to leaders from dying
much earlier than members. Meanwhile, let us evaluate the impact of this local re-
organization on the lifetime of WSNs composed of 300 and 400 nodes.
We define the lifetime as the time when the first node dies due to lack of energy. Each
node is assigned with same energy resource when it is deployed. The given resource
may let them stay alive for 60, 80 and 100 seconds respectively.
In Fig.3.25(a) and In Fig.3.25(b), it is shown that the use of local re-organization
mechanisms significantly prolong the life time of the network. For all initial energy
resources, the network life time is two to three times longer. And the more network
nodes there are, the bigger this ratio is. This is because more redundancy can be
explored by local re-organization when there are more sensor nodes. This observation
also puts forward the necessity of role scheduling in wireless ad hoc sensor network
regarding to long lifetime.
3.9.4 Energy consumption of FISCO mesh organization
The tree-like structure of FISCO can be extended into a mesh structure by executing
mesh connecting algorithm (section 3.7). In mesh organization, the maximal length
65
FISCO: An Autonomous Architecture for WSN
150 200 250 300 350 400 450 500 5507
7.5
8
8.5
9
9.5
10Active time of FISCO mesh organizations
Total nodes
Per
cent
age
of a
ctiv
e tim
e af
ter
rand
om d
eplo
ymen
t
FISCOFISCO Mesh, p=1FISCO Mesh, p=2FISCO Mesh, p=3
(a) Active time
150 200 250 300 350 400 450 500 5504
4.5
5
5.5
6
6.5
7
7.5
8Energy saving of FISCO mesh organizations
Total nodes
Per
cent
age
of e
nerg
y co
nsum
ptio
n co
mpa
iring
to a
lway
s on
FISCOFISCO Mesh, p=1FISCO Mesh, p=2FISCO Mesh, p=3
(b) Energy consumption
Figure 3.26: Impact of mesh organization on energy saving, radius=0.20
of the FISCO backbone is shortened and redundant paths are created. However, the
maintenance cost increases as the gateways need to indicate their existence for mesh
connecting. Fig. 3.26(a) and Fig. 3.26(b) illustrate this fact in terms of active time
and energy consumption.
As we expected, both the average active time and energy consumption increase when
the parameter p of mesh organization turns bigger. The additional cost comes from
the transmission of gateway broadcast messages for mesh connecting. The energy
consumption of FISCO increases about 5% for every unit increment in p. Therefore
the overhead of mesh organization is relatively low. From energy saving point of view,
FISCO mesh organization is still valuable.
3.10 Conclusions
Based on the observation that self-configuration and self-organization are two key mech-
anisms for providing an autonomous network architecture for WSNs, we have designed
the first integrated scheme to simultaneously allocate addresses and organize the net-
work. Designing them in an integrated scheme presents many advantages and im-
provements comparing to traditional approaches. FISCO provides self-organization and
self-configuration which leads to very reactive protocols. Every node can benefit from
the existing structure to get a validated address and participate in the organization.
Compared to a scheme where nodes perform a self-configuration and self-organizations,
FISCO allows newly arrived nodes to enroll into the existing network architecture in a
more efficient way. To summarize, the advantages of FISCO are:
66
Conclusions
1. An immediate configuration
By detecting the existing local structure, the newly arrived node is able to decide
of its role and get a validated address from at most two hops. While in some
self-organization schemes, new nodes are not recognized by others until the next
round of organization. Furthermore, it takes several rounds for a node to integrate
itself correctly into the network in these solutions. For instance, [45](section 2.3.2)
needs six iterations for a node to be integrated into the cluster structure.
2. Low message cost
Most self-configuration and self-organization schemes require nodes to perform
neighbor discovery explicitly or implicitly (often by using periodic HELLO mes-
sages). In FISCO, only a small part of the nodes (leaders) send periodic broad-
cast for address allocation and organization. It is observed that transmission and
reception costs in FISCO are smaller than HELLO message-based schemes.
3. Low energy consumption and long lifetime
Thanks to the avoidance of periodic HELLO messages on all nodes and to the unique
neighbor discovery process for both self-configuration and self-organization, nodes
spend more time sleeping. This yields to lower energy consumption and longer
lifetime when comparing to other schemes.
We have proposed two additional mechanisms, local re-organization and mesh con-
necting, to adapt the FISCO structure according to lifetime and redundant path re-
quirements. It is worth noting that each of them achieves its purposes by only adding
a reasonable computation and communication overhead.
We show through simulations that FISCO exhibits good properties such as low-
cardinality backbone, low message cost and low energy consumption. Especially, FISCO
provides big energy savings comparing to HELLO message-based schemes. This result
puts forward the suitability of FISCO as an autonomous network architecture for WSNs.
Based on this logic structure, in the next chapters, we will investigate on running data
centric services over FISCO. More precisely, how FISCO supports data dissemination
and data aggregation will be reported upon.
67
FISCO: An Autonomous Architecture for WSN
Publications
International conference
[IC1] Jia-Liang LU, Fabrice Valois, Dominique Barthel and Mischa Dohler. Low-
Energy Self-organization scheme for wireless ad hoc sensor networks. In Proc. of
the 4th Annual Conference on Wireless On demand Network Systems and Services
(WONS), pages 138-145, Obergurgl, Austria, January 2007. IEEE/IFIP.
[IC2] Jia-Liang LU, Fabrice Valois, Dominique Barthel and Mischa Dohler. Fully
Integrated Scheme of self-Configuration and self-Organization for ad hoc sen-
sor networks. In Proc. of Wireless Communications and Networking Confer-
ence(WCNC), pages 3370-3375, Hongkong, China, March 2007. IEEE.
[IC3] Jia-Liang LU, Fabrice Valois, Dominique Barthel and Mischa Dohler. Low-
Energy Address Allocation Scheme for Wireless Sensor Networks. In Proc. of
the 18th Annual Int’l Symposium on Personal Indoor and Mobile Radio Commu-
nications (PIMRC), pages 1-5, Athene, Greece, March 2007. IEEE.
International patent
[IP1] Jia-Liang LU, Fabrice Valois, Dominique Barthel and Mischa Dohler. Method
for organizing a network of communicating objects and the implementation of the
method. International patent, IPC H04L12/56, 2007.
68
Data Dissemination and Data Aggregation 69
Data Dissemination and DataAggregation 4
4.1 Introduction
In the previous chapter, we introduced an autonomous network architecture composed
of self-configuration and self-organization mechanisms. In our point of view, it is the
basic network architecture for WSNs, which not only achieves a quick and easy deploy-
ment but also provides an adaptive and flexible structure for communication among the
sensor nodes. Nevertheless, the majority of the application-related data communica-
tions occurs between sink nodes and sensor nodes. Different from MANET, the traffic
model in WSN is not one-to-one, but any-to-one or all-to-one. The data traffics are con-
centrated to sink nodes. In the application scenario that we mentioned in section 1.3,
the temperature sensors periodically report temperature measurements to the nodes
that connect to air-conditioning controls. Hence the temperature of the room could be
adjusted on the fly to ensure one’s comfort. In this chapter, we address this type of
communication by proposing data dissemination and data aggregation mechanisms.
In order to report the locally collected data to sink node, a node should construct a
data path to the sink node. There are two possible ways to do it: either every node
uses a unicast routing protocol to find the route to the sink (Fig. 4.1.a), or the sink
initiates the construction of a topology, based on which all sensor nodes become aware
of their paths to the sink (Fig. 4.1.b). In the first case, the identifier (address) of the
sink node should be notified to all nodes before the deployment. It is not flexible if
we consider the presence of multiple sinks in the network and it is energy-consuming
in case of mobile sink(s). In the second case, the sink should diffuse a message/query
in the network to inform its presence to all nodes. Along with the diffusion of this
message, a tree structure is constructed gradually from the sink as the root to every
end node as a leaf (used in Directed Diffusion [60]). Periodic re-construction is usually
needed to provide maintenance facing spontaneous changes. Both the construction and
the maintenance of this structure are costly.
Data Dissemination and Data Aggregation
Route Request
Sink
(b) Sink−Diffusion
Query/Data
Sink
(a) Unicast Routing
Figure 4.1: Communication in WSN
Data dissemination, considered as efficiently setting data paths, has been put forward
as one of the most important mechanisms of WSN by several prior works [60, 61, 62,
63, 64]. The aim of data dissemination is to let source nodes (sensor nodes which
have information to report) correctly deliver their data to sink nodes in the network
when minimizing the energy cost of the communication. Data dissemination is usually
composed of three phases: 1) the diffusion of data query from sink nodes, which allows
sensor nodes to be aware of sink nodes as well as their interests; 2) the data notification
phase, in which source nodes indicate the presence of data to the sink’s interests and
set up data paths to the sink node; 3) the data forwarding phase, during which source
nodes send the data to the sink nodes.
Note that multiple mobile sinks [61] bring both new advantages and challenges to
wireless sensor network. It is a core problem in data dissemination. The use of mobile
sink nodes improves the data collection efficiency in terms of shorter response time,
shorter data paths and lower energy consumption for wireless communication. The
presence of multiple sinks and mobile sinks allow to balance the searching, forwarding
and routing activities among all regions of the network. Hence the lifespan of the
network is prolonged. However, to support the mobile sinks, additional mechanisms
are needed to adapt the data paths in the WSN when sink nodes move. New challenges
are brought to WSNs such as quick localization of source nodes and sink nodes, efficient
query forwarding and data forwarding with low energy cost.
Another mechanism, data aggregation [63, 65], has also been considered as one par-
ticular technique in data dissemination, where data are combined spatially (coming
from different sources) or/and temporally (coming from different sampling cycles) on
the data dissemination structure from source nodes to sink node. As fewer quantities
70
Introduction
of data are transmitted, the data load is reduced in the network. As fewer transmitting
and relaying tasks are assigned, sensor nodes may spend more time sleeping. Intuitively,
the energy consumptions on nodes are dramatically reduced as well [64].
However, the generation of data dissemination/aggregation structures is not free-of-
charge. Tree structures are the basic structures used in data dissemination/aggregation.
Some of the existing solutions use sink flooding for the construction and periodic con-
trol message for structure maintenance [66, 63]. Geographic routing is also commonly
used while it requires that position information are frequently communicated between
neighboring nodes. Both of them cause significant message overheads as well as extra
energy consumption in a WSN.
In the previous chapter, we introduced FISCO, a low-energy, low control overhead
autonomous network architecture for WSN. Here, we consider sink-source communica-
tion as a basic application which could be developed on top of FISCO. In this chapter,
we show how FISCO structure can be transformed into a data dissemination structure
using a simple distance vector algorithm. The data dissemination structure we pro-
posed thereby takes advantage of the adaptive and localized nature of FISCO structure
to save on data communication and energy consumption, while taking benefit of FISCO
without generating additional overhead. Efficient management of multiple mobile sinks
is also a strong point of our solution.
In order to enhance the efficiency of the data collection in WSNs, a data aggregation
technique based on Adaptive AutoRegression Moving Average (A-ARMA) model is
presented. It employs low-order ARMA models to explore the temporal correlations
of samples, conforming to the sensor nodes’ constrained processing capabilities. Each
sensor node computes locally its own ARMA model parameters based on the sensed
samples. The accuracy of the model is verified whenever a number of new samples
are collected by means of a moving window technique. The model parameters are
updated if the accuracy of predicted samples within the window is poor. For each
model update, only the model parameters are communicated to the sink node, while no
data is sent when the model remains the same. The A-ARMA technique achieves both
low complexity and high accuracy with a completely localized algorithm without any
preliminary collection and computation phase. It induces significant energy savings by
reducing the data load. Moreover, based on the FISCO structure, a simple but efficient
spatial aggregation is also proposed.
This chapter is organized as follows. After reviewing the existing structures such as
TTDD [61] and Railroad [67], we propose in section 4.2 Backbone Based Data Dis-
semination (BBDD) scheme which is a distributed, low cost rendezvous system. It also
71
Data Dissemination and Data Aggregation
provides efficient management of mobile and multiple sinks in WSNs. An analysis on the
communication cost of BBDD and other structures is also given to validate the low cost
property of BBDD. In section 4.3, we propose the adaptive ARMA (A-ARMA) tech-
nique for data aggregation. We detail its complexity, accuracy and efficiency through
analysis. A framework, Self-Organization based data Dissemination and Aggregation
(SODA), which integrates data dissemination structure (BBDD) and temporal data
aggregation technique (A-ARMA) is presented in section 4.4. Its performance is evalu-
ated through simulations, while some important observations are also detailed. Finally,
we summarize our contributions in section 4.5.
4.2 Data Dissemination
Data dissemination is the way of setting up communication paths between sink nodes
and source nodes. Different from routing mechanisms in Mobile Ad-hoc Networks
(MANETs), some sensor nodes do not generate data traffic in the network. They only
relay the data packet from source nodes to sink nodes during a data collection. However,
source nodes and sink nodes are of particular importance to the network, which should
be therefore reflected in the data dissemination structure. Hence, both the dynamics of
sink nodes and sources nodes should be carefully taken into account in data dissemina-
tion scheme. The sink dynamics in a WSN comes from its movement. Therefore, data
dissemination should efficiently support sink mobility and presence of multiple sinks.
In a WSN, the source node dynamics are mostly related to the generation of data on
source nodes. The source nodes may vary over time. The data dissemination structure
should cover all sensor nodes for the construction and adjustment of data paths from
source nodes to sink nodes.
4.2.1 Overview of data dissemination schemes
Directed Diffusion [60] was one of the early works of data dissemination. It employs
the techniques of query flooding and gradual greedy reinforcement for path selection to
accommodate certain levels of network and sink dynamics. The basic idea is to let each
sink node diffuse a query to indicate the interest it wants to collect. Along with this
diffusion, the source nodes set up routes to the sink node based on gradient routing.
Greedy reinforcement was used for path selection in the first version, while energy based
path selection was proposed later in [68] to balance the energy consumption among the
intermediate nodes and to further reduce the overhead. However, in order to maintain
the paths, sink nodes should periodically diffuse their interests. It is also considered a
72
Data Dissemination
data-centric routing technique in WSNs.
Rumor Routing [69] is another application of data centric routing in data dissemi-
nation. Upon detection of an event, a source node sends out a packet named Agent.
The packet Agent will travel as far as possible from the source node in the network.
Each time it arrives at a node, it will decrease its TTL (Time To Live) value by 1. It
is dropped by the sensor node when its TTL reaches zero. The packet Agent therefore
creates a route on each node it passes toward the source node. When a sink looks for
an event, the associated query is sent in random walk in the network, until it reaches
a node having recorded the corresponding route to the source node. It is worth noting
that the data paths between the sink nodes and source nodes are much longer than the
optimal values. As a result, the unnecessary transmissions and energy consumption are
incurred in the network.
Rendezvous system is the common approach to provide data dissemination in WSNs,
especially to deal with the presence of mobile and multiple sinks. Two-tier data dis-
semination (TTDD) [61] employs a grid rendezvous system. It assumes that nodes are
aware of their locations. Each source node constructs a virtual grid to forward data to
the sink nodes. A set of dissemination nodes is selected to provide the high-tier grid for-
warding. The TTDD grid is therefore source dependent. Within each cell, geographic
routing is used. The sink mobility is divided into two categories: mobility within a cell
of grid and the mobility beyond a cell. No information update is needed for the former
one, because it is transparent to the higher-tier grid forwarding. Whereas the latter
type of mobility involves new dissemination nodes for the grid. Although the manage-
ment of sink mobility is achieved, a significant overhead is introduced for maintaining
the virtual grid.
Railroad [67] is another rendezvous based data dissemination scheme. A predefined
virtual ring structure, so called Rail is placed in the middle of the area where potential
phenomena occur. The Rail serves as an intermediate medium among sensor nodes,
sink nodes and source nodes. However, the width and the radius of this rail structure
should be planned in advance. The flexibility and scalability of this solution is therefore
limited, especially when the environment is dynamic: phenomena may travel from one
place to another (e.g. battle field monitoring such as tank tracking).
Fig. 4.2(a) - 4.2(d) summarize the data dissemination structures in WSNs. One
common idea is highlighted: all solutions apply rendezvous systems. The rendezvous
system is served as a zone where queries and data concentrate in the network. In case
of Directed Diffusion (Fig. 4.2(a)), sink nodes are the rendezvous points where all data
are intermediately concentrated to it after query flooding phase. The route created by
73
Data Dissemination and Data Aggregation
(a) Directed Diffusion [60] (b) Rumor Routing [69]
(c) TTDD [61] (d) Railroad [67]
Figure 4.2: Data dissemination: data centric routing and rendezvous systems
the packet Agent in Rumor Routing (Fig: 4.2(b)) is a rendezvous system, where the
traces of data are cached for queries. Easier to be identified, the virtual grid for TTDD
(Fig. 4.2(c)) and the Rail for Railroad (Fig. 4.2(d)), are both rendezvous systems.
4.2.2 Backbone based data dissemination
A data dissemination structure aims at creating the connection between wireless sen-
sor networks and sink nodes. FISCO, the autonomous architecture which provides
connections among wireless sensor nodes, provides a self-organized backbone and local
structures. It is a good basis for data dissemination structures, because it already copes
with all dynamics of the network changes among wireless sensor nodes. Moreover, its
generation is localized and needs no predefined parameters. In order to relate it to data
74
Data Dissemination
distribution and sink’s interests in the network, a Distance Vector (DV) algorithm is
used to direct the backbone toward the sink nodes. We named this solution as Backbone
Based Data Dissemination (BBDD).
4.2.2.1 Directed query forwarding
In BBDD, the sink node is in charge of indicating the data properties to FISCO back-
bone. When a sink node becomes active in the network by announcing a Query_Msg to
collect interests, the dependency between the sink node and the FISCO backbone is
created. At this stage, the Query_Msg is propagated through backbone and directs the
FISCO backbone (non-directed after self-organization) toward the sink node. It works
for FISCO tree backbone as well as for FISCO mesh backbone discussed in section 3.7
(Fig. 4.3).
Compared to flooding based query forwarding such as Directed Diffusion, the prop-
agation of this query will be confined to backbone nodes. Hence, it is more energy
efficient. Furthermore, a query aggregation technique is employed to further reduce the
overhead: a backbone node sends only one copy to its neighbor backbone nodes for the
queries for the same interest. Even though they are initiated by different sink nodes,
they still share the same data dissemination structure on the top of the non-directed
FISCO backbone. If several interests are required by multiple sinks, one structure
per interest in generated at the end of the query forwarding phase. Hence each data
dissemination structure is a data-centric structure.
Sink Sink
Figure 4.3: Directed query forwarding structure on FISCO mesh backbone
Each query contains InterestId (identifier of the interest) and ReportPer , in-
dicating the report period for this interest to sink nodes. Upon receiving a query,
75
Data Dissemination and Data Aggregation
a backbone node generates and records a tuple <InterestId, SinkId, UpstreamNode,
HopDistance> in its local SinkTable. It then selects the tuple with the shortest HopDis-
tance and adds a tuple <InterestId, ReportPer, UpstreamNode> with the correspond-
ing sink node into InterestTable , in order to direct the data stream back to the closest
sink node in the data forwarding phase. It also includes the InterestId in its peri-
odic one-hop control messages and broadcast them to its neighbors according to its
local time schedule. The one-hop control messages of backbone nodes (Leader Broad-
cast message in FISCO) are used to maintain the backbone structure, therefore no
additional message overhead is generated during the discovery of source nodes among
non-backbone nodes.
4.2.2.2 Data notification and data forwarding
Once a non-backbone node decides that it has the data corresponding to an interest, it
sends out the data to its backbone node using a data notification Data_Not message. If a
backbone node is a source node or it receives any data from its neighboring nodes, it will
relay the data back to the sink nodes according to the tuple <InterestId, ReportPer,
UpstreamNode> recorded during query forwarding stage. The data is thus sent back to
the sink node in Data forwarding Data_For message through the paths constructed for
each interest on the backbone (Fig. 4.4).
Sink
Multiple Sources
Event
Data
Source
Sink
Figure 4.4: Data forwarding
The directed backbone in BBDD is also responsible for storing the data of stimulus
reported by source nodes and responding to multiple mobile sink nodes. Hence in this
solution, the data dissemination is achieved on two tiers: 1) between the backbone and
the sink nodes; 2) between the backbone and the source nodes. After its generation,
76
Data Dissemination
the directed backbone is a natural distributed rendezvous system of the wireless sensor
network, through which both the queries for interests and the data may reach their
final destinations. The use of backbone avoids flooding of either queries or data in the
network, which saves significantly energy during the data communication. Comparing
to existing rendezvous system such as Railroad [67], BBDD does not require location
information, geographic routing or pre-configuration. It is localized and benefits from
self-organization to dynamically deal with network changes. The mobility and dynamics
of sink nodes are not considered by the basic self-organization scheme in the previous
chapter. It is absolutely necessary to provide a solution for managing sink mobility and
multiple sinks.
4.2.3 Management of multiple mobile sinks over BBDD
Note that FISCO construction is independent from sink location; the mobility and
dynamics of sink nodes are not considered in the basic FISCO functionalities. It is
absolutely necessary to provides a solution for managing sink mobility and multiple
sinks. In this part, we discuss how the mobile sinks and multiple sinks can be efficiently
supported over BBDD.
As sink nodes send out the Query_Msg when they require for new interest, they should
also indicate their movement in the WSN by sending out a type of messages, defined
as Noti_Msg in our proposal. It is worth noting that there is no dependency between
the Query_Msg and the Noti_Msg. The first is responsible for interest while the second
is responsible for sink mobility.
It has been proven in the previous chapter that FISCO backbone covers the en-
tire WSN field with its communication vicinities. Therefore, the presence of a sink
node is detected by backbone nodes, upon receiving a sink Noti_Msg . A backbone
node (leader or gateway) checks whether it is the first time that this sink node ar-
rives in its neighborhood. If it is the case, then it updates the HopDistance of the
corresponding tuple <InterestId, SinkId, UpstreamNode, HopDistance> in SinkTable.
It also checks if the tuple in InterestTable should be updated. It then propagates
a Sink_Update_Msg through the backbone. Each backbone node updates the HopDis-
tance and the UpsteamNode of the corresponding sink node. If shorter paths are created
for certain interests, the InterestTable should also be updated (Fig. 4.5).
Intuitively, each backbone node uses once more the distance vector algorithm upon
receiving Sink_Update_Msg in order to decide whether or not to redirect its path to the
sink node as well as the paths for related interests. Therefore, both the hop distance to
a mobile sink node and the distance to the closest sink node of an interest are evaluated
77
Data Dissemination and Data Aggregation
and updated along with the announcements of mobile sink nodes.
Changed path
Multiple paths
Unique path
Sink
Sink
Figure 4.5: Support for sink mobility
Because the backbone serves as the natural distributed rendezvous system in the
network, it is also a space for intermediate data storage in order to avoid loss of data
when the sink nodes are in the movement. The backbone node which connects to a
sink node stores the in-coming data after that sink moves away. When it receives the
path update message, it will send the stored data to the sink nodes via the new path.
4.2.4 Analysis on data dissemination
In order to evaluate the communication cost of BBDD compared to other solutions, we
adopt the models and methods in [67] for analysis. We assume that there are N sensor
nodes randomly deployed in a square field with the area of 1×1 and intensity of λ. We
denote r the communication radius of sensor nodes. In order to facilitate the analysis,
every sensor node has the same radius r.
Three types of messages are used in BBDD: query forward message, data(event)
notification message and data forwarding message. The sizes of these messages are
denoted pq, pe and pd respectively. In RailRoad [67] and TTDD [61], the same types
of messages are used as well. The size of a control message for rendezvous system
maintenance is denoted pc. We adopt the same assumption as in [67] that there are m
mobile sinks and the average number of source nodes in the network is n ([67]). Each
mobile sinks generates q queries and each source nodes generates e data in average.
Therefore, the number of queries and the number of data are mq and ne respectively.
78
Data Dissemination
The communication overheads of TTDD [61] and Railroad [67] have been analyzed
in [67]. Both of them assume the use of a greedy geographic routing protocol such
as [70]. Both queries and data messages are forwarded in the network using unicast
message. However, the communication cost related to the greedy geographic routing
protocol has not been considered. It is well known that in order to communicate
the location information (absolute or relative) to its one-hop neighbors, each node
should send periodic HELLO messages. Therefore, in addition to communication cost for
grid and rail maintenance in TTDD and Railroad, the cost of sending HELLO message
should also be taken into account. Let us assume that the frequency of HELLO message
is the same as the frequency of sink queries. Since the average number of queries
generated by a mobile sink is q, there are q rounds of 1-hop HELLO message broadcast.
The communication overhead related to greedy geographic routing is Nqpc. The total
communication overheads of TTDD and Railroad are:
CTTDD = mq [NC + λ/r · 2] pq
+λ/r · ne[
2 + (√
2/(2α))]
pd
+(n 4Nλ/r·(1/α) + Nq)pc;
(4.1)
CRailroad = neλ/r · (√
2/4)pe
+λ/r[
(ne + mq) · (√
2/4) + mq · (2√
2)]
pq
+4neλ/r · (√
2/4)pd)
+Nqpc;
(4.2)
In TTDD, α represents the size of a cell in grid, and NC is the average number of
nodes in a cell. In this analysis, NC = Nα2, because we are working on a square field
with the area of 1 × 1.
In BBDD, the total communication overhead is composed of data (event) notification
cost, query forwarding cost, data forwarding cost and FISCO control message cost. We
denote CBBDD = CDN + CQF + CDF + CFISCO.
In BBDD structure, every node is at most one-hop to the backbone. When a source
node wants to notify a data/event to backbone nodes, the data notification message is
carried at most one-hop. Therefore CDN = nepe, where NB is the number of backbone
nodes in BBDD.
During query forwarding, the query message will be propagated through the back-
bone. In the worst case where the query message needs to be forwarded to all backbone
nodes every time, we have CQF = mqNBpq.
The data forwarding cost is linked to the data path length on the backbone. Let
us define function Hλ,r(d(x, y)) = ξ d(x,y)r which represents the average number of hops
79
Data Dissemination and Data Aggregation
0 10 20 30 40 500
0.5
1
1.5
2
2.5
3
3.5x 10
6
Number of mobile sinks
Com
mun
icat
ion
over
head
TTDDRailroadBBDD
(a) n=10, m variable
0 20 40 60 80 1000
1
2
3
4
5
6x 10
6
Number of source nodes
Com
mun
icat
ion
over
head
TTDDRailroadBBDD
(b) m=5, n variable
Figure 4.6: 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 forTTDD)
between node x and y, where d(x, y) represents the Euclidean distance between node
x and y. ξ tends toward 1 when the density of the network is high. As the same, we
can define function HλB ,r(d(x, y)) = ξBd(x,y)
r for the number of hops between node x
and y via the backbone. The path length on the backbone is therefore bounded by the
Hλp,r(√
2). In the worst case, CDF = neξB
√2
r pd.
Finally, CFISCO = NB · qpc because only backbone nodes in FISCO employ periodic
1-hop control message.
CBBDD = nepe
+neξB
√2
r pd
+NBqpq
+NBqpc;
(4.3)
In section 3.8.1, we prove that the FISCO backbone has a bounded size and gives
its expression based on the network size and the node’s radius. Although we can not
give the exact numerical expression of CBBDD, we may compute its upper bound by
replacing NB by the maximal number of FISCO backbone nodes indicated in section
3.8.1.
From Fig.4.6(a) to Fig.4.7(b), the communication overheads of TTDD, Railroad and
BBDD are analyzed regarding to different parameters. In Fig.4.6(a), the number of
mobile sinks varies from 1 to 50. It is shown that the communication overhead of Rail-
road system is insensitive to the number of mobile sinks, whereas the communication
overheads of TTDD and BBDD increase. This observation signifies that the communi-
80
Data Dissemination
0 100 200 300 400 500 6000
2
4
6
8
10
12
14x 10
5
Number of data per source node
Com
mun
icat
ion
over
head
TTDDRailroadBBDD
(a) q=50, e variable
0 50 100 150 2000
0.5
1
1.5
2
2.5
3
3.5x 10
6
Number of query per sink node
Com
mun
icat
ion
over
head
TTDDRailroadBBDD
(b) e=100, q variable
Figure 4.7: Analysis of communication overhead (m=5 mobile sinks, n=10 source nodes,N=10000, r=0.1, α=0.3 for TTDD)
cation cost of Railroad is dominated by the data notification and forwarding. TTDD
has the highest communication overhead among three solutions. BBDD has the lowest
communication overhead when the number of mobile sinks is inferior to 35. Above this
number, Railroad generates lowest overhead.
The evolution of communication overhead versus the number of source nodes is given
in Fig.4.6(b) when the number of mobile sinks is fixed to 5. TTDD always generates
the highest communication overhead, while both Railroad and BBDD provide savings
on communication cost. When the number of source nodes is under 100, BBDD achieve
the lowest communication overhead.
Fig.4.7(a) and Fig.4.7(b) show the effect of average number of queries and data
messages on the communication overhead through three rendezvous systems. It is
shown that the more queries and data the network generates, the more significant a
gain BBDD provides compared to TTDD and Railroad. BBDD is therefore suitable
for long term data communication.
The BBDD achieves the goal of data dissemination, setting up data paths from
sources nodes to sink nodes, based on the FISCO autonomous architecture. The effi-
ciency of BBDD relies on the localized, flexible and low-cost natures of FISCO. Com-
paring to other data dissemination structure, very few additional control overhead is
generated. Thanks to the dominating properties of FISCO backbone, mobile and mul-
tiple sinks can be easily supported by BBDD.
In the next section, we discuss a data aggregation technique which aims at further
reducing the data load and improve the energy efficiency during a data collection. From
our point of view, data aggregation is linked to data dissemination. Because some data
81
Data Dissemination and Data Aggregation
dissemination structures can be transformed to data aggregation structures and hence
support data aggregation techniques.
4.3 Data Aggregation
The basic idea of data aggregation is to exploit data correlation among the sensor nodes
by eliminating data redundancy. Consequently, there are fewer transmissions in the
network, hence energy is saved on sensor nodes. Generally speaking, data aggregation
schemes act on top of some hierarchical logic structures, particularly tree structures.
Each sensor node receives data from its child nodes, aggregates them with the data it
senses and then sends the aggregation result to its parent node. This process is repeated
until the aggregation result is sent to the sink node.
Two questions have been thrown at data aggregation schemes: Where to perform
data aggregation? And when to perform data aggregation? Both of them address
putting an order in data communication: one based on time, one based on space.
4.3.1 Where to aggregate
One way to address where to aggregate the data in WSN is to generate an aggregation
structure. Data dissemination structures can be easily adapted as aggregation struc-
ture. For example, the sink-rooted tree generated in Directed Diffusion [60] is re-used
as a data aggregation structure in [66], because each non-leaf node may perform spatial
data aggregation, once it collects data from its child nodes.
A cluster based structure is another possible structure for running data aggregation.
CAG [63] proposes to form clusters of nodes with similar sensing values. Subsequently,
only one value per cluster is transmitted back to the sink and the readings of other sen-
sors are ignored. A query containing interests is sent by the sink. The CAG algorithm
operates in two phases: query and response. During the query phase, data-centric clus-
ters are built according to a user-specified error threshold τ . Each sensor node compares
its own value of this interest to the value in the query message. If its value is within
τ , it joins the cluster of its parent node (the node which forwards the query message
to it). Else, it adds its value in the query message. By forwarding the query message
to all its neighbors, a new cluster is created. In the response phase, CAG transmits a
single value per cluster: only the clusterheads contribute to the aggregation.
The maintenance of the data-centric clusters remains a difficult problem. The use
of local in-cluster maintenance may increase the number of clusters while each cluster
contains fewer sensor nodes for aggregation. As a result, the aggregation efficiency
82
Data Aggregation
decreases. Furthermore, the flooding phase in CAG [63], also used in [66] results in
high message overheads and maintenance costs.
Some solutions do not discuss the flooding phase, yet it is implicitly admitted. In
[71], the authors introduce snapshot queries. The idea is to let each node locally select
one or several neighbors to represent its own data, according to specific model classes.
The local selection algorithm results in a snapshot which is formed by a subset of nodes.
Only snapshot nodes respond to sink queries. Local structures are provided in snapshot
queries, whereas the global aggregation structure of the network is not discussed. It
implies that a query flooding phase is necessary in this solution and snapshot nodes
still need to use an ad hoc unicast routing protocol to deliver the data to sink nodes.
Although data aggregation is achieved by means of eliminating responding nodes using
data models, the local snapshot discovery and maintenance is costly (which should be
considered as control overheads in the work). Moreover, a node can not guarantee that
the snapshot maintenance is still localized when coping with the variation on its data.
The BBDD structure discussed in section 4.2 can be considered as a data aggregation
structure directed to the sink node, while the leader nodes in BBDD backbone become
natural aggregation points in this data aggregation structure. Therefore, we know where
to aggregate based on BBDD structure. The spatial aggregation running on each leader
is further discussed in section 4.4.1. Meanwhile, we concentrate on addressing the other
question: when to aggregate.
4.3.2 When to aggregate
One approach is to use scheduling and cross-layer design in order to set the order for
data aggregation, while employing temporal data aggregation techniques is the other
approach.
In [72], two mechanisms are used to provide scheduled data aggregation: data-aware
anycast at MAC layer and randomized waiting at application layer. Data-aware anycast
assigns additional functionalities into MAC RTS/CTS mechanism to let data packets
in one hop converge to a few nodes. Then these nodes can perform spatial aggregation
on data packets. The mechanism is repeated every time a packet gets one hop closer to
the sink node. Randomized waiting technique is used to get more data packets on each
node for aggregation. It is shown that this technique does not create control overhead
for generating and maintaining a data aggregation structure. However, the solution is
not flexible from development point of view. It is also shown that the delivery delay
becomes longer with the increase of network density.
Several temporal data aggregation techniques were proposed. TiNA [65] is a temporal
83
Data Dissemination and Data Aggregation
data aggregation solution which reduces the data loads by exploring the difference
between the current value and the last sent data value. A sensor node sends the current
sample in a data packet only when the value of the sample has changed significantly.
A sensing value can be ignored if the variation from the previous value is within the
range specified by the clause condition in a continuous query. The bigger the range is,
the more energy is saved; however, a big range also leads to low model accuracy. For
a useful level of accuracy, energy savings are thus limited, even though the complexity
of this mechanism is low.
In [73], the authors propose using an Auto-Regressive Integrated Moving Average
(ARIMA) model [74] to reduce the data load. The scheme includes two phases. In
the preliminary phase, each node sends its data at a regular report rate to the sink
node. The sink node computes for each node an ARIMA model and sends the model
parameters back to each node. In the second phase, a sensor node sends an actual
sampling data only if the deviation between the predicted value and the actual sample
is greater than a given threshold. However, in order to reach an adequate level of
accuracy, a long preliminary phase is required (the preliminary phase is set as long
as the adaptive phase in [73]). Furthermore, if the accuracy of the model decays,
then another preliminary data collection phase is needed. The inflexible nature of this
solution is related to the fact that all model parameters are computed by the sink
node. The efficiency of this solution is hence restrained by the use of a centralized
computation.
Distributed regression [75] uses regression models to reduce the transmission rate
while still retaining the structure in the data. A set of basic functions is provided by
users when the network is deployed. Each sensor node uses an algorithm based upon
kernel linear regression to compute a model with these basic functions. The coefficient
vector becomes a compressed representation of the measurements and is transmitted at
every computation. However, the regression messages for distributed regression com-
putation are composed of a squared matrix and a vector of given coefficients. It hence
requires a significant storage space. Furthermore, the regression message overhead of
order O(n2) (with n being the number of nodes in the network) is relatively high.
Note that sensor nodes are very limited on their processing capabilities as well as
storage capacities. These constraints imply that the data aggregation algorithms exe-
cuted on each sensor node should be of low complexity and low computational overhead.
No prior work related to data aggregation has met the requirement of low complexity,
high accuracy and low energy at the same time. To this end, we introduce a data
aggregation technique relying on the efficiency of an adaptive Auto-Regression Moving
84
Data Aggregation
Average (A-ARMA) model using a moving window technique.
4.3.3 Adaptive-ARMA model for data aggregation
Our A-ARMA technique employs low-order ARMA models to explore the temporal
correlations of samples, conforming to the sensor nodes’ constrained processing capa-
bilities. Each sensor node computes locally its own ARMA model parameters based on
the sensed samples. The accuracy of the model is verified whenever a number of new
samples are collected by means of a moving window technique. The model parameters
are updated if the accuracy of predicted samples within the window is poor. For each
model update, only the model parameters are communicated to the sink node, while
no data is sent when the model remains the same.
The novelty of our approach w.r.t. prior art can hence be summarized as follows:
• It achieves both low complexity and high accuracy at the same time; the use of
moving window technique in the ARMA model estimation reduces significantly
the computational complexity whilst achieving a high accuracy on the predicted
values.
• It invokes significant energy savings by reducing the data load; only ARMA model
parameters are transmitted and only when needed.
• There is no need for a preliminary data collection phase; our data aggrega-
tion technique could work as soon as sensor nodes begin to sense, i.e. no pre-
computation phases are necessary.
• The algorithm is completely localized, i.e. no centralized computation is involved.
First, we give some background knowledge on ARMA model.
4.3.3.1 ARMA model
The AutoRegression Moving Average (ARMA) model [74] is a widely-used model for
time series analysis. It uses the historical data to develop a model for the prediction
of future values. Many environmental physical quantities such as temperature and
humidity can be modeled with ARMA model. Hence, ARMA model fits well in WSN
monitoring application. It incorporates two terms, the AutoRegression (AR) term, and
the Moving Average (MA) term.
85
Data Dissemination and Data Aggregation
1. The AR term is a linear regression which represents the self-deterministic part of
the time series. It forecasts the current value Xt with p prior values:
Xt = φ0 + φ1 × Xt−1 + ... + φp × Xt−p,
An AR(p) model is characterized by the p + 1 coefficients φ0, . . . , φp.
2. The MA term captures the influence of random shocks which is independent from
autoregressive process. The model consists of random shocks on q prior elements:
Xt = θ1 × ǫt−1 + ... + θq × ǫt−q,
A MA(q) model is characterized by the q coefficients θ1, . . . , θq.
The order of an ARMA model is defined as max(p + 1, q); the higher the order is,
the higher the algorithmic complexity of the model parameter estimation.
ARMA(p, q) is not a panacea for all kinds of time series, it assumes the time series is
stationary and invertible. There are many cases of the violation of the stationary and
invertible restrictions. Some of them can be easily pre-eliminated, and ARMA model
is still useful. Differencing the time series is one of the possible method which leads to
AutoRegression Integrated Moving Average (ARIMA) models: a time series should be
differenced d times to match ARMA model. Such ARIMA(p, d, q) is denoted as:
Xt = φ0 + φ1 × Xdt−1 + ... + φp × Xd
t−p + θ1 × ǫt−1 + ... + θq × ǫt−q
where d is the degree of difference. The scope of this work is for WSN monitoring
application in which the target measurements, such as temperature and humidity, are
stationary most of the time. Therefore, the use of ARMA model is justified.
4.3.3.2 Local A-ARMA computation
The basic idea of Adaptive-ARMA (A-ARMA) is that each node computes an ARMA
model based on a fixed-size window of W consecutive samples. By merely sending the
parameters of the ARMA model to the sink node(s) and possibly further to distant
servers for rebuilding the data, the temporal correlation of these samples within each
window is exploited. The model parameters will also be used by the distant server for
data forecasting/prediction, unless it receives model updates from the sensor nodes.
Each node locally verifies the accuracy of the predicted data values w.r.t. real col-
lected samples. If the accuracy is adequate according to the given threshold, the node
86
Data Aggregation
assumes that the server can rebuild the data correctly and there is hence no need to
report the data. Otherwise, it computes a new model and communicates its parameters
to the sink node(s) for it to adjust the forecasting. In order to reduce the complexity in
the model estimation process whilst still achieving a high accuracy, a moving window
technique is introduced. This means that the verification is required every time the
window moves a step ahead. The adaptive nature of our technique is relying on the use
of this moving window.
Intuitively, the sensed data series are divided into periods by moving window tech-
nique. Furthermore, an A-ARMA model for the entire series is composed of a set of
ARMA models computed on different time intervals.
To facilitate the description of our method, several parameters are defined:
• W is the window size. The parameters of the ARMA model are computed on the
latest W samples.
• therr is the threshold of the error tolerance on the root-mean-square (RMS) error
between predicted values and actual data. It also represents the accuracy that
the A-ARMA model achieves locally over each window.
• S is the step size. Between two verifications, the window of computation moves
S samples forward.
As detailed in Figure 4.8, the A-ARMA model estimation works as follow:
1. A node builds an ARMA(p, q) model once it has collected W samples, where p+1
parameters for the AR part and q parameters for the MA part are computed. The
model parameters are communicated to the sink node.
+
W
S
RMS
new ARMA(p,q)If RMS > th_err
ARMA(p,q)
Actual samples
Predicted values
Figure 4.8: Block diagram of A-ARMA.
87
Data Dissemination and Data Aggregation
2. As time goes by, more samples are collected. Upon collecting the next S samples,
the node measures the RMS error between predicted values and the actual values
in the current window. If the difference is within therr, then the node continues
using its current ARMA model.
3. If the difference is greater than therr, then the node re-computes the new pa-
rameters for the ARMA model on the most recent W samples. From the last
computation, the window moves S samples forward. The new parameters of the
ARMA model are then communicated to the sink node(s).
The underlying idea of this model estimation based on a moving window is two-fold:
1. In order to reduce the complexity of the estimation algorithm, a low order ARMA
model is used and the estimation is based on a time series of reasonable length.
The complexity of executing an ARMA model parameter estimation process is
O(m3W ) [76], where m is the order of the model and W the length of the data
sample. In our solution, both the model order and the window size are selected
to adapt to the sensor nodes’ processing capability, which facilitates the imple-
mentation of such mechanism on wireless sensor nodes.
2. In order to achieve the accuracy of the predictive values, the model is verified along
with the progression of the moving window. Instead of checking the accuracy
of each data sample one by one (as in the case of TiNA [65]), a sensor node
validates the accuracy of the predictions every S samples (S is the step size of the
moving window technique). The moving window technique reduces the number
of verifications, thus the computational overhead.
In summary, the adaptive property of A-ARMA is a result of using the moving
window technique, which keeps the predicted output values close enough to the real
samples at a low computational complexity.
4.3.4 Analysis on A-ARMA
4.3.4.1 Accuracy and efficiency
In order to corroborate the improvement of the A-ARMA model w.r.t. the ARMA model
in terms of accuracy and efficiency, we apply three models on a real time series of 720
indoor temperature samples (provided by the ICT CAS EASINet research group [77]).
The time series is considered as the set of samples sensed by one node in the network.
We estimate two non adaptive models computed over the entire time series of 720
88
Data Aggregation
0 100 200 300 400 500 600 700 80027.4
27.5
27.6
27.7
27.8
27.9
28
28.1
28.2
28.3
Sample’s ID
Tem
pera
ture
(D
egre
e C
elsi
us)
Forecasting with ARMA and A−ARMA models
ARMA(2,2)ARMA(400,250)Sensed samplesA−ARMA(2,2): W=20, S=5, th
err=0.03
Figure 4.9: Applying ARMA and A-ARMA models on indoor temperatures.
samples, i.e. an ARMA(2, 2) model and an ARMA(400, 250) model. Furthermore, an
A-ARMA(2, 2) model is used with the following parameters for comparison: W = 20,
S = 5 and therr = 0.03.
Figure 4.9 depicts the predicted values based on different models. It clearly shows
that ARMA(2, 2) cannot fit the actual data with a satisfying accuracy. ARMA(400, 250)
is close to the real data but the computation of the parameters is far too costly. The
results of both extremes justify the need for using an adaptive technique, such as
moving window. As a result, the A-ARMA(2, 2) fits the actual data quite well. It
is worth noting that the accuracy of our solution depends on the threshold i.e., the
smaller therr is, the higher the accuracy is. However, as outlined below, accuracy and
efficiency of the aggregation mechanism should be balanced.
Figure 4.10(a) shows the evolution of the RMS error between the estimated values
and the actual samples using the A-ARMA model. We selected an A-ARMA(2, 2)
with W = 20, S = 5 and therr = 0.03, so as to obtain the same accuracy of the
ARMA(400, 250) model (in average, 0.029 degrees of RMS error per sample). 66 model
updates are necessary according to Figure 4.10(b). For each ARMA(2, 2), the commu-
nication cost is the number of the parameters (5 parameters are needed). Therefore,
the total amount of information sent by a node equals 5 × 66 = 330 parameters. In
the case of ARMA(400, 250), the communication cost equals to the number of model
89
Data Dissemination and Data Aggregation
parameters, which is 400 + 250 + 1 = 651 units. Since the same precision of model
parameters is used in both models, A-ARMA(2, 2) saves around 50% of the data loads
compared to ARMA(400, 250). Although only 5 parameters are used for ARMA(2, 2),
its accuracy is too poor for it to forecast the data correctly (c.f. Figure 4.9).
We define the efficiency of aggregation as the ratio between the quantity of infor-
mation sent to the network with and without data aggregation. In the analysis, the
same number of bits is allocated to both data sample and model parameter (for either
A-ARMA or ARMA) which avoids the transformation between different types of vari-
ables. Under this consideration, the thus defined efficiency of ARMA(400, 250) is 651720
(around 0.9); the efficiency of A-ARMA(2, 2) is 330720 (around 0.46). It is obvious that
the A-ARMA technique brings a significant improvement on data load reduction, when
comparing to a non-adaptive ARMA approach.
The analysis based on Figures 4.9 and 4.10 illustrates that the use of low-order A-
ARMA can meet all requirements on accuracy, complexity and efficiency.
4.3.4.2 Under erroneous measurements
In realistic roll-outs, errors may occur in the sensed data. They create random variations
in the sample set. We discuss in this part the impact of different errors on the proposed
A-ARMA technique in terms of accuracy and communication cost.
It shall be noted first, that the prime function of the A-ARMA model is to allow the
sink node(s) to rebuild the measurements observed by the sensor nodes even though
the samples are measured with non-stationary deviations produced by the external
00.01
0.020.03
0.040.05
5
10
15
200.024
0.026
0.028
0.03
0.032
0.034
0.036
0.038
Threshold
Impact of step size and thresholdon the accuracy of A−ARMA(2,2)
Step size
Ave
rage
RM
S e
rror
per
sam
ple
(deg
ree)
(a) RMS error
00.01
0.020.03
0.040.05
5
10
15
20
20
40
60
80
100
120
Threshold
Impact of step size and threshold onthe number of model updates in A−ARMA(2,2)
Step size
Num
ber
of m
odel
upd
ates
(b) A-ARMA model updates
Figure 4.10: Accuracy and efficiency of the A-ARMA(2,2) on 720 samples.
90
Data Aggregation
effects. However, some properties of the samples set may be conserved through the use
of A-ARMA model. A subset of non-stationary deviations could be pruned or reduced.
In order to show the impact of A-ARMA model, we validate both the accuracy of A-
ARMA model (defined as the RMS errors to the original samples) and the distance
to the sample set under non-stationary deviations (defined as the RMS errors to the
erroneous samples).
Each error is modeled as 1 degree deviation (increasing or decreasing) from a sample
in the real data series (the same used in previous analysis). We are interested in
applying two error classes:
1. Independent errors: the errors are introduced independently to the time series;
each error affects only one data sample.
2. Consecutive errors: each consecutive error is a set of correlated errors and occurs
over a time interval.
The analysis shown in the following figures are computed based on hundreds of ran-
dom distributions of errors on the entire data series.
Independent errors Figure 4.11(a) compares the impact of independent errors on the
accuracy of A-ARMA(2, 2) and ARMA(400, 250). The accuracy of both models drops
(average RMS error increases) when more and more independent errors are introduced
on the samples. However, A-ARMA(2, 2) achieves a better accuracy (with lower RMS
error) than ARMA(400, 250) does. It remains at an acceptable level (around 0.20 degree
for RMS error), considering that each error introduces 1 degree deviation. On the other
side, the predicted values of A-ARMA(2, 2) is not as close as the ARMA(400, 250)’s
to the erroneous samples. This result puts forward that the stationary properties
of samples are very important for applying A-ARMA technique. When the samples
are affected by non-stationary deviations, A-ARMA tends to correct these deviations
because of its low order, which makes the prediction values closer to the original samples.
ARMA(400, 250) is computed on the entire data series, hence its model cost remains
at 651 parameters as discussed in section 4.3.4.1. It is shown in Figure 4.11(b) that
more models are used to predict the data series when independent errors are introduced.
Almost one additional model update is necessary to deal with an error. According to
this observation, we think that it might be more interesting to just detected independent
errors in a window and send it directly, instead of changing model.
It is also worth noting that the errors at the beginning of time series have much more
impacts on the estimation. If more than 3 independent errors are introduced within the
91
Data Dissemination and Data Aggregation
0 5 10 15 20 25 300
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Impact of consecutive errors on the accuracy of A−ARMA(2,2) and ARMA(400,250)
Number of errors
Ave
rage
RM
S e
rror
per
sa
mpl
e (d
egre
e)
A−ARMA(2,2): RMS to original samplesARMA(400,250): RMS to original samplesA−ARMA(2,2): RMS to erroneous samplesARMA(400,250): RMS to erroneous samples
(a) Accuracy
0 5 10 15 20 25 30 3565
70
75
80
85
90
95
100
105
110
Number of errors
Num
ber
of m
odel
upd
ats
Impact of independent errors on the number of model updates in A−ARMA(2,2)
A−ARMA(2,2)
(b) Additional model updates
Figure 4.11: Accuracy and efficiency under independent errors
first window, then the estimation of the model fails (the predicted values do not match
the real ones). It is related to the short size of the moving window and the low order
of MA term in the A-ARMA. As a conclusion, a correct modeling at the beginning of
the time series is very important when using our technique.
Consecutive errors Let us now apply consecutive errors to the data series. For each
analysis, errors of different length are applied while the total number of errors is
fixed to 30. As shown in Figures 4.12(a), the accuracy of either A-ARMA(2, 2) or
ARMA(400, 250) does not vary with the length of errors. Once more, A-ARMA(2, 2)
achieves a better accuracy than ARMA(400, 250) (around 0.05 degree closer). On
the contrary, the distance from the erroneous samples to the predicted values of A-
ARMA(2, 2) decrease significantly when the consecutive errors turns longer. Because
the more correlation there is in the deviations, the easier these errors can be modeled
by A-ARMA.
The number of model updates shown in 4.12(b) is impacted by the way that the
deviations are modeled. The longer a consecutive error is, the less the number of model
updates increases. Hence, the additional model updates occur mainly at the edge of
correct samples and erroneous samples, as well.
We figured out the impact of erroneous measurements on the accuracy and efficiency
of A-ARMA model. Although the accuracy of A-ARMA model is impacted by the
independent errors, it still remains satisfactory. The consecutive errors also decrease
the accuracy of A-ARMA model. However, its length has few impact on the accuracy. In
both cases (independent and consecutive errors), A-ARMA achieves a better accuracy
than single ARMA does. Nevertheless, independent errors reduce the efficiency of A-
92
Data Aggregation
0 5 10 15 20 25 300
0.05
0.1
0.15
0.2
0.25
Impact of consecutive errors on the accuracy of A−ARMA(2,2) and ARMA(400,250)
Length of consecutive errors
Ave
rage
RM
S e
rror
per
sa
mpl
e (d
egre
e)
A−ARMA(2,2): RMS to original samplesARMA(400,250): RMS to original samplesA−ARMA(2,2): RMS to erroneous samplesARMA(400,250): RMS to erroneous samples
(a) Accuracy
0 5 10 15 20 25 30 3570
75
80
85
90
95
100
105
110
Length of consecutive errors
Num
ber
of m
odel
upd
ates
Impact of consecutive errors on the number of model updates in A−ARMA(2,2)
A−ARMA(2,2)
(b) Additional model updates
Figure 4.12: Accuracy and efficiency under consecutive errors
ARMA more heavily than consecutive errors do. It is also shown that the longer the
error length is, the fewer additional models are needed.
4.3.5 Highlights of A-ARMA technique
According to the above presentation, the advantages of using A-ARMA technique on
wireless sensor nodes for data aggregation are:
1. The computational complexity of the model is low enough to be run on individual
sensor nodes.
2. The use of moving windows reduces the required storage space for estimation and
further increases the accuracy.
3. The aggregation efficiency of A-ARMA on the raw data is much better than
non-adaptive techniques.
4. A-ARMA is able to reduce the impact brought by non-stationary deviations under
erroneous measurements.
We also note that the technique of the A-ARMA model has the following limitations:
1. The model fails if the number of independent errors is bigger than q of A-
ARMA(p, q) model in the first window of the series. The fault tolerance of the
technique is ensured only if sufficient samples have been analyzed at the beginning
(the number should be greater than the window size).
93
Data Dissemination and Data Aggregation
2. The introduction of errors decreases the accuracy of the model and increases the
number of model updates. Nevertheless, the accuracy of the predicted values
remains at an acceptable level: we obtain a RMS error of 0.20 degree when the
time series is affected by 30 independent errors with 1 degree of variation. And
the independent error is the most harmful error type to the efficiency of A-ARMA.
In this section, we hence validated the three properties of A-ARMA technique, i.e. low
complexity, good accuracy and high aggregation efficiency. For each sensor node, the
A-ARMA technique ensures a reduction in data loads by means of using prediction
model and moving window technique.
4.4 SODA Framework
By integrating the data dissemination structure and temporal A-ARMA technique,
we finally build Self-Organization based data Dissemination and Aggregation (SODA)
framework. As indicated by its name, the concept of SODA takes full use of autonomous
architecture FISCO to provide efficient data dissemination and aggregation. SODA
framework consists of four techniques:
1. At the lowest level, FISCO self-organization handles the network changes among
sensor nodes upon WSN deployment. It is used to support data centric structures.
2. BBDD structure is generated over FISCO structure. It is a data centric structure
which serves both data dissemination and data aggregation. It relies on FISCO
for dealing with sensor nodes’ dynamics and it provides efficient management of
mobile and multiple sinks through its backbone based rendezvous system.
3. Each node locally runs A-ARMA technique to efficiently explore the temporal
correlation of data samples at a low computational complexity and overhead.
4. Furthermore, FISCO leader nodes employ spatial packet merge technique to ag-
gregate the data coming from their member nodes during the data collection. It
is a trivial utilization of FISCO local structures.
Through SODA framework, we point out that data communication in WSN is not
an isolated problem, but involve many mechanisms and designs in networking. To
provide an efficient data communication, organizing the network is the fundamental
task. It allows the easy implementation of other mechanisms and improvement of their
performance.
94
SODA Framework
FISCO, BBDD and A-ARMA have been detailed respectively in section 3.2, 4.2.2
and 4.3.3. The last technique, spatial packet merge technique is again highly related
to FISCO organization. More precisely, it relies on the properties of FISCO local
structures. Hereby, we give the guidelines for this spatial packet merge technique.
4.4.1 Spatial packet merge on leader nodes
According to BBDD scheme, a source node should send its data toward its correspond-
ing node in the rendezvous system. The corresponding nodes are known as FISCO
leader nodes. A leader node is a local collection point and hence able to perform spa-
tial aggregation for all its members. The report cycle may be defined in the application
which is different from the data sampling cycle. During a report cycle, a leader node
receives a number of packets, if it has several source nodes as members. It is highly
profitable for it to aggregate these data packet before sending them to the sink node
through data-centric backbone.
The most trivial way for a leader node to perform aggregation is to merge the packets
from its member nodes in each report cycle (spatial packet merge technique). Although
the quantity of information in each report cycle is not reduced, only one packet is sent
and relayed through the backbone. The communication overhead is therefore reduced.
More sophisticated algorithms taking advantage of data compression techniques [78]
and data multiplexing techniques [79] are also applicable on leader nodes. Both of
them allow to exploit the correlation between data/parameters sequences in different
packets. However, these techniques require nodes to have more processing capabilities.
In SODA framework, we only implement the spatial packet merge technique.
It is also worth noting that spatial packet merge can work with A-ARMA, because
leader nodes do not need to be aware of the contents of a packet in this technique.
Instead of merging data packets, leaders merge the packets containing A-ARMA model
parameters to one packet for reporting.
4.4.2 Performance evaluation on data collection
Again, we use Scalable Wireless Ad hoc Network Simulator (SWANS) [57] to evaluate
SODA framework in data collection. It is compared to other data aggregation technique.
The simulations are set under the same condition as for performance evaluation of
FISCO presented in section 3.9.
We run simulations for different network topologies, where sensor nodes are randomly
distributed in a square area. The sink node is always placed at the center of the square.
The application is set as indoor environment monitoring, whose goal is to allow the
95
Data Dissemination and Data Aggregation
sink node to rebuild the evolution of data collected on each sensor node. The indoor
temperature samples provided by the ICT CAS EASINet research group [77] are used
as measurements for sensor nodes. Sensor nodes run the six schemes listed below during
a data collection.
1. No aggregation: Each sensor node sends a packet back to the sink node every time
it gets a new data sample. The packet is relayed to the sink node in a multi-hop
fashion. Each sensor node has set up a predefined route to the sink node.
2. Temporal packet merge: Each nodes collects 5 samples, before sending them in
one packet to the sink node.
3. Single ARMA aggregation: The technique presented in [73] is implemented. How-
ever, the ARMA model is used as a prediction model instead of the ARIMA model,
because the indoor temperature series used in the simulation is stationary. The
data collection is divided into two phases. In the preliminary data collection
phase, data samples are directly sent to the sink nodes. For each sensor node, the
sink node computes a single ARMA model and sends the model coefficients back
to the sensor nodes. In the second phase of the data collection, a sensor node
sends the current data sample only if the accuracy of the estimation value is not
adequate. In order to achieve the accuracy of the model, the preliminary phase
is set to half of the total simulation time.
4. A-ARMA: Each node uses the A-ARMA technique. The packets containing the
model parameters are sent via the same routes as in the no aggregation case.
5. Spatial packet merge: BBDD structure is generated on the top of FISCO struc-
ture. Member nodes report each data sample to their leader nodes. A leader
node merges all packets in its report cycle and sends one packet to the sink node
through BBDD structure.
6. SODA framework: every node run the complete framework including: FISCO,
BBDD, A-ARMA and spatial packet merge.
The data sampling cycle is set to 10 seconds for all schemes and the report cycle in
the temporal and spatial packet merge techniques is set to 5 times the sampling cycle.
The duration of each simulation is 2 hours, and 720 samples are collected in total by
each node. Each simulation is repeated for 30 random network topologies.
96
SODA Framework
280 300 320 340 360 380 4000
500
1000
1500
2000
2500Send message cost
Number of nodes
Num
ber
of s
end
mes
sage
sNo aggregationTemporal packet mergeSingle ARMA aggregationSODA frameworkSpatial packet mergeA−ARMA
(a) Average number of send messages
280 300 320 340 360 380 4000
500
1000
1500
2000
2500Receive message cost
Number of nodes
Num
ber
of r
ecei
ve m
essa
ges
No aggregationTemporal mergeSingle ARMA aggregationSODA frameworkSpatial packet mergeA−ARMA
(b) Average number of receive messages
Figure 4.13: Number of messages in data collection
4.4.2.1 Message cost during data collection
Fig. 4.13(a) and 4.13(b) show the average number of message sent and received. In case
of SODA and spatial packet merge, the message overhead of FISCO structure is also
taken into account. All techniques reduce the number of messages, among which the
lowest message cost is achieved with SODA framework. Although the spatial packet
merge over BBDD reduces the total number of message by half, each packet size is much
longer. The very few differences between A-ARMA and SODA also confirm that the
reduction of data quantity by spatial packet merge is not significant. In fact, spatial
packet merge targets a reduction in the communication overhead, including packet
header and MAC overhead, related to channel access. It will be further discussed in
section 4.4.2.2.
4.4.2.2 Active time during data collection
In case of intensive data communication, the behavior of MAC layer has impacts on the
active time of the radio transceivers. Therefore, the analysis of active time is divided
into two parts:
1. In the first part, we assume an ideal MAC scheduling. In case of unicast trans-
mission, only the sender and the receiver are in active model, while the other
nodes do not consume energy. In case of broadcast, the sender and all neighbors
of sender are in active mode.
2. In the second part, the BMAC [59] scheduling is used. As already presented, it
is a distributed MAC scheduling with preamble technique which does not require
97
Data Dissemination and Data Aggregation
280 300 320 340 360 380 4000
2000
4000
6000
8000
10000
12000Average active time with ideal MAC scheduling
Number of nodes
Rad
io a
ctiv
e tim
e (m
s)No aggregationTemporal packet mergeSingle ARMA aggregationSODA frameworkSpatial packet mergeA−ARMA
Figure 4.14: Average active time of sensor nodes in data collection (ideal MAC scheduling)
synchronization in the network. The simulation results over BMAC may show en-
ergy saving provided by different data aggregation technique in a realistic channel
access technique.
Based on ideal MAC scheduling In the case of ideal MAC scheduling, we assume
a perfect scheduling in which all packets are correctly received without collision and
retransmission. There is no additional channel occupation as in CSMA/CA mechanism.
Fig. 4.14 gives the average active time of sensor nodes of different techniques when
ideal MAC scheduling is assumed. The use of A-ARMA allows nodes to significantly
reduce their active time. It also shows that temporal packet merge is efficient for re-
ducing active time. Generally speaking, temporal data aggregation reduces the data
rate of each sensor node before the data load is generated in the network. The quantity
of information is reduced before being injected in the network. The efficiency of single
ARMA aggregation (technique of [73]) is not good, mostly related to the long prelimi-
nary phase where no aggregation is used. The spatial packet merge technique does not
provide a significant reduction on active time, because only the time for transmitting
packet headers are saved.
The advantage of SODA framework with ideal MAC scheduling is not very obvious
comparing to A-ARMA, because the improvement brought by spatial packet merge is
very restrained. Nevertheless, SODA provides a flexible and adaptive data aggrega-
tion structure in the WSN, whose maintenance cost is lower than flooding tree based
structure.
98
SODA Framework
280 300 320 340 360 380 4000
50
100
150
200
250
300
350
400
450
500
550Active time with BMAC
Number of nodes
Act
ive
time
(s)
No aggregationTemporal packet mergeSingle ARMA aggregationSODA frameworkSpatial packet mergeA−ARMA
Figure 4.15: Average active time of sensor nodes in data collection (BMAC)
Based on BMAC Fig. 4.15 gives the average active time when BMAC [59] is used.
The SODA framework achieves once again the lowest active time among all techniques.
Comparing to Fig. 4.14, the difference between SODA and A-ARMA is more significant.
And the spatial packet merge achieves a much better performance. Both observations
confirm that spatial packet merge has significant impact when a realistic MAC protocol
is used. It is efficient to save the channel access time. Relying on BBDD structure, a
leader relays one packet in its report cycle instead of many short packets. The channel
access is therefore lightened.
It also shows that all curves slightly increase along with the number of nodes, while
they are quasi stable in Fig. 4.14. In no aggregation, temporal packet merge, single
ARMA aggregation and A-ARMA techniques, it is related to the high density of nodes.
In average, more activities are generated around each node, and nodes are involves in
more preambling detections and channel accesses. In case of spatial packet merge and
SODA, the number of member nodes per leader increases according to the FISCO local
structure properties. Around each leader; the channel access becomes busier. In both
cases, nodes spend more time in active state to deal with medium access. However,
with ideal MAC scheduling, the access is assumed perfect, therefore the impact is not
obvious.
In order to analyze this amplification effect of BMAC on different techniques, the
factor of active time between BMAC and ideal MAC scheduling are shown in Fig.
4.16. Since there is not signal checking state in ideal MAC scheduling, the factor is
established between the sum of sending and receiving time. The active time of all
99
Data Dissemination and Data Aggregation
300 310 320 330 340 350 360 370 38035
40
45
50
55
60
65
70
75
80
85Impact of BMAC
Number of nodes
Fac
tor
Tx+
Rx
time
with
/with
out B
MA
C
No aggregationTemporal packet mergeSingle ARMA aggregationSODA frameworkSpatial packet mergeA−ARMA
Figure 4.16: Factor of active time between BMAC and ideal MAC scheduling
schemes are amplified by BMAC, while its impact varies on different techniques. The
effects of amplification are more significant on no aggregation, single ARMA aggregation
and temporal packet merge than on SODA framework, A-ARMA and spatial packet
merge. As a conclusion, both the design of BBDD structure and A-ARMA are more
advanced than other techniques when using CSMA based MAC mechanisms.
Fig. 4.17 gives further detailed distributions of transmitting, receiving and signal
checking modes among total active time in each technique using BMAC. SODA frame-
work has the lowest transmitting and receiving time ratio among all schemes. In case of
no aggregation and single ARMA aggregation, the ratio of receiving time is dominant,
because the data load in the network has not been efficiently reduced. The ratio of
receiving time of spatial packet merge represents 50% of active time, because of the
transmissions between member nodes and leader nodes.
4.4.2.3 Energy consumption during data collection
Fig. 4.18 evaluates the average energy consumption on nodes during data collection.
Almost the same observation is obtained as for average active time (Fig. 4.15). How-
ever, the impact of node number is less significant, especially for SODA, temporal
packet merge and A-ARMA. It is because that our energy model is not only based on
a linear combination of different active time and sleep time, but also takes into account
the signal checking time during which the power is low. According to the distribution
of active states in Fig. 4.17, the three techniques achieving a higher ratio on signal
100
Conclusions
300 320 340 360 3800
0.2
0.4
0.6
0.8
1No aggregation
300 320 340 360 3800
0.2
0.4
0.6
0.8
1Temporal packet merge
300 320 340 360 3800
0.2
0.4
0.6
0.8
1Single ARMA aggregation
300 320 340 360 3800
0.2
0.4
0.6
0.8
1SODA framework
300 320 340 360 3800
0.2
0.4
0.6
0.8
1Spatial packet merge
300 320 340 360 3800
0.2
0.4
0.6
0.8
1A−ARMA
Transmitting
Receiving
Singal Checking
Figure 4.17: Distribution of active modes among total active time
checking time are less impacted.
The advantage of using an adaptive technique in data aggregation is clearly stated
when comparing single ARMA and A-ARMA technique. Energy is saved every time
a set of new samples is not sent after local verification by A-ARMA. This amount of
energy includes the transmission by the source node and the transmissions/receptions
generated during the multi-hop routing or reporting via BBDD structure. Different
from the single ARMA technique, A-ARMA provides energy savings right after deploy-
ment and throughout the entire lifetime of the WSN. Therefore, it saves about 50% of
the energy compared to non adaptive ARMA solutions.
Once more SODA framework achieves the lowest energy consumption. And it con-
sumes around 80% less energy compared to no aggregation case. This result confirms
that SODA framework is not only a highly adaptive concept but also very promising
for energy saving. Therefore, SODA framework is suitable to support WSN monitoring
applications.
4.5 Conclusions
Data dissemination and data aggregation are two important paradigms in WSNs. In
this chapter, we proposed a framework consisting of data dissemination, management of
101
Data Dissemination and Data Aggregation
300 310 320 330 340 350 360 370 380 390
20
40
60
80
100
120
140
Average energy consumption
Number of nodes
Ene
rgy
cons
umpt
ion
(J)
No aggregationTemporal packet mergeSingle ARMA aggregationSODA frameworkSpatial packet mergeA−ARMA
Figure 4.18: Average energy consumption on nodes
multiple mobile sinks and in network data aggregation techniques. All these techniques
are developed around a strong idea: the use of autonomous network architecture FISCO
simplifies the mechanisms for data communication and improves their efficiency. More
generally, the idea we defend in this chapter is that self-organization provides a basis on
which data communication techniques can be easily and efficiently developed. Through
analysis and simulation results, we have clearly shown that taking advantage of self-
organization allows significant savings in terms of communication overhead and energy
consumption for data communication in WSNs.
The SODA framework is a complete framework which provides a solution from
WSN deployment to the intensive data communication in WSN applications. The self-
organization is the basis of this framework, which brings order and incorporates sensor
nodes in a structure. It allows sink nodes to easily initiate data-centric structures.
The generation and maintenance of the data-centric structures are the cores of data
dissemination mechanisms. Two types of maintenance are addressed in SODA frame-
work: the maintenance related to nodes’ dynamics is achieved within self-organization,
while the maintenance related to sinks’ dynamics is particularly discussed as the man-
agement of multiple mobile sinks over backbone based rendezvous system. Once more,
the self-organization structure ensures the effectiveness and efficiency of this manage-
ment.
We also propose an A-ARMA technique based on forecasting by means of an ARMA
102
Conclusions
model over moving windows. This technique is designed to achieve three objectives:
low complexity, good accuracy and high aggregation efficiency. The A-ARMA tech-
nique reduces the energy consumption by up to 50% compared to a non-adaptive one.
Although this technique may work without the presence of self-organization, the SODA
frame reduces the route maintenance cost and control overhead in the reporting and
provides a adaptive and flexible data centric structure which facilitate data aggregation.
The concepts and results presented in this chapter are based on FISCO tree back-
bone. We note that FISCO mesh backbone boasts several advantages compared to tree
backbone when data communication is considered: multiple data paths for lowering the
load on each path; shorter path length in the network. From the angle of data com-
munication, this is recommended. However, the control overhead of mesh backbone is
higher than tree backbone, because of more backbone nodes which should permanently
provide control information for organization and maintenance. In the near future, we
plan to quantify the impact of mesh backbone characteristics on the total cost of the
SODA framework.
103
Data Dissemination and Data Aggregation
Publications
International conference
[IC1] Jia-Liang LU and Fabrice Valois. On the data dissemination in WSNs. In
Proc. of the 3rd Int’l Conference on Wireless and Mobile Computing, Networking
and Communications (Wimob), White Plains, USA, October 2007, IEEE.
104
Entropy of Organization 105
Entropy of Organization 55.1 Introduction
In the two previous chapters 3 and 4, we discuss how to design an autonomous net-
work architecture and its advantages. Providing energy saving, using low complexity
algorithm and achieving autonomous operations are stated as the goals of our work. In
this chapter, we focus on the intrinsic properties of such autonomous logic architecture,
more precisely on the organization of a WSN. The question which we address here is:
what is a good metric for evaluating the degree of organizations generated in a network
according to different strategies of self-organization?
Several metrics are used to evaluate the self-organization strategies, which have also
been used in this work, such as:
• the protocol cost, which is generally measured as the quantity of control overhead
[80];
• the complexity of the algorithm including the memory and message complexity
[4];
• the self-stabilization properties such as the convergence time to a stable organi-
zation [81].
However, these metrics are not proposed from the angle of self-organization, and they
do not reflect how good or bad an organization in the network is.
To this end, entropy is a well established notion to characterize order in self-organized
system. It has mainly three definitions: thermodynamic definition [82], statistical
thermodynamic definition [83] and information entropy (today known as Shannon
entropy)[84]. We believe that the organization exhibits microscopic and macroscopic
levels just as thermodynamic systems do. At microscopic level, the organization is
Entropy of Organization
based on logic links, which specify local interactions between nodes: a node can com-
municate with one neighbor (existence of a logic link) or not (non-existence of a logic
link) according to a self-organization scheme. At macroscopic level, all these local in-
teractions form a virtual structure which is flexible and adaptive. Therefore, we adopt
in this work the statistical thermodynamic definition of entropy as a new metric to
evaluate the organization in ad-hoc type networks (including WSNs).
The contributions of this chapter are two-fold:
1. We are able to evaluate the intrinsic property, the order, of a network, by means
of this new entropy based metric.
2. Through the evaluation of flat and other organizations, we provide answers to
whether it is worth to organize wireless ad hoc and sensor network before running
other mechanisms.
The rest of this chapter is organized as follows: Section 5.2 reviews the original
definitions of entropy and some related works. In section 5.3, we give the formulation of
entropy to characterize the degree of the order in wireless ad hoc and sensor networks.
The first evaluation results are given in section 5.4 with the interpretations base on
entropy formulation. Section 5.5 shows the results collected on entropy variation in
different scenarii. Finally, section 5.6 opens a discussion on the evaluation of self-
organization.
5.2 Original Definitions of Entropy
The notion of entropy was first introduced by Clausius [82] as a unique measure for the
process of a reversible change in thermal energy with respect to the absolute temper-
ature. He focused on the macroscopic behavior of microscopic chemical reactions and
proposed the thermodynamic entropy. Based on this, Boltzmann defined the statistical
entropy [83] from combining microstates (giving rise to the different configurations of
a system) as S = −kB∑
i pi ln pi, where pi is the probability that a microstate i occurs
during the system’s fluctuations and kB being the Boltzmann constant; it is applica-
ble to characterize the order in the system and how the system self-organizes among
different entities.
Later, Shannon introduced the concept of information entropy H (today known as
Shannon entropy) [84]. It relates to the representation but not contents of information
by quantifying its uncertainty. This notion was used by Shannon himself to quantify
the capacity of a transmission channel and has also been extended to other problems
106
Extended Definition of Entropy
and domains: the Shannon entropy has been used in complex systems and social prob-
lems [85]; also, in [86], a social entropy has been introduced to measure the robot
group diversity (self-organization in a robotic team system); etc. Above and related
approaches, however, use information entropy as a local metric which helps guiding
self-organization to achieve optimal local decisions w.r.t. to some metrics [87, 88].
Contributions extending these concepts to describe emergent system behaviors are
scarce. In [89], the authors metaphorically describe the conditions under which coordi-
nation can emerge from individually autonomous actions in a multi-agent system; i.e.,
emergent coordination can arise through coupling the macro level (a concomitant de-
crease in entropy) to an entropy-increasing process at a micro level such as pheromone
evaporation.
5.3 Extended Definition of Entropy
In our opinion, the organization in ad-hoc type networks is similar to a thermal dynamic
system. Both of them exhibit a macroscopic behavior which is induced by microscopic
local interactions. Therefore, we adopt the definition of entropy used in statistical
thermodynamics to quantify the order in a network according to the node distribution
and channel properties under different organizational schemes. In our opinion, it is a
key metric because it exhibits the susceptibility of self-organization protocols to network
inherent parameters, such as link and node reliability.
5.3.1 Formulation of entropy
Entropy is a macroscopic description of a system taking into account microscopic in-
teractions. In ad hoc and wireless sensor networks, the use of self-organization schemes
restricts local interactions. For example, when a local minimum spanning tree (LMST)
[49] organization is applied, a node does not communicate with some of its neighbors;
the interactions among these nodes are eliminated. In order to quantify the impact
of self-organization schemes on the internal organization of a network, we apply the
concept of entropy to the links. Let us assume that a link exists between nodes u and
v with probability p(u, v); then the state of organization, i.e. the entropy of this link is
given as E = −p(u, v) log p(u, v). High values of E indicate high disorders; low values
of E high orders, i.e. a better organization.
For p(u, v) being close to zero, the entropy is close to zero; this indicates that this link
is well organized even though virtually absent; however, it is almost deterministically
absent. For p(u, v) being close to one, the entropy is also close to zero; this also indicates
107
Entropy of Organization
that this link is well organized because virtually deterministically present. Intermediate
values of p(u, v) yield larger disorder because the link may or may not be present.
Many factors have impacts on the link probabilities: the topology, the wireless chan-
nel quality, the nodes presence, the organization used, etc. We use the following model
to compute link probabilities.
1. The physical topology denoted as G(X, r). It is composed of a set of nodes X,
the communication radius r and the position of nodes in the network. In order
to facilitate the calculation, we assume that the network is a square of size 1× 1,
and the communication radius is a value smaller than 1. Therefore each node is
at a position (x, y) where x, y is between 0 and 1. Based on above information,
all possible radio links can be obtained.
2. The probability that a node appears in the network, denoted as pu and hence the
probability that it does not exist is qu = 1 − pu
3. The wireless channel is assumed to be in a good state with probability cuv, where
u and v denote the two end points of a link. We assume that the channel for each
link is independent, in order to facilitate the computation. In a real environment,
there are certain dependencies among these probabilities. However, it is highly
related to physical and radio interface adopted in the network, which is beyond
the scope of our work.
4. Based on the probability in 2. and 3., the probability that a link uv exists in the
network is p(uv) = cuvpupv
We are particularly interested in two properties of organization in ad hoc and wireless
sensor networks. The first one is the order in a network. A highly ordered network
should achieve a low entropy value which is similar to thermodynamic system. The
result of a self-organization scheme is a logic structure represented by the set of all
logic link states in the network. Therefore by summing the entropy of each link, we
are able to tell which organization achieves a better order. The global entropy of the
network is henceforth defined as:
E =∑
u,v∈X
−p(u, v) log p(u, v). (5.1)
The second property is the adaptability of a scheme during a change. A change is
a modification on physical topology, such as a link disconnection or a node departure.
When a change is introduced in the network, a self-organization automatically adapts
108
Extended Definition of Entropy
the organization to make up for the change. The variation of entropy before and after
a change in the network quantify efforts that the network spent according to the self-
organization scheme. The entropy variation is henceforth defined as:
∆E = Echange− − Echange+. (5.2)
5.3.2 Interpretations of this formulation
The phenomenon of disorder is related to interactions between the elements of a system.
In wireless ad hoc and sensor networks, nodes are considered the basic elements of the
network and links are the interactions. From the point of view of organization, the links
considered by an organization are no longer the physical capability of communication,
but the choices of the organization. Our formulation takes the statistical properties
of local interactions into account to get a global measurement of the organization of
a network. Hence, the formulation of entropy for ad hoc and wireless sensor network
conforms to its definition.
It is worth noting that this formulation allows us to give an idea on the order in a
network from an angle of organization. Although it does not allow us to draw con-
clusions on network performance such as throughput or packet deliver ratio, there are
some correlations between the entropy and these metrics. We discuss it later in section
5.4.
The entropy is not only related to the network topology and the dynamics of the
network, but also a result of self-organization schemes used in the network. It is the
self-organization schemes that bring order in the network. Hence an evaluation of
entropy makes sense only if it is applied on a network with a level of dynamics and a
running self-organization scheme. It is worth noting that our formulation of entropy is
a statistical approach, which does not address a particular resulting structure of self-
organization, but the whole set of possible resulting structures under a specific network
dynamics.
5.3.3 Application of entropy on a simple network
Let us now apply this concept. To this end, we observe that the existence of a link is
the result of:
• two independent random effects, i.e. the existence of the nodes on either end of
the link, and the availability of the wireless channel leading to a random network
state, and
109
Entropy of Organization
4
pq2 2
31q p
(3)
(2)(6)(5)
(4)
(1)
p
Figure 5.1: Flat organization.
• one deterministic effect, i.e. the way a self-organization scheme organizes the
network for a given realization of the random network state.
Let us now calculate the entropy per link and, then, per system for a flat and a
LMST [49] organization scheme, assuming a simple illustrative case as per Figures 5.1
and 5.2. The link existence probability equals to p (∀u, pu = p) and the probability of
non-existence of a link is q = 1 − p.
5.3.3.1 Flat organization scheme
As per Figure 5.1, the probability that link (1) exists is p1 = c ·(
p4 + 2qp3 + q2p2)
= c · p2.
The remaining five link probabilities (pi) can similarly be calculated, all of which lead to
the same above-given expression. The state of organization (entropy) per link is hence
−p1 log p1, which gives the entire system entropy as E = −∑6i=1 pi log pi = −6·p1 log p1.
Let us now assume that m links exist in a flat network composed of N nodes. The
value of m depends on the communication radius and the geographical distribution of
nodes. Given a density of the network ρ, we obtain m as m = ρ · N/2. The entropy of
a network using flat organization is the sum of the entropy of all links; we hence have
E = −m · cp2 · log(
cp2)
.
5.3.3.2 LMST self-organization scheme
LMST [49] is a link pruning based self-organization scheme which provides a logic
structure with fewer links than those in the flat physical structure. All nodes are kept
in the resulting structure. Each node locally computes its two-hop minimum spanning
tree (MST). For two neighborhood nodes u and v, if u is in the MST of v and v is in
the MST of u, then the link (u, v) is in the LMST: it is kept active in the network,
whereas the rest of links are dropped in the organization.
110
Extended Definition of Entropy
4
q2 2
31q p
p
p
Figure 5.2: LMST organization.
0 0.2 0.4 0.6 0.8 10
0.5
1
1.5
2
2.5
Probability of Node Existence p
Ent
ropy
E
LMSTflat
(a) Flat versus LMST for varying p; c = 1.
0 0.2 0.4 0.6 0.8 10
0.5
1
1.5
2
2.5
Probability of Channel Existence c
Ent
ropy
E
LMSTflat
(b) Flat versus LMST for varying c; p = 0.8.
Figure 5.3: Entropy value of a topology of 4 nodes
As per Figure 5.2, the probability that links (1)-(4) exist is p1 = c ·(
p4 + 2qp3 + q2p2)
and the probability that links (5) and (6) exist is p5 = c·q2p2. The entire system entropy
is hence E = − [4 · p1 log p1 + 2 · p5 log p5].
Without depicting the analytical entropy results of example topologies given in Fig.
5.1 and 5.2, the following can be observed in Fig. 5.3(a) and 5.3(b). LMST has
a lower entropy over all state probabilities and hence a higher degree of organization.
Furthermore, for low (p → 0) or high (p → 1) node existence probabilities, the difference
between both schemes is negligible. For more uncertain configurations, it is better to
use LMST. It can further be observed that for nodes occurring with increasing likelihood
(p → 1), the channel plays an increasingly important role.
111
Entropy of Organization
5.4 Entropy Evaluation
Subsequent results are obtained for four different self-organization protocols, namely
LMST, Relative Neighborhood Graph (RNG) [48], Connected Dominating Set (CDS)
[4], Minimum Independent Set based CDS (MIS-CDS) [46]. For the latter two, the
following rules are used for the links selection: i) all the links between a dominant and
its dominated are kept active, ii) we keep only one active link to connect a dominated
to only one dominant. The self-organization part of FISCO is also abstracted as the
fifth protocol. Flat organization is also evaluated as a network without being applied
any self-organizations.
The key to this evaluation is to obtain link probabilities, and hence the total entropy.
For a general network topology, this evaluation is analytically infeasible. Furthermore,
an exhaustive summation over all possible link formations is NP-complete, which re-
quires us to use simulations.
To this end, we generate a network with N nodes and let them appear with probability
p in the network. Then we apply the probability c on the existing links. On this resulting
graph, we apply for example LMST and get a subgraph, say GLMST (X, r, p, c). We
record the existing links for the particular network realization. This is then repeated
for the same topology. After sufficient runs, we obtain the average probability that
a link exists in the organization. According to equation (5.1), the entropy for this
topology can be calculated. Finally, a set of above procedures are repeated to obtain
an entropy value which is representative for the given self-organization protocol. The
standard deviations of the simulation results are within 5 % of their mean values and
hence statistically significant.
Fig. 5.4(a) compares entropy results for flat, LMST, RNG, CDS, MIS-CDS and
FISCO schemes for N = 200 nodes and a communication radius of r = 0.16. The
use of LMST/RNG/CDS/MIS-CDS/FISCO decreases the entropy and hence increases
the degree of organization in the network, under any node probabilities. This in turn
requires less links to be reconfigured and hence less energy/overhead spent in the case
of network changes. It hence complies with the objective of using self-organization, i.e.
increasing order and reducing disorder in the system.
According to our definition of entropy, LMST and RNG provide a higher order than
CDS, i.e. have a higher level of determinism in their local algorithms. This is due to
LMST and RNG keeping fewer possible interactions in their resulting logic structure,
whereas all interactions between dominating nodes in CDS are kept. We also observe
that LMST and RNG are less sensitive to changes in link existence probabilities, i.e. a
change in p will cause less change in entropy than for flat and CDS topologies. MIS-
112
Entropy Evaluation
0 0.2 0.4 0.6 0.8 10
50
100
150
200
250
300
350
400
450
500
550Entropy for 200 nodes with r=0.16
Node probability
Ent
ropy
flatLMSTRNGCDSMIS−CDSFISCO
(a) Entropy for varying p; c = 1.
0.13 0.14 0.15 0.16 0.17 0.18 0.19 0.2 0.210
100
200
300
400
500
600
Entropy for 200 nodes with p=0.8
Radius
Ent
ropy
flatLMSTRNGCDSMIS−CDSFISCO
(b) Entropy for varying r; p = 0.8.
Figure 5.4: Average entropy value on networks of 200 nodes, radius=0.16
CDS achieves a lower entropy value on low node probability but higher value than
LMST/RNG on high node probability. This is because MIS-CDS is generated via a
rooted tree algorithm, which is less efficient in a dynamic network (probability of p
between 0.3 to 0.9).
We obtain a particularly interesting observation on FISCO: its entropy value with
very low node probability (p = 0.1) is a little bigger than in the flat case. However,
it grows more slowly than any other schemes as the node probability increases. It
was discussed in chapter 3 that FISCO works better in a dense network than a spare
network. In the latter case, it is common that several disconnected partitions exist
in the network. And this organization is not stable. That is why statistically these
networks exhibit a high value of entropy, a low degree of order. With the increase of
node probability, the density of the network increases as well. Hence the entropy value
of FISCO begins to show its better performance at ordering the network than other
algorithms.
Fig. 5.4(b) shows the evolution of entropy with different radius for the same number
of nodes and constant probabilities. The entropy of flat and CDS organization increases
almost in a linear fashion; however, all other organizations seem insensitive to radius
variation in terms of order. This can be explained as follows:
• In the flat organization, with the increasing radius, nodes have more neighbors.
Hence the number of microscopic states of local interactions increases. This makes
the global network entropy increase as well.
• The cardinality of dominating backbone in CDS increase with the density, while
113
Entropy of Organization
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.10
20
40
60
80
100
120
140
160
180
200Number of transmitted messages in a broadcast
Node probability
Num
ber
of m
essa
ges
flatLMSTCDS
(a) Transmission cost in a broadcast
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.10
500
1000
1500
2000
2500
3000Number of received messages in a broadcast
Node probability
Num
ber
of m
essa
ges
flatLMSTCDS
(b) Rception cost in a broadcast
Figure 5.5: Average broadcast cost on networks of 200 nodes, radius=0.16
the local interactions within the backbone are not constrained in CDS. Therefore,
the existence of more local interactions in the backbone increases the disorder in
the network.
• In LMST/RNG/MIS-CDS/FISCO, the local interactions are more constrained by
the algorithm running on every node. Fewer interactions are obtained in resulting
structures. Therefore, despite varying radii, almost constant entropy is observed.
Although the number of logic links in MIS-CDS are small (even smaller than with
LMST and RNG), the entropy of this organization is bigger than LMST and RNG’s
in Fig. 5.4(b). This is because the algorithm MIS-CDS is not completed localized.
All algorithms used in MIS-CDS (rooted tree construction, dominating node selection
and CDS connecting) are distributed but necessarily initiated by a node. When the
network is relatively dynamic (p=0.8), statistically MIS-CDS is less efficient at limiting
the interactions to a small range. This yields a lower level of order in the network.
According to Fig. 5.4(b), FISCO achieves the lowest entropy among all self-organization
schemes. The high order in FISCO is related to both FISCO backbone and the local
structures. Different from CDS, the FISCO backbone is a tree-like backbone. The
interactions between backbone nodes are highly constrained. In the local structures,
each member node is configured to have only one possible interaction with its leader
node. For both reasons, the network is highly ordered in FISCO.
When we proposed the formulation of entropy, the primary objective was to create
a measurement with which the order in a network can be evaluated. We do not expect
to draw conclusions on network performance, yet there might be some correlations
114
Entropy Variation Evaluation
between entropy and other metrics. We show the performance of running broadcast on
the top of different organization in a network. The transmission cost is given in Fig.
5.5(a) and reception cost in Fig. 5.5(b). The broadcast is considered because it is a
mechanism that involves the participation of all nodes in the network. Flat, LMST
and CDS organization are evaluated. LMST represents link-pruning based approach
and CDS represents dominating based approach. It turns out that both transmission
and reception costs of LMST are smaller than that on CDS and flat organization under
same conditions. This is consistent to the observations on entropy in Fig. 5.4(a).
5.5 Entropy Variation Evaluation
Although the entropy is a global measurement from our point of view, the entropy
variation may quantify the increase or decrease of the orders in a network after a
change. It hence allows us to characterize the adaptation cost based on the changes of
local interactions. We discuss two scenarii of network changes: in the first one, a single
node disappears from the network; and in the second one, multiple nodes’ disappearance
is applied. It is worth noting that appearance and disappearance of nodes can be
considered a reversible transformation in a network. It is obvious that for the same
node disappearing from or appearing at the same position, we have ∆Eappearance =
−∆Edisappearance.
5.5.1 When single node disappears
First, we discuss the entropy variation during a network change: a single node disap-
pears. To this end, a network of N = 200 and r = 0.16 is evaluated. We are interested
in the entropy variation due to such a change: at instance T0 all nodes work in the net-
work, while at T1, one node disappears from the network. Flat, LMST, RNG and CDS
organizations are applied to this network. We simulate the disappearance of each node
for T0 → T1 (for all 200 nodes). It is worth noting that this change is reversible. We
can inverse the instance as T1 → T0 which models the arrival of nodes in the network.
The following entropy value is performed only for the networks which are still con-
nected (198 cases among 200 in the simulation) after the single node disappearance.
Figures 5.6(a) - 5.6(d) give the probability density function of entropy variation
values for single node disappearance. In a flat organization (Figures 5.6(a)), the
entropy variation is all positive and distributed between 0 and 9. It shows that even
in a flat organization, the disappearances of some nodes lead to a bigger difference in
entropy than in other organizations. Intuitively, nodes with high degree have more
115
Entropy of Organization
−6 −4 −2 0 2 4 6 8 10 120
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
Entropy flat
Pro
bab
ility
Den
sity
Fu
nct
ion
PDF(X) flat
(a) Pdf flat
−6 −4 −2 0 2 4 6 8 10 120
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
Entropy CDS
Pro
bab
ility
Den
sity
Fu
nct
ion
PDF(X) CDS
(b) Pdf CDS
−6 −4 −2 0 2 4 6 8 10 120
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
Entropy LMST
Pro
bab
ility
Den
sity
Fu
nct
ion
PDF(X) LMST
(c) Pdf LMST
−6 −4 −2 0 2 4 6 8 10 120
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
Entropy RNG
Pro
bab
ility
Den
sity
Fu
nct
ion
PDF(X) RNG
(d) Pdf RNG
Figure 5.6: Probability Density Function of entropy variation for simple node disappearance
influence on the flat organization. The CDS scheme (Figures 5.6(b)) has the widest
distribution among all schemes (4 times wider than LMST and RNG). It is because the
importance of nodes in the organization is amplified by the role assignment in CDS. It
has also the most negative values of entropy variation. It means that when some nodes
disappear, although CDS tries to adapt the organization to reach another equilibrium,
the network is not as well ordered as it was before.
According to Figures 5.6(c), 5.6(d), both LMST and RNG spend a relative small
cost to re-organize the network (locally). They adapt much easier to single node dis-
appearance. The impact of single node disappearance on the network organization is
reduced by LMST and RNG.
116
Entropy Variation Evaluation
0 5 10 15 20 25 300
10
20
30
40
50
60Entropy variation for 100 nodes (r=0.16)
Number of lost nodes
Ent
ropy
var
iatio
n
flatLMSTRNGCDSMIS−CDSFISCO
(a) 100 nodes
0 5 10 15 20 25 300
10
20
30
40
50
60
70
80
90Entropy variation for 150 nodes (radius=0.16)
Number of lost nodes
Ent
ropy
var
iatio
n
flatLMSTRNGCDSMIS−CDSFISCO
(b) 150 nodes
0 5 10 15 20 25 300
10
20
30
40
50
60
70
80
90
100Entropy variation for 180 nodes (radius=0.16)
Number of lost nodes
Ent
ropy
var
iatio
n
flatLMSTRNGCDSMIS−CDSFISCO
(c) 180 nodes
Figure 5.7: Entropy variation of self-organization schemes for random geometric network, ra-dius=0.16
5.5.2 When several nodes disappear
We evaluate the entropy variation as a function of the number of disappeared nodes.
We apply a random node disappearance for T0 → T1. One analysis is given per network.
It is worth noting that the connectivity is conserved after node disappearance.
Random geometric networks of different densities are studied. Four networks con-
taining respectively 100, 150, 180 nodes with radius of 0.16 are evaluated. The results
are collected with 500 simulations for each value. The following properties are observed:
1. In all organizations, the entropy variations increase in a linear fashion when more
nodes disappear from the network.
117
Entropy of Organization
2. LMST and RNG schemes adapt better (i.e. with lower variation) to the node
disappearance than other schemes do.
3. MIS-CDS and CDS schemes adapt better than flat organization, but are clearly
not as efficient as LMST, RNG or FISCO schemes.
4. Although the MIS-CDS scheme is a tree structure and its organization has fewer
logic links than other organizations, it does not achieve the lowest entropy vari-
ation. This signifies that there is no direct relation between the number of links
in the organization and the quality of organization.
5. FISCO achieves the lowest entropy variation among all schemes. However, the dif-
ference between FISCO and LMST or RNG deceases with the increase in network
density. The density of LMST and RNG resulting structure are quasi constant,
while the degree of nodes in FISCO is not uniform. Although the same amount of
entropy variation is generated for a member disappearance whatever the network
density, the disappearance of a FISCO leader is most costly in dense network than
sparse network. On one side the entropy variation for LMST and RNG is sta-
ble, one the other side it increases for FISCO. This explains why their difference
decreases in dense network.
6. We also observe that the gain of CDS and MIS-CDS compared to flat organization
are not significant when the density of the network is low.
In order to show the impact of density on the entropy variation of each scheme, Fig.
5.8(a) - 5.8(e) compare each scheme under different densities.
In order to vary the network density, we set the radius and increase the number of
nodes in the network. Fig. 5.8(a)-5.8(e) show that various densities induce different
values of entropy variation for self-organization schemes.
1. In flat organization, the entropy variation grows (w.r.t. the number of disappeared
nodes) much faster with high-density network than in low-density network. This
is because the disappearance of a node in a dense network involves more physical
links, and thus more local interactions.
2. However, the situations are opposite in LMST, RNG and especially in MIS-CDS:
removing nodes from a sparse network cause bigger changes to these organizations.
The reason behind this is that a node disappearance has much more impact on the
organization of sparse network which has small number of logic links, especially
for a tree structure as MIS-CDS.
118
Entropy Variation Evaluation
0 5 10 15 20 25 300
10
20
30
40
50
60
70
80
90
100Flat entropy variation (r=0.16)
Number of lost nodes
Ent
ropy
var
iatio
n
flat 180 nodesflat 150 nodesflat 100 nodes
(a) EV of flat organization
0 5 10 15 20 25 300
5
10
15
20
25
30
35LMST entropy variation (r=0.16)
Number of lost nodes
Ent
ropy
var
iatio
n
LMST 180 nodesLMST 150 nodesLMST 100 nodes
(b) EV of LMST organization
0 5 10 15 20 25 300
5
10
15
20
25
30
35RNG entropy variation (r=0.16)
Number of lost nodes
Ent
ropy
var
iatio
n
RNG 180 nodesRNG 150 nodesRNG 100 nodes
(c) EV of RNG organization
0 5 10 15 20 25 300
10
20
30
40
50
60MIS−CDS entropy variation (r=0.16)
Number of lost nodes
Ent
ropy
var
iatio
n
MIS−CDS 180 nodesMIS−CDS 150 nodesMIS−CDS 100 nodes
(d) EV of MIS-CDS organization
0 5 10 15 20 25 300
10
20
30
40
50
60
70CDS entropy variation (r=0.16)
Number of lost nodes
Ent
ropy
var
iatio
n
CDS 180 nodesCDS 150 nodesCDS 100 nodes
(e) EV of CDS organization
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
18FISCO entropy variation (r=0.16)
Number of lost nodes
Ent
ropy
var
iatio
n
FISCO 180 nodesFISCO 150 nodesFISCO 100 nodes
(f) EV of FISCO organization
Figure 5.8: Entropy variation of self-organization schemes under different densities
119
Entropy of Organization
3. For FISCO, the entropy variation is more impacted by the local structures. As
already explained, the disappearance of a leader involves more changes within
its local structure in dense network. Hence the entropy variation is higher in
high-density network.
4. Finally, the difference between the entropy variations of CDS ([4]) under different
density is not significant.
It is also worth noting that all self-organization schemes reduce the difference be-
tween the curves under different densities. It confirms that the use of self-organization
generally improves the adaptability of the network in the face of changes. Particu-
larly, FISCO achieves the lowest entropy variation among all self-organization schemes,
thanks to its efficient departure procedures (section 3.4). The departure procedures of
FISCO try to recover the FISCO backbone based on the information of existing orga-
nization. Intuitively, nodes try to keep the organization as close as possible to what
was before the disappearance of nodes.
5.6 Conclusions
Self-organization, which aims at generating order in a network, guides the network to
maintain certain properties and react to topology changes. We extended the metric
based on the notion of entropy to evaluate the organizational state of a network which
has then been applied to various known self-organization schemes. The utilization of
this entropy based metric clearly shows the impact of self-organization schemes on the
internal organization of the network. To our knowledge, this work is the first to evaluate
a network from an organization perspective.
This approach leads to the following quantitative and qualitative insights into the
behavior and design of self-organization protocols:
1. First, any of the chosen protocols yields a higher organizational state than a flat
topology.
2. Second, for an increasing likelihood of a node being available, the channel com-
mences to play a dominant role in the organizational state of the network.
3. Third, LMST and RNG yield the same level of organization, despite being based
on fairly different rules.
4. Fourth, FISCO is well organized when the network is not sparse.
120
Conclusions
5. Finally, whilst flat and CDS strongly depend on the communication radius, LMST,
RNG and FISCO exhibit a high degree of organizational stability.
It is worth noting that the results on entropy evaluation confirms the impacts of
FISCO autonomous architecture and its related self-organized mechanisms onto WSNs.
When setting its local interactions, a node takes into account the information on the
existing organization. Hence the interactions are more constrained compared to deci-
sions that would be purely based on physical neighborhood information. As a result,
the network exhibits a higher order.
Although this entropy based metric can not be directly applied to compute network
performance, in section 5.4, we did show some correlations between entropy and message
cost in a broadcast. This a research topic that we are continuing to investigate. We
expect to collect other network performance results such as energy cost and packet
deliver ratio, and compare with the results of entropy to exhibit more correlations.
121
Entropy of Organization
Publications
International journal
[IJ1] Jia-Liang LU, Fabrice Valois, Mischa Dohler and Dominique Barthel. Quan-
tifying organization by means of entropy. To appear in IEEE Communication
Letters, March, 2008.
122
Test-bed 123
Test-bed 6In this chapter, we describe our implementation of FISCO on a test-bed set up at
FranceTelecom Research and Development Beijing Lab. The motivation of this test-
bed is two-fold:
1. Firstly, the test-bed is used to validate the autonomous architecture that we
proposed in this work. Some properties of the architecture have been proven
and illustrated through theoretic analysis and simulations. Nevertheless, these
analysis and simulations are obtained based on a set of assumptions. The imple-
mentation of our solutions on a WSN test-bed allows us to further validate the
correctness of the proposal.
2. The test-bed is also built as a part of FranceTelecom WSN platform. The ob-
jective is to set up an extensible hardware platform with stable development
environment. It aims at demonstrating WSN applications. It is also a step from
the design to the development on real communicating objects.
The chapter is organized as follows: We first describe the test-bed in section 6.1,
including the selection of hardware, the operation system, the communication architec-
ture and the detailed architecture of MAC as well as network layer. In section 6.2, the
implementation of FISCO is detailed. Section 6.3 gives the configuration and methods
that are used in experimentations. The validation of FISCO by functional tests are
detailed in section 6.4. Finally, we conclude the results achieved on the test-bed and
discuss the future development in section 6.5.
6.1 Description of Test-bed
The test-bed is composed of 30 Imote2 nodes and 3 gateway boards from Crossbow
[90]. They are all equipped with CC2420 RF transceiver [16] (802.15.4 [20] compatible)
Test-bed
for wireless communication. On the software side, Platform X [17] (based on Linux) is
used as OS.
6.1.1 Imote2 hardware platform
We have selected the Imote2 sensor node platform to deploy our test-bed. Imote2 is
an advanced wireless sensor node platform from Crossbow [91]. This product series
was commercialized in 2006. Its design is modular and stackable with interface connec-
tors for expansion boards on both top and bottom sides. The top connectors provide
a standard set of Input/Output (I/O) signals for basic expansion boards. The bot-
tom connectors provide additional high-speed interfaces for application specific I/O. A
battery board supplying system power can be connected to either side.
The Imote2 contains the Marvell PXA271 XScale CPU. This processor can operate
in a low voltage (0.85V), low frequency (13MHz) mode, hence enabling low power
operation. The frequency can be scaled from 13MHz to 416MHz with Dynamic Voltage
Scaling (DVC). The PXA271 is a multichip module that includes three chips in a single
package, the CPU with 256kB SRAM (divided into 4 equal banks of 64 KB), 32MB
SDRAM and 32MB of FLASH memory.
The Imote2 uses the CC2420 IEEE 802.15.4 radio transceiver from Texas Instru-
ments and an on-board antenna. The CC2420 supports a 250kb/s data rate with 16
channels in the 2.4GHz band. To supply the Imote2 sensor node with all the required
voltage domains, a Power Management Integrated Circuits (PMIC) is added to the
main board. This PMIC supplies 9 voltage domains to the processor in addition to the
DVC capability. The Imote2 can be powered either by a batteries board or directly via
a mini USB port.
One of the improvements of Imote2 comparing to the Mica2/MicaZ series [11] is that
many I/O options are available with the new processor, including I2C, 3 Synchronous
Serial Ports, 3 high speed UARTs, GPIOs, SDIO, USB client and host, AC97 and
I2S audio codec interfaces, fast infrared port, PWM, Camera Interface and a high
speed bus [90]. The integration of these I/O options makes it extremely flexible to
support different sensors, user interfaces, radio options, etc. Therefore it conforms to
FranceTelecom R&D’s strong requirements on extensibility.
Imote2 is a highly flexible platform for either research projects or applications. Al-
though the commercial version of Imote2 came to the market only on 2006, this sensor
node platform has already been widely used in wireless sensor network applications
such as the structural health monitoring of the Golden Gate Bridge [21]. The Imote2
is also supported by several operating systems including TinyOS (v1.1 & v2.0) [18],
124
Description of Test-bed
Network layer
write and read Linux system call
PHY layer Hardware
driver
level
Linux
Main task layer (MTL) (FISCO;highlevel debug interface)
MAC layer (CSMA−CA, scheduling, low level debug interface)
level
application
Linux
Packet transmission layer (PTL)
Figure 6.1: Basic communication architecture
Linux [17] and SOS [92]. Regarding to the interests of FranceTelecom R&D, we choice
Imote2 platform using Linux OS (Platform X) to facilitate the implementation of var-
ious application demonstrations.
6.1.2 Basic communication architecture
Based on the Platform X, we build a basic communication architecture shown in Fig.
6.1. The interface between the MAC and PHY layers is managed directly by the CC2420
driver in Linux kernel. However, the interface between network layer and MAC layer
is totally specified in our architecture through Linux calls in different functions. Some
functionalities in MAC layer are improved such as Acknowledge of unicast transmis-
sion and low-power scheduling. FISCO is implemented on the network layer with the
standard interfaces to other modules. Due to the lack of sensors on Imote2, issues of
sensing application are on the agenda, but currently unaddressed in this architecture.
We expect to add the application layer above the network layer.
125
Test-bed
6.1.3 MAC layer based on CC2420 driver
All functionalities of the MAC layer have been implemented by us based on the CC2420
driver. CC2420 is a true single-chip 2.4 GHz IEEE 802.15.4 compliant RF transceiver
with baseband modem and MAC support. It uses standard serial protocol (SPI) and
four bits external status. It boasts programmable output power, 256 byte data buffer,
hardware MAC encryption (AES-128) and battery monitor.
CC2420 stands as a non-standard character device in Imote2 architecture. It is
composed of three layers:
1. The top layer is Linux interface layer. It adopts the Linux character driver archi-
tecture.
2. The middle layer is hardware abstraction layer. The CC2420 functions and inter-
rupt handlers are implemented.
3. The bottom layer is a hardware dependent layer.
Table 6.1: CC2420 functions
Functions Usage
Send write character stream (a message) to hardwareReceive read a character stream from hardware (blocking mode)Poll read from hardware (non-blocking mode)Channel selection Select 802.15.4 channelTransmission power control Adjust the transmission powerSet/Get MAC information (address, frequency, max data size, etc.)
Table 6.1 gives a list of functions already employed in CC2420 driver. We specify two
Linux system calls read and write to enable the communications between CC2420 and
network layer. More precisely, read and write calls are used by Packet Transmission
sub-Layer (PTL) (see, Fig. 6.1) to address CC2420 driver. This architecture makes
the implementation of network layer protocol such as FISCO totally independent from
CC2420.
6.1.4 Detailed architecture of network layer
The network layer is composed of two sub-layers. PTL, as already mentioned, is a layer
which is in charge of sending receiving message to/from MAC layer. On the top of
PTL, the network protocols are implemented in Main Task sub-Layer (MTL). Network
protocols are implemented on MTL. It contains several modules as shown in Fig. 6.2.
126
Description of Test-bed
switch on
sending
msg FIFO
received
msg FIFO
Packet transmission layer (PTL)
Main task layer (MTL)
Control interfaces
Write threadRead thread
Change FSM states
Timeout handler
FSM message handlerread messageInit
Figure 6.2: Modules description in network layer
1. The Init module is in charge of initiating hardwares (using the open system call)
and system status in network layer. It creates packet transmission threads (read
and write). Once the initialization succeeds, the execution enters into the read
message module.
2. The read message module blocks read operation from received message FIFO, so
that MTL receives the messages from MAC layer one by one. After getting a
message, the execution enters the Finite state machine (FSM) module.
3. The FSM module handles messages passed by the read message module and po-
tentially sends messages to the sending message FIFO. Network protocols should
be implemented under the form of a FSM. The execution returns to the read
message module after the message is processed.
4. The Timeout handler module handles timeout events in a protocol. Some timeout
events might change the FSM state. This module must run after the processing
of the current message in FSM, in order to avoid collisions between the message
handle event and the timeout event. The timer associated to an event is set by
FSM or renewed by the timeout handler itself.
127
Test-bed
Figure 6.3: Finite state machine of FISCO configuration
6.2 Implementation of FISCO
After setting up the communication architecture in Platform X and achieving some
modifications on CC2420 driver, we implemented FISCO in MTL as the first network
protocol on this test-bed. According to the detailed architecture of MTL, FISCO should
first be specified as a FSM. All timeout events as well as periodic events (i.e. Leader
broadcast in FISCO) should be abstracted to timeout handler. It is also important to
define the messages structures in FISCO.
6.2.1 FISCO FSM
Fig. 6.3 gives the FSM of FISCO configuration phase. It contains one-hop address
allocation (section 3.3.1), two-hop address allocation (section 3.3.2) and creation of
new partition (section 3.3.3). The changes in network topology and organization are
128
Methodology of Tests
Figure 6.4: Message header of FISCO
transformed to packet event as the inputs of FSM module. The executions which
depend on timers are transformed into time events fired by the timeout handler.
Our current experimental platform is functional but limited when compared to the
application scope of FISCO. Only the join procedures are implemented in the test-bed.
It is the result of a focused effort to produce a solution for a set of particular goals
rather than to provide a general framework.
6.2.2 FISCO message format
Fig. 6.4 presents the general message header used in FISCO. 16 Bytes are reserved
for the message header which is presented in 6 lines with 32 bits (4 Bytes) for each
line. The head length refers to the number of lines in the message, and is equal to 6
before expansion. The message type is from 1 to 11 (including the debug message),
which corresponds to different messages types in FISCO respectively. The data length
refers to the length of the sensing data after the message header. The unit is 4 Bytes.
According to FISCO design, the old address, new address and initial leader address are
not contained in the same message. Therefore we put them in the same location to
save transmission resource. We use 5 bits to specify five boolean variables of FISCO
messages in the 8-bit flag, before expansion.
Table 6.2 gives a list of FISCO messages used in the implementation including debug
messages.
6.3 Methodology of Tests
We aimed at achieving two types of tests related to FISCO on the test-bed: functional
test and performance test. However, due to the time limitation, only functional tests
which validate the operation of FISCO on Imote2 have been achieved during my work
in Beijing. Nevertheless, we have defined the methodology for both tests in advance.
129
Test-bed
Table 6.2: List of messages used in the implementation
Type Short name Messages
1 LDBR Leader Broadcast Msg2 MBNS Member Solicitation Msg3 MBNA Member Advertisement Msg4 MBAR Member Address Request Msg5 LDAA Leader Address Allocation Msg6 MBAA Member Address ACK Msg7 MGNT Merge Notification Msg8 LDAU Leader Address Update Msg9 MGBA Merge Border Ack Msg10 FDBG FISCO Debug Msg11 FDBR FISCO Debug Reply Msg
6.3.1 Funtional test
In order to validate the operation of FISCO, particularly each event-driven procedure
and timeout event, a set of functional tests were planed according to section 3.3.
• First node configuration
• A new node configures to member
• A new node fails to configure to member
• A leader node reuses of no-acked address in one-hop buffer
• A new node configures to leader
• A new node fails to configure to leader node
• A leader reuses on-acked address in two-hop address configuration
• Two-hop address configuration with a backbone (multiple leaders)
The functional tests aim at validating the individual behavior of a node running
FISCO. Therefore, the results of the tests are evaluated through output messages, local
state and local variables (indicated in FISCO debug messages) of the target node(s).
The tests are designed according to the FISCO FSM. All transitions are validated
through eight listed tests.
130
Experimental Results
(a) Hexagon Platform (b) Uniform Grid Platform
Figure 6.5: Setup for 30 Imotes test-bed platform
6.3.2 Performance test
The second test that we defined is a performance test aiming at collecting experimental
results related to FISCO, such as message cost and energy consumption. Random
network topologies are used in the simulation, while it is not feasible in experimentation
due to the place of test. In order to facilitate the deployment of WSN for test, we define
two connected topologies for performance test as shown in Fig. 6.5.
Both of them have regular form: Fig. 6.5(a) is a hexagon shape while Fig. 6.5(b) is
a uniform grid. In the first case the average degree of each node is set to 6. And for the
grid configuration, the nodes are placed to have a degree of 8. For both topology, two
orders of deployment are applied: nodes are switched on from the center to the border;
or, nodes are switched on from one border to the other border.
6.4 Experimental Results
The performance tests have not been achieved due to the time limitation. Nevertheless,
all functional tests have been validated. Hereby, we show the logs of functional tests
of first node configuration, configure to member and configure to leader. These three
tests correspond to the join procedure as follows:
1. If it detects LDBR messages, then it enters into One-hop Address Allocation (OAA)
as detailed in section 3.3.1.
131
Test-bed
2. If it does not detect any LDBR, then it considers that there is no leader in its neigh-
borhood. The new node tries to actively discover its neighborhood by sending a
Member Neighbor Solicitation (MBNS) message and waits for a Member Neighbor
Advertisement (MBNA) message during a MBNA_TIMEOUT. If it detects any MBNA, then
it enters into two-hop address allocation (TAA) procedure as detailed in section
3.3.2.
3. If the node does not detect any MBNA messages, it decides locally to create a new
partition in the network and assigns itself the first address and the first addresses
pool in the address space. This node is the first leader in the partition identified
by a random PartitionID generated locally as detailed in section 3.3.3.
.
6.4.1 First node configuration
The new node enters into the network. It does not receive either a LDBR message in
W_LDBR state or a MBNA message in NEIGHBOR_FIND state. It then changes its state from
new to first leader of the partition. It sends LDBR periodically and handles messages
as a leader. It enters in leader message handler (L_MSG_HANDLE) state. If it receives a
one-hop MBAR, then it allocates an address from its current address pool to the new
node. After sending the LDAA message, it loops on L_MSG_HANDLE state.
Functional test log: first node configuration
FISCO begin...
[State] W_LDBR
Collects zero LDBR
[State] NEIGHBOR_FIND
Collects 0 MBNA
[State] N2FL
[State] L_MSG_HANDLE
Sends LDBR periodically
Receives one MBAR
Sends one address from its address pool back
[State] L_MSG_HANDLE
132
Experimental Results
6.4.2 Configure to member
The new node enters into the network and receives LDBR messages in the W-LDBR state.
Hence it sends MBAR message to the leader (generally the one with smallest address).
The selected leader allocates an address from its address pool upon receiving the MBAR
message, and then sends the LDAA message to the new node. The new node changes
itself to a member and updates its address according to the receiving LDAA message.
It replies a MBAA message to indicate that the address allocation is achieved, and
enters into member message handle (M_MSG_HANDLER) state.
Functional test log: configure to member
FISCO begin...
[State] W_LDBR
Collects x LDBR(s)
[State] ONEHOP_ADDR_REQ
Sends MBAR to one of the x leader
Waits for the LDAA from that leader...
Receives the corresponding LDAA
[State] N2M
Sends MBAA to the leader
[State] M_MSG_HANDLE
6.4.3 Configure to leader
The new node enters into the network and does not receive the LDBR in the W_LDBR
state. It sends a MBNS message to the neighboring members/gateways. If the new
node receives MBNA in the NEIGHBOR_FIND state, it sends out MBAR message. It chooses
a gateway rather than a member for sending the MBAR message in order to reduce the
number of dominating nodes. By receiving a LDAA message, the new node enters into
N2L state. After replying a MBAA message, it makes itself leader and enters leader
message handle (L_MSG_HANDLE) state. The behaviors of a node in L_MSG_HANDLE has
been validated in section 6.4.1.
Funtional test log: configure to leader
FISCO begin...
[State] W_LDBR
Collects zero LDBR
[State] NEIGHBOR_FIND
133
Test-bed
Receives the MBNA(s)
[State] TWOHOP_ADDR_REQ
Sends MBAR to one neighbor
Waits for the LDAA from that leader...
Receives the corresponding LDAA
[State] N2L
[State] L_MSG_HANDLE
6.5 Conclusions and Discussions
The work achieved on the test-bed is divided into two parts:
1. In the first part of the work, we designed and implemented a protocol stack based
on the embedded Linux OS (Platform X). In this protocol stack, both the interface
to radio transceiver (CC2420 driver) and the interface for implementing network
protocols are specified. It clearly separates the operation of network protocols
and low layer packet transmission. This architecture is used as the standard
development environment.
2. A partial implementation of FISCO is done on the protocol stack. Despite of time
constraint, we are able to validate the self-configuration part of FISCO by a set
of functional tests. It showed that FISCO is easy to implement through a FSM
and time event handlers.
As future work, we plan to do a series of experimentation for performance test of
FISCO. Then the development of the test-bed moves to a second stage, in which the
application layer will be specified to connect the existing protocol stacks. This will
enable the implementation of versatile sensor applications. FranceTelecom R&D also
plans to integrate the WSN as an part of mobile network test-bed, in order to extend
their services in the area of intelligent housing, comfort monitoring, etc.
134
Conclusions and Discussions
Publications
International conference
[IC1] Jialu Fan, Jiming Chen, Jia-liang Lu, Yu Zhang and Youxian Sun. The imple-
mentation of a Fully Integrated Scheme of self-Configuration and self-Organization
(FISCO) on Imote2. In Proc. of the 3rd Int’l Conference on Mobile Ad-hoc and
Sensor Networks (MSN), Beijing, China, December 2007.
135
Test-bed
136
Conclusions and Persperctives 137
Conclusions and Persperctives 77.1 Contributions
The research works presented in this thesis were guided by a strong idea: In order to
be operated efficiently, a WSN needs an autonomous network architecture upon its de-
ployment. We have analyzed the needs of this autonomous network architecture from
the hardware characteristics of wireless sensor nodes, the energy constraints and the
application requirements of WSNs. We proposed FISCO, an autonomous architecture
to organize a WSN. Several functionalities have been achieved effectively and efficiently
based on this architecture such as dynamic address allocation, data dissemination and
data aggregation. It is shown that using this autonomous network architecture facili-
tates the design and implementation of other self-organized mechanisms. In addition,
it provides a better support to services and applications. We have also addressed the
evaluation of organization in a wireless multi-hop network, by proposing an entropy
based metric. It gives us some answers on how to quantify an organization in the
context of WSNs.
7.1.1 Autonomous architecture
The autonomous network architecture we proposed addresses the needs of a WSN,
obtained from either node’s hardware constraints or WSN application requirements.
Particularly, energy efficient, localized and using only local knowledge are the most
important criteria that our autonomous network architecture complies with. With
this autonomous network architecture, the organization and the communications in the
WSN are structured. It also facilitates the implementation of other mechanisms (such as
address allocation, data dissemination and data aggregation) as well as the deployment
of WSN services and application (such as data intensive monitoring applications).
The architecture is developed around a self-organization scheme. It consists of two
levels of logic structure:
Conclusions and Persperctives
1. A global structure represented as a backbone formed by leaders and gateways in
the network;
2. Local structures around each leader.
Each node takes its current local organization into consideration and integrates itself
into the organization. The information on the organization are provided by only a
subset of nodes (leaders) in the network. It leads to a significant energy saving com-
paring to HELLO message based self-organization schemes [4, 49]. We have also proven
through geometric analysis that the cardinality of FISCO backbone is bounded. And
this bound depends only on the size of service area and the communication range, but
independent from the number of nodes. Through network simulations, we show that
the control overhead of FISCO is low and it is energy efficient. Furthermore, the local
re-organization technique may significantly extend the lifetime of a WSN with the same
quantity of energy resources.
7.1.2 Self-organized mechanisms
This autonomous network architecture is not the final goal of our works. It stands as
a general platform on which other mechanisms can easily be developed while achieving
an improved efficiency. We have explored address allocation, data dissemination and
data aggregation techniques based on this architecture.
The address allocation scheme that we proposed in this work is embedded into the
autonomous network architecture. The uniqueness of address is often a requisite of
other mechanisms and it is necessary to provide address allocation as early as a node
joining the network. To achieve this, a two-level address allocation scheme is designed
according to the two levels of organization in the network. It is a new approach that
is different from current address allocation solutions (broadcast based or distributed
DHCP based). And our simulation results show that it benefits from the existing
organization and achieves a low control message overhead.
The data dissemination and data aggregation techniques that we developed in this
work are both based on the backbone generated and maintained by FISCO. A simple
distance vector algorithm is used to direct the backbone toward sink nodes. Hence the
backbone is used as a rendezvous system of the network. Multiple sinks and mobile sinks
can be efficiently managed on this rendezvous system. It is more flexible and extensible
and completely localized without any pre-configurations. Taking advantage of local
organizations, leaders act as the aggregation points in order to reduce communication
over the backbone.
138
Contributions
We also proposed a new temporal data aggregation technique based on adaptive
ARMA (A-ARMA) models. It is completely localized, has low complexity, high accu-
racy and good aggregation efficiency. It can easily be implemented in a data dissemi-
nation/aggregation structure based on FISCO.
7.1.3 Framework solution
Finally, Self-Organization based data Dissemination and Aggregation (SODA) is pro-
posed as a framework which provides a solution from its deployment to the support
of intensive data communication in WSN applications. Through this framework, we
point out that data communication in WSN is not an isolated problem, but should
involve many mechanisms and designs to meet the requirements of WSN. The network
simulation results confirm that the SODA framework achieves very promising energy
efficiency, and save up to 80% energy consumption in a data collection scenario.
7.1.4 Entropy for quantifying organization
In parallel, we have also worked on the proposition of entropy based metric in order
to give a way to evaluate self-organization schemes from the angle of order. Although
numerous metrics exist for performance evaluation of self-organization, our metric is
the first one that could give an idea on the internal organization of a network. We
extended the statistical formulation of entropy in thermodynamic system and apply
it to wireless ad hoc and sensor network. The entropy of a network running a self-
organization are computed based on the local interactions between nodes, while it
reflects the global emerging behavior of the network. The utilization of this entropy
based metric clearly quantifies the impacts of different self-organization schemes on the
internal organization of a network. For instance, any of the self-organization schemes
yields a lower entropy value than a flat organization (without self-organization). Hence,
a higher organizational state is obtained by using self-organization.
7.1.5 Experimental Imote2 test-bed
An experimental test-bed composed of 30 Imote2 nodes employing our autonomous
network architecture (FISCO) was built in FranceTelecom R&D research center in
Beijing. The objective of this test-bed is two-fold. First, we complete FISCO from its
design to its implementation. FISCO passed a set of functional test and its correctness is
validated. Secondly, the test-bed is an open platform with a basic autonomous network
architecture which supports implementation of other applications.
139
Conclusions and Persperctives
7.2 Perspectives
7.2.1 Enrich the functionalities of the architecture
The current research works on WSN generally cut the problem into pieces before study-
ing and bringing solutions. For example, there exist many solutions for routing or topol-
ogy control problems. We are taking another approach in which we build a complete
framework rather than limit our solutions on one particular problem. We believe that
this approach better adapts to the context of WSNs, because users are more sensitive
to the end behavior and the service than to the individual actions of sensor nodes.
We do believe that such an autonomous network architecture provides a general plat-
form to support other techniques and mechanisms. In this work, we have particularly
tackled address allocation, data dissemination and data aggregation techniques based
on this architecture. There are certainly other functionalities that are important for
WSNs. Most of WSN applications need to process a big volume of data. Therefore the
problems related to data such as distributed query processing, context based search and
in-network information processing should be investigated. It is worth noting that data
dissemination and data aggregation are considered as basic cases of above problems.
The ideas and results that we obtained in this work build a basis of the research in
these areas. Once again, we think that the autonomous network architecture should be
used to support these data related functionalities.
Another class of functionalities is linked with the basic task of sensor nodes: sensing.
Localization of targets and coverage problem are very important research works in
WSNs. In both problems, the sensing range is distinguished from the communication
range of wireless sensor nodes. Either the detection/localization of target objects [93] or
sensing area full/partial coverage [94] can only be achieved with the collaborative efforts
of sensor nodes. Our autonomous network architecture is built on the collaboration
of sensor nodes, although only the communication range is taken into account in its
algorithms. The architecture can be easily extended to deal with these problems. Upon
the deployment of the network, nodes may include their local sensing information in the
configuration process, so that the logic structure is generated and maintained according
to both node’s communication vicinity and sensing vicinity. In such way, the sensing
related problems can be integrated naturally into the general framework.
Besides, a collaborations has been launched with LIAMA Lab [95] on declarative
networking [IC1]. In this work, we will attempt to express FISCO (including its local
algorithms and message exchanges) through declarative modeling. Hence, executing
FISCO is transformed into processing declarative queries between nodes. It might lead
140
Perspectives
to lighter the implementations of FISCO, where enhancements and modifications are
easier.
7.2.2 Security aspects of the solutions
Like all communication systems, WSNs are exposed to security threats which, if not
properly addressed, preclude their deployment in a plethora of envisaged civilian and
military scenarios. The fact that there is no centralized element in WSNs drastically
increases the spectrum of potential security threats; this is further amplified by the
serious constraints in node battery power, thereby preventing previously known security
approaches from being deployed. Above mentioned issues require special attention
during the WSN design process, so as to facilitate a commercially attractive deployment.
Albeit traditionally associated to encryption only, security in a wireless network
encompasses three main elements, i.e. confidentiality (encryption of data), integrity
(correctness of data), and robustness (reliability of data); because of the limited power
budget in WSNs, we expect that these issues need to be tackled jointly and in conjunc-
tion with the network paradigm.
In order to address the security issues around the solutions that we proposed, we
should first identify potential roll-out phases and technology developments. We expect
to identify potential threats to confidentiality, integrity and robustness. It is worth
noting that robustness of our solutions has been addressed, considered as the consistence
of the proposition in nodes’ arrival and departure cases. In our opinion, there is a trivial
way to achieve the confidentiality and integrity on our architecture. We have proposed
dynamic address allocation to provide unique identifiers to sensor nodes during the
configuration phase in which leader nodes play essential roles. In order to establish
confidence among nodes, solutions based on the unique identifiers might be a suitable
approach. In practice, an address can be a result of a hash function. Upon the address
allocation, the node could be identified via the hash operation. In the same way, each
member of a leader can be verified. The post steps of data integrity can be then
implemented based on the confidence between a leader and its members.
The above approach is a possible way that we take for considering security problems
in WSNs. More importantly, we want to stress put our accent to say that security is
highly related to network operation and should be addressed very carefully.
7.2.3 Scaling laws
Many network parameters such as throughput and delay are known to suffer from
scalability problem. We begin to investigate on this problem by applying scaling laws
141
Conclusions and Persperctives
to give some fundamental indications on scalable and optimal structuring approaches
for large-scale wireless network [IJ1] [IC2]. This work is conducted with FranceTelecom
R&D.
We first consider an architecture scalable if the network-wide performance attribute,
scaled by the number of nodes involved, does not decrease with an increasing number
of nodes and with a non decreasing problem space. Using some well established scaling
laws from communications, i.e. Kumar & Gupta’s throughput scaling law [56], and
economics, i.e. Odlyzko Tilly’s value scaling law [96] (which is based on Zipf’s law
[97]), we have established that large scale networks generally do not scale w.r.t. some
key attributes with an increasing number of nodes.
The throughput of both unclustered and clustered architecture was shown not to be
scalable, while a clustered approach - based on some given assumptions - has shown
to exhibit a better scalability than its unclustered counterpart. We show that the
scaling properties of a clustered architecture heavily rely on the data pipes between
the clusterheads. Hence they should either be stronger or data aggregation needs to
be performed for inter-clusterhead communications. It is worth noting that SODA
framework is one of the solutions that are able to build such architecture with self-
organization, self-healing and data aggregation techniques. Hence SODA framework
achieves a better scalability than flat networking.
The results in this work are still indicative only, where different assumptions will yield
different absolute results. Nonetheless, we expect that the results expose trends which
are of use for WSN and hence for emerging and future real-world applications, such
as data collection, remote control solutions, wireless telemetry, automatic monitoring,
metering solutions and smart environments such as homes, hospitals, and buildings of
all kinds.
7.2.4 Application tunable parameters
Since the objective of our work is to provide a general platform, it should support
different classes of WSN applications. Although some applications share common re-
quirements as exposed in chapter 1, some objectives are still specified. Furthermore,
different scenarii and environments may vary the requirement of applications. How to
model these applications to a set of standard parameters through which the behavior
of autonomous network architecture can be adjusted on-the-fly is still an open issue.
It has important impact on the promotion of WSNs. Once we are able to address
this problem, the deployment of new services will significantly grow and the WSN will
finally reach the life of everyone.
142
Publications
International journal
[IJ1] Mischa Dohler, Thomas Watteyne, Fabrice Valois and Jia-Liang LU. Kumar’s,
Zipf’s and other laws - how to structure a large-scale wireless network? To be
published in Annals of Telecommunication, 2008.
International conference
[IC1] Stephane Grumbach, Jia-liang LU and Wenwu Qu. Self-organization of wireless
networks through declarative local communication. In Prof. of the Int’l Work-
shop on MObile and NEtworking Technologies for social applications (MONET),
Vilamoura , Portugal, November 2007.
[IC2] Mischa Dohler, Thomas Watteyne, Dominique Barthel, Fabrice Valois and Jia-
Liang LU. Kumar’s, Zipf’s and other Laws - how to structure an optimum large-
scale wireless (sensor) Network? Invited paper of the 13th European Wireless
2007 Conference, Paris, France, April 2007.
Conclusions and Persperctives
144
Bibliography 145
Bibliography
[1] M. Mohsin and R. Parkash. IP address assignment in a mobile ad hoc network.Proc. of IEEE Military Communications Conference, Anaheim, USA, 2:856–861,October 2002.
[2] S. Nesargi and R. Prakash. MANETconf: configuration of hosts in a mobile adhoc network. Proc. of IEEE Conference on Computer Communications, New York,USA, 2:1059–1068, June 2002.
[3] H. Zhou, L. Ni, and M. Mutka. Prophet address allocation for large scale MANETs.Proc. of IEEE Conference on Computer Communications, San Francisco, USA,2:1304–1311, April 2003.
[4] J. Wu and H. Li. On calculating Connected Dominating Set for efficient routing inad hoc wireless networks. Proc. of the 3rd Int’l Workshop on Discrete Algorithmsand Methods for Mobile Computing and Communications, Seattle, USA, pages7–14, August 1999.
[5] F. Dai and J. Wu. Distributed dominant pruning in ad hoc wireless networks.Proc. of IEEE Int’l Conference on Communications, Seattl, USA, 1:353–357, May2003.
[6] LAN MAN Standards Committee of the IEEE Computer Society. IEEE Std 802.11,1999 Edition Information technology - Telecommunications and information ex-change between systems - Local and metropolitan area networks - Specific require-ments - Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer(PHY) specifications. IEEE-SA Standards Board, 1999.
[7] Bluetooth SIG. Bluetooth Specification v2.0. 2004.
[8] Auto-ID-Center. Draft protocol specification for a 900 MHz Class 0 Radio Fre-quency Identification Tag. 2003.
[9] I. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci. Wireless sensornetworks: a survey. Computer Networks (Elsevier), 38:393–422, 2002.
[10] M. Tubaishat and S. Madria. Sensor networks: an overview. IEEE Potentials,22:20–23, April 2003.
[11] Wireless sensors from UC Berkeley and Intel researchershelp conservation biologists monitor elusive seabird in Maine.
Bibliography
http://www.berkeley.edu/news/media/releases/2002/08/05 snsor.html, August2002.
[12] E. Shih, S. Cho, N. Ickes, R. Min, A. Sinha, A. Wang, and A. Chandrakasan. Phys-ical layer driven protocol and algorithm design for energy-effcient wireless sensornetworks. Proc. of ACM Int’l Conference on Mobile Computing and Networking,Rome, Italy, pages 272–286, July 2001.
[13] T. Voigt, H. Ritter, and J. Schiller. Utilizing solar power in wireless sensor net-works. Proc. of the 28th IEEE Int’l Conference on Local Comuter Networks, Bonn,Germany, pages 416–424, 2003.
[14] S. Roundy, P. Wright, and J. Rabaey. Energy Scavenging for Wireless SensorNetworks with Special Focus on Vibrations. Boston : Kluwer Academic Publishers,2004.
[15] Chipcon AS SmartRF. CC1000 Single Chip Very Low Power RF Transceiver. US:Texas Instrument, June 2004.
[16] Chipcon AS SmartRF. CC2420 2.4 GHz IEEE 802.15.4 / ZigBee-ready RFTransceiver. US : Texas Instrument, June 2004.
[17] Platform X. http://embedded.seattle.intel-research.net/wiki/.
[18] TinyOs community forum. UC berkeley : The Rest. On: http://www.tinyos.net,visited on 01.10.2005.
[19] G. Pottie and Kaiser W. Wireless integrated network sensors. Communication ofthe ACM, 43:51–58, May 2002.
[20] LAN MAN Standards Committee of the IEEE Computer Society. IEEEStd802.15.4 for Information technology - Telecommunications and information ex-change between systems - Local and metropolitan area networks - Specific require-ments - Part 15.4: Wireless Medium Access Control (MAC) and Physical Layer(PHY) specifications for Low-Rate Wireless Personal Area Network (LR-WPANs).IEEE-SA Standards Board, 2003.
[21] S. Kim, S. Pakzad, D. Culler, J. Demmel, G. Fenves, S. Glaser, and M. Turon.Health monitoring of civil infrastructures using wireless sensor networks. Proc.of the 6th Int’l Conference on Information Processing in Sensor Networks, Cam-bridge, USA, pages 254–263, April 2007.
[22] A. Mainwaring, J. Polastre, R. Szewczyk, D. Culler, and J. Anderson. Wirelesssensor networks for habitat monitoring. Proc. of ACM Int’l Workshop on WirelessSensor Networks and Applications, Atlanta, USA, pages 88–97, September 2002.
[23] Shipboard machine monitoring for predictive mainte-nance. Newton, MA : Questex Media Group, Inc. On:
146
Bibliography
http://www.sensorsmag.com/sensors/article/articleDetail.jsp?id=314716, vis-ited on 01.04.2006.
[24] J. Jeong and J. Park. Autoconfiguration technology for IPv6 multicast services inmobile ad-hoc network. Proc. of IEEE Int’l Conference On Networks, Singapore,pages 261–265, August 2002.
[25] R. Ramanathan and J. Redi. A brief overview of ad hoc networks: challenges anddirections. IEEE Communications Magazine, 40:20–22, May 2002.
[26] M. Fazio, M. Villari, and A. Puliafito. IP address autoconfiguration in ad hocnetworks: design, implementation and measurements. Computer Networks: theInternational Journal of Computer and Telecommunications Networking, 50:898–920, 2006.
[27] J. Wang, V. Prasad, and I. Niemegeers. Self-configuration in MANETs: differentperspectives. Proc. of the 2nd Int’l Workshop on Self-Organizing Systems, the Lakedistrict, UK, pages 225–239, 2007.
[28] K. Sohrabi, J. Gao, V. Ailawadhi, and G. Pottie. Protocols for self-organizationof a wireless sensor network. IEEE Personal Communications, pages 16–27, 2000.
[29] C. Prehofer and C. Bettstetter. Self-organizaton in communication networks: prin-ciples and design paradigms. IEEE Communication Magazine, pages 78–85, July2005.
[30] N. Vlajic and N. moniz. Self-healing wireless sensor networks: results that maysurprise. Proc. of the 3rd Distributed Autonomous Network Management System,Washington, DC, USA, pages 1–6, November 2007.
[31] R. Misra and C. Mandal. Self-healing for self-organizing cluster sensor networks.Proc. of Annual India Conference, New Delhi, India, pages 1–6, Semptember 2006.
[32] P. Boonma and J. Suzuki. BiSNET: a biologically-inspired middleware architec-ture for self-managing wireless sensor networks. Elsevier Journal of ComputerNetworks, Special Issue on Middleware Challenges for Next Generation Networksand Services, 51(16):4599–4616, November 2007.
[33] L. Ruiz, T. Braga, F. Silva, H. Assuncao, J. Nogueira, and A. Loureiro. On the de-sign of a self-managed wireless sensor network. IEEE Communications Magazine,43(8):95–102, 2005.
[34] B. Karp and H. Kung. GPSR: Greedy perimeter stateless routing for wirelessnetworks. Proc. of the Sixth ACM/IEEE Int’l Conference on Mobile Computingand Networking (MobiCom ’00), Boston, USA, 2003.
[35] A. Carzaniga and A. Wolf. Content-based networking: a new communicationinfrastructure. Proc. of the NSF Workshop on an Infrastructure for Mobile andWireless Systems, Scottsdale, USA, pages 59–68, October 2001.
147
Bibliography
[36] R. Droms. Dynamic host configuration protocol (DHCP). IETF RFC 2131, 1997.
[37] S. Deering and R. Hinden. Internet protocol, version 6 (IPv6) specification. IETFRFC 2460, 1998.
[38] S. Thomson and T. Narten. IPv6 stateless address autoconfiguration. IETFRFC2462, 1998.
[39] N. Vaidya. Weak duplicate address detection in mobile ad hoc networks. Proc.of the 3rd ACM Int’l Symposium on Mobile Ad Hoc Networking and Computing,Lausanne, Switzerland, pages 206–216, June 2002.
[40] K. Weniger. PACMAN: Passive autoconfiguration for mobile ad hoc networks.IEEE Journal on Selected Areas in Communications, 23:507–519, March 2005.
[41] P. Santi. Topology control in wireless ad hoc and sensor networks. ACM ComputingSurveys, 37(2):164 – 194, June 2005.
[42] T. Camp, J. Boleng, and V. Davies. A survey of mobility models for ad hoc networkresearch. Wireless Communication and Mobile Computing (WCMC), Special Issueon Mobile Ad Hoc Networking: Research, Trends and Applications, 2(5):483–502,2002.
[43] A. Ephremides, J. Wieselthier, and D. Baker. A design concept for reliable mobileradio networks with frequency hopping signaling. Proc. of the IEEE, 75:56–73,January 1987.
[44] W. Lou and J. Wu. A cluster-based backbone infrastructure for broadcasting inMANETs. Proc. of the 17th Int’l Symposium on Parallel and Distributed Process-ing, Nice, France, pages 22–26, April 2003.
[45] O. Younis and S. Fahmy. HEED: a hybrid, energy-efficient, distributed clusteringapproach for ad hoc sensor networks. IEEE Transactions on Mobile Computing,3(4):366–379, 2004.
[46] P. Wan, K. Alzoubi, and O. Frieder. Distributed construction of connected dom-inating set in wireless ad hoc networks. Proc. of IEEE Conference on ComputerCommunications, New York, USA, 3:1597–1604, March 2002.
[47] C. Adjih, P. Jacquet, and L. Viennot. Computing connected dominated sets withmultipoint relays. Journal on Ad Hoc and Sensor Wireless Networks, 1:27–39,2005.
[48] G. Toussaint. The relative neighborhood graph of finite planar set. Pattern Recog-nition, 12:261–268, 1980.
[49] J. Cartigny, F. Ingelrest, D. Simplot-Ryl, and I. Stojmenovic. Localized LMSTand RNG based minimum energy broadcast protocols in ad hoc networks. Ad HocNetworks, 3(1):1–16, January 2005.
148
Bibliography
[50] A. Qayyum, L. Viennot, and A. Laouiti. Multipoint relaying for flooding broadcastmessages in mobile wireless networks. Proc. of the Hawaii Int’l Conference onSystem Sciences, Hawaii,USA, 9:298–297, January 2002.
[51] J. Pan, Y. Hou, L. Cai, Y. Shi, and S. Shen. Topology control for wireless sensornetworks. Proc. of the 9th Int’l Conference on Mobile Computing and Networking,San Diego, USA, pages 286–299, September 2003.
[52] F. Theoleyre and F. Valois. A virtual structure for hybrid networks. Proc. of IEEEWireless Communications and Networking Conference, Atlanta, USA, pages 1035–1040, March 2004.
[53] R. P. Brent. Uniform random number generators for supercomputers. Proc. of the5th Australian Supercomputer Conference, Melbourne, Australia, pages 95–104,December 1992.
[54] I. Stojmenovic and J. Wu. Mobile Ad Hoc Networking, chapter Broadcasting andactivity scheduling, pages 205–229. Piscataway, NJ : IEEE/Wiley, 2004.
[55] V. Balakrishnan. Theory and problems of combinatorics. New York : McGraw-Hill,1998.
[56] P. Gupta and P.R. Kumar. The capacity of wireless networks. IEEE Transactionson Information Theory, 46(2):388–404, March 2000.
[57] B. Rimon. Handbook on Theoretical and Algorithmic Aspects of Sensor, Ad hocWireless, and Peer-to-Peer Networks, chapter Scalable Wireless Ad hoc NetworkSimulation, pages 297–311. Boca Raton : CRC Press, 2005.
[58] Freescale Semiconductor. MC13192 / MC13193 2.4 GHz Low Power Transceiverfor the IEEE 802.15.4 Standard: Technical Data, April 2005.
[59] J. Polastre, J. Hill, and D. Culler. Versatile Low Power Media Access for Wire-less Sensor Networks. Proc. of ACM Conference on Embedded Networked SensorSystems, Baltimore, USA, pages 95–107, November 2004.
[60] C. Intanagonwiwat, R. Govindan, and D. Estrin. Directed diffusion: a scalable androbust communication paradigm for sensor networks. IEEE/ACM Transactions onNetworking, 11:2–16, 2000.
[61] F. Ye, H. Luo, J. Cheng, S. Lu, and L. Zhang. A two-tier data disseminationmodel for large-scale wireless sensor networks. Proc. of the 8th Int’l Conferenceon Mobile Computing and Networking, Atlanta, USA, pages 148–159, September2003.
[62] D. Petrovic, R. Shah, and J. Rabaey. Data funneling: routing with aggregationand compression for wireless sensor networks. Proc. of IEEE Int’l Workshop onSensor Network Protocols and Applications, Seattle, USA, pages 156– 162, May2003.
149
Bibliography
[63] S. Yoon and C. Shahabi. Exploiting spatial correlation towards an energy effi-cient clustered aggregation technique (CAG). Proc. of IEEE Int’l Conference onCommunication, Seoul, Korea, 5:3307–3313, May 2005.
[64] T. Lin and P. Huang. Sensor data aggregation for resource Inventory applica-tion. Proc. of IEEE Wireless Communications and Networking Conference, NewOrleans, USA, 4:2369–2374, March 2005.
[65] J. Beaver, M. Sharaf, A. Labrinidis, and K. Chrysanthis. TiNA: a scheme for tem-poral coherency-aware in-network aggregation. Proc. of IEEE Int’l Workshop onSensor Network Protocols and Applications, Seattle, USA, pages 69–76, September2003.
[66] S. Nath, P. Gibbons, S. Seshan, and Z. Anderson. Synopsis diffusion for robustaggregation in sensor networks. Proc. of the 2nd Int’l Conference on EmbeddedNetworked Sensor Systems, Baltimore, USA, pages 250–262, November 2004.
[67] J. Shin, J. Kim, K. Park, and D. Park. Railroad: virtual infrastructure for datadissemination in wireless sensor networks. Proc. of the 2nd ACM Workshop onPerformance Evaluation of Wireless Ad Hoc, Sensor and Ubiquitous Networks,Montreal, Canada, pages 168–174, October 2005.
[68] A. Marcucci, M. Nati, C. Petrioli, and A. Vitaletti. Directed diffusion light: lowoverhead data dissemination in wireless sensor networks. Proc. of IEEE 61st Vehic-ular Technology Conference Spring, Stockholm, Sweden, 4:2538– 2545, May 2005.
[69] D. Braginsky and D. Estrin. Rumor routing algorthim for sensor networks. Proc. ofthe 1st ACM Int’l workshop on Wireless sensor networks and applications, Atlanta,USA, pages 22–31, September 2002.
[70] B. Karp and H. Kung. GPSR: Greedy perimeter stateless routing for wirelessnetworks. Proc. of the 6th ACM/IEEE Int’l Conference on Mobile Computing andNetworking, Boston, USA, pages 243–254, August 2000.
[71] Y. Kotidis. Snapshot queries: towards data-centric sensor networks. Proc. of IEEEInt’l Conference on Data Engineering, Tokyo, Japan, pages 131– 142, April 2005.
[72] K. Fan, S. Liu, and P. Sinha. On the potential of structure-free data aggregationin sensor networks. Proc. of IEEE Conference on Computer Communications,Barcelona, Spain, pages 1–12, April 2006.
[73] C. Liu, K. Wu, and M. Tsao. Energy efficient Information Collection with theARIMA model in wireless sensor networks. Prof. of IEEE Global CommunicationsConference, St. Louis, USA, 5:5–10, November 2005.
[74] C. Chatfield. The Analysis of Time Series. Boca Raton : CRC Press, 2004.
150
Bibliography
[75] C. Guestrin, R. Thibaux, P. Bodik, M. Paskin, and S. Madden. Distributed re-gression: an efficient framework for modeling sensor network data. Proc. of IEEESymposium on Information Processing on Sensor Networks, Berkeley, USA, pages1–10, April 2004.
[76] K. Deng, A. Moore, and M. Nechyba. Learning to recognize time series: combiningARMA models with memory-based learning. Proc. of IEEE Int’l Symposium onComputational Intelligence in Robotics and Automation, Monterey, USA, 1:246–250, 1997.
[77] Easinet research group. On http://www.easinet.cn, visited on 08.08.2007.
[78] N. Kimura and S. Latifi. A survey on data compression in wireless sensor networks.Proc. of Int’l Conference on Information Technology: Coding and Computing, LasVegas, USA, 2:8–13, April 2005.
[79] A. Das, F. Lewis, and D. Popa. Data-logging and supervisory control in wirelesssensor networks. Proc. of the 7th ACIS Int’l Conference on Software Engineer-ing, Artificial Intelligence, Networking, and Parallel/Distributed Computing, LasVegas, USA, pages 330–338, June 2006.
[80] H. Guo, F. Ingelrest, D. Simplot-Ryl, and I. Stojmenovic. Performance evaluationof broadcasting protocols for ad hoc and sensor networks. The Fourth AnnualMediterranean Ad Hoc Networking Workshop (Med-Hoc-Net), Ile de Porquerolles,France, June 2005.
[81] F. Theoleyre and F. Valois. About the self-stabilization of a virtual topology forself-organization in ad hoc networks. Proc. of the 7th Int’l Symposium on SelfStabilizing System, Barcelone, Spain, pages 214–228, October 2005.
[82] R. Clausius. Annalen der Physik, (79):368–397, 500–524, 1850.
[83] L. Boltzmann. Uber die Beziehung zwischen dem zweiten Hauptsatz der mecha-nischen Warmetheorie und der Wahrscheinlichkeitsrechnung respektive den Satzenuber das Warmegleichgewicht. Wien. Ber.76, 373-435, Band II, pages 164–223,1877.
[84] C.E. Shannon. The mathematical theory of communication. Urbana : Universityof Illinois Press, 1949.
[85] K. Sneppen, A. Trusina, and M. Roswall. Measuring information networks. Pra-mana - Journal of Physics, 64:1121–1125, 2005.
[86] T. Balch. Hierarchic social entropy: an information theoretic measure of robotgroup diversity. Autonomous Robots, 8:209–237, 2000.
[87] T.H. Labella and F. Dressler. A bio-inspired architecture for division of labourin SANETs. Proc. of the 1st Int’l Conference on Bio inspired models of network,
151
Bibliography
information and computing systems, Cavalese, Italy, pages 209–228, December2006.
[88] K. Sneppen, A. Trusina, and M. Rosvall. Hide-and-seek on complex networks.Europhysics Letters, pages 853–859, March 2005.
[89] V. Parunak and S. Brueckner. Entropy and self-organization in multi-agent sys-tems. Proc. of the 5th Int’l Conference on Autonomous agents, Montreal, Canada,pages 124–130, May 2001.
[90] Crossbow. Crossbow Imote2 datasheet. San Jose, Clifornia : Crossbow Technology.
[91] Crossbow Mica Motes/ Imote2. San Jose, Clifornia : Crossbow Technology. Onhttp://www.xbow.com, visited on 06.03.2006.
[92] C. Han, R. Rengaswamy, R. Shea, E Kohler, and M. Srivastava. A dynamicoperating system for sensor nodes. Proc. of the 3rd Int’l Conference on MobileSystems, Applications, and Services, Seattle, USA, pages 163–176, June 2005.
[93] T. Roosta, M. Mishra, and A. Ghazizadeh. Robust estimation and detection in adhoc and sensor networks. Proc. of IEEE Int’l Conference on Mobile Ad-hoc andSensor System, Vancouver, Canada, pages 236–245, October 2006.
[94] Z. Huang and C. Shen. Distributed topology control mechanism for mobile ad hocnetworks with swarm intelligence. Proc. of the ACM Int’l Symposium on MobileAd Hoc Networking and Computing, Annapolis, USA, pages 21–22, June 2003.
[95] LIAMA Laboratory. http://liama.ia.ac.cn/wiki/.
[96] B. Briscoe, A. Odlyzko, and B. Tilly. Metcalfe’s law is wrong. IEEE SpectrumMagazine, July 2006.
[97] G. Zipf. Some determinants of the circulation of information. Amer. J. Psychology,59:401–421, 1946.
152