ENERGY EFFICIENT AND ADAPTIVE SERVICE ADVERTISEMENT, …pages.cs.aueb.gr/~chris/C.N.Ververidis-PhD Thesis-2008.pdf · A Dissertation Submitted to the Faculty of the Department of

ENERGY EFFICIENT AND ADAPTIVE SERVICE ADVERTISEMENT,

DISCOVERY AND PROVISION FOR MOBILE AD HOC NETWORKS

A Dissertation

Submitted to the Faculty

of

the Department of Computer Science,

of the Athens University of Economics and Business

by

Christopher N. Ververidis

In Partial Fulfillment of the

Requirements for the Degree

of

Doctor of Philosophy

July 2008

Athens University of Economics and Business

Athens, Greece

ii

iii

DEDICATION

I would like to dedicate my Phd Thesis to my mother Kaiti, to my sister

Anna, to my grandmother Anna, and to the memory of my father Nikos.

iv

v

ACKNOWLEDGMENTS

I would like to thank my advisor George Polyzos for the invaluable guidance

throughout the years. He taught me what is scientific research and showed

me the way into academic life. He has been an inspiring teacher and a caring

mentor for me.

I would also like to express my gratitude to professor Michalis Vazirgiannis

for giving me the opportunity to work with him and for supporting my research

right from the start of my Phd. It has been an honor for me to work with

George Stamoulis as his teaching assistant, and a privilege to be able to discuss

research issues with him. I would also like to express my appreciation to the

other members of my committee, Professors Vana Kalogeraki, Vasilis Siris,

Ioannis Stavrakakis, and George Xylomenos.

It has been a pleasure to work together, to share problems and to have

nice discussions with all my friends of the MMLAB and NES-LAB; the “vet-

erans” Panayotis Antoniadis, Manos Dramitinos, Elias Efstathiou, Thanasis

Papaioannou and of course Sergios Soursos, and also the “newer generation”

Pantelis Fragoudis, Kostas Kalogiros, Ntinos Katsaros, Vasilis Kemerlis and

George Thanos. We had a really great time being together in the lab. Special

thanks go to Nafsika Kokkini, Lina Kanelopoulou and Eleftheria Nifli not only

for solving many day-to-day problems but also for being there for me as good

listeners in hard times.

Finally I would like to thank all my close friends and family for their love,

understanding, patience, endless support and encouragement that made this

Thesis a reality; they have been my source of strength and energy.

vi

vii

ABSTRACT

Mobile Ad Hoc Networks (MANETs) are networks of mobile computing nodes

(e.g. portable computers, PDAs etc.) equipped with wireless interfaces and

communicating with each other without relying on any infrastructure. In these

networks each mobile node may act as a client, a server and a router. MANETs

have emerged to fulfill the need for communication of mobile users in locations

where deploying a network infrastructure is impossible, or too expensive, or

simply is not available at that time. Characteristic scenarios for MANETs are

disaster relief operations, battlefields and locations where infrastructure-based

WLAN coverage (also called hotspots) is not provided and wireless WANs (e.g.

GPRS/UMTS) are too expensive to use or too slow.

Most of the research on MANETs has focused on issues dealing with the

connectivity between mobile nodes in order to cope with the dynamism of such

networks and the arising problems thereof. This dynamism is due to the mo-

bility of nodes, the wireless channel’s adverse and fast changing conditions and

the energy limitations of mobile nodes, all of which lead to frequent disconnec-

tions and/or node failures. These research efforts have led to the creation of a

sound technical basis for dealing with the aforementioned problems regarding

node connectivity in MANETs (mainly through routing protocols, link layer

protocols etc.).

However, solving the problems of connectivity alone is not sufficient for

the adoption of MANETs. Since their basic role is to allow mobile users to

exchange data and use each other’s services, there is also a need for architec-

tures, mechanisms and protocols for Service Discovery and Provision. Service

Discovery is defined in general as the process allowing networked entities to:

i) advertise their services, ii) query about services provided by other entities,

iii) select the most appropriate services and iv) invoke the services.

viii

Service Discovery has been mainly addressed in the context of wired net-

works in the past. In that context, clients and especially servers are typically

powerful and resource-rich machines connected to the wired network. Service

discovery and advertisement for wired networks generally follows a central-

ized or semi-centralized architecture assuming that well-known and “always

on” service registries (or directories) exist for matching service requests to

available services. Most of these architectures also rely on inefficient flooding

techniques for service discovery and advertisement since resource scarcity is

not a key issue. All the characteristics mentioned above, render those archi-

tectures inappropriate for service discovery in MANETs. In this context, the

requirements are radically different than those of wired networks. In MANETs

both clients and servers are more lightweight devices with limited resources.

The assumption of (possibly dedicated and) always-available nodes serving as

service directories is no longer valid. Also, frequent disconnections are a ma-

jor issue affecting service availability. Disconnections may happen due to node

mobility, due to depletion of the energy resources of the nodes, or even due

to switching-off of some nodes. It is rather straightforward that in such envi-

ronments it is of utmost importance to have energy efficient service discovery

mechanisms.

In this thesis we develop two energy efficient service discovery protocols

integrated into the routing process in order to avoid redundant network mes-

saging. Both developed protocols are distributed in nature and employ both

proactive and reactive information dissemination techniques. The performance

of the developed protocols is thoroughly investigated through extensive simu-

lations in terms of energy consumption, service discoverability and achievable

service availability under a wide range of settings (e.g. mobility, network den-

sity, channel characteristics, service replication, traffic patterns). The first

(basic) protocol, called Extended Zone Routing Protocol (E-ZRP) is used in

our study for comparing cross-layer with application-layer based service dis-

covery protocols. Although integrating service and route discovery has been

already proposed in the past, previous work regarding the energy gains of im-

ix

plementing service discovery at the routing layer as compared to implementing

it at the application layer (as a separate process) is problematic due to the

following reasons: i) most approaches either deal with energy only implicitly

(by measuring the overhead in number of packets) or employ non-realistic en-

ergy consumption modeling, ii) the majority of proposed approaches involve

performance analysis comparisons of the developed route and service discovery

protocols against unrealistic application layer based service discovery protocols

which are based on global flooding and hence are not suitable to MANETs, and

iii) none of those approaches takes into consideration the impact of protocol

message sizes on protocol performance. In our work we have performed full-

stack simulations employing a realistic energy consumption model, accounting

for the actual energy consumption of the nodes, in order to reveal the real gains

from integrating the service discovery process with the routing process. The

integrated service discovery protocol, named E-ZRP proves to be much more

efficient compared to a similar but application layer based service discovery

protocol. More importantly, we developed an even more sophisticated service

discovery protocol, called Adaptive SerVicE and Route Discovery ProTocol

for MANETs (AVERT). AVERT demonstrates superior efficiency than the

best alternatives, mainly due to its built-in capability of adapting to a volatile

environment such as a MANET. Using a monitoring mechanism for traffic seen

locally, a node may adapt the operation of AVERT to best match the fluctu-

ating traffic and mobility conditions typically found in MANETs.

In parallel to the aforementioned research effort, we developed a novel

mechanism for service providers to select to serve those clients from a MANET

that can maximize their profits. Optimized service provisioning is a challenging

problem in dynamic environments such as MANETs. We consider the nodes in

MANETs to be independent, rational agents trying to maximize their profits

through service provision. We model this problem as a Generalized Assignment

Problem (GAP). We adopt a pay-as-you-go model, where clients pay for the

service as long as they are receiving the service, since a pay-in-advance model

would be unfair especially in MANETs where connection loss is very proba-

x

ble. We introduce into the proposed profit maximization algorithm expected

payoffs based on estimates of server-to-client connectivity. Those estimations

can be used for computing the actual payoff that will be received from any

client that is selected by the service provider. We experimentally study cases

with non-cooperative and cooperative servers and investigate the gain of the

estimate based profit maximization algorithm versus a classic profit maximiza-

tion algorithm, which does not take into account the network’s dynamics that

affect server-to-client connectivity. We especially study the duration of con-

nectivity (irrespectively of path changes) between two nodes of a MANET.

Previous work, both analytical and experimental, has focused only on esti-

mating the duration of a single path between a client and a server without

considering changes to a path’s length caused by node mobility. The connec-

tivity between two nodes irrespectively of changes in paths or the path length

has not been investigated. In this dissertation we derive an approximation of

the connectivity duration, taking into account the network’s density (number

of nodes, terrain size, wireless transmission range), the node speed and the

initial distance (in number of hops) between two nodes (one representing a

client and the other a server). Simulations show that our approach achieves

up to three-fold improved server profits compared to the classical one and is

especially suited for MANETs with high-mobility and/or low density, which

verifies that the proposed model accurately captures the effects of server to

client connectivity on the overall performance.

Summing up, in this thesis we propose energy efficient highly adaptive

service discovery protocols integrated with hybrid routing protocols. We im-

plemented and experimentally evaluated those protocols to show their energy

and service discoverability/availability gains under different MANET condi-

tions and against similar application layer-based as well as routing layer-based

protocols. We also developed and evaluated an innovative mechanism for al-

lowing mobile service providers to maximize their profits when offering their

service in MANETs, by estimating their connectivity to candidate clients.

xi

TABLE OF CONTENTS

Page

ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

TABLE OF CONTENTS . . . . . . . . . . . . . . . . . . . . . . . . xi

LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii

LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . xiv

VITA AND PUBLICATIONS . . . . . . . . . . . . . . . . . . . . . . xvii

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Overview and Motivation . . . . . . . . . . . . . . . . . . . . 1

1.2 Thesis and Contributions of the Dissertation . . . . . . . . . 4

1.3 Outline of the Dissertation . . . . . . . . . . . . . . . . . . . 6

2 Previous Work on Service Advertisement and Discovery Approachesfor MANETs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.1 Service Discovery Architectures . . . . . . . . . . . . . . . . 12

2.1.1 Directory-based Architectures . . . . . . . . . . . . . 12

2.1.2 Directory-less Architectures . . . . . . . . . . . . . . 17

2.1.3 Hybrid Architectures . . . . . . . . . . . . . . . . . . 21

2.1.4 Critique of Service Discovery Architectures . . . . . . 21

2.2 Service Discovery Modes . . . . . . . . . . . . . . . . . . . . 23

2.2.1 Reactive Mode . . . . . . . . . . . . . . . . . . . . . 23

2.2.2 Proactive Mode . . . . . . . . . . . . . . . . . . . . . 23

2.2.3 Hybrid Mode . . . . . . . . . . . . . . . . . . . . . . 24

2.2.4 Critique of Service Discovery Modes . . . . . . . . . . 24

2.3 Cross Layer Service Discovery . . . . . . . . . . . . . . . . . 25

3 Integrating Service Discovery with Hybrid Routing Protocols . . . 35

3.1 The E-ZRP Service and Route Discovery Protocol . . . . . . 35

3.1.1 Concept and Motivation . . . . . . . . . . . . . . . . 37

3.1.2 Design . . . . . . . . . . . . . . . . . . . . . . . . . . 40

xii

Page

3.1.3 Experimental Evaluation . . . . . . . . . . . . . . . . 45

3.2 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

3.3 The AVERT Service and Route Discovery Protocol . . . . . 71

3.3.1 Concept and Motivation . . . . . . . . . . . . . . . . 72

3.3.2 Design . . . . . . . . . . . . . . . . . . . . . . . . . . 73

3.3.3 Experimental Evaluation . . . . . . . . . . . . . . . . 77

3.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

4 Profit Maximization Mechanism for Service Provision in MANETs 85

4.1 Background and Applications for Charged Service Provision inMANETs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

4.2 Background on Connectivity Estimation . . . . . . . . . . . 88

4.3 System Description . . . . . . . . . . . . . . . . . . . . . . . 91

4.4 Problem Formulation and Analysis . . . . . . . . . . . . . . 93

4.5 Approximation of Duration of Connectivity . . . . . . . . . . 96

4.6 Performance Analysis of the Profit Maximization Algorithm 100

4.6.1 Cooperative Servers . . . . . . . . . . . . . . . . . . . 102

4.6.2 Non-Cooperative Servers . . . . . . . . . . . . . . . . 105

4.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

5 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

5.1 Smart adaptation of AVERT using service information . . . 107

5.2 Connectivity duration estimation formulas under various mo-bility models . . . . . . . . . . . . . . . . . . . . . . . . . . . 108

5.3 Interoperability . . . . . . . . . . . . . . . . . . . . . . . . . 108

5.4 Benchmarking . . . . . . . . . . . . . . . . . . . . . . . . . . 109

6 Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . 111

LIST OF REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . 114

ABBREVIATIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123

xiii

LIST OF TABLES

Table Page

3.1 Symbols for the energy model . . . . . . . . . . . . . . . . . . . 49

3.2 Byteload per packet type (N is the average number of a node’s1-hop neighbors) . . . . . . . . . . . . . . . . . . . . . . . . . . 51

3.3 Simulation Settings . . . . . . . . . . . . . . . . . . . . . . . . . 52

3.4 Simulation Settings for Evaluating the impact of Broadcast Timers 56

3.5 Impact of Density on SAD and Service Sessions . . . . . . . . . 65

3.6 Average energy consumption vs. E-ZRP zone radius . . . . . . . 69

4.1 Network Density Values for square terrain. . . . . . . . . . . . . 98

xiv

LIST OF FIGURES

Figure Page

2.1 Service Discovery Architectures. . . . . . . . . . . . . . . . . . . 13

2.2 Taxonomy of Cross Layer and Application Layer Service DiscoveryApproaches based on Control Packet Overhead. . . . . . . . . . 33

3.1 NDP packet format (for ZRP). . . . . . . . . . . . . . . . . . . 40

3.2 IARP packet format (for ZRP). . . . . . . . . . . . . . . . . . . 41

3.3 IERP packet format (for ZRP). . . . . . . . . . . . . . . . . . . 42

3.4 Routing Zones and the Bordercasting Process. . . . . . . . . . . 42

3.5 APS packet format. . . . . . . . . . . . . . . . . . . . . . . . . . 46

3.6 Energy gains from using E-ZRP vs. APS. . . . . . . . . . . . . 52

3.7 Investigation of the impact of APS packet size on energy consump-tion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

3.8 Investigation of the overhead of Bordercasting vs. the overhead ofFlooding. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

3.9 Investigation of the impact of Broadcast Timers on service discov-erability and energy consumption in a static context. . . . . . . 57

3.10 Investigation of the impact of Broadcast Timers on service discov-erability and energy consumption in a mobile context. . . . . . . 58

3.11 E-ZRP: Avg. Service Session Duration PDF vs. Speed (Low-Medium SAD). . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

3.12 E-ZRP: Avg. Service Session Duration PDF vs. Speed (Medium-High SAD). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

3.13 Mean Service Availability Duration (SAD) vs. Speed. . . . . . . 61

3.14 Avg. Number of Service Sessions Discovered vs. Speed. . . . . . 63

3.15 E-ZRP: Service Duration Distribution vs. Density. . . . . . . . 64

3.16 Density impact on SAD. . . . . . . . . . . . . . . . . . . . . . . 65

3.17 Density impact on Service Sessions. . . . . . . . . . . . . . . . . 66

3.18 Channel and mobility impact on SAD. . . . . . . . . . . . . . . 67

3.19 Channel and mobility impact on Service Sessions. . . . . . . . . 67

xv

Figure Page

3.20 E-ZRP’s Zone Radius Impact on SAD. . . . . . . . . . . . . . . 68

3.21 E-ZRP’s Zone Radius Impact on Service Sessions. . . . . . . . . 68

3.22 Delay for out-of-zone services discovery. . . . . . . . . . . . . . . 70

3.23 The routing protocol design space for MANETs. . . . . . . . . . 72

3.24 Effect of IARP and NDP broadcast interval size on IARP and IERPoutgoing traffic. . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

3.25 Effect of the T parameter on the Percentage of Completed Servicesfor AVERT. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

3.26 Effect of the T parameter on the Energy Consumption per nodefor AVERT. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

3.27 Effect of the T parameter on σ for AVERT. . . . . . . . . . . . 81

3.28 Performance gains of AVERT against IZR and SPIZ. . . . . . . 81

3.29 Performance of SPIZ under different service request frequencies. 82

4.1 VANET with mobile servers. . . . . . . . . . . . . . . . . . . . . 92

4.2 Connection Duration vs. Density and Number of Hops. . . . . . 99

4.3 Connection Duration vs. Density and Speed. . . . . . . . . . . . 100

4.4 Impact of server capacities and node density on Profit Gains bycooperative servers using E-GAP vs. using GAP. . . . . . . . . 102

4.5 Impact of mobility on Profit Gains by cooperative servers usingE-GAP vs. using GAP. . . . . . . . . . . . . . . . . . . . . . . . 104

4.6 Loss in Total Profit when servers are non-cooperative (Server Ca-pacity = 5 units, Servers use the E-GAP (oracle)) . . . . . . . . 106

xvi

VITA AND PUBLICATIONS

2002 - 2008 PhD, Department of Computer Science, Athens Universityof Economics and Business

2000 - 2001 MSc, Department of Computer Science, Athens Universityof Economics and Business

1996 - 2000 BSc, Department of Computer Science, Athens Universityof Economics and Business

Academic Experience

• Research Assistant at the Mobile Multimedia Laboratory, AUEB, under

Prof. G.C. Polyzos, February 2000 to 2008.

• Systems and Network Administrator at the Educational Laboratory (CSlab)

of the CS Dept., AUEB, February 2000 to February 2001.

• Systems and Network Administrator at the Mobile Multimedia Labora-

tory, AUEB, February 2001 to 2008.

• Teaching Assistant, AUEB, February 2002 to 2008, in the following

courses: Wireless Networks and Mobile Communications (graduate and

undergraduate), Computer Networks (graduate), Topics in Multimedia

Systems (graduate) and Multimedia Technology (undergraduate) classes

under instructors Prof. George C. Polyzos, Prof. George D. Stamoulis

and Dr. George Xylomenos.

• Instructor in Computer Applications in the Continuing Education and

Training Center of AUEB, Fall 2003.

xvii

Professional Activities

• Reviewer for academic journals and conferences, including: IEEE Trans-

actions on Mobile Computing, European Transactions on Telecommuni-

cations (Wiley), Wireless Communications and Mobile Computing (Wi-

ley), Computer Communications (Elsevier), IEEE Wireless Communi-

cations Magazine, IEEE INFOCOM, ACM MOBICOM, IEEE GLOBE-

COM, IEEE ICC, IEEE VTC, IEEE LANMAN, IEEE WoWMoM, IFIP

Networking, IEEE/IFIP WiOpt.

• Technical Program Committee Member of IEEE WCNC 2008.

List of Publications

Journal Article

• Christopher N. Ververidis, George C. Polyzos, “A Routing Layer Based Ap-proach for Energy Efficient Service Discovery in Mobile Ad Hoc Networks,”Wireless Communications and Mobile Computing, Willey, 2008 (in press),DOI: 10.1002/wcm.618.

Magazine Article

• Christopher N. Ververidis, George C. Polyzos, “Service Discovery for MobileAd Hoc Networks: A Survey of Issues and Techniques,” IEEE Communica-tions Surveys and Tutorials, No 3, 2008 (in press).

Refereed Conference and Workshop Papers

• Christopher N. Ververidis, George C. Polyzos, ”AVERT: Adaptive SerVicEand Route Discovery ProTocol for MANETs,” in Proceedings of the 4th IEEEInternational Conference on Wireless and Mobile Computing, Networking and(WiMob 2008), Avignon, France, October 2008.

• Christopher N. Ververidis, George C. Polyzos, “Service Provision Optimiza-tions for Mobile Ad Hoc Networks,” in Proceedings of the IEEE InternationalSymposium on a World of Wireless, Mobile and Multimedia Networks (WoW-MoM 2008), Newport Beach, California, June 2008.

• Sotirios E. Athanaileas, Christopher N. Ververidis, George C. Polyzos, “Opti-mized Service Selection for MANETs using an AODV-based Service DiscoveryProtocol,” in Proceedings of the 6th Annual Mediterranean Ad Hoc Network-ing Workshop (MEDHOCNET 2007), Corfu, Greece, June 2007. (1 citation)

xviii

• George C. Polyzos, Christopher N. Ververidis, Elias C. Efstathiou, “ServiceDiscovery and Provision for Autonomic Mobile Computing,” in Proceedings ofthe 2nd IFIP Workshop on Autonomic Communication (WAC 2005), Athens,Greece. Springer Lecture Notes in Computer Science Vol. 3854 (LNCS), pp226-236. (2 citations)

• Christopher N. Ververidis, George C. Polyzos, “Extended ZRP: a RoutingLayer Based Service Discovery Protocol for Mobile Ad Hoc Networks,” inProceedings of the 2nd Annual International Conference on Mobile and Ubiq-uitous Systems: Networking and Services (MobiQuitous 05), pp. 114-123, SanDiego, California, July 2005. (7 citations)

• Christopher N. Ververidis, George C. Polyzos, “Routing Layer Support forService Discovery in Mobile Ad Hoc Networks,” in Proceedings of the 3rdIEEE International Conference on Pervasive Computing and CommunicationsPerCom 2005 - Pervasive Wireless Networking Workshop, PERCOM Work-shops 2005, Page(s):258 262, Kauai Island, Hawaii, 08-12 March, 2005. (10citations)

• Christopher Ververidis, Elias C. Efstathiou, Sergios Soursos, George C. Poly-zos, “Context-Aware Resource Management for Mobile Servers,” in Proceed-ings of the 10th HP-OVUA Annual Workshop, Geneve, Switzerland, 2003.

• E. Valavanis, C. Ververidis, M. Vazirgiannis, G. C. Polyzos, K. Norvac, “Mo-biShare: Sharing Context-Dependent Data & Services from Mobile Sources,”in Proceedings of the 2003 IEEE/WIC International Conference on Web In-telligence (WI 2003), Halifax, Canada, October 2003. (29 citations)

• C. Ververidis, E. Valavanis, M. Vazirgiannis, G. C. Polyzos, “An Architecturefor Sharing, Discovering and Accessing Mobile Data and Services: Locationand Mobility Issues,” in Proceedings of the IST-Location Based Services Clus-ter Workshop, Mykonos, Greece, 2002. (12 citations)

• Margaritis Margaritidis, Christopher N. Ververidis, George Xylomenos, GeorgeC. Polyzos, “A Differentiated Services QoS Scheme Preventing Malicious FlowBehaviour in Mobile Ad hoc Networks,” in Proceedings of the European Wire-less 2006 Conference, Athens, Greece, April 2006.

• Emmanuel A. Panaousis, Christopher N. Ververidis, George C. Polyzos, “Op-timizing the Channel Load Reporting Process in IEEE 802.11k-enabled WLANs,”in Proceedings of the 16th IEEE Workshop on Local and Metropolitan AreaNetworks (LANMAN 2008), Cluj-Napoca, Transylvania, Romania, September2008.

• C. Ververidis, G. C. Polyzos, “Mobile Marketing Using Location Based Ser-vices,” in Proceedings of the 1st International Conference on Mobile-Business2002, Athens, Greece, July 2002. (14 citations)

xix

Book Chapter

• Christopher N. Ververidis, George C. Polyzos, “Location Based Services inthe Mobile Communications Industry,” published in the Encyclopedia of E-Commerce, E-Government and Mobile Commerce, ISBN: 1-59140-799-0, IdeaGroup Reference, March 2006.

Papers with Poster Presentations

• Christopher N. Ververidis, George C. Polyzos, “Impact of Node Mobility andNetwork Density on Service Availability in MANETs,” in Proceedings of the14th IST Mobile & Wireless Communications Summit 2005, Dresden, Ger-many, June 2005.

• G. Chalkiadakis, S. Politis, C. Ververidis, G. C. Polyzos, “Mobile Multime-dia Portal,” in Proceedings of the 8th HP OpenView University AssociationWorkshop (HPOVUA 2001), Berlin, Germany, June 24-27, 2001.

xx

1

1. INTRODUCTION

This chapter serves as a general overview of the dissertation. Section 1.1

discusses the motivation behind this research. Section 1.2 presents the thesis

and key contributions of this dissertation and Section 1.3 presents the structure

of the dissertation.

1.1 Overview and Motivation

Mobile Ad Hoc Networks (MANETs) were introduced as a networking

alternative for cases where traditional infrastructure-based networks were un-

available either due to prohibiting costs (e.g. in remote and hard to reach

locations) or prohibiting situations (e.g. after a devastating disaster like an

earthquake). Initially the main application scenarios for MANETs, included

emergency response team operations and military missions. Maturing fur-

ther, MANET technology began to be considered also for cooperative working

environments, for group communications (e.g. conferences), for peer-to-peer

applications (e.g. gaming, file sharing etc.), for sensor networks and for vehicle-

to-vehicle communications [1]. The great potential of MANETs has attracted

the interest of many researchers and the evolution of MANET technology has

been remarkable. Initially, most of the research on MANETs has focused on

issues dealing with the connectivity between mobile nodes in order to cope

with the dynamism of such networks and the arising problems thereof. This

dynamism is due to the mobility of nodes, the wireless channel’s adverse condi-

tions and the energy limitations of mobile nodes, all of which lead to frequent

disconnections and/or node failures. These research efforts have led to the cre-

ation of a sound technical basis for dealing with the aforementioned problems

regarding node connectivity in MANETs (mainly through routing protocols,

link layer protocols etc.). However, solving the problems of connectivity alone

2 Introduction

was not sufficient for the adoption of MANETs. Since their basic role is to

allow mobile users to exchange data and use each other’s services, there was

also a need for developing architectures, mechanisms and protocols for Service

Discovery in such networks.

Service Discovery has been addressed in the context of wired networks in

the past. However, in the context of MANETs new challenges arise:

• Energy limitations of portable devices.

• Node mobility, affecting service availability.

• Frequent disconnections of the server, the client or intermediate nodes

breaking or changing the path and the service selection parameters.

• Channel variability, leading to significant communication characteristics

variability (data rate, delay etc.).

Considering the aforementioned challenges researchers reached a consensus

that traditional service discovery protocols for infrastructure-based networks,

were not suitable for MANETs and that radically different approaches should

be employed. Traditional approaches were developed based on the assump-

tion of a resource rich environment (in terms of devices’ energy resources and

available bandwidth), which was also relatively stable (no node mobility, no

large delay jitter, no fluctuating data rates). Hence, the majority of the de-

veloped solutions used centralized architectures with well-known service reg-

istries, which served as the rendezvous points between service providers and

service requesters (or clients). In these approaches, service providers pub-

lished their service description to the service registries, and nodes in need of

a service had to issue their requests to the service registries. In case of a

matching between a service request and a service description, the service reg-

istry would return to the requesting client the address of the service provider

that actually hosted the service. Techniques based on flooding or multicast-

ing were employed for service information (service advertisements or service

requests) dissemination over the wired network, since bandwidth and energy

Overview and Motivation 3

were not considered to be constrained resources. However, the major and

radically different requirements for service discovery approaches for MANETs

were: to allow nodes to discover services in the MANET without consuming

large amounts of energy (global flooding has prohibitive cost in MANETs),

to be distributed in nature (not relying on any fixed/well-known service reg-

istries), and also to flexibly adapt to a MANET’s changing conditions. The

aforementioned requirements served as our motivation for developing the en-

ergy efficient and highly adaptive service discovery protocols that are presented

in detail in Chapter 3.

In the same context of service provisioning in MANETs, we also identified

that the basic assumption made by researchers was that the service providers

were willing to provide their services for free. Removing this assumption we

have been led to cases where mobile service providers offer their services to the

rest of the nodes in the MANET, for profit. Assuming that a method for pay-

ments among the members of a MANET is in place, we focused on designing

a mechanism that would allow a mobile service provider to optimally select

which clients to serve. The basic problem in the selection of candidate clients

(assuming a finite serving capacity of the mobile server, which is less than the

demand) is how to select the clients that will return the maximum profits. Es-

pecially considering that clients may have different valuations on the provided

service and that the connectivity between a server and candidate clients is

not guaranteed (and hence payments are also not guaranteed), the problem of

optimal client selection when service interruptions are possible, becomes more

complex. We assume that clients will keep paying for the received service only

as long as they are connected to the service provider. Then, our proposed

mechanism is based on deriving probability estimates on the duration of the

connection between client and server, which are then used by a server in order

to compute which clients are expected to bring in the greatest profits. Our

approach is presented in Chapter 4.

4 Introduction

1.2 Thesis and Contributions of the Dissertation

In this dissertation we develop service discovery protocols integrated with

routing protocols. The purpose of integrating the routing process with the

service discovery process is to save energy by avoiding excessive and possibly

redundant information dissemination in the network. By integrating the two

processes a node is capable of acquiring information about the services pro-

vided in the MANET and at the same time (using the same message) to be

informed about the routes toward the service providers. In non-integrated, ap-

plication layer based service discovery protocols, a node would first discover the

available service providers and then would initiate the route discovery process

in order to find routes toward the discovered providers. Although integrat-

ing the service with the routing process has been proposed in the past, this

dissertation has the following original contributions in this field of research:

• We identify the need to derive the gains in energy consumption by inte-

grating service discovery into a hybrid routing protocol and comparing

it to an approach implementing a similar but application layer based

service discovery protocol. Related work in this area does not reveal the

realistic gains of the integrated approach, since comparisons are made

only against inefficient flooding based protocols implemented at the ap-

plication layer, which are not optimized for use in MANETs.

• We provide an analysis (experimental and theoretical)of the gains in

terms of energy consumption from using the proposed integrated service

discovery approach instead of a similar but application layer based ap-

proach (requiring the existence of a separate routing protocol). In our

theoretical model we take into account the characteristics of the IEEE

802.11 MAC layer and the related probability of collisions, while energy

consumption models for service discovery protocols found in the litera-

ture make the assumption of a perfect MAC.

Thesis and Contributions of the Dissertation 5

• Special focus has been given into environments where many different

nodes can provide the same service. It is important for service discov-

ery protocols running in such environments to be able to keep the client

connected to at least one such provider for the maximum time possi-

ble. Previous to this thesis, studies regarding service discovery protocols

and their performance in terms of service discoverability, only involved

the achieved service hit ratio metric. However, this metric is largely

dependent on the frequency of issued queries and does not measure the

availability of services in terms of time in a straightforward and accurate

manner. We propose the use of a new metric called Service Availability

Duration (SAD), which is defined as the length of time that elapses from

the moment the service is discovered until that time when access to the

service is lost, as a result of mobility or interference. It should be noted

that if the path to the original service provider is lost, but the protocol

has discovered another node providing the same service, then the service

is still considered ‘alive’. Only when all routes from a node to all the

available providers of the service are lost, this particular service is consid-

ered not to be available any more to that node. In this research effort the

SAD probability distribution has been extensively investigated through

simulations (both for E-ZRP and an application layer based service dis-

covery protocol) under different client and server densities and mobility

patterns, providing good insights into how these parameters affect the

performance of the service discovery protocol.

• We design and implement a hybrid integrated service discovery proto-

col, which differs from other similar protocols found in the literature in

that it allows nodes to adjust the frequency of advertising routes and

services in their vicinities based on monitoring traffic seen locally on

their interfaces. Experimental results show that the proposed protocol

outperforms all the other protocols belonging to the same class. For this

performance evaluation we introduce a novel metric combining both the

6 Introduction

service success ratio and the related energy consumption per successfully

invoked service.

Related to the profit maximization problem for service providers in MANETs,

the contributions of this dissertation are:

• We model the problem of profit maximization for capacity constrained

service providers in MANETs as a generalized assignment problem (GAP).

The innovative part of the proposed model is that we introduce into the

GAP model, estimations about the client to server connectivity.

• Our work on estimating client to server connectivity is also novel, since

similar work is done only on a per path basis and does not consider the

case of multiple paths between a client and a server or paths of changing

lengths during the lifetime of a connection.

1.3 Outline of the Dissertation

The remainder of this dissertation is organized as follows. Chapter 2 re-

views the literature on service discovery protocols for MANETs and especially

presents work on integrated service and route discovery approaches. Chapter 3

presents the two proposed hybrid integrated service discovery protocols and the

application layer based service discovery approach; the chapter also provides

detailed analytical and experimental evaluation of the proposed approaches.

Chapter 4 introduces our approach for modeling the problem of profit max-

imization for service provision in MANETS, then it presents our model for

estimating client to server connectivity and concludes by comparing through

extensive simulations the proposed approach against a similar approach not

employing estimations for client to server connectivity. Chapter 5 presents

open research issues in this field giving ideas and directions for future work,

and Chapter 6 provides the summary and the conclusions of the dissertation.

7

2. PREVIOUS WORK ON SERVICE

ADVERTISEMENT AND DISCOVERY

APPROACHES FOR MANETS

In this chapter we provide a comprehensive review of the state of the art

regarding service discovery approaches for MANETs. In the following sections

we make a categorization of service discovery approaches according to the

mechanisms they utilize and their features. We also highlight features that

are not fully developed yet and are addressed by the approaches proposed in

this thesis. The structure of the remainder of the chapter is as follows: Section

2.1 describes the basic service discovery architectures, Section 2.2 presents the

possible modes for service discovery and Section 2.3 highlights approaches

based on cross-layer optimizations.

Before continuing with the literature review regarding research efforts on

service discovery for MANETs it is worth briefly presenting the pioneering

service discovery approaches developed and adopted by the industry, namely

Jini1 [2], Salutation2 [3], UPnP3 [4], Bluetooth SDP4 [5], SLP5 [6] and Bon-

1Sun Microsystems.2Salutation Consortium members: Canon Inc., Consumer Electronics Association (CEA),Continental Automated Buildings Association (CABA), Fuji Xerox Co., Ltd., HewlettPackard, Infrared Data Association (IrDA), Institute of Certified E-Commerce Consultants(ICECC), International Business Machines, Konica Minolta Holdings, Inc., Kyocera MitaCorporation, National Institute of Standards and Technology (NIST), Oki Data Corp., Ri-coh Company, Ltd., Seiko Epson Corp, SISCO, Sun Microsystems.3UPnP Forum’s Steering Committee members: Broadcom Corporation, Cable TelevisionLaboratories, Inc., Intel Corporation, LG Electronics, Microsoft Corporation, Motorola,Inc., Nokia Corporation, Panasonic, Philips Consumer Electronics, Pioneer Research, Ri-coh Company, Ltd., Samsung Electronics Company, Ltd., Siemens AG, Sony Corporation,Thomson Inc.4Bluetooth Special Interest Group members: Agere Systems, Ericsson Technology LicensingAB, Intel Corporation, Lenovo, Microsoft Corporation, Motorola, Inc., Nokia, and ToshibaCorporation.5Internet Engineering Task Force (IETF) standard.

8Previous Work on Service Advertisement and Discovery Approaches for

MANETs

jour6 [7].

Jini

Jini is a service discovery architecture specifying how service discovery

and service invocation is to be performed among Java-enabled devices (a Java

Virtual Machine is mandatory). A central component of a Jini deployment

is a Lookup server. Lookup servers act as directories. They store services

published by service providers and they also reply to client queries. Lookup

servers announce their presence in response to requests multicasted by ser-

vice providers or clients. Service providers register their services with Lookup

servers by sending service objects along with their attributes. Service objects

are actually proxies written in Java and serve as interfaces for clients to ac-

cess a remote service. Clients receive those proxies (usually RMI stubs) from

Lookup servers upon successful match of their requests. Requests may include

the type of the requested service as well as other attributes. Jini also supports

leases, which means that services are registered for a specific amount of time

and if they do not get updated they are erased from Lookup servers. Another

characteristic of Jini is that through Java remote events, clients can be notified

upon changes in the status of a remote service. Finally, Jini provides security

through the Jini Security Framework.

Salutation

Salutation was primarily designed for home and enterprise environments.

Its architecture allows devices, services and applications to advertise their

capabilities, discover and access each other. Capabilities are expressed as at-

tribute sets. A basic component of the architecture is Salutation Managers

(SLM), who are responsible for storing attribute sets. Every device has a local

SLM with descriptions of its own services. However SLMs at different devices

communicate with each other via the Salutation Manager Protocol in order

to discover services available in other devices. Communication protocol inde-

6Apple Inc.

9

pendence is achieved through a transport independent layer between an SLM

and the Salutation Transport Manager (TM), which implements the transport

functionality. One SLM may have many TMs in order to operate over differ-

ent network technologies (e.g. IR, Bluetooth etc.). Service availability can be

checked by setting a local SLM to periodically query a remote SLM about the

needed service. Regarding security Salutation supports only password-based

authentication. For small footprint devices a less demanding version of Salu-

tation, Salutation-Lite [8], has also been developed.

UPnP

Like Salutation, UPnP was also proposed for use in small office and home

environments and mainly targets device and service discovery. Through UPnP,

devices first advertise their presence in a network and upon request they also

present their capabilities using XML for service descriptions. The basic enti-

ties considered in UPnP are control points (acting as service directories) and

devices. Control points are optional so if there is no control point, devices

may also listen to service advertisements directly. Service discovery in UPnP

is based on the Simple Service Discovery Protocol, which operates using HTTP

over multicast and unicast UDP. It is worth noting that UPnP can be deployed

only over TCP/IP networks and generally operates better over reliable net-

works. A special feature is that through AutoIP, UPnP devices automatically

receive an IP address even when a DHCP server is absent. Unfortunately due

to the extensive use of multicasting (multicasting is used both for service ad-

vertisements and service requests) UPnP cannot scale well7. Also it does not

support attribute-based querying for services. UPnP provides many mecha-

nisms for securing service discovery and access through UPnP Security [9].

Bluetooth SDP

Bluetooth SDP is a service discovery protocol for Bluetooth enabled de-

7Scalability problems when using multicasting in MANETs are discussed in more detail insection 2.1.2


MANETs

vices. Bluetooth SDP addresses only service discovery and does not address

service advertising, service caching in registries or service access. Every service

is described by a service record consisting of a set of attribute-value pairs each

of which describes a service characteristic. Bluetooth SDP defines two meth-

ods for discovering services, namely ‘service searching’ and ‘service browsing’.

With the first method a client formulates a query containing desired service

attributes and those are matched against service records at the provider and

the result is returned. The latter method allows a client to send a generic

query and get a list of all services of a specific provider. We should note that

Bluetooth SDP supports only 1-hop discovery and hence its discovery capa-

bility is limited to the immediate proximity of a device.

SLP

The Service Location Protocol (SLP) is an IETF standard and has been

embedded in many commercial products (by Hewlett Packard, IBM etc.). SLP

addresses only service discovery and leaves service invocation unspecified. Ser-

vice descriptions consist of unique URLs (for locating the service in the net-

work) and a set of attribute-value pairs. Clients may query for services using

their type or some combination of their attributes utilizing SLP’s capability

of substring matching. SLP also allows grouping services in scopes. Service

browsing is also allowed if a client requests to see all available services. SLP

can work in a totally distributed manner using only User Agents (UA) on

client devices and Service Agents (SA) on service providers. Communication

among them takes place through multicasting. If directories exist, then they

are represented by Directory Agents (DA). In the directory-based operation

of SLP when a DA enters the network it multicasts a beacon and any SA that

hears it must register its service to this DA. UAs that receive this message

unicast their queries to the DA. If no DA is present UAs multicast queries and

all receiving SAs with matching descriptions respond using unicasts. Regard-

ing security, SLP provides only a PKI-based mechanism for signing service

advertisements. Finally, we should note that a more lightweight version of

11

SLP, called SLPManet, was proposed in [10], excluding features like optional

SLP messages, DAs and authentication.

Bonjour

Bonjour is a technology developed by Apple to provide service and device

discovery among computers, electronic appliances and other networked devices

(e.g. printers, faxes etc.). Bonjour runs over the IP protocol and also has the

capability of automatically assigning IP addresses to networked devices, even

without the help of a DHCP server. Bonjour’s core is a service discovery

protocol entirely based on the Multicast DNS Service Discovery - MDNS-SD.

Actually MDNS-SD extends MDNS [11] so that hosts in an ad hoc network can

resolve, in addition to host names also service names to IP addresses without

relying on DNS servers. In MDNS-SD clients multicast their DNS-like queries

specifying (as defined in DNS-SD [12]) the service type they are looking for, the

domain where the service resides and the preferred communication protocol.

Service providers respond to those queries by DNS service records. However, a

new provider coming into the network may make a multicast announcement so

that other devices become aware of its presence. The service records are cached

on client devices for a limited time and if not updated (by querying again) they

are deleted. However, multicasting everything creates a significant amount of

traffic. Bonjour tries to address this by employing “exponential back-off” for

increasing the gap between queries and announcements in order to minimize

traffic while keeping the user’s view as fresh as possible [7].

All the aforementioned approaches were mainly designed for administered

networks (even if ad hoc), some requiring fixed-well known directories, others

making extensive use of broadcasting and multicasting (hence not scaling for

large ad hoc networks) and others not supporting mobility. However, they have

served as a solid base and source of inspiration for developing new protocols

oriented to pure ad hoc environments.


MANETs

2.1 Service Discovery Architectures

Regarding service information dissemination, there are three basic archi-

tectures that a service discovery approach may adopt (see Figure 2.1). We

will present each of them and refer to representative approaches found in the

literature.

2.1.1 Directory-based Architectures

In this type of architecture there are three possible roles for a mobile node.

A node can be a server (service provider, offering one or more services to other

nodes), a client (service requester, requesting services from other nodes) or a

service directory (facilitating communication between providers and clients).

Service providers register their services to service directories and service re-

questers are informed about the available services in the network only through

these directory nodes. A directory can be implemented as centralized (hosted

by a single node) or can be distributed among several nodes. Centralized

approaches were primarily adopted by service discovery protocols in wired

networks or in wireless local area networks where one or more fixed hosts take

up the role of a directory (e.g. UDDI [13]). A simple centralized directory,

however, is not a good solution for an ad hoc network, since no node is always

reachable. A centralized directory also represents a single point of failure and

is not well suited for such volatile environments. Scalability is also another

problem, since in MANETs the nodes are resource-poor and a single node act-

ing as a directory would not be able to handle responses for a large number of

nodes. Distributed directories are thus more suitable for MANETS.

A basic question is whether global service discovery is to be provided (i.e.

to be possible for every node to learn and invoke any service provided in the

ad hoc network). One approach is to use full replication for directory nodes

in order for every directory to store all services available in the MANET,

irrespectively of their location.

Service Discovery Architectures 13

Fig

.2.

1.Ser

vic

eD

isco

very

Arc

hitec

ture

s.


MANETs

A classic distributed directory approach is Jini, where a few nodes, named

Lookup servers, act as directories. However, there is no communication among

Lookup servers and it is at the discretion of service providers to publish their

service to more than one directory node and keep them updated. In the case

that automatic replication is not provided, a service may be known only locally,

around the directory node that originally hosts it and remote MANET nodes

will not be able to easily discover it. Global discovery is hence not supported,

since services are advertised only in the area where the Lookup servers reside.

In more elaborate distributed directory approaches, nodes acting as direc-

tories are in constant communication with each other to disseminate and also

replicate service information among them. Such approaches are based on pro-

tocols that create and maintain a backbone of directory-enabled nodes. For

example in [14] a backbone of directory nodes is formed using a Minimum

Dominating Set algorithm. Servers advertise their services to one or more

members of the backbone. However, despite the fact that service replication is

not inherently provided, global discovery is possible since backbone members

disseminate to each other service discovery requests that could not be satisfied

locally. This way a service requester and a service provider connected to the

opposite edges of the formed backbone can still discover each other and com-

municate. A better way to forward requests to neighboring backbone members

(instead of doing it randomly) was proposed in [15]. There, backbone members

frequently exchange directory profiles guaranteeing that service requests are

forwarded to nodes that are likely to cache the description of the requested

service.

An alternative to the aforementioned backbone based approaches for im-

plementing distributed directories are clustering approaches. A representative

technique is “Service Rings” [16]. In Service Rings a number of clusters are

formed. Each cluster (called a ring) of service providers is formed based on

physical proximity and semantic proximity of the descriptions of provided ser-

vices. Every ring has its own Service Access Point (SAP), which is responsible

for handling service registrations and service requests (operating as a direc-


tory). SAPs also communicate with each other and exchange summaries about

all the services they are aware of in their own ring. This way higher-level rings

are also formed iteratively. Global discovery is possible since if a node’s re-

quest cannot be satisfied by its local SAP, then this SAP forwards the request

to neighboring SAPs (and eventually to higher level SAPs) that are possibly

capable of satisfying the request based on the service summaries they have

previously sent. A similar approach is also adopted in [17] with the difference

that the hierarchy of clusters is strictly dependent on a common service on-

tology. At the bottom level of this hierarchy, clusters group devices offering

services described by the same leaf term of the ontology and being within radio

range of each other. Moving up the hierarchy, every level consists of groups of

clusters of their respective lower level. The higher the level, the more general

the semantic descriptions become (always in alignment with the generaliza-

tion of categories performed as we move up the ontology tree). In [18] each

cluster groups nodes with similar mobility patterns. In each cluster one of

the nodes (called clusterhead) stays awake permanently and answers discov-

ery requests. The rest of the nodes periodically wake up to provide the actual

services and also to inform the clusterhead about their presence and services.

The clusterheads are re-elected periodically to avoid draining a single node’s

battery.

Another solution to the global discovery problem, when replication is not

provided (either by servers or directories), is to use Distributed Hash Table

based techniques along with location information. Such approaches are de-

scribed in [19] and [20]. The network topology is divided into geographical

regions, where each region is responsible for a set of keys representing the ser-

vices of interest. Each key is mapped to a region based on a hash-table like

mapping scheme. A few elected nodes within each region are responsible for

storing these keys (in [20] all nodes inside the region store these keys), thus

acting as directories. Global discovery is possible since a node requesting a

service, uses the same hashing function (as the one used by service providers)

and finds the directory-location where its description is stored. The service


MANETs

request is then routed, using this location information, toward that directory.

Location information is also used in [21] for creating clusters of nodes based

on physical proximity. Every cluster has a gateway, which is responsible for

handling routing and discovery requests and for storing service descriptions

from nodes located within its region. Inter-gateway request forwarding is also

possible for global service discovery and is done on a region-covering basis

(actually requests are routed to neighboring regions, where another gateway

will be present and will try to answer them). This approach however differs

from other backbone or cluster based approaches in the sense that cluster lead-

ers (gateways in this case) are elected or de-selected automatically based on

location information and do not need to keep contact with each other.

The election of nodes for taking the role of a directory or for participating

in a directory structure (backbone) is a very crucial issue for a directory-based

service discovery approach. In [17] and [16] an election mechanism is not

provided, while in [19] a directory node election is performed randomly. In

more elaborate approaches, like [14] and [15], criteria like average packet loss

rate, effective degree (for connectivity to neighbors) and capacity are used to

choose one among a set of candidate nodes.

It is important to note here that directory based approaches imply ad-

ditional communication costs in the network for maintaining the directory

structure and also for exchanging data among the members of a distributed

directory for preserving service consistency and for replicating service infor-

mation. If maintenance and consistency procedures are not well tuned, then

either too much traffic will be generated, causing congestion and hence render-

ing the whole MANET useless, or inconsistencies in service information and

directory structure (due to insufficient updating) will degrade the performance

of the service discovery process.


2.1.2 Directory-less Architectures

This type of architecture differs from the previous type in that there are

no service directories to mediate communication between service providers and

service requesters. It is much simpler from directory-based architectures since

there is no need for directory selection and maintenance mechanisms. Service

providers broadcast service advertisements and service requesters broadcast

service requests. Both processes may take place at the same time in the

network. In the early approaches of this type only servers could reply to

service requests. Later, intermediate nodes (located along the paths between

servers and requesters) were also allowed to reply to service requests based on

the information they had cached locally by overhearing past server replies.

A basic problem in those non directory-based approaches is how to deter-

mine the frequency of service advertisements in order to reduce network load

and avoid redundant transmissions. Scheduling and prioritization was one of

the first techniques proposed to deal with the problem. For example in [22]

servers periodically broadcast service advertisements to their 1-hop neighbors.

These advertisements contain services provided locally by the sending node

and also services that the sending node has learned from its neighbors, which

are then stored as service records in the receiving node’s local cache. Servers,

whose services are about to expire8 or have expired, are assigned a greater

probability to make the next broadcast. An exponential back-off algorithm reg-

ulates the periodicity of broadcasts depending on server priority and changes

in the network (i.e. new servers). A provider postpones advertising its services

and backs-off for a fixed amount of time if it receives an advertisement that

contains its own services (this means that nodes in its vicinity are aware of

its service and up-to-date). Very close to this concept is also the mechanism

proposed in [23] where a provider listens to other’s broadcasts and when it is

its turn to broadcast, it only broadcasts service information (if any) that has

8Each service record has a Time To Live (TTL) field. This TTL continuously decreaseswith time until it reaches zero (except if a new advertisement is received and the TTL isrefreshed). When the TTL becomes zero the corresponding service is considered expiredand its record should be deleted from the node’s cache.


MANETs

not expired and has not been seen recently in previous broadcasts. A similar

approach is employed in [24] where providers aborts from sending a reply to

a received service request if it hears another provider’s answer to the same

request.

Another way proposed for lowering the load imposed on the network by

flooding service advertisements and/or service discovery requests was to use

multicasting. In [25] authors assume that the network supports multicast-

ing so that servers can multicast their advertisements on a fixed multicast

group and service requesters can multicast their service queries. It is assumed

that the multicast messages reach the whole network, however authors do not

make any specific comments on the multicast routing protocol used. It is

known from the literature [26] that multicast protocols tailored to MANETs,

and which are based on maintaining a tree topology, experience unacceptable

packet delivery ratios (≤35%, when flooding can achieve 99% packet delivery)

even for networks of low to medium mobility (≥3km/h). Mesh-based multi-

cast protocols seem to cope better with mobility (≥60% packet delivery ratio)

due to the existence of more alternative paths for routing multicast packets,

but suffer from exponential growth in control traffic due to mobility increases.

High control traffic is also created when the mesh-based protocol requires the

construction of one multicast-mesh per sender. Moreover, it is shown that

for such networks, flooding approaches can be more efficient both in terms

of overhead and packet delivery ratios. The main reason for the performance

problems of multicasting in MANETs is that mobility causes rapid changes in

network topology, which makes it difficult for a host to maintain timely and

accurate multicast-related state information. To elaborate more, maintaining

accurate multicast state requires both increased storage capacity (for buffering

packets, or simply for storing the state) and also extra control packet overhead

for updating state information.

Covering the whole network using either broadcasting or even multicasting

techniques is very costly. This is why many approaches use various other

techniques:


• Advertisement range bounding/scoping.

• Selective, probabilistic and intelligent (advertisement/request) forward-

ing.

• Peer-to-peer (P2P) information caching.

• Intermediate node responding to service requests.

Several approaches taking advantage of these ideas and techniques are de-

scribed next. Many approaches use an advertisement range measured in num-

ber of hops specifying when the advertisement message will be dropped. In

the Group-based Service Discovery (GSD) protocol [27] and the Alliance-based

Service Discovery (Allia) protocol [28] such a technique is adopted. However,

in order to allow most of the nodes in the network to eventually become aware

of the advertised services, these two approaches also include a technique called

peer-to-peer (P2P) information caching for nodes to merge services heard by

others and re-advertise them (using again a range) along with their own ser-

vices. Eventually, most nodes will become aware of all services in the network,

but at a lower cost since service merging is performed.

The two approaches mentioned above also employ selective forwarding of

service requests to further reduce the load of service discovery. Selective for-

warding means that a node receiving a service request that it cannot fulfill will

forward the request only to those of its neighbors that are known to host the re-

quested service, or similar services. Besides selective forwarding [29] and [30]

propose that the overhead of GSD can be further reduced by an additional

mechanism called Broadcast Simulated Unicast (BSU). Instead of forwarding

the same query in unicasted packets toward selected neighbors, with BSU the

message is forwarded once using broadcast. Only the selected neighbors will

further process this packet since it contains a list with the intended recipients.

If a neighboring node receives such a packet and does not find itself in the

receiver’s list, it will just discard the packet. However, a significant amount of

bandwidth will have been saved.


MANETs

The P2P information caching technique is also used in [31] along with

probabilistic forwarding instead of selective forwarding. In this case a node

receiving a service request that it cannot fulfill, forwards it with a probability

that decreases with the number of hops that the request has already traveled.

Intelligent forwarding can also be used for spreading service advertisements,

as done in [32]. In [32] every node continuously monitors its 2-hop neighbor-

hood. In order to avoid duplicate packet forwarding and also to cover every

node, each server initially sends its advertisements only to those nodes in its

1-hop neighborhood (called brokers) through which all its 2-hop neighbors can

be reached. In the next advertising round new brokers will be formed for the

2-hop neighbors of the originating server thereby expanding the service cov-

erage in the same way (by forwarding the advertisement only to a subset of

their 1-hop neighbors through which all their 2-hop neighbors can be reached).

Costly broadcasting is hence replaced by a few unicasts in every advertising

round.

Another way to reduce the load imposed by service discovery requests and

advertisements is to allow intermediate nodes to respond to service requests.

Intermediate nodes may have been informed about the existence of some ser-

vices either by receiving and forwarding service advertisements, or because

they themselves have requested these services in the past. Hence a service

request may not need to travel all the way to the service provider, since it can

be answered by an intermediate node located closer to the service requester.

In [33] intermediate nodes are allowed to answer service requests. However

in order not to decrease the number of discovered services the authors propose

that intermediate nodes must be informed of all the services matching the

issued requests. This is because dropping requests at intermediate nodes that

already know one out of many matching services may decrease the service

discoverability of the protocol. It is proposed that when answers come to a

service requester from different servers and different paths, intermediate nodes

and servers along those paths are updated to become aware of all the services

that were returned to the requester. Thus, when they receive another request


for the same kind of service from another node, any server or intermediate

node will be able to reply with all the matching services they became aware

of by informing each other in previous requests.

Finally in order to totally avoid broadcasting or multicasting and the asso-

ciated costs, the use of location information has also been proposed in [34] for

sending service advertisements and service requests. In this protocol, servers

periodically send their advertisements along cross-shaped trajectories. At each

node in the trajectory, a backwards pointer is set up establishing paths leading

to the service provider. Any service requester need simply send a query along

a path that intersects with the advertisement path. Requests are answered

by nodes at the intersection (intermediates) of the advertising and requesting

trajectories.

2.1.3 Hybrid Architectures

In these architectures service providers register their services with service

directories if they locate any in their vicinity (if not they simply broadcast

service advertisements). Service requesters send their queries to the service

directories they are aware of. If they are not aware of any service directory,

they broadcast them to the whole network. Service replies may come both

from service providers and service directories.

2.1.4 Critique of Service Discovery Architectures

Despite the multitude of publications on each of the service discovery ar-

chitectures described in this section, researchers have not reaached a general

consensus on which architecture is better. The basic criteria for evaluating the

effectiveness of service discovery architectures are service availability, messag-

ing overhead and latency. The reason that makes it difficult to come to a gen-

eral conclusion on which architecture is more suitable for a MANET is that it

depends on many factors, some of which relate to the MANET’s characteristics


MANETs

(e.g. server and client density, node mobility and service request frequency)

and others to tunable parameters of the discovery architecture employed (e.g.

flooding / broadcasting scopes, directory node density, service registration and

announcement frequency). For example, for a MANET with a high degree of

mobility and a low service request frequency, a distributed architecture without

caching could prove to be more efficient than a directory-based architecture,

since the latter would either suffer from stale service information in directories,

or would demand much overhead for maintaining service information integrity

and coping with mobility. However, if the same MANET of the previous exam-

ple faced a very high service request frequency, a directory-based architecture

could be more efficient. In this case, a directory-less architecture would de-

mand that clients frequently flood the whole network with their queries. This

traffic would most probably outweigh the traffic created in a directory-based

architecture for maintaining consistent directories and for unicasting queries

to directories only, instead of flooding them to the whole network. The above

would hold in the general case where both architectures’ parameters are tuned

similarly. However, there exist certain values for the tunable parameters that

could affect an architecture’s performance so severely, that a seemingly bet-

ter/matching architecture could prove to be worse than the other.

Generally, none of the three architectures can outperform the other two in

all of the above mentioned performance criteria. Even the underlying routing

protocol (especially when integrated with the service discovery process) may

have an impact on the performance of a service discovery architecture as shown

in [35] and [36]. In [35] simulations show that in proactively routed MANETs

the hybrid architecture outperforms in terms of service availability the other

two architectures. However, the directory-less architecture outperforms the

other two architectures in messaging overhead. A more recent work [36] for

reactively routed MANETs, however, shows by simulations that a directory-

less architecture may outperform a hybrid one, both in terms of higher service

availability and lower message overhead, while having almost the same delays.

Service Discovery Modes 23

2.2 Service Discovery Modes

Irrespectively of the service discovery architecture there are three possible

ways or modes of operation for a service requester to acquire service informa-

tion: the Reactive Mode, the Proactive Mode and the Hybrid Mode.

2.2.1 Reactive Mode

In this mode a service requester issues a query in an on demand basis to

directory nodes or directly to service providers. There are many variations of

this mode, some of which have been discussed in the service discovery archi-

tectures section of this chapter. To name a few options for service requesters,

they may choose to set a limited TTL so that they do not flood the whole

network when there are no directories. They may expand their search step-

by-step by gradually increasing the hops that a service request is allowed to

travel. They may utilize mechanisms to selectively forward their requests to

specific neighbors only, instead of sending them to every neighbor. They also

may unicast, multicast or broadcast a query to one or more directories or to

one or more servers.

2.2.2 Proactive Mode

In this mode service providers advertise their services (either to service

directories or directly to potential service requesters) on discrete time inter-

vals. The same holds for advertisements originating from directory nodes.

Servers and directories have also the option to use ranges for the advertise-

ments instead of flooding the whole network. A basic tunable parameter is

how frequently those advertisements should be sent, since it greatly depends

on the level of dynamism of the MANET (mobility, failures, congestion).


MANETs

2.2.3 Hybrid Mode

In this mode both proactive and reactive communication between service

requesters, service providers and service directories is possible. For exam-

ple servers may proactively advertise their services to service directories, but

clients may issue requests to service directories only reactively (on demand).

As explained in [37], several strategies can be employed by clients and servers

in order to discover services. For example in a “greedy” strategy all servers

may advertise services to all nodes and all clients query all nodes in the net-

work in order to discover services, while in a “conservative” strategy servers

may advertise services to a random set of nodes and clients may also query

only a random set of nodes. Other more complex strategies include incremen-

tal increase of the advertisement and querying sets and memorizing previously

queried nodes in order to avoid querying them again in next rounds, for the

case that a service has not been discovered in the previous round. As expected,

authors conclude that “greedy” strategies offer higher success rates and lower

delays than “conservative” strategies, but produce much higher overheads.

However, they also note that depending on factors such as success rate re-

quirements, delay tolerance, overhead tolerance, node memory constraints,

network dynamism (expressed as mobility and underlying routing protocol -

proactive or reactive) the preferred strategy is different.

2.2.4 Critique of Service Discovery Modes

Mohan et al. in [38] present a simulation analysis of the proactive and

reactive modes in their simplest form, which involves global flooding. Accord-

ing to these results the proactive mode outperforms in terms of latency and

overhead the reactive mode when the number of servers is significantly lower

than the number of clients. The opposite happens when the available servers

are significantly more than the clients in the network. A hybrid scheme is pro-

posed to give on the average better results in terms of overhead and latency

Cross Layer Service Discovery 25

for most combinations of number of servers to number of clients. This hybrid

scheme is enhanced by a mechanism allowing servers (respectively clients) to

determine network congestion before deciding to send an advertisement (re-

spectively a query). If the congestion is over a given threshold, the senders

(either clients or servers) exponentially back-off in order to avoid congesting

the network further. However, careful selection of this threshold is difficult in

such a dynamic environment as a MANET.

In a MANET, with a proportion of clients to servers close to 50% the

preferred approach depends on the actual demand for discovering services.

It is intuitive that in such MANETs, if service discovery requests are rather

rare, a reactive approach would be more efficient (at least in terms of control

overhead) than a proactive or hybrid approach. Of course in cases where ser-

vice discovery is performed frequently, a proactive scheme would prove to be

preferable (provided that services are advertised in appropriate time intervals,

matching the demand). This is also backed by experimental results in [39],

where the authors provide a thorough analysis (both theoretical and experi-

mental), on the performance of reactive and proactive service discovery modes

investigating the impact of several factors (mobility, traffic patterns, message

aggregation, use of caching). They conclude that the actual service context is

what determines which mode is most efficient and that a hybrid mechanism

able to adapt to service demand is the preferred choice. They especially in-

vestigate the impact of the underlying routing protocol and also its coupling

with the service discovery process on the performance of each service discov-

ery mode. This coupling leads to a special case of service discovery protocols,

namely the cross layer service discovery protocols, which we examine in the

following section.

2.3 Cross Layer Service Discovery

In contrast to traditional application layer based service discovery, there are

approaches that employ cross layer techniques in order to benefit from infor-


MANETs

mation available at lower layers of the protocol stack. Most of these approaches

are based on integrating the routing process with the service discovery process.

The motivation for integrating routing and service discovery stems from the

fact that any service discovery protocol implemented above the routing layer

will always require the existence of some kind of routing protocol for its own

use. Hence, two message-producing processes must coexist: the first one com-

municates service information among service providers and service requesters;

the second one communicates routing information among them. As a result,

a node is forced to perform multiple times the battery-draining operation of

receiving and transmitting (control) packets. Cross layer service discovery

exploits the capability of acquiring service information along with routing in-

formation (from the same message) by piggybacking service information onto

routing messages. This way, transmissions of service discovery packets at the

application layer are avoided and energy is saved. Henceforth, we will refer to

those cross layer service discovery protocols as integrated protocols.

The idea of providing routing layer support for service discovery was first

introduced by Koodli and Perkins in [40]. They argue that for proactively

routed MANETs, a service reply extension added to topology updating rout-

ing messages is enough for providing both service discovery and route discov-

ery concurrently. In reactively (or on-demand) routed MANETs, the service

discovery process follows the traditional route discovery process by using its

message formats for route requests (RREQ packets) and route replies (RREP

packets) extended to carry also a service request, or a service reply, respec-

tively. In [41] the authors have extended the Ad hoc On-Demand Distance

Vector (AODV) routing protocol with service discovery functionality and have

experimentally compared it with NOM [42] (a pure application based service

discovery protocol for MANETs). Their findings show that the integrated pro-

tocol produces 30% to 50% less control overhead and has 2 to 7 times lower

service acquisition latency than the application layer based protocol (depend-

ing on simulation parameters). The authors of [43] and [44] have provided


additional extensions to the integrated AODV protocol to also support QoS

aware service selection.

In [45] AODV and Dynamic Source Routing (DSR) are extended (named

SD-AODV and SD-DSR) to support service discovery and are compared in

terms of traffic overhead against an application layer service discovery pro-

tocol based on global flooding and also against a protocol with global knowl-

edge. The global knowledge protocol uses an oracle to determine which service

providers are available in the network and to select the closest one for commu-

nication. Once again it is experimentally shown that both integrated protocols

outperform the application layer based approach under any network density,

request frequency and speed. SD-DSR is also shown to be more efficient than

SD-AODV, since it allows its nodes to update their routing information and

maintain a consistent view of routes and services by overhearing other nodes’

transmissions. Regarding the global knowledge application layer service dis-

covery protocol, when using AODV at the network layer, SD-AODV presents

comparable performance, but requires that services are cached for short peri-

ods of time so that stale service information (e.g. due to node movement) is

erased and service provider selection is nearly optimal (i.e. the closest server

must be selected). SD-DSR compared to the global knowledge protocol with

DSR at the network layer performs slightly better or worse depending on net-

work conditions. The global knowledge protocol always tries to contact the

closest provider. If the path to the closest provider is unreliable the global

knowledge protocol will keep trying several times before choosing the second

nearest provider. In DSR a provider is tried only once and if there is no

response the second nearest provider is contacted. The weak point in the re-

search presented in [45] is that the authors have chosen to compare integrated

protocols with an application layer based protocol that uses global flooding.

This flooding protocol suffers from the broadcast storm problem, rendering it

practically useless in medium to high density networks. It is more interesting

to compare the integrated protocols against application layer protocols tai-


MANETs

lored for MANETs, such that the actual performance gains from combining

routing and service discovery into one process are shown.

A similar study on DSR and AODV integrated protocols was conducted

in [46], where the authors propose the use of a module at the link layer,

which is responsible for assembling and disassembling packets, to embed service

information from an application layer service discovery protocol and at the

same time routing information from a network layer protocol. This way routing

protocols are not extended or modified in any way. The authors experimentally

show that cross layer service discovery using the link layer module and AODV

(respectively DSR) produces about 15% (respectively 90%) less traffic than

an application layer based service discovery protocol (using global flooding)

without the module and using AODV (respectively DSR) at the network layer.

In [47] in order to compare a reactive routing service discovery protocol and

a proactive routing and service discovery protocol, DSR and the Destination-

Sequenced Distance Vector protocol (DSDV) are extended to provide service

discovery functionality. Those approaches are compared against SLP, imple-

mented at the application layer. The extended DSR protocol proves to have

the least messaging overhead among the three, with second best the extended

DSDV protocol. However, SLP with its extensive use of multicast messages

(flooded to the network) was not designed for MANETs and is more suitable

to managed/administrated Wireless Local Area Networks (WLANs) rendering

its comparison to integrated protocols unfair.

DSDV is not the only proactive routing protocol extended with service

discovery functionality. In [48] and [49] researchers have also extended the

Optimized Link State Routing (OLSR) proactive routing protocol to support

service discovery, but no comparisons with other integrated or application layer

protocols are presented.

Another category of routing protocols, namely multicast routing protocols

has been used for service discovery. The authors in [50] and [51] extend the On-

Demand Multicast Routing Protocol - (ODMRP) to support service discovery

functionality. According to this approach each server and its possible clients


form a multicast group. Each server multicasts an advertisement encapsulated

in an ODMRP join query packet. Any client, interested in the advertised ser-

vices, stores the advertisement and sends a service awareness reply encapsu-

lated in an ODMRP join reply packet. Once the multicast group between a

server and all interested clients has been formed, the server will re-send ad-

vertisements only if its service changes. Otherwise it waits for explicit queries

from clients. In [14] the authors show that an AODV-based integrated proto-

col performing service discovery using anycasting9 is much better in terms of

delay and control packet overhead compared to an ODMRP-based integrated

protocol constructing requester-based multicast meshes for performing service

discovery. However, the ODMRP-based integrated protocol has a significantly

higher service hit ratio than the AODV-based integrated protocol especially

in highly mobile environments. However, according to studies on the perfor-

mance of ODMRP (e.g. [52] and [53]), ODMRP does not scale as the number

of senders increases since it builds one multicast mesh structure per sender

and has to periodically flood join packets to maintain the mesh connectivity.

A more radical approach is adopted in [54], where the authors do not

integrate service discovery with a well-known protocol, but build their own

multicast routing protocol named HESED, which also supports service discov-

ery. In HESED multicast routing is used both for service requests and service

responses. Intermediate nodes locally cache service reply information but do

not use it to reply to requests. When requesting a service, a client first searches

its local cache, and if it finds a matching service record, it calculates the prob-

ability that the path to the service is still valid. The routing part of HESED

uses a beaconing mechanism allowing nodes to know their 2-hop neighbors.

9In anycasting, a virtual server node is defined that is uniquely identified by the IP anycastaddress, for which only the actual server nodes have routing entries. In the anycast-1 servicediscovery scheme, every node receives the service advertisements from the different serviceinstances and stores only one single entry in its routing table, the one toward the neighborwhich sent the advertisement with the smallest hop count value. Therefore, a query isalways sent to the neighbor that is the closest to any service instance. The major drawbackof this simple anycast implementation is its lack of robustness. Due to its single entry perservice, it fails to deliver a query when any of the links on the path to the service becomesunavailable.


MANETs

Depending on the change rate of their neighbors, nodes calculate the proba-

bility that a route to a server is valid. The proposed protocol shows significant

gains (up to 80% less delay and control packet overhead) over a flooding based

protocol implemented at the application layer. This is especially true for high-

density scenarios, since HESED employs an intelligent forwarding mechanism

similar to the one proposed in [32].

Finally, another benefit provided by cross layer approaches is the exploita-

tion of routing information for restoring service sessions, or making handovers

from provider to provider. This idea was implemented in [55], where the

authors integrated the GSD protocol with routing. The integrated protocol

additionally provides automatic redirection to another service provider when

the route to the selected service provider fails. Comparisons of the integrated

protocol with the simple GSD protocol over AODV showed increased service

success ratio of up to 50% for the integrated protocol.

In opposition to the aforementioned integrated approaches using either a

proactive or a reactive routing protocol, in this thesis we investigate the per-

formance of integrated protocols based on hybrid routing protocols. In hybrid

routing protocols each node proactively advertises the routes and services it

is aware of by sending control messages to its neighbors up to a fixed number

of hops away (this is called the node’s zone). Information for routes or ser-

vices outside this zone may be gathered only upon request (reactively). We

have developed two protocols, namely the Extended Zone Routing Protocol

(E-ZRP) and the Adaptive SerVicE and Route Discovery ProTocol (AVERT).

E-ZRP is based on the Zone Routing Protocol (ZRP) for its operation.

Similar work includes the CARD protocol proposed in [56], which is based on

ZRP and was proposed as a resource discovery protocol suitable for very short

transactions in very large scale MANETs (from a few hundred to tens of thou-

sands of nodes). In the CARD protocol, the central idea is to not to use the

bordercasting mechanism of the ZRP protocol but a mechanism that allows

every node to maintain connections to nodes (named contact nodes) located

outside its proactive routing zone. The contact nodes are utilized in order to


make efficient searches in case resources are not located inside the proactive

zone of a node. They have compared their contact node-based mechanism to

bordercasting and to flooding. However, due to the cost of maintaining connec-

tions to distant contact nodes the CARD protocol showed inferior performance

to bordercasting (and also to flooding) for small to medium scale MANETs

(up to a couple of hundreds of nodes). The performance was measured in

terms of control packet overhead and was shown that for low to medium sized

networks (less than 250 nodes) and medium to low per node resource request

traffic for resources out of zone, the contact node-based mechanism is worse by

more than 20% than bordercasting or flooding. Also the authors do not report

on packet sizes and formats regarding all the aforementioned protocols neither

do they employ a realistic energy consumption model (since they assume no

collisions in the MAC layer), as we do in our analysis in chapter 3. Further-

more, they mostly focus on very large scale MANETs and on the process of

discovering distant resources for engaging very short transactions, while we fo-

cus on practical MANETs of low and medium scale, extensively exploring the

performance gains regarding intra-zone service discovery and also the benefits

of using E-ZRP instead of a similar service discovery protocol implemented in

the application layer.

As already mentioned, E-ZRP is based on ZRP, which assumes that all

nodes have the same zone radius, and this is fixed according to the character-

istics of the MANET, which must be known a priori. Our second protocol,

AVERT, is differentiated from E-ZRP in that it is based on the Independent

Zone Routing framework which supports a different zone radius per node. In

IZR each node monitors current traffic and adapts its own zone size such that

the total overhead (for reactively and proactively) routed packets is minimized.

In [57] we find a hybrid integrated protocol, named Hybrid Adaptive proto-

col for Integrated Discovery (HAID),with adaptable zone size for proactively

sending routing and service information. The basic differences of that protocol

and AVERT (for details see Section 3.3.2) are: i) in HAID only the service

providers send proactive advertisements about services and routes, while in


MANETs

AVERT all nodes send proactive traffic, ii) in HAID every service advertise-

ment contains information only about the service of the advertising node, while

in AVERT the proactively sent messages contain information on all the ser-

vices the node is aware of in its 1-hop vicinity (such that ultimately all nodes

become aware of all the services available in their zone), iii) HAID provides no

mechanism for performing efficient propagation of queries as in AVERT, but

resorts to global flooding for query propagation, and iv) the adaptation of the

zone radius of providers in HAID is based on service usage frequencies and not

on local traffic monitoring as done in AVERT. Finally, in [58] another hybrid

integrated protocol is presented, where an autonomous and adaptive zone ra-

dius determination mechanism (based on service popularity) is provided. In

this protocol, named Service Advertisement/Discovery protocol with Indepen-

dent Zones (SPIZ), besides adjusting the zone radius based on IZR’s inherent

mechanism for doing so (for details on this mechanism see Section 3.3.2), the

authors propose that service providers should enlarge their proactive zones

when the popularity of their service is increased, and decrease it when the

service popularity falls. The basic mechanism of IZR for zone adaptation is

also used in AVERT, but AVERT additionally uses a mechanism for allowing

all nodes (providers or clients) to adjust the frequency of sending proactive

messages, such that when a node is not actively requesting or providing any

service and also does not forward any such traffic it can save energy by low-

ering its broadcasting rate. In 3.3.2 AVERT and SPIZ are compared against

each other and AVERT is shown to outperform SPIZ both in terms of service

success ratio and energy conservation.

In Figure 2.2 we make a rough categorization10 regarding which type of

routing protocol (reactive, proactive or hybrid) is more efficient (in terms of

messaging overhead) when integrated with a service discovery protocol based

on simulation results collected and combined from the following papers: [56],

[47], [57], [14], [58], [45], [46], [59], and [60].

10Since direct comparisons are difficult due to lack of specific compatible performance datafor the various protocols, only a rough taxonomy can be provided.


Fig. 2.2. Taxonomy of Cross Layer and Application Layer Ser-vice Discovery Approaches based on Control Packet Overhead.


MANETs

35

3. INTEGRATING SERVICE DISCOVERY WITH

HYBRID ROUTING PROTOCOLS

In this chapter we provide a quantitative analysis of the energy consumption

gains that can be achieved by implementing service discovery at the routing

layer, instead of the application layer. Our approach is to implement service

discovery in the routing layer by piggybacking the service information into the

routing protocol control messages, thus enabling the devices to acquire both

service and routing information simultaneously. This way a node requesting

a service (henceforth called service requester), in addition to discovering the

service, is simultaneously informed of the route to the service provider.

3.1 The E-ZRP Service and Route Discovery Protocol

We propose the piggybacking of service information in routing messages,

in order to decrease communication overhead, save battery power and mini-

mize discovery delays. This way, besides these savings, we can also achieve

smooth service discovery adaptation to severe network conditions (e.g. network

partitions). Smooth adaptation occurs because service availability is tightly

coupled with route availability to serving nodes. Hence when all routes toward

a node fail, this is immediately translated to a loss of service availability for

the services that this node provides.

We demonstrate the benefits of our approach (i.e. routing layer based

service discovery) versus a similar but application layer-based service discovery

protocol for MANETs, by extending the Zone Routing Protocol (ZRP) so that

it is capable of encapsulating service information in its messages. ZRP is a

hybrid routing protocol, i.e. proactive for a number of hops around a node

(this is called the node’s zone) and reactive for requests outside this zone.

We also study and give a detailed insight on the availability of services as

36 Integrating Service Discovery with Hybrid Routing Protocols

an indication of the capability of the discovery protocols under comparison

to construct a consistent view of the available services in the MANET. We

measure the implications of several factors impacting the service availability

such as node mobility, network density and interference. In order to measure

the service availability we define a new metric called SAD (Service Availability

Duration), which is defined as the length of time that elapses from the moment

the service is discovered until that time when the service is lost as a result of

mobility or interference. It should be noted that if the path to the original

service provider is lost, but there exists another provider for the same service-

type in the node’s routing table, then the service is still considered available.

Only when all the routes from a node to all the available providers of the

service are lost, this particular service is considered not to be available any

more to that node. We should note here that the SAD metric is applicable

when the service discovery approach utilizes proactive service advertisements.

In those approaches, service providers periodically send their advertisements

in order to keep clients informed on the availability (and possibly the status)

of the services that they provide. This way clients keep a soft state for every

service they become aware of. A service entry is deleted from a clients cache if

the client has not received any related service advertisement for a given period

of time. Since in E-ZRP we use scoped service advertising, when a node

does not have any path toward any provider of a given service, this does not

necessarily mean that the service is no longer available in the MANET (or

that a network partition has occurred), but it may also signify that there is no

provider for that service inside the (proactive) zone of that particular node. If

a node is interested in monitoring the status of a service irrespectively of the

distance to the service provider, then it could periodically issue requests for

that service using the reactive part of E-ZRP. This way in case of a successful

answer to the request that arrives before the expiration of the related cache

entry, the node could refresh the status of the service and keep the associated

entry in its cache (thus prolonging the service’s availability duration as far as

the service is still accessible). In the literature [61] [62], a similar metric,

The E-ZRP Service and Route Discovery Protocol 37

called Path Duration has been widely used to measure the impact of mobility

on routing protocols for MANETs. However these studies mainly focus on

reactive routing protocols and do not consider service discovery. Moreover,

they focus on node availability and not service availability, which is a different

concept. In general a good discovery protocol should be able to adapt to

different network conditions in order for nodes to effectively discover as many

long-lived services as possible. The proactive maintenance of routing zones also

helps improve the quality and survivability of discovered routes, by making

them more robust to changes in network topology. Once routes have been

discovered, the routing zone offers enhanced, real-time, route maintenance.

Link failures can be bypassed by storing multiple paths for destinations within

the routing zone. Similarly, sub-optimal route segments can be identified and

traffic re-routed along shorter paths.

The remainder of this chapter is organized as follows. In Section 3.1.1

we outline the concept and motivation behind the use of E-ZRP, in Section

3.1.2 we describe the protocol packet formats and the operation of E-ZRP and

in Section 3.1.3 we perform both theoretical and experimental evaluation of

E-ZRP against a similar application layer based approach.

3.1.1 Concept and Motivation

Our motivation for seeking a routing layer solution for service discovery

stems from the fact that any service discovery protocol implemented above the

routing layer will always require the existence of some kind of routing protocol

for its own use. Hence, two message-producing processes must coexist: the first

one communicates service information among service providers and service re-

questers; the second one communicates routing information among them. As

a result, a node is forced to perform multiple times the battery-draining oper-

ation of receiving and transmitting (control) packets. Our approach exploits

the capability of acquiring service information along with routing information

(from the same message) by piggybacking service information onto routing


messages. This way, redundant transmissions of service discovery packets at

the application layer are avoided and energy is saved.

The idea of providing routing layer support for service discovery was first

introduced by Koodli and Perkins in [40]. They argue that for proactively

routed MANETs, a service reply extension added to topology updating rout-

ing messages is enough for providing both service discovery and route discov-

ery concurrently. In reactively (or on-demand) routed MANETs, the service

discovery process follows the traditional route discovery process by using its

message formats for route requests (RREQ packets) and route replies (RREP

packets) extended to carry also a service request or reply respectively. In

this chapter we present the E-ZRP protocol. E-ZRP is based on the Zone

Routing Protocol - ZRP which is appropriately extended to support the si-

multaneous discovery and advertisement of services and of routes toward the

service providers that host those services. Furthermore, we evaluate E-ZRP

comprehensively through extensive simulations.

With the Zone Routing Protocol (ZRP) a node is proactively informed

about routes toward nodes located in its vicinity. If a route toward a more

distant node is needed, then the node reactively initiates the route discovery

procedure. This feature of ZRP makes it a perfect candidate for extending

it with service discovery capability. This is because services provided in a

mobile ad hoc network will most probably have a local nature (especially

when requiring physical interaction–imagine for example a user in need of a

printing service) and also services far away in the MANET from the requester

are very likely to disappear (causing severe service disruptions) due to the

mobile wireless network’s dynamics and hence are not worth the effort of

being continuously advertised to distant clients. In other words, continuous

monitoring and state maintenance of services residing at very distant nodes

will incur high cost and, on the other hand, interaction with such services is

risky since it is highly likely that the connection will fail before the service

interaction has been completed. Considering the above issues we have chosen


the Zone Routing Protocol (ZRP) for adding service discovery functionality

since:

• ZRP proactively and continuously maintains (routing and with our ex-

tensions also service) information available in the vicinity of a node

(through the notion of zones further described later on) in a highly dy-

namic and energy efficient way, and

• ZRP may reactively discover and collect information available at distant

network areas through the use of intelligent request forwarding instead

of global flooding (explained later on).

Additionally we should note that the proactive maintenance of routing and

service information for a node’s zone allows flexible adaptation to changes due

to mobility, because of the existence of multiple paths inside a node’s zone.

Finally, ZRP was our selection for performing routing layer based service

discovery also for another reason. In contrast to classic (monolithic) routing

protocols for MANETs, ZRP can be also seen as a routing framework con-

sisting of one reactive and one proactive part. Any existing purely reactive

routing protocol (e.g. AODV or DSR) can be used as the IERP (Inter Zone

Routing Protocol) and any existing purely proactive protocol (e.g. DSDV)

can be used as the IARP (Intra Zone Routing Protocol). Also, depending

on ZRP’s zone radius, ZRP can be transformed to a purely reactive protocol

(when the zone radius equals 0) or to a purely proactive protocol (when the

zone radius is equal to infinity, or the network diameter). Hence, ZRP may

be considered as the best candidate for routing layer based service discovery

(and in some sense a framework for a parameterizable class of protocols). In

the next section we describe the basic operation of ZRP and the extensions

we have introduced in order to enhance it with service discovery capabilities.


3.1.2 Design

In the following paragraphs we describe the basic operation of ZRP and

then we define the extensions made to ZRP’s messages formats in order to

create the integrated route and service discovery protocol, namely E-ZRP.

ZRP

First we proceed to describe the structure and operation of ZRP. ZRP

actually consists of three sub-protocols, namely:

• The Neighbor Discovery Protocol (NDP), through which every node peri-

odically broadcasts a “hello” message to denote its presence. Since there

is no official specification of the exact contents of an NDP packet we have

implemented it as shown in Figure 3.1. The fields ‘Source Address’ and

‘Sequence Number’ are self-explanatory, while the field ‘Interval Time’

denotes the time when the next packet will be sent, and the ‘Validity

Time’ denotes the period after which the node should be deleted from

the receiver’s neighbor list if the originator node has not sent any other

“hello” packet.

Fig. 3.1. NDP packet format (for ZRP).

• The Intra Zone Routing Protocol (IARP), which is responsible for proac-

tively maintaining route records for nodes located inside a node’s routing

zone (for example records for nodes located up to 2-hops away). We have

implemented the IARP packet format as specified1 in the now expired

1In our implementation, we have only excluded the optional field ‘Link Destination SubnetMask’.


IETF draft draft-ietf-manet-zone-iarp-02 and field descriptions are avail-

able from [63]. The IARP packet format is depicted in Figure 3.2.

Fig. 3.2. IARP packet format (for ZRP).

• The Inter Zone Routing Protocol (IERP), which is responsible for re-

actively creating route records for nodes located outside a node’s rout-

ing zone (e.g. records for nodes located further than 2-hops away).We

have implemented the IARP packet format as specified in the now ex-

pired IETF draft draft-ietf-manet-zone-ierp-02 and field descriptions are

available from [64]. The IERP packet format is depicted in Figure 3.3.

The proactive component of ZRP, named the Intra-zone Routing Protocol

(IARP), uses a timer-based link state protocol. With this protocol each node

broadcasts periodically its link state to all nodes up to R - 1 hops away. The

link state is formed by periodic hello beacons broadcasted by each node to its

one-hop neighbors (this the NDP protocol). The vicinity of a node, defined by

all nodes located up to R hops away, is called a node’s routing zone and the

nodes that are located at exactly R hops away are called the node’s peripheral

or border nodes (shaded nodes H, F and G for node A in Figure 3.4).


Fig. 3.3. IERP packet format (for ZRP).

Fig. 3.4. Routing Zones and the Bordercasting Process.

Through IARP every node is aware of the routes in its routing zone and also

has identified its border nodes. For discovering routes toward nodes outside

its zone, a node uses the Inter-Zone Routing Protocol (IERP). When a node

needs such a route, it unicasts an IERP route query only to its border nodes

(bordercasting). These nodes check their routing tables to see if the requested

node is located in their own zones, and if not they re-bordercast the query

to their own border nodes. This way global flooding is substituted with a

few unicasts and the query propagates the network much more efficiently.

Of course, query termination techniques do exist to avoid loops of the route


queries. For example, in Figure 3.4, assuming zones of 2 hops for all nodes,

when node A searches for a route toward node Q, it will first query its border

nodes H, F and G, and they, in their turn, will query their border nodes. The

result is that node F’s border node P knows a route toward Q and may reply

to the query. Following the reverse route, the reply will eventually reach node

A, which can now use the discovered route toward node Q.

E-ZRP

In order to add service discovery capabilities to ZRP we embedded an extra

1-byte field in NDP “hello” messages for storing service IDs. We use the con-

cept of Unique Universal Identifiers (UUIDs) instead of service descriptions2,

keeping packet lengths small for the routing messages and minimizing the ef-

fects on the network (the bigger the messages the larger the delays and the

possibility of transmission errors). Such an approach implies that all nodes

know a-priori the mappings between services (service types3) offered in the

MANET and UUIDs. This is a common assumption and is justified by the

fact that most MANETs are deployed for certain purposes where there is lack

of fixed communication infrastructure (e.g. a battlefield or a location of phys-

ical disaster). In such environments, the roles of every participating node are

concrete and can be easily classified in types of services. For example, in a

battlefield one node may offer radar information to the rest, while another one

may offer critical mission update information. In the case of a disaster such

as an earthquake, an on-site relief team usually consists of members having

different missions (e.g. one may be able to provide information about trapped

people under ruins, another may provide information about terrain stability,

and others may try to find and provide valuable structural information about

the collapsed buildings etc.). In such environments the mapping of services to

2Lengthy service descriptions (e.g. using XML in [25], or OWL in [65]) are only foundin application layer based service discovery approaches, where only the routing protocolsunderneath retain small packet size.3Henceforth, the term service and service type will be used interchangeably, both denotinga certain type of service.


UUIDs is more than sufficient for service discovery. Semantic matching of rich

service descriptions is of no particular use in these cases, not to mention that

these techniques lead to increased energy consumption (a scarce and valuable

resource in the above scenarios). Thus, by extending “hello” messages with

service UUIDs, a node is able to denote both its presence and the services it

provides. We also argue that the purpose of an integrated service discovery

protocol is to allow nodes to discover or advertise their services in a highly

efficient way and let higher layer protocols to deal with service invocation and

specific service descriptions. Under this point of view, we argue that the pur-

pose of an integrated energy efficient service discovery protocol is only to pass

to upper layer protocols information on the available types of services pro-

vided in the network and also their location (addresses of the available service

providers). Once this information is made available (avoiding excessive energy

consumption for doing so) and possibly with the required intervention of the

user the processes of service selection and service invocation can be handled

at an upper layer.

ZRP was further extended in order to include service information in ev-

ery routing entry of the IARP and IERP routing messages and tables. For

including service information in routing messages we have used 1 out of the 3

available ‘Reserved’ Fields in the IARP and IERP headers such that the packet

length remains the same. This allows service UUIDs to be able to express 256

different service types4. The IARP protocol gathers information available from

NDP messages, updates its link state table and then periodically broadcasts

this table to its neighbors through IARP packets. A node broadcasting an

IARP packet sets the Time To Live (TTL) field in these packets equal to R

- 1 (hops). This way each node knows the routes to all the nodes in its zone

and also the services that these nodes offer; thus adding the service discovery

capability to the proactive part of ZRP. IERP is responsible for routing toward

4If a hierarchical format is preferred UUIDs can take a type-subtype format depending onthe bits assigned for each level of the hierarchy (for example if a two level service typehierarchy is used the 1-byte UUID could express 64 different service types each having upto 4 sub-types).


resources that are not available in a node’s zone. When a node cannot find a

desired service in its local IARP table then an IERP message with a NULL

destination address and a service field with the service requested is border-

casted. When a node receives such a message it first checks if it provides the

requested service or if it is aware of another node that provides the service;

and if it does, it generates an IERP reply message. Otherwise it re-bordercasts

the message adding its own address to the previous hop list, so that a reverse

route to the requester can be established and used when the requested service

is found.

In the following section we present our simulation results from testing E-

ZRP under a variety of settings for a MANET.

3.1.3 Experimental Evaluation

In this section we will compare the performance of our integrated approach

which combines routing and service discovery (using E-ZRP) against a tradi-

tional non-cross layer approach employing two separate protocols (ZRP for

routing) and (APS for service discovery). We will especially focus on the per-

formance evaluation of the proactive part of E-ZRP, since as mentioned in the

literature review there do not exist similar detailed investigations. What fol-

lows is a description of the application layer based service discovery approach

(APS over ZRP) against which E-ZRP will be compared.

The APS protocol

As already mentioned in Section 2.3 our aim was to compare E-ZRP to a

similar but application layer based service discovery protocol. For that rea-

son we implemented a protocol named APS (APplication layer-based Service

discovery protocol). APS works in a proactive manner similar to IARP. The

format of an APS packet is presented in Figure 3.5. In APS, periodically, every

node sends an APS message setting the ‘F’ field to 0, which means that the


Fig. 3.5. APS packet format.

message is an advertisement, and includes the UUID of the provided service

in the ‘ServiceID’ field. The message is propagated up to R hops away from

the originator node (the area defined by R hops is called the node’s scope).

Any node that receives the message checks the ‘HopCount’ field and if this

is greater than 0 it decreases it and re-broadcasts it. This means that APS

is a scoped protocol and service providers do not employ global flooding for

advertising their services to the whole network5. The major difference of APS

with IARP is that an APS message does not contain information about any

other service but only about the service provided by the originating node.

This decision on the design of APS was taken in order to implement an ap-

plication layer service discovery protocol, which is as lightweight as possible.

The same packet format is also used for service requests, but in that case the

service requester sets the ‘F’ field equal to 1, the ‘HopCount’ field to 0 and

the ‘ServiceID’ to the UUID of the requested service.

In the next section, we analytically obtain the energy gains of E-ZRP

against the APS approach for proactive service discovery.

Theoretical Analysis of Energy Consumption

Before computing the energy costs for both approaches we proceed in defin-

ing a model for the multihop wireless ad hoc network. We are especially inter-

ested in a model capturing the characteristics of broadcasting in IEEE802.11

networks, since all the packets of E-ZRP, ZRP or APS are broadcast. Such a

model was developed in [66], where the authors perform an analysis of the per-

5Only when a node requests a service, that is not provided by any node in its local scope,then it floods its query to the whole network.


formance of the IEEE802.11 broadcasting scheme with numerical techniques.

Next, we briefly describe this model and then use it for developing the energy

consumption model that will be used to evaluate the performance of the E-ZRP

and the APS-based approach. According to [66], the following assumptions

are made:

• All nodes in the network are two-dimensionally Poisson distributed with

density λ, i.e., the probability p(i, A) of finding i nodes in an area of size

A is given by p(i, A) = (λA)ie−λA

i!.

• All nodes have the same transmission and receiving range, which is de-

noted as Rt. N is the average number of neighbor nodes within a circular

region of radius Rt. Therefore, we have N = λπRt2.

• A node transmits a frame only at the beginning of each time slot. The

size of a time slot, τ , is the duration including transmit-to-receive turn-

around time, carrier sensing delay and processing time.

• The transmission time or the frame length is the same for all nodes.

• When a node is transmitting, it cannot receive simultaneously.

• A node is ready to transmit with probability p. Let p′ denote probability

that a node transmits in a time slot. If p′ is independent at any time

slot, it can be defined to be p′ = p · Prob (Channel is sensed idle in a

slot) ≈ p · PI , where PI is the limiting probability that the channel is

sensed to be idle.

• The carrier sensing range is assumed to vary between the range [Rt, 2Rt].

By modeling the channel as a two state Markov chain (idle state and busy

state), and assuming that that the time spent in the idle state is τ and the

time spent in the busy state is δdata, p′ can be obtained by:

p′ =τ · p

δdata · (1 − e−p′·N) + τ(3.1)


Using this equation and modeling the broadcast process of IEEE802.11 as

a three state Markov chain (idle state, success state and collision state), the

authors in [66] obtain the following formula for the probability of successful

transmission of a broadcast packet:

PS =(p′ · e−p′·(N+Nph·(2·δdata+τ))) · δdata

τ + p′ · (δdata + τ)(3.2)

where Nph is the number of potential interfering stations to the broadcasting

node (actually those are the hidden nodes). 2δdata + τ is the duration of the

period for which all potential interfering stations must not transmit in order

for the broadcasting node to make a successful transmission (i.e. the packet

must be received by all its 1-hop neighboring nodes).

However, since energy will also be consumed in the case of an unsuccessful

broadcast (collision), using the Markov chain model of [66], we also derive the

probability of collision:

PC =(p′ − p′ · e−p′·(N+Nph·(2·δdata+τ))) · δdata

τ + p′ · δdata

(3.3)

Taking into account the probabilities for successful and unsuccessful broad-

cast we can now define the detailed energy consumption model. The model

captures the energy consumption of the node that sends a broadcast message

and also the energy consumption of the nodes in its 1-hop neighborhood that

receive the message (both for the case of collision and success). In Table 3.1

we present the symbols that will be used in the analysis.

As explained in the previous section, when a node uses APS over ZRP and

assuming broadcasting timers of the same periodicity for all the protocols6, a

6The assumption of same periodicity for all the protocols was made in order to establish afair comparison basis. However, in our simulation-based evaluation of APS and E-ZRP wealso test different settings for the broadcasting timers.


Symbol MeaningPSE The probability of successful broadcast when using E-ZRP.PCE The probability of collision when using E-ZRP.PSA The probability of successful broadcast when using APS over ZRP.PCA The probability of collision when using APS over ZRP.BI The length in bytes of an IARP packet when using ZRP or E-ZRP.BN The length in bytes of an NDP packet when using ZRP.BN ′ The length in bytes of an NDP packet when using E-ZRP.BA The length in bytes of an APS packet.r The ratio of the energy required for receiving 1 byte of data to the

energy required for sending one byte of data.R The data rate of the channel.EE The energy consumption per broadcast cycle for E-ZRP.EA The energy consumption per broadcast cycle for APS.p The probability that a node is ready to transmit.p′ The probability that a node transmits in a time slot.N The average number of 1-hop neighbor nodes.Nph The number of potential interfering stations to the broadcasting node.

Table 3.1Symbols for the energy model

node has to transmit one NDP, one IARP and one APS packet every x seconds.

In the case of E-ZRP the node has to transmit modified (i.e. service extended)

versions of NDP and IARP every x seconds. Without loss of generality we

will assume that in each case we will transmit a continuous stream of bytes

containing the byteload of all the packets that are to be broadcast in every

period of x seconds, as defined by the broadcasting timers. Setting δdata equal

to this byteload divided by the channel’s rate we obtain the various broadcast

success and collision probabilities for the APS-based and the E-ZRP-based

approaches as follows:

PSE =(p′ · e−p′·(N+Nph·( BI+B

N′R +τ))) · BI+BN′

R

τ + p′ · (BI+BN′R + τ)

(3.4)

PCE =(p′ − p′ · e−p′·(N+Nph·(2·BI+B

N′R +τ))) · BI+BN′

R

τ + p′ · BI+BN′R

(3.5)

PSA =(p′ · e−p′·(N+Nph·( BI+BN+BA

R +τ))) · BI+BN +BA

R

τ + p′ · (BI+BN+BA

R + τ)(3.6)


PCA =(p′ − p′ · e−p′·(N+Nph·(2·( BI+BN+BA

R )+τ))) · BI+BN+BA

R

τ + p′ · BI+BN+BA

R

(3.7)

For our energy model and in order to match the realistic energy consump-

tion specifications of a wireless network interface as reported in [67], we assume

that broadcasting 1 byte of data costs 1 unit of energy and receiving 1 byte

of data costs r units of energy (0 ≤ r ≤ 1). We obtain the following formu-

las for computing the energy consumption experienced by the nodes in one

broadcasting cycle (either successful or not) for each protocol:

EE = PSE · ((BI + BN ′) + r · N · (BI + BN ′)) + PCE · ((BI + BN ′) + r · N · (BI + BN ′

2))

= PSE · (1 + r · N) · (BI + BN ′) + PCE · (2 + r · N) · (BI + BN ′

2) (3.8)

EA = PSA · ((BI + BN + BA) + r · N · (BI + BN + BA)) + PCA · ((BI + BN + BA)+

+ r · N · (BI + BN + BA

2))

= PSA · (1 + r · N) · (BI + BN + BA) + PCA · (2 + r · N) · (BI + BN + BA

2) (3.9)

So, for the case of a successful broadcast cycle, the broadcasting node using

E-ZRP will expend BI + BN ′ units of energy (respectively BI + BN + BA

units of energy for a broadcasting node using APS over ZRP), while its N

1-hop neighbors will expend r · (BI + BN ′) units of energy (respectively r ·(BI + BN + BA) units of energy for a broadcasting node using APS over

ZRP) for receiving the broadcasted data. For the case of a broadcast cycle

with collisions, the broadcasting node using E-ZRP will expend BI + BN ′

units of energy (respectively BI + BN + BA units of energy for a broadcasting

node using APS over ZRP), while on the average its N 1-hop neighbors will

expend r ·(BI+BN′2

) units of energy (respectively r ·(BI+BN+BA

2) units of energy

for a broadcasting node using APS over ZRP) for receiving the broadcasted

data. The formula must take into account the energy consumption induced

to a receiver only due to the reception of (part of) the specific packet that

is broadcasted. This is because when an interface is in receive mode, it does

not expend more energy to simultaneously receive colliding data from more


than one nodes. If we assume that the reference broadcast packet collides

with the reception of another packet (all packets are of equal length) then the

average duration of conflict is approximately half the packet size as indicated

by the equation (3.10). Hence the energy consumption for receiving only the

reference packet is equal to the time required for receiving half of the packet

multiplied by the energy spent per received byte. This means that for the case

of a collision the formula takes into account the average time that a neighbor

node (having a collision) is receiving only the (part of the) reference broadcast

packet.

∑Psizei=1 (Psize − i)

Psize − 1, (3.10)

where Psize is the packet size expressed in IEEE802.11 slots.

In Table 3.2 we present the length in bytes for each packet type. We use

the packet sizes as they are specified in sections 3.1.2 and 3.1.3.

Packet ByteloadBA 12BBI 12B+Nx12BBN 8BBN ′ 9B

Table 3.2Byteload per packet type (N is the average number of a node’s1-hop neighbors)

In order to compare the approach employing APS over ZRP with the ap-

proach employing only E-ZRP for route and service discovery we obtain the

fraction of their energy consumption formulas as follows:

EE

EA=

PSE ·(1+r·N)·(BI+BN′ )+P

CE ·(2+r·N)·(BI+BN′

2)

PSA ·(1+r·N)·(BI+BN+BA)+P

CA ·(2+r·N)·(BI+BN +BA2

)

(3.11)


We evaluate the energy consumption model by contrasting it to simulation

results. For the simulations we have used the Qualnet Simulator [68] and

performed full stack simulations including the functionality of every layer from

the physical layer up to the application layer. At the physical and MAC layer

we have used the IEEE 802.11 module. We have also implemented the E-ZRP

and APS protocols. The settings for both the formula and the simulations are

shown in Table 3.3.

Parameter ValueNode Range 380mτ 50 ∗ 10−6s (equal to a DIFS period in

IEEE802.11b)Data Rate 2 ∗ 106bits/sBroadcast Timer Period 10sAverage Number of Neighbors [2...17]r 0.65Mobility Model RWP with constant speed of 1m/s

Table 3.3Simulation Settings

Fig. 3.6. Energy gains from using E-ZRP vs. APS.

Figure (3.6) reveals that simulation results closely follow our analytical

model. Both curves follow the same trend and the small difference observed


between the two curves are mainly due to the fact that in our analytical

model we do not account for node mobility (which causes the average number

of neighbors per node to fluctuate during the simulation) and we also take

the assumption of a single packet with a payload equal to the sum of pay-

loads of all protocol packets that are utilized. However, it is evident that the

analytical model can quite satisfactorily capture the performance gains in en-

ergy consumption for the two protocols, since the average percentage error is

approximately 8.3%. The fact that the two protocols tend to have the same

energy consumption in highly dense networks is due to the fact that the IARP

messages (either for E-ZRP or ZRP) grow in size and hence dominate in the

total energy consumption computation. Hence, the energy consumption by

NDP and APS packets becomes insignificant as the IARP messages grow in

size and the energy gains from using E-ZRP decrease. We further investigate

the impact of the size of the APS packet on the relative energy gains of E-ZRP

versus APS. In this investigation the packet size for APS ranges from 12 to

100 bytes. The results are presented in Figure 3.7. As expected the more

heavyweight the application layer service discovery protocol is, the more the

energy gains from using E-ZRP.

Fig. 3.7. Investigation of the impact of APS packet size onenergy consumption.


All the above results concerned in-zone proactive service discovery. Re-

garding out-of-zone service discovery, it is rather straightforward that E-ZRP

outperforms APS in terms of energy consumption, since APS uses global flood-

ing for service request dissemination while E-ZRP uses bordercasting. As

explained in Section 3.1.2 with the mechanism of bordercasting the energy-

hungry flooding of a query to all neighbors is replaced by a few broadcasts

along a spanning tree connecting the originator of the query to its border

nodes (and then the border nodes to their respective border nodes etc.). It

would be unfair to compare the two protocols in terms of energy consumption

for out-of-scope/zone service discovery, since APS does not use any optimiza-

tion in its mechanism for forwarding service requests. However, we consider

it useful to obtain the theoretical control packet overhead ratio imposed by

a service request when using IERP and when using global flooding. For our

analysis we will assume that the spanning tree connecting any node to its bor-

der nodes has a degree equal to n and that each node in the network has on

average N 1-hop neighbors. It is intuitive that n can never be greater than N.

We obtain the following ratio:

Y =OverheadE−ZRP

OverheadAPS=

∑Di=1 ni−1

∑Di=1 N i−1

, (3.12)

where D is the distance in hops to the provider hosting the requested ser-

vice. Figure 3.8 shows the value that this ratio takes for various combinations

of n, N and D.

As it was expected, bordercasting is much more efficient than global flood-

ing, especially when the average number of 1-hop neighbors is large, the span-

ning tree degree is small and the provider that hosts the requested service is

located many hops away from the requesting node.

In the former theoretical and experimental analysis we have assumed that

both the examined approaches employ for their proactive operation broadcast-

ing timers of the same periodicity. We now proceed in investigating through

simulations the possibility of employing larger update intervals at the ap-


(a) D=1

(b) D=2 (c) D=4

(d) D=10

Fig. 3.8. Investigation of the overhead of Bordercasting vs. theoverhead of Flooding.

plication layer protocol compared to those used in the routing layer, hence

minimizing the energy consumption of APS as much as possible. In all the


following experiments the nodes that run the APS protocol, use the plain ZRP

protocol for routing (with equal zone radius to the scope of APS). Table 3.4

summarizes the settings for the first set of experiments.

Parameter ValueNode Range 340mData Rate 2 ∗ 106bits/sNumber of Nodes 250Mobility Model NoneSimulation Time 1000sE-ZRP Zone Radius 3hopsAPS Scope 3hopsIARP Broadcast Timer 10sService Deletion Timer if not updated for IARP 4*IARP Broadcast TimerAPS Broadcast Timer (A) 200sAPS Broadcast Timer (B) 160sAPS Broadcast Timer (C) 80sAPS Broadcast Timer (D) 40sAPS Broadcast Timer (E) 20sAPS Broadcast Timer (F) 15sAPS Broadcast Timer (G) 10sService Deletion Timer if not updated for APS 4*APS Broadcast Timer

Table 3.4Simulation Settings for Evaluating the impact of Broadcast Timers

For the cases that the APS Broadcast Timer is greater than 10 seconds

(which is the value used for the IARP Broadcast timer) the APS protocol

sends messages in larger time intervals and hence decreases the energy con-

sumption. However this comes at the cost of decreased capability of discovering

services. We use the term ‘service discoverability’ to refer to the metric that

measures the capability of discovering services (expressed in number of dis-

covered services). The purpose of these experiments was to show the optimal

configuration of an application based service discovery scheme (as represented

by APS), so that service discoverability is equal or better to that achieved by

a similar but routing layer based approach (as represented by E-ZRP). The

results of these experiments are presented in Figure 3.9. Each point on the


curve corresponds to different parameter settings for the update and service

deletion intervals (those presented in Table 3.4) for the APS protocol. The

vertical and horizontal dotted lines denote the average energy consumption

per node and the average number of services discovered per node respectively,

for E-ZRP with a broadcast interval of 10 seconds.

Fig. 3.9. Investigation of the impact of Broadcast Timers onservice discoverability and energy consumption in a static con-text.

It is evident that the application layer based service discovery scheme

(APS) may perform better than the routing layer based scheme in terms of

service discoverability for broadcast intervals lower than 40 seconds. However,

this comes at the cost of energy consumption, which is increased by 30% or

more compared to the routing layer based scheme with the original broadcast

interval of 10 seconds. This is again explained by the fact that the messages

of the application layer based scheme are much shorter (in order to be more

economic) and hence less informative than those of the routing layer based

scheme. Service discoverability is reduced by reducing the number of broad-

casted messages (bigger intervals means fewer messages transmitted, hence

every node receives less information about services).

We further test the performance of the two schemes in a mobile context. All

the parameters regarding node mobility, the APS broadcast interval and the

service deletion interval are the same as those used in the previous experiments.


We study 2 extreme cases of mobility. The first case is for low mobility, where

nodes move according to the random waypoint mobility model with minimum

speed 0,1m/s, maximum speed 0,5m/s and pause time 30 seconds. The second

case is for high mobility, where the mobility parameter of maximum speed

changes to 12,5m/s. Figure 3.10 depicts the results for service discovery and

energy consumption respectively in this mobile context.

Fig. 3.10. Investigation of the impact of Broadcast Timerson service discoverability and energy consumption in a mobilecontext.

Each point on both curves corresponds to different parameter settings for

the update and service deletion intervals for APS (those presented in Table

3.4). As shown in Figure 3.10 the application layer based service discovery

scheme performs better in terms of energy consumption (compared to the

routing layer based scheme - dotted lines) when the broadcast interval is equal

or more than 160 seconds (point B) saving 3% more power but discovering 43%

fewer services for low mobility cases and 22% fewer services for high mobility

cases. The energy gains observed for APS for large broadcast interval are due

to the increased total byteload produced by E-ZRP. For example when the APS

Broadcast Timer is equal to 160 seconds while the ZRP and E-ZRP Broadcast

Timers are equal to 10 seconds, this means that during the 160 seconds the E-

ZRP approach creates more messages of greater byteload than those produced

by the APS over plain ZRP approach. Hence, we observe the energy gains for


the application layer based approach. However, the cost in terms of service

discoverability is too high to justify those energy savings and we can come to

the conclusion (assuming similarity of the two schemes) that an application

layer based scheme cannot be more efficient in service discovery compared to

a routing layer based one. We also note, that realistically the application

layer based scheme in our case could not be much more lightweight, since we

already have sacrificed the possibility of advertising all the services that a node

is aware of in its vicinity instead of only its own.

Investigation of Service Availability Duration

In order to evaluate the quality of discovered services using E-ZRP we

also conducted the following experiments. We assumed that each node hosts

one out of three possible service types and runs E-ZRP as its routing and

discovery protocol. The selection of any of these 3 service types has the same

probability for any node, hence at the end of the allocation 1/3 of the node

population hosts a service of the first type, another 1/3 hosts a service of the

second type and the last 1/3 hosts a service of the third type. In this context

the service discoverability is measured as the ‘number of discovered service

sessions’ and not as the ‘number of discovered services’. Measuring the number

of discovered service sessions is more meaningful in an environment where each

service provider does not host a unique service, but a service belonging to a

common set of service types. A service session begins from the moment a node

discovers one or more service providers of a given service type until the moment

it loses communication with all the service providers of that specific service

type (i.e. as long as there is at least one service provider of the requested service

type visible to the node, the session for the specific service is considered alive).

In this context the Service Availability Duration (SAD) metric measures the

service session lifetime. Through the following experiments we investigate the

conditions under which service discoverability and at the same time the SAD

of discovered services are maximized, for E-ZRP or APS.


We simulated a network comprising 20 nodes uniformly dispersed in a

4000x4000 meters square area. We used a random waypoint mobility model.

First, we tested the sensitivity of Service Availability Duration (SAD) at dif-

ferent speeds. We simulated five different scenarios. In the first scenario each

node’s speed (in meters/second) was distributed between 1 and 3,5m/s (low

mobility), in the second scenario between 1 and 7m/s (medium mobility), in

the third scenario between 1 and 9m/s (medium mobility), in the fourth sce-

nario between 1 and 11m/s (high mobility) and in the last scenario between

1 and 14m/s (high mobility). The zone radius for E-ZRP was set to 3 hops.

In order to capture the effects of mobility only on the performance of E-ZRP

and APS we have used a perfect channel (at the end of the section we also

evaluate the two protocols under a noisy channel). The simulation duration

was 2000 seconds in every experiment (each scenario was run 10 times with

different simulation seeds and the results represent averages).

Fig. 3.11. E-ZRP: Avg. Service Session Duration PDF vs.Speed (Low-Medium SAD).

Figures 3.11 and 3.12, depict the Probability Density Function for the

average (per node) service availability duration of discovered services for E-

ZRP. As Figures 3.11 and 3.12 show, it is more probable for E-ZRP to discover

short-lived services in highly mobile environments (due to node mobility and

service rediscoveries), while more long-lived services can be discovered only

in low mobility cases. This is explained by the fact that when the nodes are


Fig. 3.12. E-ZRP: Avg. Service Session Duration PDF vs.Speed (Medium-High SAD).

highly mobile, paths are difficult to be maintained and hence far-away services

tend to last for a very short amount of time since the probability for a path

break is larger when nodes move faster. When nodes move slower these paths

tend to be more stable and hence services tend to be available for a longer

time. However, it is not obvious from these figures when we can achieve the

maximum average SAD, which is a metric of great importance in analyzing the

quality of discovered services. The values of mean SAD over low, medium and

high mobility are presented in Figure 3.13, where we also present the mean

SAD for APS given the same settings.

Fig. 3.13. Mean Service Availability Duration (SAD) vs. Speed.


The lines connecting the 5 spots in the figure do not correspond to results

for speeds other than the five defined above, but are drawn for better viewing.

We have also implemented a tracking protocol, which measures the realistic

connectivity between the nodes in the network taking into account only their

Euclidean distances. This protocol, called Tracker, checks the physical dis-

tances of nodes on the terrain and calculates the connectivity graph. Then,

knowing the types of services offered by the nodes it calculates the realistic

service duration time for all nodes of the graph. In order to allow the same

service disconnection tolerance followed by APS and E-ZRP (40 seconds), the

Tracker protocol considers a service active if connectivity to any of its providers

has been detected at least once during the last period of 40 seconds. In case

that no such connectivity has been detected it removes the service from the

node’s cache and keeps a record of its duration. Under the given density and

the perfect channel assumption, both protocols closely follow the Tracker pro-

tocol and hence accurately reflect the realistic connectivity among nodes. It

is also evident from this figure that the average SAD actually decreases when

the speed increases both for E-ZRP and APS. However, it would not be fair to

compare the performance of the protocols with respect to service duration only.

The number of service sessions discovered is also important, since it is usually

preferable for a node to discover a small number of service sessions with long

durations, throughout its lifetime, instead of a high number of service sessions

with small durations. Since in the experiments we assume that every node is

a service provider and there are 3 service types, the optimal results would be

to have 2 service sessions per node each having duration of 2000 seconds (sim-

ulation duration). This implies that each node has discovered all the service

types (excluding its own) and has kept connectivity to them until the end of

the simulation. In Figure 3.14 we show the average number of service sessions

discovered per node in case of low, medium and high mobility.

As expected, the high mobility case (maximum speed = 14m/s) outper-

forms all the other in the number of service sessions discovered both for E-ZRP

and APS. So, there is a trade-off between average SAD and number of service


Fig. 3.14. Avg. Number of Service Sessions Discovered vs. Speed.

sessions. In order to evaluate when a protocol performs better, we should be

aware of the average transaction duration (ATD) between a node and any ser-

vice. So, for high ATD, the discovery protocol would perform better in a low

mobility setting. This is explained by the fact that the additional service ses-

sions discovered in higher mobility settings would be of no use, because their

low average SAD would be inadequate to complete a transaction. However

the discovery protocol would perform well even in a high mobility setting for

low ATD.

For our investigation related to density, we simulated four scenarios. In the

two first scenarios we change the network density by changing the number of

participating nodes, while in the second set of scenarios we keep the number of

nodes fixed and change the terrain size (on which they are allowed to move).

The first scenario included 20 nodes moving on a terrain of 2000x2000 meters,

following the random waypoint model with speed ranging from 1m/s to 14m/s

(no pause time). The zone radius for E-ZRP was set to 3 hops. The second

scenario (half density scenario) was identical to the previous one but included

only 10 nodes. Both scenarios had duration of 2000 seconds each (each scenario

was run 10 times with different simulation seeds and the results represent

averages). The results are shown in Figure 3.15, where it is obvious that by

reducing network density to one half, the number of long-lived service sessions


in the half-density case is significantly smaller than the number of the long-

lived service sessions found in the full-density case. This is due to the fact

that re-discoveries of services are more frequent in a denser environment.

Fig. 3.15. E-ZRP: Service Duration Distribution vs. Density.

One would expect that in the denser environment services would tend to

last longer, since there are more alternative paths to a service provider and also

more alternative service providers. Hence, a failure of one or more paths does

not necessarily mean that the node cannot access the given service. Simulation

results presented in Table 3.5 validate this. Actually, when density increases,

due to the existence of multiple paths and providers, the average service du-

ration is increased. The total number of service sessions discovered, however,

decreases as network density increases (Table 3.5). Increased density leads to

more stable connections between servers and clients (since more alternative

paths do exist) and hence the average SAD increases followed by a decrease in

number of service sessions discovered (since most nodes can establish sessions

once with all of the provided service types (3 in our case) in the network and

keep them until the end of the simulations).

The third scenario included 20 nodes moving in one case on a terrain of

2000x2000 meters (high density case) and on a second case on a terrain of

4000x4000 meters (low density case), following the random waypoint model


Full Density(20 nodes) Half Density(10 nodes)Avg. SAD/node 963s 352sAvg. Service Sessions/node 3,97 8,87

Table 3.5Impact of Density on SAD and Service Sessions

with maximum speed ranging between 3,5m/s and 14m/s (minimum speed

is still 1m/s). Both E-ZRP and APS are evaluated under these two different

densities using a zone (respectively scope) range of 3 hops. Figure 3.16 presents

the performance of E-ZRP and APS regarding SAD for the two aforementioned

densities under varying speeds and Figure 3.17 presents the two protocols’

performance regarding the average number of service sessions discovered per

node.

Fig. 3.16. Density impact on SAD.

Both protocols provide increased SADs for environments with narrower

terrain size and tend to discover a lower number of service sessions for such

environments, which is explained by the fact that better connectivity is pro-

vided and fewer service session breaks occur.


Fig. 3.17. Density impact on Service Sessions.

As stated earlier, the above simulations used a perfect channel in order

to reveal the effects of mobility on the performance of both protocols. In

the following experiment we assume a realistic (affected by noise) channel

in order to also see the effects of the channel on the performance of the two

protocols. For this we have simulated a network consisting of 20 nodes moving

on a terrain of 2000x2000 meters, following the random waypoint model with

maximum speed ranging between 3,5m/s and 14m/s (minimum speed is 1m/s).

The zone (respectively scope) range is set to 3 hops. In Figure 3.18 and 3.19

we present the results. It is evident that E-ZRP performs better than APS

under realistic situations (noisy channel). This is due to the fact that an

IARP message encapsulates more information regarding the services available

in the neighborhood of a node, compared to the information carried by an

APS message, which only informs the receiving node about the service of one

of its neighbors.

Hence, in a realistic (noisy) environment with packet losses, losing an APS

packet costs more in constructing an accurate view of the available services

as compared to losing an IARP packet (since IARP packets from different

neighbors contain overlapping information for the zone of the receiving node).

Combining the results presented in Figure 3.18 and 3.19 this is validated,


Fig. 3.18. Channel and mobility impact on SAD.

Fig. 3.19. Channel and mobility impact on Service Sessions.

since APS is shown to discover more short-lived service sessions than the E-

ZRP. Also in the case of APS the fact that there exists an additional and

separate packet sending process at every node’s routing layer, that of the

routing protocol (traditional ZRP in our case) worsens the channel conditions.


From the simulator’s packet traces we identified increased packet collisions due

to the existence of both protocols.

Another issue worth investigating is the impact of E-ZRP’s zone radius

both on SAD and number of discovered service sessions. We evaluate E-ZRP’s

performance using zones of 1 up to 20 hops for 20 nodes moving on a terrain

of 2000x2000 meters for 2000 seconds both for perfect and realistic channels.

As the Figures 3.20 and 3.21 depict, increasing the zone radius more than

Fig. 3.20. E-ZRP’s Zone Radius Impact on SAD.

Fig. 3.21. E-ZRP’s Zone Radius Impact on Service Sessions.


a threshold7 (in our case it is 2 hops) does not provide any significant extra

gains but leads to highly increased average energy consumption per node as

shown in Table 3.6 below.

Zone Radius (in hops) Avg. Energy (in mWhr)1 0.1857552 0.4857553 2.2237154 3.44455 4.3207357 5.12224520 10.49182

Table 3.6Average energy consumption vs. E-ZRP zone radius

As noted in Section 3.1.3 it would be unfair to compare the two protocols in

terms of energy consumption for out of zone service discovery, since APS does

not use any optimization in its mechanism but uses global flooding. However,

it would be interesting to compare the two service discovery schemes in terms

of latency for discovering an out-of-zone service. In Figure 3.22 we provide

our experimental results showing the delays imposed by both protocols under

different zone settings and different number of hops-to-provider. To be more

precise, in order to take into account the effects of bordercasting, for each

selected distance D between client and server we run IERP D-1 times (each

time increasing the zone radius by one hop, i.e. the zone radius takes the

values of 1 up to D-1 hops); then we obtain the average latency achieved by

IERP over all zone radius settings for that particular client-server pair.

Each point on the diagram is an average obtained over 20 service discovery

requests between different node pairs having the same hop distance. Giving

delays to discover a service in the area of 10 to 50 milliseconds, it is clear

that E-ZRP outperforms APS, where using the latter a node needs from 200

milliseconds up to 800 milliseconds to discover a service. Of course for both

7This threshold actually depends on the network density, the node mobility and the degreeof replication of the available service types among nodes


Fig. 3.22. Delay for out-of-zone services discovery.

protocols the further the requested service is located (in number of hops) the

larger the delay to discover it. The main reason for this is that APS instructs

forwarding nodes to delay the forwarding of a query for a period randomly

chosen between 0 and 100ms. This delay is mandatory in order to avoid ex-

cessive collisions when nodes re-forward the query packet to their respective

neighbors. The delay period is much smaller in the case of IERP (ranges be-

tween 1 and 10ms) since the forwarding nodes are only the nodes located on

the spanning tree of a node toward its border nodes, and hence the chances

of collisions by simultaneous forwarding are dramatically decreased. The val-

ues for the forwarding delay period range were selected to be the minimum

allowing both protocols to make successful requests and are highly dependent

on network density. The denser the network the larger the forwarding delay

period range should be to guarantee that the query will finally reach the node

that hosts the requested service.

3.2 Conclusions

In the previous sections we have (both experimentally and analytically)

presented the achievable gains in energy consumption and service discover-

ability from implementing an integrated service and route discovery protocol,

compared to an application layer protocol for service discovery running over a

The AVERT Service and Route Discovery Protocol 71

similar (to the integrated) routing protocol. The developed analytical model

for computing the energy consumption of both approaches revealed that the

most important factors that affect the relative performances of the approaches

are the advertisement message size and the network density. Also, this model

can be used for computing the energy consumption of other (service and/or

route discovery) protocols that are based on periodic broadcasting of infor-

mation in 802.11 networks. Furthermore, our investigation revealed that even

the most lightweight application layer based service discovery scheme cannot

outperform a similar integrated approach in terms of energy consumption and

at the same time in service discoverability. Even if the application layer based

scheme is tuned to be much more conservative in service advertising than the

integrated scheme, the insignificant energy gains of 3% come at the cost of

dramatically lower discoverability (22% to %43, depending on mobility). We

have also studied the impact of network density, node mobility and channel

characteristics of a MANET in the duration of service sessions for both ser-

vice discovery schemes. We have shown that under a perfect channel both

schemes can achieve similar service session availability (the application layer

based scheme being more expensive in energy consumption). However, under

a realistic channel the application layer based scheme, increasing the amount

of traffic in the network (and hence the collisions), achieves shorter service

session durations. This phenomenon is expected to be even more profound

in cases of high network density where the increase in transmission collisions,

due to the increased overhead from using two separate processes for service

and route discovery, is more severe.

3.3 The AVERT Service and Route Discovery Protocol

In this section we propose AVERT, a hybrid service and route discovery

protocol that differs from other service discovery protocols based on ZRP, in

that it not only allows adaptation of zone radius but also adaptation of the

rate of proactive messages sent by nodes, based only on local traffic monitoring


on each node. In the following sections we present our motivation for creating

AVERT and also evaluate AVERT in terms of energy consumption, contrasting

it to similar service discovery protocols.

3.3.1 Concept and Motivation

Hybrid routing protocols have been proven to operate more efficiently than

proactive or reactive protocols in MANETs, the main reason being their flex-

ibility to adapt to changing network conditions. This was the case for ZRP,

which uses a proactive protocol for local routes and a reactive protocol for

global routes. Actually the proactive protocol serves as a basis for the global

reactive protocol to discover distant routes more effectively.

In Figure 3.23 we show the design space with node mobility and call rate as

the two dimensions, and also the general regions where the proactive and the

reactive protocols perform well [69]. Protocol hybridization means to com-

bine the best features from the two aforementioned protocols to guarantee

best performance independently of the underlying network characteristics. In

its most effective setup, hybridization allows both kinds of protocols to run

simultaneously, but for different scopes.

Fig. 3.23. The routing protocol design space for MANETs.

In order for ZRP (respectively E-ZRP) to operate efficiently in any MANET,

it is implied that all nodes fix their zone radius to the best value, assuming

that they know a priori the call rate and the mobility rate in the MANET.


Moreover, all nodes following ZRP have to use the same zone radius. In order

to remove these assumptions the authors in [69] have proposed IZR. IZR is a

sophisticated version of ZRP, which allows every node to have a different zone

radius and also to dynamically tune it on-the-fly by monitoring local traffic.

In the next section we briefly describe the operation of IZR and the extensions

made to it in order to build the AVERT route and service discovery protocol.

3.3.2 Design

The basic difference of IZR compared to ZRP is its mechanism for adapting

a node’s zone radius to changing conditions in the MANET. In the following

paragraphs we briefly explain the zone radius adaptation mechanism of IZR.

The traffic produced by either IERP or IARP is largely dependent on

the zone radius. The larger the zone the more IARP traffic is created (for

updating a larger set of nodes) and the less IERP traffic is needed, since more

destinations are inside the local zone and there are less queries for out-of-zone

nodes. Actually in [70] it is experimentally shown that the total traffic (IERP

and IARP) is a convex function of the zone radius. Taking this into account

two distributed zone configuration algorithms are used by IZR.

The first algorithm, called Min-searching, is utilized for finding the zone

radius, which corresponds to a local traffic minimum (which is also a global

minimum due to convexity). Periodically each node measures the amount of

routing traffic (IERP and IARP) that passes through it and chooses to increase

or decrease the zone radius. If the decision in a previous measurement period

was to increase (respectively decrease) the zone radius, and the amount of

routing traffic measured in the next period has decreased then the node further

increases (respectively decreases) the zone radius. If the traffic has increased

compared to the previous period, then the zone radius in changed in the inverse

direction of the one that was followed in the previous period. The algorithm

stops when the traffic of the previous period is less than the traffic of the period

before that and also less than the current period. This is considered to be the


best zone radius for achieving a short-term minimum routing traffic overhead.

However, this minimum corresponds to current network conditions and should

be continuously adapted. This adaptation is done using the Adaptive Traffic

Estimation (ATE) algorithm described in the next paragraph.

Having reached a temporary minimum, ATE takes control and tries to

adapt the zone radius by increasing or decreasing it in order to match the

changing network conditions. Having as a reference the ratio IERP traf-

fic/IARP traffic corresponding to the minimum discovered by Min-searching,

it periodically measures the current traffic ratio and if it is increased by more

than a factor of H, then it chooses to increase the zone radius by one hop. An

increased ratio means that the IERP traffic dominates the routing traffic and

hence the zone radius is smaller than it should be, given the current network

conditions, and must be increased. Increasing the zone radius would lead to

more efficient bordercasting and less IERP traffic. In the opposite case, where

the current ratio is measured to be less by a factor of more than H than the

reference ratio, then ATE decides to decrease the zone radius. A decreased

traffic ratio means that the IARP traffic is now dominating and hence the zone

radius is bigger than it should be. In case that a very large change is detected

in the current IERP/IARP traffic ratio, then the Min Searching mechanism is

re-initiated in order to find the new optimal zone radius (and the respective

IERP/IARP traffic ratio). The re-invocation of the Min Searching mechanism

can also be done periodically.

Experiments in [69] have shown that the IZR with dynamic zone radius

configuration leads to more than 60% reduction in routing control traffic com-

pared to the optimal setting of regular Zone Routing [70]. The basic difference

between those two approaches is that in [70] all nodes have the same zone ra-

dius, which does not change, while in [69] each node may have a different zone

radius and also adapt it using the two aforementioned algorithms.

The extensions made to the packets used in IZR in order to include service

discovery information are essentially the same to the extensions done in the

ZRP packets (see Section 3.1.2). However as we said in the introduction of this


section, AVERT builds on IZR in order to be even more efficient than E-ZRP

but also introduces an adaptation mechanism for controlling the frequency

of proactive traffic, in order to achieve higher energy savings when possible.

To be more precise, the adaptation mechanism, called Broadcasting Frequency

Optimizer (BFO), adapts the frequency with which the service aware NDP and

IARP messages are broadcasted. The basic idea of BFO is that nodes that are

not currently engaged in service invocation, discovery or provision (either as

clients, providers or intermediates), can decrease their rate of sending proactive

traffic (namely NDP and IARP packets) in order to conserve energy. In the

following paragraph we describe BFO in detail.

BFO runs periodically on every node. In each period the node measures

only the data traffic8 that passes through it and compares it to the data

traffic passed through it in the previous period. If the current traffic is found

to be lower than the traffic of the previous period, the node increases the

time intervals between two subsequent broadcasts of IARP and NDP packets

by T seconds. In the opposite case, it decreases both these intervals by T

seconds. In order to avoid the two extremes of setting the broadcast interval

to arbitrarily high values or to zero, a maximum allowable and a minimum

allowable value for the broadcast interval are taken also into account such

that they are never violated. When nodes increase their IARP broadcast

interval the routing entries become stale more easily, hence nodes have to

issue IERP queries more frequently in order to find routes and services. The

obvious outcome is that such nodes that seem to decrease their involvement in

service discovery-invocation and routing, decrease their outgoing traffic, thus

saving energy to themselves and to their neighbors. Also, the increase of the

IARP broadcast interval is accompanied by an increase in the NDP broadcast

interval, which means that fewer changes in the 1-hop neighborhood of nodes

are detected and hence the amount of expedited9 IARP messages can also be

decreased, thus further decreasing the total proactive traffic. Now in case that

8By data traffic we mean the IARP or IERP packets that carry service invocation data andnot the IARP advertisement packets or the IERP query and reply packets9Expedited IARP messages are broadcasted if a node senses a change in its link state.


the node becomes more involved in creating or receiving of relaying traffic,

BFO decreases the broadcast interval of both IARP and NDP messages, such

that the node informs its neighbors frequently about its state. This is done

because it is crucial for itself and for the connected nodes to maintain accurate

connectivity information.

We could say that AVERT uses MinSearching and ATE to decrease the

total traffic in the network, and also employs BFO in order for every node

to decrease its own outgoing traffic (sending packets costs more energy than

receiving). Also BFO tries to do this in a harmless way for other nodes. It

decreases the proactive traffic in cases that the node seems not to be too

involved in sending or receiving traffic. In Figure 3.24 we show how the broad-

cast intervals for proactive traffic affect a node’s outgoing traffic. It is intu-

Fig. 3.24. Effect of IARP and NDP broadcast interval size onIARP and IERP outgoing traffic.

itive that IARP outgoing traffic is a non-increasing function of the IARP and

NDP broadcast intervals, since increasing these intervals means sending pack-

ets more sparsely. On the other hand this has the effect that IARP entries and

link state get outdated more easily, hence a node may increase the usage of

IERP requests for finding routes. Hence IERP traffic is a non-decreasing func-

tion of IARP and NDP broadcast intervals. Now, BFO tries to optimize the

outgoing traffic by changing the broadcast interval accordingly, and it does so

with the aim of not disrupting the service and routing processes that currently

go through the node.


In general, we could say that BFO is not affected by MinSearching or ATE

since it measures only the data traffic on every node and not the control traffic.

However, it may affect those two mechanisms, since it controls the amounts

of proactive traffic through the adaptation of the broadcasting intervals. In

the next section we experimentally investigate the effects of coexistence of the

BFO mechanism with the MinSearching and ATE mechanisms, as revealed by

the service success ratios and the energy consumption achieved when using

those two mechanisms with and without BFO. We also investigate the energy

savings obtained from using AVERT against using IZR alone.

3.3.3 Experimental Evaluation

The Qualnet [68] simulation environment was used to simulate AVERT.

The initial broadcast intervals for IARP and NDP have been set to 10 sec-

onds, which is also the lowest allowable interval, and the maximum allowable

intervals have been set to 100 seconds. The MinSearching and ATE mech-

anisms are run every 200 seconds and the BFO mechanism runs every 100

seconds. The simulation time for every experiment is set to 10000 seconds.

The initial zone range has been set to 5 hops for all nodes.

In the following experiments the network consists of 20 nodes uniformly

spread over an area of 2000 x 2000 meter2. All nodes move following the

Random Waypoint Mobility model (RWP) with constant speed of 3,5m/s and

no pause time. The wireless transmission range is set to 380 meters. Each

server may host only one out of three possible service types (for the case of

3 servers in the network, service assignment to servers is done such that each

type of service is hosted by exactly 1 server). Every 100 seconds, each client

selects with probability 1/3 one out of the three available service types and

tries to establish a service session with anyone of the servers that hosts the


requested service type. Each service session involves the transfer of one item

of 200KB size using FTP10.

The AVERT protocol is compared with the IZR and the SPIZ protocols.

The primary performance metric that concerns us is the energy consump-

tion. However, and in order to reveal the possible costs especially of the BFO

mechanism we should also take into account the completed service sessions.

Defining the correct performance metric is however not trivial. As a first ap-

proach, we assume that the optimal operating point (in terms of broadcasting

frequency) for all protocols is at the point where the ratio of Success Ratio

(successfully delivered services) to the Total Energy expended is maximized.

We have observed through simulations that this simple metric is maximized

(for all protocols) when they set their broadcast timers to the maximum al-

lowable (100 seconds). At this point however the protocols are confined to

unacceptable success ratios (52%-65% of the maximum achievable, which can

be much less than 100%). The problems stem from the fact that an increase by

x% in the success ratio for any of the tested protocols requires a much larger

than x% increase in the expended energy. This also means that using the

aforementioned simple metric, a protocol that can achieve higher success ratio

compared to another at a reasonable extra energy cost could be characterized

as less performant. To cope with this situation a better performance metric

would take into account not the success ratio and the total energy expended

but the success ratio and the number of successfully delivered services per unit

of energy expended. In this context, we derive the following metric (service

efficiency σ) for comparing the three aforementioned protocols:

σ = Percentage of Completed Services · Completed Services

TotalEnergy(3.13)

This formula is helpful in characterizing the performance of the service discov-

ery protocol, since it accounts simultaneously for the success ratio achieved and

10We assume that once the FTP session with a server has been established, then if theserver gets disconnected form the client before the completion of the item transfer, theclient cannot transfer the session to another server


also accounts for the extra effort (in terms of energy consumption) required

for achieving acceptable success ratios (65%-85% of the maximum achievable).

In our experiments we compare the AVERT protocol with allowable broad-

cast intervals of 10 seconds up to 100 seconds against IZR and SPIZ with

broadcast intervals ranging from 10 to 100 seconds. We conduct three sets of

experiments; in the first set we assume a scenario (Scenario-1) with high client

to server ratio (5.6 to 1), in the second set we assume a scenario (Scenario-2)

with medium client to server ratio (1 to 1) and in the third set we assume a

scenario (Scenario-3) with low client to server ratio (0.17 to 1). We use these

different scenarios in order to see the impact of the data traffic on the per-

formance of BFO. The data traffic is higher as the ratio of clients to servers

increases (keeping the node population fixed), since there exist more clients

in the network requesting services. The results represent average values ob-

tained over 10 runs for each experiment. We should note here that this is also

the first (implicit) performance comparison of SPIZ against that of IZR, since

in [58] the authors of SPIZ compared the performance of SPIZ only against

a service extended (non-adaptive) ZRP-based service discovery protocol using

zone radius of 1 or 2 hops.

Before proceeding to the experiments mentioned above we investigate (see

Figure 3.25 and 3.26) what is the optimal value for T, which represents how

many seconds the BFO will increase or decrease the broadcast intervals each

time. Adjusting the broadcast intervals with larger step T leads to possibly

greater energy gains but at the cost of decreased success ratios for AVERT.

This is because if during a service session all routes toward the destination

expire due to infrequent broadcasting, the client or server must try to discover

again the route toward each other. In the mean time the application’s toler-

ance may be exhausted and the service session may break before completion.

However, changing the broadcast intervals by smaller values (e.g. when T=1

second up to T=5 seconds), the nodes may gradually reach the optimal broad-

casting interval based on current conditions. This means that in case of wrong

estimations by nodes, before the change begins to affect in undesired ways


Fig. 3.25. Effect of the T parameter on the Percentage ofCompleted Services for AVERT.

Fig. 3.26. Effect of the T parameter on the Energy Consump-tion per node for AVERT.

the current service sessions in the network, the nodes are given the chance to

re-adapt the broadcasting intervals. In the case that T has large values this

is more difficult to happen, since increasing (or decreasing) the broadcast in-

tervals a lot may severely impact routes before a node can react to correct the

situation. Also smaller T means that more fine-grained adaptation can take

place. What validates this is that the value of σ decreases as T increases, as

shown in Figure 3.27. Also it is worth mentioning that as the ratio of clients to

servers decreases the service success ratio increases since there are more avail-

able servers, possibly located closer to the requesting clients. Moreover, when


the number of clients is low, there is less congestion in the network since the

data traffic due to service invocation is less and also localized around nodes.

Fig. 3.27. Effect of the T parameter on σ for AVERT.

Fig. 3.28. Performance gains of AVERT against IZR and SPIZ.

Proceeding to the comparison of AVERT against SPIZ and IZR we fix

T to 1 second, and compare the protocols based on the achieved σ ratios.

In Figure 3.28 the y-axis represents the relative gains in the σ achieved by

AVERT against the σ of the other protocols (σ(AV ERT )/σ({IZR, SPIZ})),while the x-axis represents different broadcast intervals for IARP and NDP in

the range of 10 to 100 seconds. Under the two scenarios tested, AVERT, using

the BFO mechanism for determining a near optimal value for the NDP and


IARP Broadcast intervals, shows performance gains of up to 35% (depending

on the values for the NDP and IARP Broadcast intervals chosen by IZR and

SPIZ). Since the two latter protocols cannot adapt their rate of broadcasting

IARP and NDP messages to the conditions in the MANET, they are confined

to use a “hard coded” value for these rates. However, this value cannot achieve

the maximum performance under all MANET scenarios, or even within the

same scenario assuming that network conditions change dramatically during

the lifetime of the MANET. For example, in Figure 3.29 we plot how the

broadcast intervals of NDP and IARP impact the performance of IZR as the

frequency of requesting services changes (under Scenario-3). It is obvious from

the figure that as the service usage frequency increases, the performance is

optimized using shorter broadcast intervals. Returning to the results shown in

Figure 3.28, the performance of AVERT is presented to be slightly worse than

that of SPIZ and IZR only in Scenario-3. Actually in this scenario the BFO

mechanism of AVERT does not have adequate feedback from data traffic (data

traffic is low) and hence cannot tune the broadcasting frequency optimally.

This is reflected especially when comparing AVERT with IZR and SPIZ when

the latter protocols use relatively low broadcasting intervals.

Fig. 3.29. Performance of SPIZ under different service request frequencies.

Conclusions 83

3.4 Conclusions

Considering the results above we conclude that choosing a small value

of T is more effective, allowing smooth adaptation of the NDP and IARP

broadcast intervals to current network conditions. Comparing AVERT to the

non adaptive protocols IZR and SPIZ shows that employing a method for

determining the optimal NDP and IARP broadcast intervals in real time can

lead to significant performance improvement both in terms of successful service

invocations and energy consumption.


85

4. PROFIT MAXIMIZATION MECHANISM FOR

SERVICE PROVISION IN MANETS

Optimized service provisioning is a challenging problem in dynamic environ-

ments such as Mobile Ad Hoc Networks (MANETs). Most of the existing

approaches assume an environment, where service provision is free (and dic-

tated) and servers do not have an incentive to maximize their benefit. In

this chapter we consider the nodes in MANETs to be independent, rational

agents trying to maximize their profits through service provision. We model

this problem as a Generalized Assignment Problem (GAP). We adopt a pay-as-

you-go model, where clients pay for the service as long as they are receiving the

service, since a pay-in-advance model would be unfair especially in MANETs

where connection loss is very probable. We introduce into the proposed profit

maximization algorithm expected payoffs based on estimates of server-to-client

connectivity. Those estimations can be used for computing the actual payoff

that will be received from any client that is selected by the service provider.

We experimentally study cases with non-cooperative and cooperative servers

and investigate the gain of the estimate based profit maximization algorithm

versus a classic profit maximization algorithm, which does not take into ac-

count the network’s dynamics that affect server-to-client connectivity. The

results show that our approach achieves up to three-fold improved server prof-

its compared to the classical one and is especially suited for MANETs with

high-mobility.

4.1 Background and Applications for Charged Service Provision in

MANETs

A basic assumption made by researchers when MANETs were in their

infancy, was that all nodes of the network were willing to cooperate. In ad-

86 Profit Maximization Mechanism for Service Provision in MANETs

dition, the view of the participating nodes in the network was that they were

peers of equal (and possibly restricted) capabilities. At the same time the

assumption of unconditional cooperation among nodes in a MANET began

to seem unrealistic and issues arising from malicious, misbehaving and selfish

(non-cooperative) node behavior came to the surface. In this context, the de-

ployment of MANETs for commercial purposes (e.g. spontaneous markets for

charged exchange of services and information) was considered a far vision.

A first step toward the “commercialization” of MANETs was to consider

them as physical extensions (a term commonly referred to as extended hotspots)

of infrastructure networks that allow service providers (e.g. mobile network

operators) to extend their coverage to the members of the MANET; assuming

that the MANET is essentially connected through a gateway to the provider’s

backbone network. Such a system was proposed in [71] where subscribers of

a certain provider may be granted access to the network’s services in places

where there is no fixed network coverage (base stations) but coverage is sup-

ported by a MANET formed by such nodes and rooted to a special gateway

node. The proposed system in [71] utilized the Polynomial-Assisted Ad Hoc

Charging Protocol (PACP) [72] for Charging and rewarding nodes either for-

warding data or receiving service through the MANET. PACP is based on the

creation of proofs for forwarding packets, which are cryptographically secured,

and which eventually reach the provider’s accounting system, such that the

provider may offer discounted services for packet forwarders and also charge

the users of the provided services.

However, the increasing heterogeneity of wireless mobile devices capable of

forming spontaneous networks led the researchers to differentiate the nodes of a

MANET based on their hardware (also software) and particular characteristics

(e.g. memory, storage capacity, battery etc.) and consider that some of the

nodes could actually take up the roles of mini mobile service providers. In this

context, where there is no connection to an infrastructure, service provision

but also the most elementary operation of a MANET, routing, depends on the

willingness of the participants to forward other nodes’ traffic. To deal with

Background and Applications for Charged Service Provision in MANETs 87

this problem, significant research efforts ( [73], [74], [75] and [76]) have been

devoted on developing mechanisms based on the exchange of virtual money,

or credit payments from sender nodes to forwarding nodes, such that packet

forwarding has the right incentives.

This evolution path has brought us closer to commercial-purpose MANETs

either connected through gateways to one or more infrastructure-based net-

works (or directly to the internet) or standalone. In the former case, granting

access to all the members of the MANET to the backbone may not be possible,

especially if the interconnection link is of low capacity. Imagine for example

the scenario of an extended hotspot (e.g. for a festival taking place in a remote

location with insufficient fixed network coverage, or an extended hotpot cover-

ing a crowded beach), where the members of the MANET wish to transfer (or

even stream!) audio and video from their location or chat with their friends or

family (possibly over the internet). Now in the case of a standalone MANET,

mini-mobile service providers may offer their services to the rest of the nodes

in the MANET, but certainly their serving capabilities are too limited to be

able to grant all service requests. For both cases it is essential for the service

providers to be able to select to serve (without violating their capacity con-

straints) only those clients that would be willing to pay the most for getting

the service. In this chapter we target such commercial-purpose MANETs and

investigate how service providers can maximize their profits by accounting for

the volatility of a MANET environment.

In the next Section we present related work. In Section 4.3 we describe

in detail the system under consideration. In Section 4.4 we briefly describe

the Generalized Assignment Problem (GAP) and we formulate the problem of

server profit maximization as such. In Section 4.5 we develop our model for

approximating the connectivity between a client and a server and in Section

4.6 we present simulation results.


4.2 Background on Connectivity Estimation

As already stated in the previous section, in MANETs service breaks can be

much more frequent due to increased link and path failures. In [77] and [61]

it is shown that for medium and high node speeds the path (availability)

duration decreases exponentially and hence only short-lived communications

requiring at most a few tens of seconds will be completed before a path break

occurs. In Section 3.1 we introduced the Service Availability Duration (SAD)

metric. Based on measurements using various speeds and densities and (even)

assuming a high server to client ratio, we showed that a service does not

typically have a SAD larger than a couple of hundreds of seconds. If the

server to client ratio is less, then the SAD will drop to a few tens of seconds.

Also if the service state cannot be transferred when switching servers, then

the SAD actually drops to the path duration mentioned earlier. Moreover, in

the most severe cases, path breaks may also lead to network partitions that

separate clients from their prospective servers.

In the literature, there have been quite a few approaches trying to deal with

those severe cases. Their main aim is to identify when a partition is going to

happen so that the requested service is replicated in advance and connectivity

to it can be guaranteed. In [78] each client monitors the set of disjoint paths

between itself and the server and computes a metric. If this metric falls below a

certain threshold then a potential partition is identified and server replication

is initiated. This method assumes that services are such that client nodes may

also bear to host them and also no considerations on server profits are taken.

In [79] the authors propose a distributed localized algorithm for detecting

critical nodes. Critical nodes are those nodes that if they get disconnected,

a network partition will occur. Using this algorithm, servers can predict the

partition and replicate the service to another node, which is part of the right

future partition. In [80] a service backbone formation mechanism is proposed

to handle network partitions. This backbone is such that every node can

contact at least one of its members in at most r hops. Every node monitors the

Background on Connectivity Estimation 89

number of nodes that are in its vicinity (r -hops away) and do not have access

to a server. The node with the highest number of such neighbors must get a

service replica. Once again, servers and clients are considered equally powerful

and service provision is free. In [81] a partition prediction model is proposed

based on grouping nodes according to their position and speed. Every client

sends its coordinates and velocity to the server. The server groups nodes based

on a pattern-matching algorithm. Having this global knowledge, the server can

predict future partitions and be replicated accordingly. The same assumptions

of replication capability for any node are also made here. Replication is also

used in [82], where an algorithm based on the partition detection mechanism

of the Temporally Orderered Routing Algorithm (TORA) is used along with

an optimized replica deployment scheme. Similar schemes are also proposed

in [83], taking also into account link failure probabilities during data replication

and trying to balance data accessibility and query delay.

Closely following the fundamental mechanism for data replication in mobile

Ad Hoc networks, initially proposed by T. Hara [84], all the aforementioned

approaches assume that service replication can always be carried out and do

not consider servers as business entities seeking to maximizing their profits.

They mainly focus on service continuation despite network partitioning. The

strong assumption that a service can always be replicated cannot hold in most

MANETs, either due to client device constraints, or to the nature of the ser-

vices. Imagine for example that a server node is providing live stock price

feeds obtained over its 3G connection. Such a service cannot be replicated to

any node since a 3G connection is necessary and might not be available.

In our approach, we consider non-replicable services and focus on the eco-

nomic potential from providing services in MANETs. A similar point of view

was adopted in [85]. However, the authors in [85] neglect that traditional so-

lutions to profit maximization for service providers in fixed networks are not

suitable to MANETs since they do not account for the volatility of a MANET,

assuming that services will be provided in full and hence payments will be re-


turned in full to service providers. The main factors that differentiate service

provisioning over MANETs from service provisioning over fixed networks are:

• Service provision in MANETs is opportunistic

– There are no fixed, well-known service providers.

– Any node (individual) can be a service provider for her own benefit

and for as long as she participates in the MANET, or for as long as

she desires to be a service provider.

• For-profit (charged) service provisioning in MANETs is in its infancy

and at this stage it is better for clients to make payment offers and let

servers decide whether they want to accept them or not. (Pull-based

model)

• MANET communications are significantly more unreliable and error-

prone than fixed network communications. Hence, service provision in

MANETs (especially if services are charged and a server’s goal is profit

maximization) must account for this.

• Server capacity is much more constrained in MANETs (because servers

are typically smaller in order to be light and mobile, on battery power

etc.).

– Communication costs are significant due to devices’ energy con-

straints.

– Client selection is a major issue especially if servers want to maxi-

mize their profit.

• Solutions for server profit optimization must be computationally efficient

due to the processing constraints of mobile devices.

We show that neglecting the volatility of a MANET leads service providers to

suboptimal client set selection. In our approach, we take into account path

failure probabilities in the proposed algorithms for enabling servers to select

System Description 91

the optimal client set that will maximize their expected profits. Also in [85]

the authors do not address cases where servers are non-cooperative and as a

result more than one server selects to serve the same client (which would result

in loss of profit for all but the server finally chosen by the client). We study

such cases and additionally we experimentally show that in MANETs the total

value for service provisioning can only be obtained if servers cooperate.

4.3 System Description

As already stated in the previous sections of this chapter, we study commercial-

purpose mobile Ad Hoc networks. Those networks are comprised of mobile

clients and mobile servers acting selfishly in that they try to maximize their

own profit. As it is done in [73], we assume the existence of management points

acting as central banks and controlling the flow of payments among clients,

servers and forwarders. Also, virtual money (possibly exchangeable to real

money) is used by clients for paying servers for the provision of a service. In

the experiments and for the obtained results, we assume that all intermediate

nodes on the path from a client to a server will not deny forwarding1.

We also assume that servers participating in the network have a finite

capacity. In our context capacity refers either to server memory or processing

capacity constraints or both. Periodically, servers announce their presence and

wait for client requests. After their announcement, servers begin collecting

client requests (including the client’s bid for the requested service) for a given

period. Upon the end of this period, the servers construct a schedule for serving

those clients that maximize their profit while not exceeding their capacity. We

should note here that there are the following two cases:

• If there are more than one servers belonging to the same owner, then

it is natural to assume that there is some form of communication (e.g.

1To some extent we could address some forms of stochastic denials of forwarding in ourmodeling of a probability of path breaks.


using a side channel) so that the servers can collaboratively decide on

the best client allocation among them (common knowledge required).

• If there are more than one servers, but they belong to different providers,

then each one would try to maximize its own profits. Here, two servers

may select the same client in their respective optimal client sets. In this

case, the client chooses the server that it estimates to be the most stable

(reachable longer).

Once they have finished providing the requested services to the current

client set, the servers enter their next announcement period. A last assumption

is that clients will only pay for the amount of service they have received. This

means that if a service provision is terminated earlier than its expected normal

termination, then the client will have paid only for the amount of the service

received until the termination happened. For example consider the scenario,

Fig. 4.1. VANET with mobile servers.

depicted in Figure 4.1, of a Vehicular Ad Hoc Network (VANET) where a car

Problem Formulation and Analysis 93

is capable of offering a navigation service to other cars in the network. A car

using this service will be paying for the service as long as it can reach the

serving vehicle via the multihop network (e.g. cents/packet).

4.4 Problem Formulation and Analysis

Based on the description given, the problem of server profit maximization

can be modeled as a Generalized Assignment Problem (GAP). The GAP is

defined as follows. There are n items x1 through xn and m bins. Each item has

a weight aij (weight of item j if assigned to bin i) and a value cij (value of item

j if assigned to bin i) and every bin has a capacity bi. The problem is to find the

optimal assignment of items into the bins such that the capacity constraints

of the bins are not violated and the total value obtained is maximized. The

mathematical formulation of the GAP is the following:

maximizem∑

i=1

n∑j=1

xij · cij (4.1)

subject to:n∑

j=1

xij · aij ≤ bi, i = 1, ..., m (4.2)

m∑i=1

xij ≤ 1, j = 1, ..., n (4.3)

xij ∈ {0, 1}, i = 1, ..., m, j = 1, ..., n (4.4)

This model is directly applicable to our problem if we assume that items

are clients, that aij is the amount of resources consumed at server j if it selects

to serve client i (also called requested capacity), that cij is client i’s payment

to server j and that bins are servers with serving capacities bi.

Solving the GAP is NP hard and it is even APX-hard to approximate it

[86]. However, there exist polynomial time approximation algorithms (having

approximation guarantee equal to (1−1/e−ε), for any ε > 0), further analysis

of which is out of the scope of this thesis (the interested reader is referred to [87]


and [88]). Especially since the size of commercial MANETs cannot grow more

than a few tens or hundrends of nodes and thus the instances of GAP are

small, the approximation algorithms (and even greedy algorithms) perform

satisfactorily.

Solving the classic GAP problem would lead to a solution vector xij , where:

xij =

⎧⎨⎩

1, if client j has been assigned to server i

0, if client j has not been assigned to server i

It is true that given the cij and solving the GAP, the xij vector is selected

so that we obtain the maximum∑m

i=1

∑nj=1 xij · cij.

However, if we consider that service breaks may occur and that clients pay

providers only for the part of the service received until the break happened,

then the produced xij vector may not lead to the actual profit maximizing

solution. The proof is given in the following:

Assume:

xij is the computed “optimal” solution vector,

∑mi=1

∑nj=1 xij · cij is the “optimal” profit obtained,

pij is the portion of the service actually received from client j when being

served by server i.

If there is one client l, allocated to server k, in the optimal allocation,

and another allocation of one client n to server k, not included in the optimal

allocation, for which:

akn = akl, ckn ≤ ckl and pkn ≥ pkl (4.5)

such that:

ckn · pkn ≥ ckl · pkl (4.6)

Problem Formulation and Analysis 95

then using a pay-as-you-go model and taking into account that link failures

may happen we get:

m∑i=1

n∑j=1

xij · cij · pij ≤m∑

i=1

n∑j=1

x∗ij · cij · pij (4.7)

where:

xkl = 1, xkn = 0 and x∗kl = 0, x∗

kn = 1 (4.8)

Hence, the computed solution using the classic GAP model is not always

the optimal one considering that service failures may occur. For getting close

to optimality in an error prone environment as a MANET, we enhance the

GAP model with estimates for the profits (instead of fixed profits) of service

provisioning. The model can then be re-formulated as follows:

maximizem∑

i=1

n∑j=1

xij · cij · pij (4.9)

subject to:n∑

j=1

xij · aij ≤ bi, i = 1, ..., m (4.10)

m∑i=1

xij ≤ 1, j = 1, ..., n (4.11)

xij ∈ {0, 1}, i = 1, ..., m, j = 1, ..., n (4.12)

Since we do not consider the case where a server can overbook, we do not

replace equation (10) with:

n∑j=1

xij · aij · pij ≤ bi, i = 1, ..., m , (4.13)

since if a server under-estimates pij it may admit more clients than it can

serve. The parameter pij is an estimate of the proportion of service that can

be delivered to client j by server i. This parameter is directly related to the

path failure probability from the client to the server due to channel conditions

or mobility. It is worth noting here that any algorithm for solving the GAP


problem can be directly applied to this model if instead of cij ’s, the cij · pij

values are taken into account. Also, the pij can be based on sophisticated

estimates on client-server connectivity as we will show in the next section.

It is obvious that solving the GAP requires server cooperation and can

be applied in MANETs where service providing nodes belong to the same

owner. Actually every server must inform her team-member servers about

her connectivity estimates pij to all clients, the client’s bids cij and their

requested capacities aij along with the server’s own maximum capacity. All

this information can be encoded into small vectors and be efficiently distributed

over a side-channel (e.g. 3G connection) among servers.

In case that competitive teams of service providers (i.e. teams of service

providers belonging to different owners) exist in the MANET, then each team

can solve a GAP and the solutions among teams may have overlaps. An

overlap means that a given client is chosen by more than one server teams. In

this case the client will select to get the service from the team whose server

is estimated to have the best connectivity to it. The connectivity is measured

as the proportion of the service that can be delivered while the client is in

contact with the server. In this case there will be a loss of profit for all the

other teams that have also chosen that particular client.

4.5 Approximation of Duration of Connectivity

As described in the previous section, servers announce their presence in

the network at regular intervals and wait for clients’ bids for the next serving

period. Then, based on those bids, the requested capacities and the estimation

of the proportions of service that each client will receive, the servers make

their selection about which clients to serve. In this section we will propose

an algorithm for estimating the duration of the connectivity, i.e. the expected

lifetime of network connection between a client and a server.

It is widely known that the duration of a path between two nodes in a

MANET is dependent on network density, on node speed, on the number of

Approximation of Duration of Connectivity 97

hops separating the two nodes and on the transmission range. Returning to

our case, a server must be able to obtain an estimate of the connectivity time

with any client that has requested to be served for the next serving period.

In the literature, there exist many approaches that try to estimate the du-

ration of a path between two nodes. The main tool in the hands of researchers

for analyzing the properties of path duration has been simulation. Simulation

results of [62] show that for paths of 4 hops or more the duration can be ap-

proximated using an exponential distribution, and in [89] an analytical model

was developed to validate these results. However, in reality, the volatility of

MANETs usually renders paths longer than 4 hops impractical [90]. Using

empirical data the authors of [91] have shown that the mean residual lifetime

of routes depends on the number of hops as well as on the mean link duration.

Also, Tseng et al. in [92] provide an analytical study on this issue.

However, analytical models on path duration in MANETs have also been

proposed in [93], [94]. A key assumption in those models is that the path

becomes invalid as soon as one of its links is broken. In our case we are in-

terested in the duration of a connection between two nodes irrespectively of

the change in intermediate links and paths. This is a fundamentally different

metric than path duration, since it essentially involves many possibly different

paths and also changes in path length throughout the lifetime of a connection.

Due to the large complexity for analytically obtaining approximations for es-

timating connection duration (including many degrees of freedom), we obtain

estimates based on our empirical results under various network settings. We

will assume that all nodes move according to the Random Waypoint (RWP)

mobility model at constant speed and without pause time, and also a perfect

MAC (actually connectivity relies only on physical proximity).

Regarding the impact of the initial distance (in number of hops) between

client and server on the expected connection duration, it is intuitive that the

fewer the hops, the better the connectivity between the two nodes. In the

following we show how density impacts the connection duration when the client

is at an initial distance of up to 4 hops from the server. Our simulation-based


measurements show that average connection duration follows certain patterns

and can be accurately estimated. The results were obtained assuming serving

periods of 100 seconds. For node mobility we use the Random Waypoint

mobility model with constant speed (maximum speed takes the values 3.5m/s,

7m/s and 14m/s) and no pause time. We used the Qualnet simulator [68]

for obtaining the connection duration measurements. Under these settings

we have tested several scenarios (see Table 4.1) by modifying the number of

participating nodes and the terrain sizes (node range was fixed at 380 meters).

We compute network density using the following formula:

D =N ·π·R2

t

TerrainSize,

where N is the number of clients and Rt is the transmission range.

Side of (square) Terrain Number of Nodes Density2000m 11 1,38162000m 17 2,13521500m 11 2,45612000m 22 2,72631250m 11 3,53681500m 17 3,79591500m 22 4,91231250m 17 5,46611000m 11 5,52641250m 22 7,07371000m 17 8,5408

Table 4.1Network Density Values for square terrain.

In Figure 4.2 we present the results. Each point in the diagram represents

an average connection duration obtained from experiments with 90 seeds and

having duration of 4000 seconds each (this corresponds to 3600 serving peri-

ods). From Figure 4.2 we see that for a fixed speed and serving period the

average duration of connection with initial distance between client and server

of 1, 2, 3 and 4-hops, can be well approximated (R2 > 98%) as a logarithmic

function of density that increases as density increases (due to the change in

the number of nodes or the terrain size).

Approximation of Duration of Connectivity 99

Fig. 4.2. Connection Duration vs. Density and Number of Hops.

In Figure 4.3 we present how those logarithmic trends change as speed

increases. Figure 4.3 presents the results for 1 hop initial distance but the

impact on the pattern is similar when having an initial distance of 2, 3 and

4-hops also. Modeling how speed impacts the coefficients of the logarithmic

functions used to approximate the connection durations we discovered that

they can be well approximated (R2 > 98%) by another set of logarithmic

functions.

Hence we can derive analytical formulas, which take into account node

speed and density and estimate the connection duration between a client and

a server when their initial distance is 1-hop, 2-hops, 3-hops or 4-hops. The

formulas are of the following form: Fx = (ax · ln(speed) + bx) · ln(Density) + (cx ·

ln(Speed) + dx), where x is 1 to 4 and corresponds to number of hops.

Hence, if we assume that nodes also include their position coordinates and

speed along with the bids they sent to servers, servers can actually compute

network density and average speed. Servers also know the number of hops

required to reach every client and hence by using the approximation formulas


Fig. 4.3. Connection Duration vs. Density and Speed.

derived from the experimental results, they could estimate the pij values as

follows:

pij = min{Serving period duration, Fx between client j and server i}Serving period duration

4.6 Performance Analysis of the Profit Maximization Algorithm

For computing the solution to the GAP when servers are cooperative we

have used a branch and bound algorithm based on linear programming that

always finds the optimal solution. We compare the results of the algorithm

when it is run:

C-GAP : Without considering payment estimates for solving the classic GAP

(C-GAP algorithm),

OE-GAP: Considering payment estimates and accurate knowledge of client-

server connectivity (E-GAP Oracle algorithm),

Performance Analysis of the Profit Maximization Algorithm 101

AE-GAP: Considering payment estimates with approximations obtained for

client-server connectivity based on density and speed (E-GAP Ap-

proximation algorithm).

For the non-cooperative server case we consider single server non-cooperative

teams. In this case each server solves a specific version of GAP where there is

only one server (no info on other servers). This is actually a Single Knapsack

(SK) problem, which we also solve with a branch and bound algorithm that

always finds the optimal solution. Once again we compare the results of the

algorithm when it is run:

C-SK : Without considering payment estimates for solving the classic SK

(C-SK algorithm),

OE-SK: Considering payment estimates and accurate knowledge of the path

failure probability (E-SK Oracle algorithm),

AE-SK: Considering payment estimates with approximations obtained for

client-server connectivity based on density and speed (E-SK Ap-

proximation algorithm).

Using Qualnet [68], we have simulated a MANET consisting of 2 servers and

20 clients. The node range has been set to 380 meters. Server cooperation

and non-cooperation scenarios have been investigated. In order to take into

account the effects of node speed on our algorithms, all nodes move following

the Random Waypoint Mobility model with constant speed (experiments for

a variety of different speeds have been conducted). We also investigated the

effects of density by using various square terrain sizes ranging from sides of

1250 meters to 2500 meters.

For simplifying the analysis of our experimental results we assume that

aij=cij=1 unit (requested capacity and payment unit respectively) for all i’s

and j’s2, and that the 2 servers have equal capacities (either 5 or 25 units).

2This way we actually can talk about client allocations to servers, which is more compre-hensible. Experiments with aij following a uniform distribution led to the same conclusionsand are omitted for the shake of better readability.


When servers have 5 units of capacity this means that each server can provide

only part of the requested services (since there are 20 clients), while when

servers have 25 units of capacity this means that each server may serve all

requests on his own.

4.6.1 Cooperative Servers

Figure 4.4 presents simulation results considering cooperative servers hav-

ing capacity of 5 units (hence each server may serve up to 5 clients in every

serving period). Nodes move according to the Random Waypoint model with

constant speed of 3.5m/s. The results are average values obtained over 20

experiments with different seeds, each having duration of 4000 seconds (40

serving periods per experiment). The y-axis presents the percentage of to-

tal profit gain/loss using the E-GAP solving algorithm (taking into account

client-server connectivity) instead of the classic GAP solving algorithm. The

x-axis presents different densities.

Fig. 4.4. Impact of server capacities and node density on ProfitGains by cooperative servers using E-GAP vs. using GAP.

Performance Analysis of the Profit Maximization Algorithm 103

In this simple case, since all clients have the same requests and pay the

same amount of money to any server, if servers solve the classic GAP problem

then the client to server assignment is actually done randomly by the C-GAP

algorithm. However, since path failures occur, such a random assignment can-

not lead to profit maximization for the 2 servers due to the fact that clients

may not be assigned to the server they are best connected to (i.e. the server for

which the client-server connection has the maximum expected lifetime among

all connections to servers). Using the E-GAP algorithms (oracle or approxi-

mation), path failures are taken into account and the allocation of clients to

servers is done optimally (i.e. every client is allocated to the server it is best

connected to), hence leading to a maximization of the total profits obtained.

The sparser the network and the higher the mobility, the greater is the need

for carefully choosing the clients to be served, since path failures are more

prevalent and the random assignment will most likely lead to frequent con-

nection failures and to decreased profits (assuming the pay-as-you-go model).

This fact is validated by the results presented in Figure 4.4, where it is ev-

ident that the profit gain increases with decreasing density. Also, the profit

gains are larger when the servers cannot satisfy all the demand, since then it

is even more crucial to select to serve only the best connected subset of clients.

Finally, it is shown that the approximation algorithm closely follows the or-

acle algorithm, which has perfect knowledge of the connectivity between any

client-server pair, hence leading also to much better performance than that of

the C-GAP algorithm.

In Figure 4.5 we show the experimental results regarding the investigation

of the impact of mobility on the profit gains obtained using the E-GAP algo-

rithms for cooperative servers. The general simulation setup is the same but

now nodes are placed on terrains of different density. We have selected to eval-

uate the impact of mobility allowing the nodes to move on a 2500x2500 meters

square terrain for one scenario and on a 1750x1750 meters square terrain for

a second scenario. Simulation traces prove that the second (higher density)

scenario allows longer (by 20%) client to server connection durations. In this


higher density scenario connectivity is generally maintained for extended pe-

riods of time when node mobility is low. The problems in connectivity rise

as the mobility increases. Higher mobility means that paths break more often

or more easily. This is why the E-GAP algorithms show significantly more

gains than the C-GAP algorithm when the speed is set to 15m/s as compared

to the gains achieved when the speed is set to 5m/s. For the lower density

scenario mobility does not have a significant impact on the gains from using

the E-GAP algorithm. This is because the profits are already decreased in this

lower density scenario for the GAP algorithm due to the sparser topology, and

its suboptimal selection of clients (selection of clients that are likely to loose

connectivity to the server). In this scenario, increasing mobility may nega-

tively impact the GAP algorithm but not as severely as in the high density

scenario, where the context-agnostic client selection by the GAP algorithm

has more chances to be as good (i.e. connections to selected clients have long

durations) as the selection performed using the E-GAP algorithm only for low

mobility due to the denser topology.

Fig. 4.5. Impact of mobility on Profit Gains by cooperativeservers using E-GAP vs. using GAP.

Conclusions 105

4.6.2 Non-Cooperative Servers

The results of the simulation when the servers are non-cooperative are ac-

tually the same as the ones observed in Figures 4.4 and 4.5, the only difference

being that in cases where the servers have capacities of 25 units (each server

can satisfy the whole demand on its own) the profit gains for using E-GAP

instead of GAP are zero. This is explained by the fact that either GAP or

E-GAP will result in the solution that each server selects to serve every client

in order to maximize its profit. The same conflicts will exist and hence the

total profits will be the same irrespectively of the algorithm.

What is more interesting is that the total profit obtained in the non-

cooperative case is on the average lower. This is due to the fact that when

servers are non-cooperative the client-to-server allocations may include con-

flicting sets of clients, since the servers do not try to optimally “share” clients.

A conflict means that a client has been selected by more than one server.

Since this client will be finally served by one of the servers, this results to loss

of profit for the other server that had selected that same client in its alloca-

tion. In Figure 4.6 we show that being non-cooperative is bad for the total

profit obtained especially for high density scenarios. In such scenarios there

are more chances for servers to select the same clients and hence have conflicts,

which results in decrease of the total profits. However, in sparser networks,

the servers will most likely select non-conflicting client sets consisting of nodes

in their respective vicinities. Hence, the total profits obtained will not differ

compared to the case of having cooperative servers.

4.7 Conclusions

In this chapter we have extensively studied the problem of optimized ser-

vice provisioning in MANETs. We argued that when service provision is not

free and capacity constrained service providers seek to maximize their profits

by providing services to other nodes in the MANET, it is crucial that opti-


Fig. 4.6. Loss in Total Profit when servers are non-cooperative(Server Capacity = 5 units, Servers use the E-GAP (oracle))

mization mechanisms exist for selecting the best client-set to be served. We

have modeled this problem as a Generalized Assignment Problem (GAP) and

showed the benefits of taking into account server-to-client distances and es-

timates for the connection failure probabilities in the algorithms for solving

it.

We have studied cases with non-cooperative and cooperative servers. Our

simulations showed that the proposed estimate based algorithms are valuable,

especially for sparse MANETs. Their performance compared to a classic GAP

algorithm was better by as much as a factor of 3 in terms of accumulated

profits for the servers. Further experiments revealed also that the sum of

individual profits for competitive servers can be larger if the servers choose to

cooperatively select their respective client sets.

107

5. FUTURE WORK

In the following sections we identify directions for future research especially

regarding AVERT and also the problem of profit maximization from service

provision in mobile Ad Hoc networks. We also present more general open

issues related to service discovery in mobile Ad Hoc networks that present an

open field for further investigation and research.

5.1 Smart adaptation of AVERT using service information

As described in Chapter 3, in AVERT every node uses the BFO mechanism

in order to adapt its frequency of sending proactive traffic based on the data

traffic volume observed on its network interface. In the context of this disser-

tation the BFO mechanism performance has been investigated assuming some

sort of homogeneity in the provided services (in the sense of the amount of

transferred data upon a service invocation). It would be interesting to study

the BFO mechanism under increased heterogeneity by assuming that there

exist various services, the invocation of which involves different amounts of

data volumes to be transferred between client and server. This would make

it necessary to enhance BFO with the capability of identifying the type of

service the locally seen data belongs to. For example, in the case that a node

observes that the service data traffic has decreased compared to the previous

period it would decrease the rate of sending proactive traffic. This is a correct

decision if the change in data traffic is due to changes in the network topol-

ogy (e.g. due to mobility, availability of paths etc.). However, if the case is

that this decrease in service data traffic is due to the fact that nodes have

started to invoke services involving smaller amounts of data transfers (assum-

ing no topology changes), then decreasing the rate of sending proactive traffic

could prove a bad choice. In this case the BFO mechanism should be able to

108 Future Work

make more informed decisions by taking also into account the characteristics

of the invoked services (further increasing the interaction between layers in the

protocol stack).

5.2 Connectivity duration estimation formulas under various mo-

bility models

In Chapter 4 we underlined the need of connectivity estimation formulas

when capacity constrained service providers try to maximize their profits by

offering their services to other members of a mobile Ad Hoc network. The

derived formula for estimating the duration of connectivity between client and

servers presented in section 4.5 is valid when the nodes of the MANET move

according to the random waypoint mobility model. Future research around

connectivity duration estimation could involve the derivation (if possible) of

such formulas when the nodes follow other well-known mobility models, like the

Manhattan model or the group-based mobility model. Under those mobility

models we do not know if connectivity estimation formulas can be derived or

what is their form (e.g. continuous, linear, exponential).

5.3 Interoperability

Considering the multitude of service discovery standards, architectures and

protocols and taking also into account the ubiquitous and pervasive nature of

future environments, interoperability in service discovery will be a major issue

requiring attention (to avoid building a ‘Tower of Babel’). It is clear that re-

quiring all devices to support all service discovery protocols is far from being

realistic. To the contrary interoperation seems to be the way forward. Despite

a few efforts [95], [96], [97], [98], [99], [100] much remains to be done toward

this. It is out of the scope of this section to analyze in detail the approaches

proposed for service discovery protocol interoperability; however, it is worth

outlining their basic characteristics and weaknesses. Some of the approaches

Benchmarking 109

try to make direct translations from one protocol to another, while others try

to translate all protocols to a common protocol. It is obvious that not ev-

ery protocol provides the same functionality and some mappings are simply

not achievable (e.g. UPnP service status notifications cannot be mapped to

any SLP function). On the other hand defining a common protocol which

can support all possible functionality ranging from service description meth-

ods to service invocation to security provision to context-awareness etc. is too

optimistic if not impossible. Furthermore, some approaches, called explicit, re-

quire that client applications make calls to the common protocol implemented

as middleware. Other approaches, called transparent, implement middleware

that accepts any service discovery protocol call issued by legacy clients and

transforms it appropriately depending on the protocol provided in the network.

However, all of those approaches (transparent or explicit) require that transla-

tion modules for all possible protocols are available in the middleware, or that

nodes are always available to make translations (either fixed bridges or mobile

clients), or that thin client devices can deal with the complexity of the code

required for identifying the used service discovery protocols and for making

protocol translations. Considering the above, further work is needed for devel-

oping truly scalable interoperability solutions for service discovery, matching

the requirements as well as the restrictions posed by MANET environments.

5.4 Benchmarking

One of the major problems in the research area of service discovery for

MANETs is that little attention has been given in standardizing the evalua-

tion of service discovery protocols. In effect discovery protocols found in the

literature are often incomparable since different settings and assumptions have

been made during their evaluation.

Developing a universal evaluation framework for service discovery protocols

would allow fair and direct comparisons among protocols. Maybe a first effort

toward such a framework is the one proposed in [101], where authors present

110 Future Work

the BenchMANET. This benchmark specifies a number of tests associated with

realistic service discovery applications and scenarios for MANETs. However,

this framework does not take into account important parameters e.g. related to

the frequency of advertisement and querying or the specifics of the underlying

routing protocol.

Besides simulation an evaluation can use analytical models too. Regard-

ing analytical modeling and evaluation of service discovery protocols there

have been proposed two models in [102] and [103]. The model developed

in [102] uses a M/G/c/c queue to model and predict the behavior of the

service cache on a node. Using the model the average timeout of a service

description can be determined given the protocol’s parameters (e.g., advertise-

ment frequency, service cache size etc.). Also the optimal timeout for service

descriptions can be calculated for achieving a non-fluctuating average num-

ber of services discovered (equilibrium). Unfortunately the aforementioned

model, being abstract, cannot take into account radio link behavior and node

mobility as well as of other specifics of the service discovery protocol (e.g.

forwarding policies). Moreover the model was designed only for evaluating

proactive directory-less service discovery. Directory-based service discovery

was modeled in [103], where a queuing based model for service caches was also

developed. However, this model is more elaborate since it accounts for ser-

vice provider movement and update message failure rate. Through this model

the optimal service advertisement update rate can be determined in order to

optimize system performance in terms of success rate and network overhead.

However important factors as the mobility of clients and the network density

have not been taken into account in the proposed model. Analytical models

can prove to be valuable tools in the evaluation and optimization of developed

service discovery protocols, but require more sophistication to cover a broader

range of approaches (e.g. hybrid discovery architectures) taking into account

all/most of the important factors affecting the performance of service discovery

protocols in MANETs.

111

6. SUMMARY AND CONCLUSIONS

This chapter briefly summarizes the research conducted in this dissertation

and presents the main conclusions.

In the first part of this research the focus has been on studying the energy

savings from implementing service discovery in the routing layer instead of the

application layer. Similar work in the literature involved comparisons of inte-

grated approaches to application layer based approaches using global flooding.

However, since global flooding is not a viable option for communication in

energy and bandwidth constrained networks like MANETs, the results found

in the literature do not have a realistic base. In this work we have obtained

both analytical and experimental results from comparing similar service dis-

covery approaches (suitable to MANETs) but implemented at different layers.

In this context we have proposed, designed and implemented two innovative

integrated protocols combining service and route discovery capabilities.

The first integrated protocol, named E-ZRP, is a hybrid protocol and was

compared to a similar protocol implemented at the application layer, namely

the APS protocol. For our study we first developed a simple yet accurate

enough analytical model for obtaining the energy savings experienced when

using E-ZRP instead of APS. This model’s innovation stands in the fact that

it captures the characteristics of broadcasting in a MANET (collisions) and

takes into account message sizes, the network density and also the rate of send-

ing advertisement packets (parameters that have been neglected in analytical

models found in the literature when comparing service discovery protocol and

which severely affect the energy consumption).

We have also provided an assessment of the traffic overhead required for

reactive service discovery using flooding versus using bordercasting. The major

finding of this assessment is that depending on network topology bordercasting

112 Summary and Conclusions

may require orders of magnitude less overhead to propagate a service query

inside a MANET. This is especially true for dense MANETs (where nodes

have a high average number of 1-hop neighbors), where the degree of the

spanning tree connecting nodes to their border nodes is low. In those cases,

using bordercasting, a node may “scan” the network for the requested service

very efficiently using only a very small fraction of the messages that would be

required for a search using global flooding.

We have also performed extensive simulations for showing the performance

of the developed service discovery schemes and have extended our evaluation

to take into account the service discoverability. We showed that even if the

application layer based scheme is tuned to be much more conservative in ser-

vice advertising than the integrated scheme, the insignificant energy gains of

3% come at the cost of dramatically lower discoverability (22% to %43, de-

pending on mobility). We have also studied the impact of network density,

node mobility and channel characteristics of a MANET in the duration of ser-

vice sessions for both service discovery schemes. We have shown that under

a perfect channel both schemes can achieve similar service session availability

(the application layer based scheme being more expensive in energy consump-

tion). However, under a realistic channel the application layer based scheme,

by increasing the amount of traffic in the network (and hence the collisions),

achieves shorter service session durations. This phenomenon is expected to

be even more profound in cases of high network density where the increase in

transmission collisions, due to the increased overhead from using two separate

processes for service and route discovery, is even higher.

Continuing our research on energy efficient hybrid integrated route and

service discovery protocols we have developed AVERT. AVERT is a sophisti-

cated protocol able to adapt its proactive and reactive operation based on the

characteristics of a MANET. It is also capable of adapting the rate of sending

proactive traffic by monitoring the traffic pattern seen on the local interface

of a node. Comparing AVERT to other non adaptive protocols of the same

class shows that employing a method for determining the optimal proactive

113

traffic broadcasting intervals in real time can lead to significant performance

improvement both in terms of successful service invocations and energy con-

sumption.

In the second part of this research we have extensively studied the problem

of optimized service provisioning in Mobile Ad Hoc Networks (MANETs). We

argued that when service provision is not free and capacity constrained service

providers seek to maximize their profits by providing services to other nodes in

the MANET, it is crucial that optimization mechanisms exist for selecting the

best client set to be served. We have modeled this problem as a Generalized

Assignment Problem (GAP) taking into account server-to-client distances and

estimates for the connection failure probabilities in the algorithms for solving

it. We have studied cases with non-cooperative and cooperative servers. Our

simulations showed that the proposed estimate based algorithms are valuable,

especially for sparse mobile Ad Hoc networks. Their performance compared to

a classic GAP algorithm was better even by a factor 3 in terms of accumulated

profits for the servers. Further experiments revealed also that the sum of

individual profits for competitive servers can be larger if the servers choose to

cooperatively select their respective client sets, especially in dense networks.

114

LIST OF REFERENCES

115

LIST OF REFERENCES

[1] M. Gerla, “From battlefields to urban grids: new research challenges inad hoc wireless networks,” Pervasive Mobile Computing, vol. 1, no. 1,pp. 77–93, 2005.

[2] K. Arnold, B. O’Sullivan, R. W. Scheifler, J. Waldo, and A. Wollrath,The Jini Specification. Addison Wesley, 1999.

[3] Salutation Consortium, Salutation Architecture Specification. Availableat: http://web.archive.org/web/20030623193812/www.salutation.org/,1999, (The Salutation Consortium was disbanded on 30 June 2005).

[4] Microsoft Corporation, Universal Plug and Play: Background. Availableat: http://www.upnp.org/resources/UPnPbkgnd.htm, 1999.

[5] B. S. I. Group, Specification of the Bluetooth System. Available at:http://www.bluetooth.com, 1999.

[6] E. Guttman, C. Perkins, J. Veizades, and M. Day, Service Location Pro-tocol, Version 2. IETF RFC 2608, 1999.

[7] Apple Inc, Bonjour Technology White Paper. Available at:http://images.apple.com/macosx/pdf/MacOSX Bonjour TB.pdf, 2007.

[8] B. Pascoe, Salutation-Lite: Find-and-Bind Technologiesfor Mobile Devices. Salutation Consortium, Available at:http://web.archive.org/web/20031023072133/www.salutation.org/whitepaper/Sal-Lite.pdf, June, 1999.

[9] C. Ellison, “Home network security,” Intel Technology Journal, vol. 6,pp. 37–49, 2002.

[10] M. A. E. Saoud, T. Kunz, and S. Mahmoud, “SLPManet: service lo-cation protocol for MANET,” in Proceedings of the International Con-ference on Wireless Communications and Mobile Computing, IWCMC2006, Vancouver, British Columbia, Canada, July, 2006 (S. Onoe,M. Guizani, H.-H. Chen, and M. Sawahashi, eds.), pp. 701–706, ACM,2006.

[11] S. Cheshire and M. Krochmal, Multicast DNS (Internet Draft). Ap-ple Computer, Inc., Available at: http://files.multicastdns.org/draft-cheshire-dnsext-multicastdns.txt, 2004.

[12] S. Cheshire and M. Krochmal, DNS-Based Service Discovery (Internet-Draft (work in progress)). Apple Computer, Inc., Available at:http://files.dns-sd.org/draft-cheshire-dnsext-dns-sd.txt, 2006.

[13] Universal Description Discovery and Integration Platform. Availableat: http://www.uddi.org/pubs/Iru UDDI Technical White Paper.pdf,2000.

116

[14] U. C. Kozat and L. Tassiulas, “Service discovery in mobile ad hoc net-works: an overall perspective on architectural choices and network layersupport issues,” Ad Hoc Networks Journal, vol. 2, no. 1, pp. 23–44, 2004.

[15] F. Sailhan and V. Issarny, “Scalable service discovery for MANET,” inthe Proceedings of the 3rd IEEE International Conference on PervasiveComputing and Communications, 2005, PerCom 2005, Kauai Island,Hawaii, March, 2005, pp. 235–244, IEEE Computer Society, 2005.

[16] M. Klein, B. Konig-Ries, and P. Obreiter, “Service rings - A semanticoverlay for service discovery in ad hoc networks,” in the Proceedings ofthe 14th International Workshop on Database and Expert Systems Ap-plications, DEXA ’03 Workshops, pp. 180–185, IEEE Computer Society,2003.

[17] M. Klein and B. Konig-Ries, “Multi-layer clusters in ad-hoc networks -an approach to service discovery,” Lecture Notes in Computer Science(Revised Papers from the NETWORKING 2002 Workshops on Web En-gineering and Peer-to-Peer Computing), vol. 2376, pp. 187–201, 2002.

[18] G. Schiele, C. Becker, and K. Rothermel, “Energy-efficient cluster-basedservice discovery for ubiquitous computing,” in the Proceedings of the11th ACM SIGOPS European Workshop, Leuven, Belgium, September,2004, ACM, 2004.

[19] K. Seada and A. Helmy, “Rendezvous regions: A scalable architecturefor service location and data-centric storage in large-scale wireless net-works,” in the Proceedings of the 18th Interantional Parallel and Dis-tributed Processing Symposium (IPDPS’04) - Workshop 12, Santa Fe,New Mexico, April, 2004, p. 218a, IEEE Computer Society, 2004.

[20] S. Sivavakeesar, O. Gonzalez, and G. Pavlou, “Service discovery strate-gies in ubiquitous communication environments,” IEEE Communica-tions Magazine, vol. 44, no. 9, pp. 106–113, Sept. 2006.

[21] J. Tyan and Q. H. Mahmoud, “A comprehensive service discovery solu-tion for mobile ad hoc networks,” ACM/Springer Mobile Networks andApplications (MONET) Journal, vol. 10, no. 4, pp. 423–434, 2005.

[22] M. Nidd, “Service discovery in deapspace,” IEEE Personal Communica-tions, [see also IEEE Wireless Communications], vol. 8, no. 4, pp. 39–45,August, 2001.

[23] C. Lee, S. Helal, and W. Lee, “Gossip-based service discovery in mobilead hoc networks,” IEICE Transactions on Communications, vol. 89-B,no. 9, pp. 2621–2624, 2006.

[24] C. Campo, M. Munoz, J. C. Perea, A. M. Lopez, and C. Garcıa-Rubio,“PDP and GSDL: A new service discovery middleware to support spon-taneous interactions in pervasive systems,” in the Proceedings of the 3rdIEEE International Conference on Pervasive Computing and Communi-cations, 2005, PerCom Workshops 2005, Kauai Island, Hawaii, March,2005, pp. 178–182, IEEE Computer Society, 2005.

[25] S. Helal, N. Desai, V. Verma, and C. Lee, “Konark - a service discoveryand delivery protocol for ad-hoc networks,” IEEE Wireless Communica-tions and Networking, WCNC 2003, vol. 3, pp. 2107–2113 vol.3, March,2003.

117

[26] S.-J. Lee, W. Su, J. Hsu, M. Gerla, and R. Bagrodia, “A performancecomparison study of ad hoc wireless multicast protocols,” in the Pro-ceedings of the 19th Annual Joint Conference of the IEEE Computer andCommunications Societies (INFOCOM 2000), Tel-Aviv, Israel, March,2000, pp. 565–574, 2000.

[27] D. Chakraborty, A. Joshi, Y. Yesha, and T. W. Finin, “Toward dis-tributed service discovery in pervasive computing environments,” IEEETransactions on Mobile Computing, vol. 5, no. 2, pp. 97–112, 2006.

[28] O. Ratsimor, D. Chakraborty, A. Joshi, and T. Finin, “Allia: alliance-based service discovery for ad-hoc environments,” in the Proceedings ofthe Second ACM International Workshop on Mobile Commerce (WMC-02), New York, USA, September, 2002, pp. 1–9, ACM Press, 2002.

[29] Z. guo Gao, L. Wang, M. Yang, and X. Yang, “CNPGSDP: An effi-cient group-based service discovery protocol for MANETs,” ComputerNetworks, Elsevier, vol. 50, no. 16, pp. 3165–3182, 2006.

[30] Z. guo Gao, L. Wang, X.-Z. Yang, and D. Wen, “PCPGSD: An enhancedGSD service discovery protocol for MANETs,” Computer Communica-tions, Elsevier, vol. 29, no. 12, pp. 2433–2445, 2006.

[31] Z. guo Gao, X.-Z. Yang, T. yi Ma, and S.-B. Cai, “RICFFP: An efficientservice discovery protocol for MANETs,” in the Proceedings of the In-ternational Conference on Embedded and Ubiquitous Computing, EUC2004, Aizu-Wakamatsu City, Japan, August, 2004 (L. T. Yang, M. Guo,G. R. Gao, and N. K. Jha, eds.), vol. 3207 of Lecture Notes in ComputerScience, pp. 786–795, Springer, 2004.

[32] A. Nedos, K. Singh, and S. Clarke, “Service*: Distributed service ad-vertisement for multi-service, multi-hop MANET environments,” in theProceedings of the 7th IFIP International Conference on Mobile andWireless Communication Networks (MWCN’05), Marrakech, Morocco,September, 2005.

[33] S. Motegi, K. Yoshihara, and H. Horiuchi, “Service discovery for wirelessad hoc networks,” in Proceedings of the 5th International Symposiumon Wireless Personal Multimedia Communications, Honolulu, Hawaii,October, 2002, vol. 1, pp. 232–236.

[34] J. B. Tchakarov and N. H. Vaidya, “Efficient content location in wirelessad hoc networks,” in the Proceedings of Mobile Data Management 2004(MDM’04), Los Alamitos, CA, USA, p. 74, IEEE Computer Society,2004.

[35] G. E. Guichal, “Service location architectures for mobile ad hoc net-works,” Master Thesis, Georgia Institute of Technology, 2001.

[36] P. Engelstad and Y. Zheng, “Evaluation of service discovery architec-tures for mobile ad hoc networks,” in the Proceedings of the 2nd An-nual Conference on Wireless On-demand Network Systems and Services(WONS ’05), St. Moritz, Switzerland, 2005, pp. 2–15, IEEE ComputerSociety, 2005.

[37] H. Luo and M. Barbeau, “Performance evaluation of service discoverystrategies in ad hoc networks,” in the Proceedings of the 2nd Annual Con-ference on Communication Networks and Services Research (CNSR),

118

Fredericton, N.B., Canada, May, 2004, pp. 61–68, IEEE Computer So-ciety, 2004.

[38] U. Mohan, K. C. Almeroth, and E. M. Belding-Royer, “Scalable ser-vice discovery in mobile ad hoc networks,” in the Proceedings of the3rd International IFIP-TC6 Networking Conference NETWORKING2004, Athens, Greece, May, 2004 (N. Mitrou, K. P. Kontovasilis, G. N.Rouskas, I. Iliadis, and L. F. Merakos, eds.), vol. 3042 of Lecture Notesin Computer Science, pp. 137–149, Springer, 2004.

[39] J. Hoebeke, I. Moerman, B. Dhoedt, and P. Demeester, “Analysis of de-centralized resource and service discovery mechanisms in wireless multi-hop networks,” Computer Communications, Elsevier, vol. 29, no. 13-14,pp. 2710–2720, 2006.

[40] R. Koodli and C. E. Perkins, “Service discovery in on-demand ad hocnetworks,” IETF Internet Draft, draft-koodli-manet-servicediscovery-00.txt.

[41] A. Garcia-Macias and D. A. Torres, “Service discovery in mobile ad-hocnetworks: Better at the network layer?,” in the Proceedings of the 34thInternational Conference on Parallel Processing (ICPP ’05) - Workshopon Wireless and Sensor Networks, Oslo, Norway, June, 2005, pp. 452–457, IEEE Computer Society, 2005.

[42] D. Doval and D. O’Mahony, “Nom: Resource location and discovery forad hoc mobile networks,” in the Proceedings of the 1st Annual Mediter-ranean Ad hoc Networking Workshop, Med-hoc-Net 2002, Sardegna,Italy, September, 2002.

[43] S. Athanaileas, C. N. Ververidis, and G. C. Polyzos, “Optimized serviceselection for manets using an aodv-based service discovery protocol,”in the Proceedings of the 6th Annual Mediterranean Ad Hoc NetworkingWorkshop (MEDHOCNET 2006), Corfu, Greece, June, 2007.

[44] Z. Fan and E. G. Ho, “Service discovery in ad hoc networks: Performanceevaluation and qos enhancement,” Wireless Personal Communications,vol. 40, no. 2, pp. 215–231, 2007.

[45] G. P. Halkes, A. Baggio, and K. Langendoen, “A simulation study ofintegrated service discovery,” in the Proceedings of the 1st EuropeanConference on Smart Sensing and Context (EuroSSC 2006), Enschede,Netherlands, October, 2006 (P. J. M. Havinga, M. E. Lijding, N. Mer-atnia, and M. Wegdam, eds.), vol. 4272 of Lecture Notes in ComputerScience, pp. 39–53, Springer, 2006.

[46] L. Gavrilovska, “Cross-layering approaches in wireless ad hoc networks,”IEEE Wireless Personal Communications, vol. 37, no. 3, pp. 271–290,2006.

[47] P. Engelstad, Y. Zheng, R. Koodli, and C. Perkins, “Service discoveryarchitectures for on-demand ad hoc networks,” International Journal ofAd Hoc and Sensor Wireless Networks, vol. 2, no. 1, pp. 27–58, 2006.

[48] J. L. Jodra, M. Vara, J. M. Cabero, and J. Bagazgoitia, “Service discov-ery mechanism over OLSR for mobile ad-hoc networks,” in the Proceed-ings of the 20th International Conference on Advanced Information Net-working and Applications (AINA 2006), Vienna, Austria, April, 2006,pp. 534–542, IEEE Computer Society, 2006.

119

[49] L. Li and L. Lamont, “A lightweight service discovery mechanism for mo-bile ad hoc pervasive environment using cross-layer design,” in the Pro-ceedings of the 3rd IEEE International Conference on Pervasive Com-puting and Communications, 2005, PerCom Workshops 2005, Kauai Is-land, Hawaii, March, 2005, pp. 55–59, IEEE Computer Society, 2005.

[50] L. Cheng, “Service advertisement and discovery in mobile ad hoc net-works,” in the Workshop on Ad hoc Communications and Collaborationin Ubiquitous Computing Environments, in conjunction with the ACM2002 Conference on Computer Supported Cooperative Work, New Or-leans, LA, USA, November, 2002.

[51] W. Ma, B. Wu, W. Zhang, and L. Cheng, “Implementation of alightweight service advertisement and discovery protocol for mobile adhoc networks,” in the Proceedings of the IEEE Global Telecommunica-tions Conference 2003 (GLOBECOM ’03), San Fransisco, USA, Decem-ber, 2003, vol. 2, pp. 1023–1027, IEEE Communications Society, 2003.

[52] K. Obraczka, G. Tsudik, and K. Viswanath, “Pushing the limitsof multicast in ad hoc networks,” in the Proceedings of the 21thIEEE International Conference on Distributed Computing Systems (21thICDCS’2001), Phoenix, AZ, USA, April, 2001.

[53] S.-J. Lee, W. Su, and M. Gerla, “On-demand multicast routing protocolin multihop wireless mobile networks,” ACM/Springer Mobile Networksand Applications (MONET) Journal, vol. 7, no. 6, pp. 441–453, 2002.

[54] H. S. Hassanein, Y. Yang, and A. Mawji, “A new approach to servicediscovery in wireless mobile ad hoc networks,” International Journal ofSensor Networks, vol. 2, no. 1/2, pp. 135–145, 2007.

[55] D. Chakraborty, A. Joshi, and Y. Yesha, “Integrating service discoverywith routing and session management for ad-hoc networks,” Ad HocNetworks Journal, vol. 4, no. 2, pp. 204–224, 2006.

[56] A. Helmy, S. Garg, N. Nahata, and P. Pamu, “CARD: A contact-based architecture for resource discovery in wireless ad hoc networks,”ACM/Springer Mobile Networks and Applications (MONET) Journal,vol. 10, no. 1-2, pp. 99–113, 2005.

[57] C.-S. Oh, Y.-B. Ko, and Y.-S. Roh, “An integrated approach for effi-cient routing and service discovery in mobile ad hoc networks,” in theProceedings of the 2nd IEEE Consumer Communications and Network-ing Conference (CCNC 2005), Las Vegas, Nevada, USA, January, 2005,pp. 184–189.

[58] D. Noh and H. Shin, “Spiz: an effective service discovery protocol formobile ad hoc networks,” EURASIP Journal on Wireless Communica-tion and Networking, vol. 2007, no. 1, pp. 27–27, 2007.

[59] C. N. Ververidis and G. C. Polyzos, “Extended ZRP: a routing layerbased service discovery protocol for mobile ad hoc networks,” in the Pro-ceedings of the 2nd Annual International Conference on Mobile and Ubiq-uitous systems: Networks and Services (Mobiquitous 2005), San Diego,CA, USA, July, 2005, pp. 65–72, IEEE Computer Society, 2005.

120

[60] A. Helmy, “Contact-extended zone-based transactions routing forenergy-constrained wireless ad hoc networks,” IEEE Transactions onVehicular Technology, vol. 54, no. 1, pp. 307–319, January, 2005.

[61] N. Sadagopan, F. Bai, B. Krishnamachari, and A. Helmy, “PATHS: anal-ysis of PATH duration statistics and their impact on reactive MANETrouting protocols,” in the Proceedings of the 4th ACM International Sym-posium on Mobile Ad Hoc Networking and Computing (MobiHoc 2003),Annapolis, MD, USA, June, 2003, pp. 245–256, ACM, 2003.

[62] F. Bai, N. Sadagopan, B. Krishnamachari, and A. Helmy, “Modelingpath duration distributions in MANETs and their impact on reactiverouting protocols,” IEEE Journal on Selected Areas in Communications,vol. 22, no. 7, pp. 1357–1373, 2004.

[63] Z. J. Haas, M. R. Pearlman, and P. Samar, “The intrazone routingprotocol (iarp) for ad hoc networks,” IETF Internet Draft (expired),Available at: http://people.ece.cornell.edu/haas/wnl/Publications/draft-ietf-manet-zone-iarp-02.txt.

[64] Z. J. Haas, M. R. Pearlman, and P. Samar, “The interzone routingprotocol (ierp) for ad hoc networks,” IETF Internet Draft (expired),Available at: http://people.ece.cornell.edu/haas/wnl/Publications/draft-ietf-manet-zone-ierp-02.txt.

[65] D. Chakraborty, A. Joshi, Y. Yesha, and T. W. Finin, “GSD: a novelgroup-based service discovery protocol for MANETS,” in the Proceedingsof the 4th IEEE Conference on Mobile and Wireless CommunicationsNetworks (MWCN ’02), Stockholm, Sweden, September, 2002, pp. 140–144, IEEE, 2002.

[66] J.-M. Choi, J. So, and Y.-B. Ko, “Numerical analysis of IEEE 802.11broadcast scheme in multihop wireless ad hoc networks,” in the Proceed-ings of the International Conference on Information Networking (ICOIN2005), Jeju Island, Korea, January, 2005 (C. Kim, ed.), vol. 3391 of Lec-ture Notes in Computer Science - Convergence in Broadband and MobileNetworking, pp. 1–10, Springer, 2005.

[67] L. M. Feeney, “An energy consumption model for performance analysisof routing protocols for mobile ad hoc networks,” ACM/Springer MobileNetworks and Applications (MONET) Journal, vol. 6, no. 3, pp. 239–249,2001.

[68] Scalable Network Technologies Inc., Qualnet Simulator,http://www.scalable-networks.com, 2005.

[69] P. Samar, M. R. Pearlman, and Z. J. Haas, “Independent zone routing:an adaptive hybrid routing framework for ad hoc wireless networks,”IEEE/ACM Transactions on Networking, vol. 12, no. 4, pp. 595–608,2004.

[70] M. Pearlman and Z. Haas, “Determining the optimal configuration forthe zone routing protocol,” IEEE Journal on Selected Areas in Commu-nications (JSAC), vol. 17, no. 8, pp. 1395–1414, Aug 1999.

[71] M. Almeida, R. Sarro, J. P. Barraca, S. Sargento, and R. L. Aguiar,“Experimental evaluation of the usage of ad hoc networks as stubs formultiservice networks,” EURASIP J. Wirel. Commun. Netw., vol. 2007,no. 1, pp. 39–39, 2007.

121

[72] J. P. Barraca, S. Sargento, and R. L. Aguiar, “The polynomial-assistedad hoc charging protocol,” iscc, vol. 00, pp. 945–952, 2005.

[73] S. Zhong, J. Chen, and Y. R. Yang, “Sprite: A simple, cheat-proof,credit-based system for mobile ad-hoc networks,” in the Proceedings ofthe 22nd Annual Joint Conference of the IEEE Computer and Commu-nications Societies (INFOCOM 2003), San Fransisco, CA, USA, April,2003.

[74] L. Buttyan and J.-P. Hubaux, “Enforcing service availability in mo-bile ad-hoc WANs,” in the Proceedings of the 1st Annual Workshop onMobile Ad Hoc Networking and Computing (MobiHoc 2000), Boston,Massachusetts, USA, August, 2000, pp. 87–96, ACM, 2000.

[75] Z. Ji, W. Yu, and K. J. R. Liu, “An optimal dynamic pricing frame-work for autonomous mobile ad hoc networks,” in the Proceedings ofthe 25th IEEE International Conference on Computer Communications(INFOCOM 2006), Barcelona, Catalunya, Spain, April, 2006.

[76] M. J. Neely, “Optimal pricing in a free market wireless network,” in theProceedings of the 26th IEEE International Conference on ComputerCommunications (INFOCOM 2007), Anchorage , Alaska , USA, May,2007, pp. 213–221.

[77] A. McDonald and T. Znati, “A path availability model for wireless ad-hoc networks,” in the Proceedings of the IEEE Wireless Communica-tions and Networking Conference 1999 (WCNC 1999), New Orleans,LA, USA, September, 1999, vol. 1, pp. 35–40.

[78] D. S. M. Hauspie and J. Carle, “Partition detection in mobile ad hocnetworks,” in the Proceedings of the 2nd Mediterranean Workshop onAd-Hoc Networks (MEDHOC-NET-03), Mahdia, Tunisia, June, 2003.

[79] M. Sheng, J. Li, and Y. Shi, “Critical nodes detection in mobile ad hocnetwork,” in The IEEE 20th International Conference on Advanced In-formation Networking and Applications (AINA), Vienna, Austria, April,2006, pp. 336–340, IEEE Computer Society, 2006.

[80] B. Liang and Z. Haas, “Virtual backbone generation and maintenancein ad hoc network mobility management,” in the Proceedings of the19th Annual Joint Conference of the IEEE Computer and Communica-tions Societies (INFOCOM 2000), Tel-Aviv, Israel, March, 2000, vol. 3,pp. 1293–1302.

[81] K. Wang and B. Li, “Efficient and guaranteed service coverage in parti-tionable mobile ad-hoc networks,” in the Proceedings of the 21st AnnualJoint Conference of the IEEE Computer and Communications Societies(INFOCOM 2002), New York, USA, June, 2002.

[82] A. Derhab, N. Badache, and A. Bouabdallah, “A partition predictionalgorithm for service replication in mobile ad hoc networks,” in the Pro-ceedings of the 2nd Annual Conference on Wireless On-demand Net-work Systems and Services (WONS ’05), St. Moritz, Switzerland, 2005,pp. 236–245, IEEE Computer Society, 2005.

[83] L. Yin and G. Cao, “Balancing the tradeoffs between data accessibil-ity and query delay in ad hoc networks,” in the Proceedings of the 23rd

122

International Symposium on Reliable Distributed Systems (SRDS’04),Florianpolis, Brazil, October, 2004, pp. 289–298, IEEE Computer Soci-ety, 2004.

[84] T. Hara, “Effective replica allocation in ad hoc networks for improvingdata accessibility,” in the Proceedings of the 20th Annual Joint Confer-ence of the IEEE Computer and Communications Societies (INFOCOM2001), Anchorage, Alaska, April, 2001, vol. 3, pp. 1568–1576.

[85] J. Zhai, Q. Li, and X. Li, “Data caching in selfish MANETs,” in the Pro-ceedings of the 3rd International Conference on Networking and MobileComputing (ICCNMC 2005), Zhangjiajie, China, August, 2005 (X. Luand W. Zhao, eds.), vol. 3619 of Lecture Notes in Computer Science,pp. 208–217, Springer, 2005.

[86] C. Chekuri and S. Khanna, “A PTAS for the multiple knapsack prob-lem,” in the Proceedings of the 11th Annual ACM-SIAM Symposium onDiscrete Algorithms, San Fransisco, CA, USA, January, 2000, pp. 213–222.

[87] L. Fleischer, M. Goemans, V. Mirrokni, and M. Sviridenko, “Tight ap-proximation algorithms for maximum general assignment problems,” inthe Proceedings of the 17th Annual ACM-SIAM Symposium on DiscreteAlgorithms, Miami, Florida, USA, January, 2006, pp. 611–620.

[88] R. Cohen, L. Katzir, and D. Raz, “An efficient approximation for thegeneralized assignment problem,” IPL: Information Processing Letters,vol. 100, 2006.

[89] Y. Han, R. J. La, A. M. Makowski, and S. Lee, “Distribution of pathdurations in mobile ad-hoc networks - palm’s theorem to the rescue,”Computer Networks, Elsevier, vol. 50, no. 12, pp. 1887–1900, 2006.

[90] S. Jiang, D. He, and J. Rao, “A prediction-based link availability estima-tion for mobile ad hoc networks,” in the Proceedings of the 20th AnnualJoint Conference of the IEEE Computer and Communications Societies(INFOCOM 2001), Anchorage, Alaska, April, 2001, vol. 3, pp. 1745–1752.

[91] S. Cho and J. Hayes, “Impact of mobility on connection stability in adhoc networks,” vol. 3, pp. 1650–1656.

[92] Y.-C. C. Yu-Chee Tseng, Yueh-Feng Li, “On route lifetime in multi-hop mobile ad hoc networks,” IEEE Transactions on Mobile Computing,vol. 2, no. 4, pp. 366–376, 2003.

[93] Y.-C. Tseng, Y.-F. Li, and Y.-C. Chang, “On route lifetime in multi-hop mobile ad hoc networks,” IEEE Transactions on Mobile Computing,vol. 2, no. 4, pp. 366–376, 2003.

[94] A. Trivino-Cabrera, J. G. de la Nava, E. Casilari, and F. J. Gonzalez-Canete, “An analytical model to estimate path duration in MANETs,” inthe Proceedings of the 9th International Symposium on Modeling Anal-ysis and Simulation of Wireless and Mobile Systems (MSWiM 2006),Torremolinos, Malaga, Spain, October, 2006 (E. Alba, C.-F. Chiasserini,N. B. Abu-Ghazaleh, and R. L. Cigno, eds.), pp. 183–186, ACM, 2006.

123

[95] J. Allard, V. Chinta, S. Gundala, and G. G. R. III, “Jini meets UPnP:An architecture for jini/UPnP interoperability,” in the Proceedings of the2003 International Symposium on Applications and the Internet (SAINT2003), Orlando, Florida, USA, January, 2003, pp. 268–275, IEEE Com-puter Society, 2003.

[96] A. Friday, N. Davies, N. Wallbank, E. Catterall, and S. Pink, “Support-ing service discovery, querying and interaction in ubiquitous computingenvironments,” Wireless Networks, vol. 10, no. 6, pp. 631–641, 2004.

[97] P. Grace, G. S. Blair, and S. Samuel, “ReMMoC: A reflective middlewareto support mobile client interoperability,” in CoopIS/DOA/ODBASE(R. Meersman, Z. Tari, and D. C. Schmidt, eds.), vol. 2888 of LectureNotes in Computer Science, pp. 1170–1187, Springer, 2003.

[98] P. Raverdy, V. Issarny, A. de La Chapelle, and R. Chibout, “A multi-protocol approach to service discovery and access in pervasive environ-ments,” in the Proceedings of the 3rd Annual International Conferenceon Mobile and Ubiquitous Systems: Networks and Services, (Mobiquitous2006), San Jose, California, USA, July 2006, pp. 1–9.

[99] Y.-D. Bromberg and V. Issarny, “INDISS: Interoperable discovery sys-tem for networked services,” in Middleware (G. Alonso, ed.), vol. 3790of Lecture Notes in Computer Science, pp. 164–183, Springer, 2005.

[100] T. Koponen and T. Virtanen, “A service discovery: A service brokerapproach,” in the Proceedings of the 37th Hawaii International Confer-ence on System Sciences (HICSS-37 2004), Big Island, Hawaii, USA,January, 2004, IEEE Computer Society, 2004.

[101] M. A. E. Saoud, T. Kunz, and S. Mahmoud, “Benchmanet: An evalu-ation framework for service discovery protocols in manet,” in the Pro-ceedings of the 3rd Annual IEEE Communications Society Conferenceon Sensor and Ad Hoc Communications and Networks (SECON 2006),Reston, Virginia, USA, June, 2006, vol. 3, pp. 860–865.

[102] D. Chakraborty, A. Shenoi, Y. Yesha, Y. Yesha, and A. Joshi, “A Queue-ing Theoretic Model for Service Discovery in Ad-hoc Networks,” in theProceedings of the Communication Networks and Distributed SystemsModeling and Simulation Conference 2004 (CNDS 2004), San Diego,California, USA, January, 2004.

[103] T. Wu and G.-S. Kuo, “An analytical model for centralized service dis-covery architecture in wireless networks,” in the Proceedigns of the 64thIEEE Vehicular Technology Conference 2006 (VTC-2006 Fall), Mon-treal, Canada, September, 2006, pp. 1–5.

124

ABBREVIATIONS

3G Third Generation

AE-GAP Approximation Enhanced Generalized Assignment Prob-

lem

AE-SK Approximation Enhanced Single Knapsack Problem

AODV Ad hoc On Demand distance Vector

APS APplication layer-based Service discovery protocol

ATD Average Transaction Time

ATE Adaptive Traffic Estimation

AVERT Adaptive serVicE and Route discovery proTocol

OWL Web Ontology Language

BFO Broadcasting Frequency Optimizer

Bluetooth SDP Bluetooth Service Discovery Protocol

BSU Broadcast Simulated Unicast

CARD Contact-based Architecture for Resource Discovery

C-GAP Classic Generalized Assignment Problem

C-SK Classic Single Knapsack

DA Directory Agent

DHCP Dynamic Host Configuration Protocol

DIFS Distributed Inter Frame Space

DNS Domain Name System

DSDV Destination-Sequenced Distance Vector

DSR Dynamic Source Routing

E-ZRP Extended Zone Routing Protocol

GAP Generalized Assignment Problem

GPRS General Packet Radio Service

GSD Group-based Service Discover protocol

125

HAID Hybrid Adaptive protocol for Integrated Discovery

HESED High Efficiency SErvice Discovery protocol

HTTP Hypertext Transfer Protocol

IARP Intra-zone Routing Protocol

IERP Inter-zone Routing Protocol

IP Internet Protocol

IR Infrared

IZR Independent Zone Routing protocol

MAC Medium Access Control

MANET Mobile Ad-hoc NETwork

NDP NEighbor Discovery Protocol

ODMRP On-Demand Multicast Routing Protocol

OE-GAP Oracle Enhanced Generalized Assignment Problem

OE-SK Oracle Enhanced Single Knapsack

OLSR Optimized Link State Routing protocol

P2P Peer to Peer

PDA Personal Digital Assistant

PKI Private Key Infrastructure

QoS Quality of Service

RMI Remote Method Invocation

RREP Route REPly

RREQ Route REQuest

RWP Random WayPoint

SA Service Agent

SAD Service Availability Duration

SAP Service Access Point

SLM SaLutation Manager

SLP Service Location Protocol

SREP Service REPly

TCP Transmission Control Protocol

TM Salutation Transport Manager

126

TORA Temporally Ordered Routing Algorithm

TTL Time To Live

UA User Agent

UDDI Universal Description Discovery and Integration

UDP User Datagram Protocol

UMTS Universal Mobile Telecommunications System

UPnP Universal Plug n’ Play

URL Uniform Resource Locator

UUID Universal Unique IDentifier

VANET Vehicular Ad-hoc NETwork

WLAN Wireless Local Area Network

XML eXtensible Markup Language

ZRP Zone Routing Protocol

Documents

ENERGY EFFICIENT AND ADAPTIVE SERVICE ADVERTISEMENT, …pages.cs.aueb.gr/~chris/C.N.Ververidis-PhD Thesis-2008.pdf · A Dissertation Submitted to the Faculty of the Department of