Performance Analysis of Opportunistic Content Distribution via …1256074/... · 2018. 10. 29. · Introduction Where should I start? Start from the statement of the problem. György

Performance Analysis of Opportunistic ContentDistribution via Data-Driven Mobility Modeling

LJUBICA PAJEVI∆

Doctoral ThesisStockholm, Sweden, 2018

TRITA-EECS-AVL-2018:76ISBN 978-91-7729-980-6

School of Electrical Engineeringand Computer Science

KTH, Stockholm, Sweden

Akademisk avhandling som med tillstånd av Kungl Tekniska högskolan framläggestill o�entlig granskning för avläggande av doktorexamen onsdagen den 21 november2018 i Hörsal F3, Kungl Tekniska högskolan, Lindstedtsvägen 26, Stockholm.

© Ljubica PajeviÊ, November 2018

Tryck: Universitetsservice US AB

AbstractFor the most part of the developed world, ubiquitous connectivity has

become a norm in human daily lives. Globally, the access to informationon the Internet and telecommunication services has been recognized as oneof the accelerators of human progress, promoting the right to education andfacilitating access to globally shared knowledge. The increasing capacity anduser experience demands on the one side, and the need for increased net-work coverage and the availability of a�ordable telecommunication serviceson the other side, pose significant challenges for the current network archi-tectures. While the next generation wireless technologies and the pervasivedeployment of the currently available communication systems are poised tosolve the above challenges, the main inhibitors of immediate action are stillsocio-economical factors. Due to the pervasiveness of end-user mobile devices,device-to-device communication (D2D) can be leveraged as an intermediate—and a complementary—solution.

D2D communication enables networking in an opportunistic manner. Anopportunistic network is formed by co-located mobile users in order to ex-change data via direct wireless links when their devices are within transmis-sion range, without relying on the use of fixed network infrastructure. Inthis thesis we investigate the capabilities of these opportunistic networks andcover two main areas: data-driven modeling of user mobility and analyticperformance evaluation of location-aware opportunistic content distribution.

The first part of the thesis focuses on mobility modeling. We collectand analyze a dataset of user associations in a university wireless network.The dataset was recorded at the KTH Royal Institute of Technology, anddocuments the prevalence of smartphone network users, as compared to laptopusers. We characterize the mobility of users in this dataset both from thenetwork and from the user perspective. From the network perspective, wemodel the aggregate mobility and access patterns to di�erent parts of thenetwork. To characterize individual mobility, we explore how mobile the usersare, and how accurately movements can be predicted in the near future. Wecombine the findings on the access patterns with the analysis of several othermobility traces to propose a mobility model for populations with churn, thatis specifically tailored for the evaluation of opportunistic content distribution.

In the second part of the thesis, we evaluate the performance of op-portunistic content distribution in urban environments. Our focus is onephemeral, location-aware networks where content is stored only on the userdevices within the locale of interest. We develop a framework for modelingthe performance of content distribution; the framework allows modeling of thespread of information as a stochastic process and accurate capturing of thestochastic fluctuations in the number of distributed content items. We utilizethe system model to study the feasibility of opportunistic content distributionand by means of stochastic stability analysis we assess how the mutable sys-tem parameters can be engineered to ensure content persistence. Our resultsdemonstrate that the content persistence strongly depends on the density ofusers, and that the requirements for user resources are relatively low alreadyfor moderate densities. This finding can help incentivize users to participatein opportunistic networking applications.

SammanfattningFör de större delarna av den industrialiserade världen har oupphörlig

nätverkstillgänglighet blivit regel i människors liv medan, globalt sett, såhar internet och telekommunikationstjänster erkänts som en accelerator avsamhällsutveckling, främjande rätten till utbildning och möjliggörande till-gången av världsomspännande information. De växande kraven på kapacitetoch användarupplevelse å den ena sidan och behoven av nätverkstäckning ochtillgängligheten av prisvärda telekommunikationstjänster å andra sidan gerupphov till signifikanta utmaningar för dagens nätverksarkitekturer. Dessautmaningar angrips i dagsläget av nästa generationens trådlösa teknik ochomfattande utbyggnad av fasta kommunikationssystem. Trots detta häm-mas de omedelbara genomslagen av i huvudsak socioekonomiska faktorer. Dåtillgången till mobiltelefoni är omfattande kan enhet-till-enhets kommunika-tion Device-to-Device (D2D) användas som en alternativ och komplementärlösning.

D2D-kommunikation möjliggör opportunistiska nätverk som bildas mel-lan samlokaliserade mobilanvändare. Via direkta trådlösa anslutningar kandata delas utan att förlita sig på fast nätverksinfrastruktur. Denna avhan-dling undersöker potentialen för opportunistiska nätverk inom två områden:datadriven modellering av användares mobilitet respektive analysmetoder föratt utvärdera platsmedveten opportunistisk informationsfördelning.

Den första delen av avhandlingen fokuserar på mobilitetsmodellering. Visamlar in och analyserar en datamängd av associationer mellan användare iett universitetets trådlösa nätverk. Datamängden samlades in på KungligaTekniska Högskolan (KTH) och dokumenterar prevalensen av smartphonean-vändare i nätverket jämfört med bärbara datorer. Vi karaktäriserar mobilitetav användara i denna datamängd både från ett nätverks- och användarper-spektiv. Från perspektivet av nätverket så modellerar vi sammaslagen mo-bilitet och åtkomstmönster till olika delar av nätverket. För att karaktäriseraindividuell mobilitet undersöker vi hur mobila mobilanvändarna är och hurförutsägbara rörelser är på kort sikt. Vi kombinerar insikterna om åtkomst-mönster med analysen av flera rörelsespår för att föreslå en mobilitetsmodellför populationer med bortfall, som är skräddarsydd för att utvärdera oppor-tunistisk informationsfördelning.

Del två av avhandlingen utvärderar hur opportunistisk informationsfördel-ning presterar i stadsmiljöer. Vårt fokus är kortlivade platsmedvetna nätdär information endast lagras på användarenheter inom ett studerat om-råde. Vi utvecklar ett ramverk för att modellera prestationen av informations-fördelning. Ramverket tillåter att modellera informationsspridningen som enstokastisk process och fångar noggrannt de stokastiska fluktuationerna i an-talet fördelade informationssposter. Vi använder denna systemmodell för attutforska potentialen för opportunistisk informationsfördelning och med hjälpav stokastisk stabilitetsanalys härleder vi hur reglerbar systemparametrar kanväljas för att garantera kvarlevnad av information. Våra resultat visar att in-formationens kvarlevned kraftigt beror på koncentrationen av användare samtatt kraven på användarresurser är relativt låga redan för måttliga koncentra-tioner. Detta resultat kan vara till hjälp för att uppmuntra användare attdelta i opportunistiska nätverkstillämpningar.

Contents

Contents iii

Acknowledgments v

1 Introduction 11.1 Thesis Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . 41.2 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Understanding and Modeling Human Mobility 72.1 Foundations of Mobility Modeling . . . . . . . . . . . . . . . . . . . 82.2 Human Mobility Through the Space-Time Prism . . . . . . . . . . . 92.3 Historical Overview of Approaches to Mobility Modeling . . . . . . . 142.4 Scenario-Based Modeling . . . . . . . . . . . . . . . . . . . . . . . . 15

3 Collecting and Analyzing Mobility Data 193.1 Trace Acquisition and Source Origins . . . . . . . . . . . . . . . . . . 193.2 WLAN Mobility Traces . . . . . . . . . . . . . . . . . . . . . . . . . 233.3 Mobility and Network Simulators . . . . . . . . . . . . . . . . . . . . 25

4 Opportunistic Networking: Applications, Platforms and Perfor-mance Evaluation 274.1 Mobile Ad Hoc Networks and Delay Tolerant Networking . . . . . . 274.2 Data Delivery in Opportunistic Networks . . . . . . . . . . . . . . . 294.3 Location-Based Opportunistic Networks . . . . . . . . . . . . . . . . 314.4 Opportunistic Communication Systems: Prototypes and Applications 334.5 Mathematical Tools for Modeling Content Dissemination and of Per-

formance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

5 Summary of Original Work 39

iii

iv

6 Conclusion and Future Work 45

Bibliography 49

Paper A: Revisiting the Modeling of User Association Patterns ina University Wireless Network 63

Paper B: Predicting the Users’ Next Location from WLAN Mobil-ity Data 81

Paper C: Epidemic Content Distribution: Empirical and AnalyticPerformance 99

Paper D: Modeling Opportunistic Communication with Churn 115

Paper E: Ensuring Persistent Content in Opportunistic Networksvia Stochastic Stability Analysis 143

AcknowledgmentsMy gratitude extends beyond the limits ofmy capacity to express it.

Iain M. Banks, The Player of Games

Although shorter than the next introductory, technical chapters of the thesis,this chapter has taken proportionally more time to write and has involved deeperthought with a considerable amount of reflection. Incidentally, a productive mo-ment of reflection occurred while I was listening to Evgeny Kissin’s live performanceseveral months ago. This is when I started phrasing the following sentences: theyflow straight from the heart like Debussy’s notes thorough the pianist’s fingers.

First and foremost, I would like to thank my main supervisor, Prof. ViktoriaFodor for her unwavering support and exceptional availability during the last twoyears of my studies. Collaboration with a person of such perspicacity, determina-tion, and a cheerful personality has been an invaluable, prolific and fun experience.I wish every student researcher had such a devoted, supportive and helpful super-visor, and I firmly believe that professors like her make the academic communitymore productive and a better place.

I am grateful to my second supervisor, Prof. Gunnar Karlsson, for giving methe opportunity to join his research group at the former Laboratory of Communi-cation Networks (LCN), for introducing me to the world of academic research andfor numerous fruitful discussions which resulted in formulating research problemsaddressed in this thesis. My sincere gratitude goes to Prof. Håkan Hjalmarssonwho provided the necessary guidance and support when my research path starteddeviating from its destination. I am greatly indebted to Prof. Jörg Ott for welcom-ing me at the Chair of Connected Mobility of the Technical University of Munich,where a substantial part of this thesis was written.

I would like to thank Prof. Thrasyvoulos Spyropoulos for accepting to actas an opponent to this dissertation, and to Prof. Björn Landfeldt, Prof. KarinAnna Hummel, and Dr Aline Carneiro Viana for taking time to act as committeemembers.

v

vi

I wish to express my appreciation to the KTH School of Engineering for theirfinancial support of my studies through the Program of Excellence.

At LCN, as well as at the Chair of Connected Mobility, I had a pleasure tolearn from and work with a number of excellent researchers. To all the membersof the two research groups—professors and students, current and past: thank youfor creating a stimulating work environment and also for being a part of my PhDjourney! While living in Stockholm, I had a fortune to meet some wonderful peopleand make new friends; I could not possibly omit mentioning a few of them who havebeen exceptionally supportive in one way or the other: Vladimir, Ognjen, Sla�ana,Ivana and Cenny, thank you!

I am endlessly and forever grateful to my mother Zorica and my father Radiöa fortheir unconditional love and support, and for always trusting in me and encouragingme to pursue my goals. I dedicate this thesis to them. I wish to thank the rest of myfamily, especially my brother Aleksandar for providing motivation with a healthydose of competitiveness, and to my grandfather Radenko, for his enthusiasm andcuriosity about academic life in a far-away country. A demanding research worknecessitates occasional breaks, and I am thankful to my brother’s family: Jelena,Jovana, and Simona, for careless and joyful family moments, to my parents-in-law,Hannele and Seppo, for providing the most relaxing atmosphere during my holidaytrips and weekend breaks, and to my close Serbian friends, from Belgrade and fromabroad, for staying in touch and for their remote support.

Finally, it is not an exaggeration to say that completing this thesis would nothave been possible without the inexhaustible encouragement, understanding, andsupport of my better-half, Teemu. He has been my (system-oriented) academiccompanion on this journey, an advice-giver, and the constant inspiration and reasonfor continuous personal and professional development. I consider myself incrediblyfortunate to have him by my side.

Ljubica PajeviÊMunich, September 2018

Chapter 1Introduction

Where should I start?

Start from the statement of the problem.

György Pólya, How to Solve it

Ubiquitous connectivity has now become a norm in our daily lives while theaccess to information on the Internet has been recognized as one of the basic hu-man rights [1], promoting the right to education and globally accelerating humanprogress. The digital divide, the gap between those with access to the Internet andthose still without or with limited access, started to shrink thanks to the enormouse�orts to bridge this gap by deploying the necessary telecommunication infrastruc-ture, especially in rural regions, and by extending communication services. Yet, theprogress of these e�orts is slow and greatly limited by socio-economic factors. Atone end of the digital divide are people without a�ordable network access and with-out this access to globally shared knowledge, communication services and healthcare; at the other end users with growing demands for faster networks and largerdata plans.

The current network architectures will struggle to keep pace with the increasingdemand for capacity and user experience demands on one end, and the need forincreased network coverage and the availability of a�ordable telecommunicationservices on the other end. Whereas the fifth generation (5G) networks hold thepromise of meeting the capacity requirements through network densification, i.e.,through the deployment of more base stations, deploying such solutions poses greatfinancial and operational burdens on mobile network operators, while reaching onlya small portion of users in the initial phase. However, the lack of availability of thenetwork infrastructure is still one of the key barriers to high-speed Internet accessin developing regions.

1

2 Chapter 1. Introduction

Due to the pervasiveness of end-user mobile devices, device-to-device commu-nication (D2D) can be leveraged for a more immediate solution to the above chal-lenges. Mobile devices such as smartphones, digital tablets and readers are todayequipped with ample computing resources and often support several communicationtechnologies (Wi-Fi/Wi-Fi Direct, Bluetooth), which makes these devices suitablefor o�oading processing and networking tasks from the edge network to end de-vices. D2D communication can be exploited in cellular networks to increase usageof radio resources by utilizing both cellular and D2D links [2], to extend networkcoverage [3], and to achieve communication with extremely high bit rates and lowlatencies, thanks to the proximity of devices. Aside from the infrastructural mode ofcommunication, e.g. to alleviate the tra�c load by o�oading a portion of it to theend users, a pure D2D ad hoc communication mode has seen even more promisesfor potential applications. This mode is particularly well-suited when the infras-tructure is absent e.g., in remote rural areas, disrupted due to natural disasters orotherwise unavailable being subjected to censorship by oppressive authorities.

D2D communication enables a particularly interesting type of mobile networkstermed opportunistic networks, formed by co-located mobile users in order to ex-change data via direct wireless links when their devices are within transmissionrange. The concept of opportunistic networking is not entirely new—some of thefundamental principles are borrowed from the research on mobile ad hoc networks(MANETs) and delay tolerant networks (DTNs) as conceptual predecessors. In aMANET, network topology is created from mobile devices with a primary goal tomimic infrastructure networks and ensure end-to-end connectivity between a sourceand a destination device, in a purely wireless domain. This principle of autonomyand independence on the infrastructure has been adapted to opportunistic commu-nication. Acknowledging that MANETs are prone to frequent link breakages due tothe node mobility, DTNs are envisioned to cope with frequent network reconfigura-tions and partitioning. Mobile DTNs—opportunistic networks being one instanceof such—consider an alternative approach to exploit intermittent connectivity ofdevices; the nodes’ mobility is seen as a feature rather than a limitation, and itis utilized to transfer messages in a store-carry-forward manner. Opportunisticnetworks and DTNs may not rely on the existence of end-to-end paths and there-fore do not provide guarantees of the message delivery. Furthermore, there maybe one or more potential recipients rather than a dedicated single recipient. Forthese reasons, opportunistic networking is well suited for content dissemination ona best-e�ort basis and for applications that do not have strict latency requirements.

Use-casesLet us review a few use-cases that may significantly benefit from using opportunisticcommunications.

In opportunistic social networks users interested in exchanging delay tolerantdata, such as media files and personal profiles, or in messaging and decentralizedmobile networking, can form user groups based on their mutual interests and social

3

ties. Examples are university students in the same class, communities of friends,conference attendees, and people attending massive public events or business meet-ings. This type of social networking is also seen as advantageous with respect toprivacy concerns when compared to social networks that require infrastructure andsupport of some kind of centralized entity [4].

Distributed vehicular data sensing is another representative example of oppor-tunistic use-cases. Crowdsourcing is becoming a valuable way of gathering infor-mation, both for users and service providers. One of the currently most exploitedsources of data is floating car data (FCD), which includes positioning data such asGPS localization, information about cellular hando�s of mobile devices, and cov-erage of Wi-Fi networks. FCD together with user-provided updates can be usedto improve real-time tra�c estimates. Currently, there exists many infrastructure-based solutions, however, they require centralized data processing, accompaniedwith significant computational complexity, and non-negligible latencies. In thiscontext, opportunistic vehicle-to-vehicle communication can be beneficial for fastergathering, processing and disseminating tra�c updates in a distributed fashion.

Similarly to the previous case, data sensing in smart cities is a scenario in whichmobile devices can be used to collect sensing data beyond positioning and networkconnectivity information. Smartphones are already equipped with a rich set ofpowerful (and inexpensive) embedded sensors, such as camera, microphone, GPS,accelerometer, gyroscope, while new features including pressure, temperature andhumidity sensors are being added. These features can be exploited for generatingfine-grained maps of city areas with measured levels of noise, reporting infrastruc-ture problems and public safety issues, measuring flows of pedestrian and crowdsfor the purpose of urban planning or positioning inside buildings, shopping-mallsand airport terminals.

Location-based opportunistic services lie at the intersection of the previous cases.Irrespective of the type of content being distributed (sensing data, tra�c updatesor social media updates), the mobility modality (pedestrian or vehicular mobility),and the underlying social structure (connected communities or random travelers),the common feature of these services is that the content is only of local and perhapsof time relevance, while considered irrelevant outside of the given time and spatialwindows.

In contrast to urban scenarios where opportunistic networking can be used toaugment already existent infrastructure, opportunistic services in developing andrural regions can be employed to enable communication and provide asynchronousInternet access in regions that are characterized by very sparse or non-existent net-working infrastructure. A few initiatives have been launched to provide “connectiv-ity for everyone”, namely Google’s project Loon [5] or Facebook’s Free services [6].Their development is still ongoing and it is yet unclear how much interest the biginitiator companies will have to keep the projects running, or even how (socially)beneficial these projects are for their users [7].


1.1 Thesis Contributions

In this thesis we investigate the capabilities of opportunistic networks by meansof analytic modeling and performance evaluation. Towards this end, we cover twomain areas. The first area is mobility modeling, which is essential for performancestudies since mobility patterns determine how opportunistic networks perform. Thesecond area concerns modeling and analysis of opportunistic networks with a fo-cus on one specific case—location-aware opportunistic content distribution. Thecontributions in these areas are the following.

Modeling Mobility of Populations with Churn

Despite a plethora of research in the field of mobility modeling for opportunisticcommunication, a specific property of populations of mobile users has been over-looked: the population turnover—also known as churn—where users can join thesystem and depart from it, thus the number of nodes in the system is changing.In practice, such dynamics realistically depicts scenarios in small urban areas, suchas a grid of streets, city square, or subway station where users arrive to the area,move inside for a while and then exit (or leave the system). In this research area,we combine analysis and synthesis to devise a mobility model.

• First, in Paper A we take a top-down approach and analyze aggregate proper-ties of the user mobility in a campus environment. We do so by collecting andmining a trace of user associations in a university wireless network. The traceanalysis reveals the prevalence of mobile users, as well as—from the modelingperspective—an advantageous property that the arrivals to network areas canbe treated as Poisson arrivals, which holds valid for certain time windows.

• From this point, we move onto an analysis of individual mobility of usersfrom the same trace, presented in Paper B. We first characterize how mobilethe users are, and then we investigate how accurately their movements canbe predicted in the near future. This is inspired by the concern whether thepredictability of user mobility should be considered as a feature to be takeninto account when devising future mobility models.

• We extend our search for mobility properties with the aim to devise a mobil-ity model for a specific scenario—localized epidemic content dissemination.Thus, in Paper C, we consider several real-world traces and we evaluate theperformance of epidemic content spreading with the underlying mobility pat-terns. These mobility patterns are representative of conference and o�cebuilding settings. We are particularly interested in understanding what arethe implications of the assumption of the users’ homogeneous mobility foraccurately capturing the content spreading process.

• The homogeneous mobility of users entails homogeneous contact patternsamong them. We infer that these simplifying assumptions may be inade-quate for the examined mobility traces, due to the existence of social ties

1.2 Thesis Outline 5

among users. Therefore, in Paper D we investigate a new set of mobilityscenarios capturing a variety of contexts: movements on city streets and opensquares, in a subway station and in buildings. This second set of scenarios ischaracterized by more random contact patterns as compared to the traces inPaper C. We show that even though the nodes exhibit heterogeneous mobilitypatterns, a moderate heterogeneity can be approximated by a homogenousmodel. Combining these results, we propose a model that incorporates churnand that is able to capture the observed heterogeneity.

Performance Modeling and Feasibility Analysis of OpportunisticContent Distribution

In the second sub-area of this thesis, we investigate the feasibility and evaluatethe performance of content dissemination in urban environments. Our focus is onephemeral, location-aware networks where content is stored only on the user deviceswithin the locale of interest—a scenario that can be seen as an infrastructure-lessmobile cloud storage. We develop a framework for modeling the performance ofcontent dissemination; the framework is configurable to capture di�erent communi-cation and user participation scenarios. Specifically, we employ the model of userswith limited contribution to the content spreading process and assess the minimumrequirements for the viable system.

• In Paper D we introduce the model for opportunistic content spreading, byoverlaying the epidemic content dissemination scheme on the proposed open-population mobility model. We model the spread of information as a stochas-tic process, showing that the model is able to reflect stochastic fluctuationsin the number of disseminated content items. Furthermore, we confirm thatthe requirement of capturing heterogeneity can be relaxed, and the e�ect ofheterogeneous node interactions can be approximated by an averaged e�ect.

• In Paper E we use the stochastic model of the system to study the feasibil-ity of opportunistic content dissemination and by means of stochastic sta-bility analysis we assess how the mutable system parameters which can beengineered—in this setup the forwarding time—can be tuned to ensure thatthe content can be sustained in the area exclusively relying on the presenceof nodes carrying and forwarding the content item, and without relying onany infrastructural support.

1.2 Thesis Outline

The remainder of this thesis is structured as follows. Chapter 2 gives an overviewof the mobility modeling endeavors, from the earliest e�orts to the current state-of-the-art models. The motivation for our contribution in the area is also stated.In Chapter 3 we argue about the necessity to revisit the established results, inthe light of newly available, massive amounts of data capturing human mobility.


Chapter 4 describes our vision of an opportunistic communication scheme, based onthe publish/subscribe paradigm and principles of epidemic content spreading; wealso survey similar concepts and other proposed, mature opportunistic networkingsolutions. Chapter 5 summarizes the contributions of papers included in this thesis.With Chapter 6, we conclude this thesis by discussing its main results, and bypresenting challenges and possible directions for future studies.

Chapter 2Understanding and Modeling

Human Mobility

To understand is to perceive patterns.

Isaiah Berlin

Modeling the mobility of users in wireless networks enables the evaluation ofnetwork protocols, services and applications, and the calibration of the protocolparameters in the development phase. In opportunistic and mobile networks, mo-bility models are necessary to assess the system feasibility and to evaluate systemperformance. The choice of a mobility model, however, can have a significant e�ecton the performance investigation and even put the credibility of the study in ques-tion [8, 9, 10, 11]. Therefore, a careful choice of the model is essential for a soundanalysis and evaluation.

Historically, mobility models have been divided into two main categories: ana-lytic models, that can be explained by a set of mathematical formulas, and trace-based models comprising of the movement traces collected from experiments, whichare directly fed into simulations. The latter group represents real mobility, howeverit is restricted to specific scenarios where the traces were obtained; moreover, thescenarios often represent very specific contexts such as a university campus, an en-terprise building, a part of a city or a theme park, or taxis in a particular city. Asopposed to this, analytic models—also referred to as synthetic models—representonly a simplified abstraction of mobility but they are often necessary to formallyanalyze the impact of human movements on the design of network protocols. Whilethe first used synthetic models bore little resemblance to any realistic movement,many researchers had to resort to using them due to the lack of deeper understand-ing of mobility and the unavailability of real traces. The emergence of locationand communication technologies such as Global Positioning System (GPS), Blue-tooth and Wi-Fi-based tracking, streamlined the acquisition of large mobility traces,

7

8 Chapter 2. Understanding and Modeling Human Mobility

which further opened new opportunities for redesigning synthetic models and forincorporating statistical properties observed from real traces. Consequently, theveracity gap between the two categories of models started to shrink.

In this chapter, we are concerned with analytical approaches to mobility mod-eling; we consider data-driven approaches in the next chapter. Thus, herein weprovide an overview of the current classifications, explore the space of the state-of-the-art models and position our work in this space.

2.1 Foundations of Mobility Modeling

One of the earliest models, the random walk model [10] was first introduced intoscience by mathematician Karl Pearson in a letter to Nature [12] in 1905:

A man starts from a point O and walks l yards in a straight line; he thenturns through any angle whatever and walks another l yards in a straight line.He repeats this process n times. I require the probability that after these nstretches he is at a distance between r and r + ”r from his starting point O.In the same volume of the journal, Lord Rayleigh provided the solution, noticing

the analogy between the given problem and the problem he solved 25 years earlier,while considering superposition of sound waves of equal frequency and amplitude.When applied to the random walk problem, the sought probability distribution isa function of the number of steps n and the radius r from the origin O, which canbe expressed as P (r) ¥

2r

ne≠r

2/n for a very large n. The probability is therefore

highest in the area around the origin1. Yet some decades later, Albert Einsteinmathematically described this problem, today also known as Brownian motion. Tothis day, random walk models remain relevant in various scientific and engineeringdisciplines.

In the early studies of mobile networks, mobility models considering randommovements were the most commonly used models, due to their simplicity. A va-riety of random models, termed as generic and loosely based on observations ofreal mobility and its properties (for example node speed or group mobility), havebeen proposed [10]. Perhaps the most well known random mobility model is therandom waypoint (RWP) model, an extension of the random walk model withpauses between changes in direction or speed. Mobile ad hoc network protocolsand algorithms were routinely evaluated based on their performance under simu-lated generic mobility models. The underlying mobility model, however, is knownto greatly a�ect the evaluation results [8]. These findings, as well as the strive todevise more realistic models, have motivated many research e�orts in the mobilenetworking area, over nearly two decades. In other scientific disciplines, however,human mobility has already been attracting considerable attention for a longerperiod of time.

1Anecdotally, Prof. Pearson noticed that “[...] the lesson of Lord Rayleigh’s solution is thatin open country the most probable place to find a drunken man who is at all capable of keepingon his feet is somewhere near his starting point!”

2.2 Human Mobility Through the Space-Time Prism 9

!"#$%&'

()*+$

,- ./01

02

1

34#$5/34#$5.

6478

Figure 2.1: Space-time paths for two nodes meeting at their common location of work.

2.2 Human Mobility Through the Space-Time Prism

Understanding human mobility has been an objective of multiple research disci-plines: from public health (epidemiology) to urban planning and transportationresearch to, most recently, mobile networking. As it is usually the case with in-terdisciplinary fields, many concepts get rediscovered or introduced in parallel, butalso there is a disarray in terminology. The obtained knowledge is inevitably dis-persed, and thus there is a need to relate interdisciplinary models. In this sectionwe attempt to bridge the gap by introducing the terminology used in transportationresearch and relating it to the terminology used in networking research.

In transportation research, the dominant approach to examining mobility is theactivity-based approach. This approach is based on the concept of space-time geog-raphy, proposed by Hägerstrand in 1970’s [13] to demonstrate how human spatialactivity is often governed by limitations, and not by independent decisions of spa-tially or temporally autonomous individuals. In this conceptual framework, humanmobility is constrained by the individual’s capability, by the coupling of individuals,and by the restrictions imposed by authorities over them. Capability constraintsrefer to the limitations on human movement due to physical or biological factors,such as the walking speed, access to cars or public transportation and the need toreturn to a given location after a journey. Coupling constraints refer to the needto be in one particular place for a given length of time, often in interaction withother people. Authority constraints relate to the limitations and control of accessexerted on nodes to restrict their possible mobility, such as tra�c or safety rules(cars are not allowed to drive through pedestrian zones, people are not normallypermitted to enter a military base).

Examples of space-time paths are given in Fig. 2.1, which shows paths of two in-dividuals, a pedestrian and a driver, during a work day. These two individuals havea common work location. Authority over them imposes a restriction on work hours,locating them in vertical segments of their paths, while individuals’ capabilities—access to cars, walking or driving speed—determine the positions of other segmentsof their paths in the space-time graph. While an individual path represents the


Table 2.1: Examples of constraints on mobility at di�erent levels of abstraction.

Level Capability Coupling AuthorityStrategic Available time Work locations;

areas for shopping

Regulated work hours;

shopping hours

Tactical Means of transport Transport routes Schedules of public

transportation

Operational Speed of walking

or driving

Queuing and crowding Speed limits;

tra�c rules

actual path taken by the pedestrian or by the driver, a set of all points that can bereached by an individual, given a maximum possible speed from a starting pointto an ending point in the space-time graph determines the individual’s space-timeprism. Any kind of mobility, either trace-based or synthetic, can be formalized inthis manner. Consider synthetic mobility in the form of RWP: It is characterizedby mobility of a fixed number of nodes in a closed area with a convex boundary andwithout internal obstacles. The capability is given by the distribution of the speedof the flights the nodes make and the distribution of their pause times. There arenot any coupling constraints since nodes do not have a physical size or any otherform of interaction, and there are no authority-imposed restrictions that limit themovements other than the boundary of the simulation area.

In addition to constraints on mobility, there are travel behavior levels. Humanmobility can be seen as consisting of three levels [14]: At the strategic level humansdecide the activities they would like to perform such as going to work, shopping, ortaking a walk in a park as well as the schedule of those activities—these decisionsdefine their daily movements. The tactical level considers the implementation ofa strategic decision, such as choosing a mode and a way of travel, taking intoconsideration which is the shortest or fastest path as given by environmental factorslike obstacles and congestion. At the operational level, the physical process ofhuman movement is considered, including the walking or driving speed, the physicalsize of nodes and the interaction with other tra�c due to queuing for avoidingcollisions. Table 2.1 exemplifies the constraints and capabilities at the three levelsof abstraction.

The above terminology is more relevant and common in, for example, trans-portation research. In the networking research, models are usually classified withrespect to the scale at which mobility is observed. Thus, there are microscopic,mesoscopic or macroscopic mobility models. Microscopic mobility describes in de-tail individual behavior and interaction with other nodes and the environment.Phenomena observed in such detail pertain to tactical and operational levels. Atthe macroscopic scale, a model concerns system behavior in its entirety and with-out any reference to its underlying microscopic nature. Flows of vehicles on ahighway, or streams of pedestrians on a street are examples of macroscopic models.An individual is represented only as a contribution to the mean density, and cannot be followed from its origin to its destination. This corresponds to mobilityat the strategic level. At the intermediate position between the macroscopic and


microscopic representations is the mesoscopic approach, which aggregates severalindividuals, with the same properties and moving in the same environment, intogroups [15]. This means that mesoscopic models do not attempt capture mobilityof every single individual, but rather group mobility, where every group has itsown rules of behavior. Similarly to the microscopic models, this level of abstractionconcerns tactical and operational levels; however, since individual behavior is deter-mined by the group properties, the operational level is represented with less detailthan in micro-mobility. The mesoscopic models in this regard represent a compro-mise between macro- and microscopic models, aiming to achieve the computationale�ciency and analytical tractability of macroscopic models, but at the same timebeing able to represent some of the diversity and individuality of real users. Micro-scopic mobility is the most precise abstraction of how individuals move, but the useof such models is limited to analysis of small areas (or small number of nodes) andto agent-based simulations. Besides, for the majority of practical applications towireless networking, namely in cellular and traditional wireless local area networks,modeling an individual’s behavior in a high level of detail has limited usefulness,since the details of interaction between individuals are not considered essential forthe system performance [16]. Mesoscopic models, therefore, should be preferred tomicroscopic models, for finer granularity representations, however they are insu�-cient when describing an entire system consisting of heterogeneous mobility groups,moving in larger spaces and observed during long temporal scales. In such cases,the use of macroscopic models is necessary, since this category represents mani-festations of strategic and tactical decisions such as the route choice and activityscheduling, given certain constraints.

To summarize, the structure of mobility can be considered in di�erent aspects:in the sense of constraints and capabilities, levels of abstraction and the scale. Nextwe consider measurable properties of mobility.

Metrics of Mobility

In the wide spectrum of existing mobility models, many of these models were tai-lored to a specific networking application and resulted from di�erent modelingobjectives. Due to a possibly significant overlap across the models, e�ectively com-municating the models and modeling approaches is beneficial to systematize theknowledge in the field. Despite numerous e�orts to introduce a standardized ter-minology to this end, there has not been an unanimously accepted classification.Nevertheless, there is some consistency in the metric space that can be used as areference. These metrics can be classified in three categories based on the numberof nodes considered for the metric: individual, pairwise and community metrics.The provided listing of the metrics is only illustrative; we use the introduced met-rics to exemplify di�erent modeling approaches and to define metrics that will beused in the contributed papers.


Individual Metrics

Individual mobility metrics are used to describe behavior of a single node, that is,its trajectory in space and time. These metrics consider the nodes as independent,but aim at capturing the diversity of the individual patterns. To this end, metricsin this class are usually represented via probability distributions over possible setsof parameter values that the overall node population may exhibit.

• Visiting time, also pause or sojourn time, is the time that a user spends in agiven location, before moving to the next location.

• Flight length or jump size is a measure of the length of each segment in auser’s path.

• Number of visited locations [17] describes the spatial dimension of mobilityby measuring how much of the simulation space individuals cover.

• Location visiting preference and periodic re-appearance at a single locationcapture the distribution of the time that a user spends at a given locationand the frequency of returning to a given location. These two metrics have asignificant role in distinguishing realistic movement from random motion.

• Entropy and entropy rate represent the measures of uncertainty of the futurelocations that the user will visit. The entropy considers only the distributionof probabilities of visited locations, ignoring the order in which the user visitedthese locations. Given the sequence of visited locations L, the number ofdistinct locations in the sequence N , and the probability pi of being at alocation i, the entropy is measured as H = ≠

qN

i=1pi log2 pi. The entropy

rate observes the temporal correlation between visits to di�erent locations andis defined as a conditional entropy of the last visited location given the past,Hr= limnæŒ H(Xn|Xn≠1, Xn≠2, ..., X1) where the user’s location at time i isassumed to be generated by a random variable Xi, i = 1, ..., n.

Pairwise Metrics

This group is also referred to as encounter, as well as connectivity metrics. We canbroadly define contact as the basic unit of interaction in mobile ad hoc networks;a contact captures information about a pair of nodes interacting and the time andduration of interaction.

• Contact duration is the time that two nodes spend in each other’s transmissionrange. In opportunistic communication, this metric relates to throughput: theshorter the duration, the fewer messages can be transmitted.

• Inter-contact time is the time between two consecutive contacts for a pair ofnodes. In some use-cases, the inter-any contact time is relevant to measurehow often a node interacts with any other node.


• Contact rate is defined as the total number of contacts that a node establishesduring its lifetime normalized by the node’s lifetime, i.e., it is the number ofcontacts per unit time. Since our studies are focused on systems with churn(open populations), we consider contact rate to be a more suitable metricfor such settings than the number of contacts, which is usually measured inclosed, fixed population systems.

Community Metrics

At the macroscopic level of mobility, pairwise patterns get aggregated and we canobserve the emergence of the group or community patterns. Community met-rics come from the network theory, which is currently mostly developed for staticnetworks, thus somewhat disconnected from our main focus. These metrics areespecially interesting for designing social opportunistic networks.

• Node degree represents a cumulative metric of the number of contacts a userencounters. Studying the distribution of node degrees in the network helpsidentify the type and the structure of the network, i.e. whether it is a small-world, scale-free or a random network. The distribution of node degree is animportant property, since it helps answering, based on this simple classifica-tion, questions about processes occurring on a network, for instance how fastan epidemic would spread on it.

• Centrality measures the importance of a node in the network and there aredi�erent types of centrality: degree centrality, directly related to the previousdegree metric, which assesses how many direct neighbors a node can a�ect.Betweenness centrality is a measure of the extent to which a node is connectedto other nodes that are not connected to each other—in other words, thismetric measures how critical the node is to keep the entire network connected.One application is in finding the shortest paths between two nodes in thenetwork. Some advanced opportunistic routing protocols leverage the node’sbetweenness centrality in order to find bridge connections to otherwise poorlyconnected network partitions.

Developing a realistic mobility model involves determining which metrics areessential for the target scenario and fitting the model accordingly. However, byadjusting the model to exhibit observed statistical properties of metrics in onegroup, for instance individual metrics, the properties of the model in the same orin the two other groups, pairwise or group metrics, may be a�ected. Consequently,with such alteration of metrics, the model may yield either undesirable properties(e.g. too short or missed contacts) or infeasible parameters (unrealistically highspeeds of movement). The former case requires refinement of the metric space, whilethe latter case mandates re-evaluation and prioritization of the most importantmetrics that need to be captured.


2.3 Historical Overview of Approaches to MobilityModeling

Over the last two decades, mobility modeling has been a prolific area. Herein wedo not survey or classify particular models, instead we provide an overview and atimeline of relevant surveys; each survey, at the time of its compilation, documentedthe state of the art models and progress in the area.

A first classification of random walk models has been proposed by Bettstet-ter [18] in 2001. Camp at el. [10] survey and discuss several synthetic models,which are mostly based on random node movement and classified as either entityor group models. This work has been formative both for its historical value as beingone of the earliest surveys of models used for simulation of ad hoc networks andfor raising concerns about the reproducibility of performance results when di�erentmobility models, or even di�erent parameters for the same models, are used. Baiand Helmy [8] further investigated the reproducibility of results for routing proto-cols performance, documenting a stark variation in the ordering of protocols basedon the underlying mobility models.

The first survey that addresses social ties in mobility models is provided in [19].The authors give an overview of three model categories: random movement-basedmodels, which they denote as purely synthetic, the category of models that incorpo-rate some properties observed from real traces, such as the social network structure,and the third category of models based on the real world traces. The survey alsopresents a few pointers to network and connectivity simulation tools, and raisesimportant—yet still not entirely answered—questions regarding the necessity forstandardized benchmarks for protocol and system evaluation and for standardizedformats of mobility datasets.

As trace collection e�orts caught up the momentum, by 2011 many publicly ac-cessible libraries of mobility traces have become available. Aschenbruck et al. [20]survey these libraries and traces, and categorize them by the primary purpose forcollection, i.e., for analysis or for derivation of statistical properties. In this article,the authors also classify models in terms of model dependencies and restrictionsof node movement. With a particular interest in opportunistic networking scenar-ios, Karamshuk et al [21] observe manifestations of human movements and projectthe extracted properties along the temporal, spatial and social dimensions. In thesame networking domain, Thakur and Helmy [9] propose a framework that includesa multi-dimensional mobility metric space to investigate the adequacy of existingmobility models, but with the di�erence that metrics are projected on the individ-ual, pairwise and collective (communal) axes. These all are protocol-independentmetrics used for evaluation and comparison of protocols performance in oppor-tunistic setting. The fundamental idea behind the framework [9] is to explicitlysynchronize events leading to nodes visiting the same location at a given time,hence forcing the emergence of contact opportunities and explicitly targeting theapplication scenario. Remark that this principle of event synchronization imple-ments coupling constraints from Section 2.2.

2.4 Scenario-Based Modeling 15

Considering mobility modeling from a broader perspective, that is, for mobilenetworking research in general and not only limited to opportunistic networking,two recent surveys aim to establish systematic approaches to developing and com-municating mobility models. The survey by Treurniet [22] focuses on microscopicmodels and introduces classification of models with respect to implemented model-ing decisions: target and path selection (strategic and tactical mobility decisions),motion determination and obstacle avoidance (operational mobility decisions), andgroup dynamics. Hess et al. [23] devise a framework for developing and validatingnew models and provide guidelines for choosing among already available models.

Finally, with the proliferation of machine learning techniques for mobility pat-tern mining, numerous new methods have been devised to extract information fromlarge-scale data and multiple sources. These methods are surveyed in [24]. The ref-erence also proposes a taxonomy for the methods based on the application of theanalysis: whether it is for user, place or trajectory modeling.

2.4 Scenario-Based Modeling

This section proposes a concise roadmap with the main steps for devising or choos-ing a mobility model to fit the target application. Herein we also highlight the needfor a specific model to be used for the performance evaluation of opportunisticcontent distribution.

A scenario gives a description of node behaviors in a given setting. Thesebehaviors must be reflected in the corresponding mobility model. When developinga mobility model, or choosing among existing ones, a modeler first needs to considerthe specifics of the physical area where the communication takes place. The scaleof the area determines whether the population should be considered open or closed:in smaller areas, the mobility of nodes is unlikely to be confined within the areaboundary, and the node arrival process must be included in the mobility model. Inaddition, the function of the space can be instrumental for characterizing the arrivalprocess and the dynamics of an open population. These properties are importantto be captured since they are likely to a�ect the performance of the communicationscheme, e.g., bursty arrivals may a�ect the performance di�erently from arrivalswith smaller variations. If the area of interest is large, for example if the goalis to model movements of an entire city population, a single model may not beable to capture mobility in su�cient detail. Multiple models can then be appliedhierarchically. On a larger (macro) scale, the primary mobility model governs theselection of the city region [25, 26] where the nodes travel to in order to performcertain activities. The secondary, lower lever mobility model is next used to movenodes inside the chosen region, until the primary model recalls the node by selectinga new region. Depending on the intended level of detail, there may even be need forthe third-level mobility model responsible for the operational mobility. As a result,the composite model would comprise one model for each of the structural levelsof mobility (see Table 2.1). This is especially true for generating opportunistic


networking scenarios, where each of the levels can impact the network performancedi�erently. Namely, the strategic and tactical mobility determine the length ofinter-contact times, while the operational mobility may a�ect the occurrence ofcontacts and their durations.

Upon deciding on the structural levels of mobility that should be captured inthe targeted scenario, the next step concerns the exploration of the parameter spaceand the choice of mobility parameters. Helgason et al. [16] conclude that the inter-action between nodes caused by mobility and the structure of the space in a givenscenario have larger e�ect on the system performance and are hence more importantthan the detailed estimation of the mobility parameters, namely, the node arrivalprocess, speed distributions and walking behavior of the nodes. However, for theperformance of applications built on top of mobility models, the particularities of ascenario such as the internal structure of the space, may not be crucial, as we showin Paper D. That is, accurately capturing connectivity patterns is more importantthan recreating the movement caused by the physical constraints of the space wherecommunication takes place. In di�erent networking scenarios, di�erent propertiesof mobility must be captured. Below we explore a few motivating examples.

Opportunistic Networks

The main objective of generating opportunistic networking scenarios is to capturecontact events among the nodes. The structure of an opportunistic network canbe represented via temporal graph with nodes as vertices and intermittent contactsbetween two nodes represented by graph edges. Depending on the use-case whereopportunistic communication takes place, several sub-cases can be distinguished.An important aspect of opportunistic social networks [27, 28, 29] is the underlyingsocial graph, which should emerge from the recreated mobility patterns. As opposedto this, ephemeral opportunistic networks serving to disseminate contents in urbanareas [30] or one-time events [31] do not rely on the social connections, in fact,social structure may not even exist and the mobility model only needs to reflectthe dynamics of the network graph changes.

Modeling Dynamic Mobile User Populations

One of the use-cases for opportunistic communication that we described previouslyis location-aware content distribution in urban areas. In this context, the populationof mobile users is subjected to frequent changes in the number of users currentlyresiding in the locale. It is therefore important to take churn into account, that is,the mobility model must incorporate the process of users joining the population,i.e., the user arrival process and the process that explains how users leave thesystem and are no longer considered relevant for communication, the user departureprocess. Motivated by the need to evaluate the performance of this particular typeof opportunistic applications, we have identified the lack of simple, analyticallytractable models for populations with churn (also termed open populations). The

2.4 Scenario-Based Modeling 17

contribution in this respect is provided in Papers D and E where we propose,parametrize and validate a model based on queuing theory.

Future Networking ApplicationsPlanning deployments of future generation mobile networks requires rethinking howuser mobility will a�ect connectivity and handover. In 5G technologies, for instance,the promise of high data rates can be e�ciently realized by reducing the coverageareas, that is the radius of network cells. Moreover, the key technologies in 5G,millimeter wave wireless transmission technologies, further reduce the transmissiondistance between the cell tower and the users. The impact of the user mobilitytherefore increases as the cells shrink, and to support seamless connectivity in spiteof frequent handovers, mobility should be considered at a finer scale then in thecurrent cellular networks. The main parameters to be modeled accurately are thepatterns of user arrivals to the cells, the cell visiting times, also termed pause times,and the transition probabilities between neighboring cells.

Chapter 3Collecting and Analyzing Mobility

Data

Data is a precious thingand will last longer thanthe systems themselves.

Tim Berners-Lee

The purpose of collecting location data is manifold: to observe mobility andits driving forces, to parametrize mobility models, to replicate mobility scenariosby importing the traces to mobility simulators, and most recently, to perform ad-vanced data mining, for example for activity recognition or mobility prediction ofusers. In this chapter, we scan the space of methods for collecting mobility data,survey several well-known traces and trace libraries, and summarize the statisticalproperties extracted from trace analysis.

3.1 Trace Acquisition and Source Origins

Understanding and modeling human mobility, as we have seen in the previouschapter, has a long history. For demystifying the laws and forces that drive humanmovement, empirical data—gathered both at the population and at the individuallevel of mobility—has been an indispensable resource of information.

The first attempts to systematically describe human movement can be tracedto even before Pearson’s mathematical model for a random walker, to the 1885publication The Laws of Migration [32]. This study used census data, one of thefirst and still relevant data sources for modeling human mobility at populationlevels [33, 34]. Complementary to this, survey data about individuals’ activitieshas been analyzed in [35, 36]. Census and survey data from these examples sharethe commonality of the data gathering procedure—which is that they do not rely on

19

20 Chapter 3. Collecting and Analyzing Mobility Data

any technological means to record human mobility. Another such source of data istransportation and travel surveys. Representative examples are the Travel TrackerSurvey1 for the metropolitan area of Chicago, documenting detailed trips (includingtrip purpose, transport mode, departure and arrival times) for each of the membersof more than ten thousand households, and a similar survey conducted for the cityof Los Angeles2 [37].

In the last decades, the availability of communication and localization tech-nologies as well as the pervasiveness of human-carried devices such as smartphoneshave been the main enablers for collecting large datasets of human movements.Carried by their owners throughout the day, smartphones are generally seen asproxies for the personal identity and location of their owners. A game-changingmethodology for uncovering mobility patterns was introduced with the analysis ofCall Data Records (CDRs) [38]. CDRs contain attributes about telecommunica-tion transactions, such as a voice call or message exchange, including, among others,identifiers and locations of the two participating end devices. In addition to po-sitioning information provided by cell towers, mobile phones utilize measurementsof Global Positioning System (GPS) and the locations of nearby Wi-Fi networks.Being collected by devices, these measurements are considered as the ground truthfor tracking the users’ locations. Services and application interfaces, such as GoogleLocation Services [39] further facilitate accurate localization by leveraging multiplesources of user’s digital footprints.

Next we survey common methods for collecting traces, from smaller, controlledexperiments to large collection campaigns. Two principal methods to collect tracesare [20]: 1) by recording user locations directly from their devices by means ofdedicated or built-in tools, and 2) by collecting records that users leave in commu-nication systems—digital artefacts. We discuss these methods in more detail.

Recording Device Location

Recording GPS coordinates of users devices is the most widely used way to obtainlocation data. However, GPS depends on the quality of satellite signals and factorsincluding weather conditions, physical obstacles such as buildings, which all limitits outdoor accuracy to several meters at best, while performing poorly indoors.For accurate indoor positioning, alternative localization techniques are available, forinstance, techniques that utilize measurements from Wi-Fi signals or infer proximityfrom Bluetooth Low Energy (BLE) enabled devices.

While monitoring location directly from devices—when accurate—is beneficial,since it gives the ground truth, running large scale experiments is challenging eitherbecause it is cumbersome to carry out, e.g., if it requires recruiting participants anddeploying dedicated devices, or raises privacy concerns, when participants are us-ing their personal devices. Taking necessary measures to insure the privacy of user

1Available: http://www.cmap.illinois.gov/travel-tracker-survey.2Available: http://www.scag.ca.gov/travelsurvey.

http://www.cmap.illinois.gov/travel-tracker-survey

http://www.scag.ca.gov/travelsurvey

3.1 Trace Acquisition and Source Origins 21

identities and data, several research groups have conducted experiments employinghundreds of participants and made the collected traces publicly available: NokiaMobile Data Challenge [40, 41], Geolife [42], and SensibleDTU [43]. By collectingmeasurements from multiple sources such as device sensors (for proximity detec-tion) and phone usage records, these studies sought to understand life patterns ofindividual participants, for example, what places they visited, as well as the socialinteractions among the participants, the number of which ranged from one hundredto nearly one thousand; the duration of experiments spanned from several monthsto two years. Participants in these campaigns were mostly student volunteers andresearchers who were running the experiment, thus the studies gave a biased viewon human mobility in general. Still limited to a specific scenario, but in a di�erentcontext, mobility traces were obtained from theme parks and country fairs [44, 45].For the Locaccino study [46], Cranshaw et al. conducted an experiment in whichthey collected locations of 489 users and their social network ties, both in physicalworld and in location-based online social networks. This study aims to identifythe implicit social link between physical interaction and online connections amongusers; the dataset, however, is not publicly available. There also exist studies onmovement patterns following an emergency and disaster situations, such as afteran earthquake or natural disasters (storms, flooding) [47, 48].

The specifics of the networking scenario which we aim to recreate determinethe sampling resolution for obtaining the location, as well as the format of thetrace and what additional information is necessary for the analysis and mobilitymodel development. For instance, observing an individual’s location independentlyof other users su�ces for use-cases when services are provided to that individualonly, but is probably less interesting for studying social interactions. As opposedto this, understanding contact patterns is critical for designing and evaluating op-portunistic communication schemes. While the contact information, for a pair orgroup of individuals, can be inferred for example, from geographical coordinates,this requires high spatial and temporal granularity of the examined trace, andassumes that two users are able to communicate if located within a certain dis-tance, which may not be the case due to physical obstacles or interference. Thus,some measurement studies combine recording both users’ locations and contacts be-tween users. Contact measurements provide information about people co-located intime and space and as such, this type of measurements is especially interesting foropportunistic communication. Beyond applications in networking, capturing andcharacterizing contact patterns has an important application in other disciplines,e.g., epidemiology, for monitoring and controlling the spread of infectious diseases.The exact definition of a contact varies for di�erent applications. In opportunis-tic networking, we are concerned with user proximity within their devices’ wirelesstransmission range, which depends on the used technology, whereas in epidemiologyonly contacts within certain (short) distances matter. Di�erent objectives of theexperimenters have resulted in capturing human proximity at various scales: fromseveral meters—a typical range of Wi-Fi and Bluetooth communication [49, 50, 51,52, 53, 54], to few meters long distances [55] and close face-to-face interactions [56].


Mining Cellular and Wireless Networks Data

The second method exploits data from existing communication systems and net-works, including massive communication records from cellular networks (previouslymentioned CDRs), but also smaller data sources from university and enterprisewireless networks. Since the data can be pulled from a single, or a few points in thenetwork, this approach is more practical for acquiring large-scale data, in terms ofthe number of users and the trace duration. However, the location accuracy of suchtraces depends on the density of network elements, i.e., base stations in the cellu-lar network or access points (APs) in the wireless network; moreover, it is limitedby the assumption that users are connected to the nearest AP or a base stationwhich may not always be true. Another large source of mobility information isuser-contributed content, for instance, geo-tagged photos and tweets, and check-insto location-based social networks [57]. As the measurement from a wireless networkwas the primary method to collect traces analyzed in this thesis, we will continueand extend the discussion in Section 3.2.

CDRs contain timestamped geographical coordinates for a user for every callmade and text message sent or received. These records represent a vast amountof data about every mobile network user, automatically generated by telecommu-nication systems and archived for billing purposes and network troubleshooting.The great advantage of using CDRs is the amount of data, both in terms of thenumber of users and in terms of number of communication events. One of the firstCDR-based studies were [58, 59]; Onnela et al. [58] used the dataset that coversseven million users to reconstruct social structure and infer the strength of commu-nication ties between the users in the set. Gonzalez et. al. [59] investigated a set ofhundreds of thousands of users to infer individual’s trajectories and calling activi-ties. The dataset analyzed in [60] is probably the largest CDR trace—and perhapsmobility trace in general—comprising records of more than 60 million people overthree months in 2013. However, CDRs on their own do not provide su�cientlyhigh granularity for strong conclusions about users movement since they exhibitvast uncertainty of users’ whereabouts in-between calls. Calabrese et al. surveyexisting filtering and processing techniques to extract insights from cellular networkdata [61], while Chen et al. [62] propose adaptive techniques to reduce temporalsparsity in the CDR traces.

Databases and Trace LibrariesWhile most of the collected traces were meant to serve for specific studies and, dueto the nature of the traces and privacy concerns, have not been publicly disclosed,there have been initiatives to gather datasets and make them available to the en-tire research community. One such initiative resulted in the CRAWDAD project[63], a community archive for wireless network data, housing around 120 datasetscollected for purposes such as network tra�c characterization, monitoring of signalstrength, bit-error rates and energy e�ciency in wireless local and sensor networks,positioning of pedestrians and vehicles.

3.2 WLAN Mobility Traces 23

Aside from CRAWDAD, several stand-alone repositories also provide open data:a sizable and significant contribution is the multi-source datasets provided by Tele-com Italia [64] comprising two months of telecommunications, weather, news, socialnetworks and electricity data from the population of the city of Milano. Other cam-paigns, for instance Orange Data for Development (D4D) [65, 66] were open onlyfor limited time and upon request, but their significance is in the volume of thedata and the fact that they comprise large populations of users—up to 9 millionmobile network subscribers. Cellular network datasets are di�cult to keep publiclyavailable since the data contains not just locations of the users but also call andmessaging details, and records of user data tra�c.

Concerns About Data Privacy and Anonymity

Conducting research on users’ private data, such as location, has to deal withthe contrasting goals of reproducibility and disclosure of sensitive data–which iswhy there are very few publicly shared datasets. There are two ways to attackthis challenging issue, depending on the the trace collection method. First, if thedata has been collected experimentally, with consenting participants, it may bepossible to release at least a part of it, depending on the agreements and thelevel of anonymity in the released subset. Furthermore, participants must be ableto understand the implications of sharing their data, but also have the opt-outpossibility, if they wish to stop their data to be used in some future studies, forwhich they have not consented to at the time of trace collection. The secondcase concerns datasets for which obtaining consent from all involved individualswould be impossible, examples are CDRs and WLAN traces. Thus the first steptowards obfuscating users private data is to anonymize the users’ identifiers, butthis is insu�cient, since the data providers should guarantee that users cannot bere-identified with even anonymized data. The problem of re-identification is alsoknown as the “power of four”, that is, only four data points is all that is neededto uniquely identify, with a high probability any individual [67, 68]. Fortunately,security research has been very active in this field and several solutions have beenproposed including: k-anonymity for user trajectories [69], a method which “hides”a user in a cluster of users with similar movements, location obfuscation [70], i.e.,irreversibly altering users locations so that they do not represent real locations anda solution achieving di�erential privacy [71] by generating a synthetic model thatclosely reflects a real CDR trace.

3.2 WLAN Mobility Traces

As a special case of large datasets from communication networks, we considerwireless local area network (WLAN) measurements and mobility inferred fromthese data. The context where the data acquisition takes place is usually a singleinstitution—such as a university or a company—providing access to its residents.


The merit of WLAN measurements is that they can be relatively easy to acquire,for instance if the access data is regularly logged for statistical purposes, such as inpublic networks providing Eduroam roaming services [72]. A few main advantagesof collecting and processing these traces compared to records from cellular networksshould be noted. First, the cell size in mobile networks can be at least an orderof magnitude larger than the coverage area of access point in WLANs, thus usingmobile network records may result in coarser location granularity as compared toWLAN traces [62]. The second is the possibility to directly collect the traces fromthe network, and to protect the privacy of users by anonymizing their identities, lo-cation and tra�c data during the acquisition procedure. The same procedure maybe much more di�cult to carry out in the case of cellular network data: due to thebusiness-sensitive nature of this type of data, mobile operators are often reluctantto share the records with researchers. Below we survey to main groups of WLANmobility traces, aquired in campus WLANs and in public hot-spots.

Campus Traces

The procedure for acquiring traces of user access and tra�c patterns from WLANscan take several forms: from the simplest one consisting of pulling access and ac-counting records from the authorization server, to the more complicated procedureswhich assume setting up syslog event triggers, polling APs through Simple NetworkManagement Protocol (SNMP), and the implementation of tcpdump sni�ers [73].Even though the latter procedures are cumbersome due to the vast amounts ofgenerated data, the goal of obtaining measurements (possibly with the assistanceof network administrators) is attainable. Thus, majority of the analyzed WLANtraces come from research institutions and university networks. One of the earli-est studies, from 2000, was published by Tang and Baker [74] who analyzed thebehavior of users of a wireless network in an o�ce building at Stanford Univer-sity, in particular patterns of how often and at what locations users access thenetwork. Reference [75] studied the user behavior and network performance in acampus environment, however in the context of users accessing a public WLANnetwork since the trace captures activity of a more diverse set of users—conferenceattendees. Chincilla et al. [76] analyzed wireless information locality and accesspatterns at the University of North Carolina. Kotz [77] and Henderson [78] ana-lyzed, to this day, one of the most popular wireless network traces—the Universityof Dartmouth trace. Tuduce and Gross proposed a model of users moving insidea campus [17]. The above references incorporate trace analysis of both handhelddevices and laptops; the latter type of devices however is clearly not best suitedfor inferring the users’ mobility with high temporal granularity. Focusing on thewireless network traces generated by users’ personal digital assistants (PDAs) [79],the authors pursued to understand and model mobility traces and usage patterns ofusers with hand-held devices, contrasting those and comparing with—at the time—dominant laptop users. All the above pioneering works established the methodology

3.3 Mobility and Network Simulators 25

for recording and analyzing users movement and network usage patterns. As a con-tribution in this thesis, the traces collected from our university campus network atKTH are analyzed in Papers A and B.

Many researchers continued to examine traces from campus networks [44, 50,80, 81, 82]. The major limitation of university wireless network traces is that thescenario is limited to users on the campus, e.g., students and employees, who maynot even be on the campus all the time during the day and if they are on thecampus, there is a high degree of repeatability of their behaviors.

City-Wide Traces

Public WLANs capture a wide range of user profiles, thus traces from these networkscan provide a more general view of the user mobility and network usage than theuniversity traces. The analyses of such traces have implications for network tra�cmodeling and provisioning, and investigating users movement patterns, in particularfor considering hot-spot occupancies.

Afanasyev et al. [83, 84] analyzed tra�c and usage of the Google’s metropoli-tan Wi-Fi network deployed in Mountain View. Ile Sans Fils [85] is a dataset ofwireless sessions collected for six years from Wi-Fi hotspots deployed in the cityof Montréal, Québec. The trace contains session records for around 115 thousandusers connected at 352 hot-spots over a period of six years. The trace is exploredin di�erent contexts: Isaacman et al [86] use the traces to explore the design spacefor networks operating with constrained connectivity, Bao and Liang [87] to modeluser transitions between hot-spots by means of queuing network analysis, while [88]examined the trace to predict locations that users will visit next. Ghosh et al [89]analyze and model tra�c and association patterns of users at hot-spots in two largeUS cities, and highlight the di�erences for the modeled quantities across di�erentvenues (e.g., co�ee shops versus enterprises).

3.3 Mobility and Network Simulators

The two main applications of mobility traces are in analytic and in simulation-based studies. As previously explained, the aim of analytic studies is to inferproperties of human mobility—either universal (e.g., flight lengths, speed distribu-tions) or scenario-specific (visiting time at a given location). In simulation-basedstudies, mobility traces are used to replicate real-world mobility in a simulationenvironment, which is a crucial first step in the testing and development of networkprotocols and in the evaluation of mobile systems performance. This section ad-dresses the use of mobility and integrated (mobility and network) simulators in thedomain of opportunistic communication. We limit the focus on the two simulatorsthat we use in this thesis: the first one is a pure mobility simulator, and the secondone can be used to simulate both mobility and networking operations.


Pedestrian Mobility SimulatorRecognizing that opportunistic communication strongly depends on microscopicmobility and aiming to fill the gap in the space of high resolution mobility traces,VukadinoviÊ et al. propose the use of a pedestrian mobility simulator for generatingmobility patterns of users in urban scenarios, such as people walking on a street.The traces were generated by Legion Studio [90], an agent-based pedestrian mo-bility simulator used for commercial applications and design of large public spacessuch as subway and railway stations, stadiums and airports. The simulator is in-tended to recreate very realistic operational mobility of individual nodes, takinginto account di�erent behavioral models and walking speeds, thus implementingthe capability constraints, as well as the coupling constraints manifested throughthe interaction with other nodes (due to queuing or congestion) or with physicalobstacles in the environment. The same approach towards modeling microscopicmobility was applied in [16] to study the impact of pedestrian operational mo-bility on connectivity of users in two scenarios: pedestrians in a grid of streetsand passengers in a subway station. We use the same traces in Papers D andE to parametrize our open population model and to evaluate the performance ofcontent spreading schemes in these urban scenarios. We note that pedestrian mo-bility simulators are a valuable source of realistic mobility traces since they allowfor generating customized, scenario-specific mobility of arbitrary large user pop-ulations, while faithfully recreating both social and physical forces driving users’mobility. On the down side, sophisticated commercial solutions may not be avail-able to most researchers, for example Legion Studio has been superseded by LegionSpaceWorks, a software no longer open for academic purposes. Nevertheless, elab-orate open-source frameworks such as Vadere [91] are also emerging in the space ofpedestrian micro-mobility simulators.

Opportunistic Network SimulatorThe above discussed mobility simulators are essential for generating realistic andeasily configurable mobility scenarios. However, for the purpose of testing the op-eration on an entire opportunistic mobile system, we need yet another type of asoftware tool—a network simulator. Several such simulators are available and weuse an integrated mobility and network simulator, the Opportunistic NetworkingEnvironment (ONE) [92], to obtain contact patterns from the aforementioned Le-gion mobility traces. The ONE simulator is a simulation environment capable ofgenerating movements of nodes and handling delay-tolerant networking operationsincluding event generation, message exchange, DTN routing and application pro-tocols, visualization and analysis. The generated events include connection events(link up/down) and messaging events (message received/transferred/deleted). Thesimulator is built as an extensible framework with interfaces for importing and ex-porting mobility traces, connectivity events and messages. From the perspective ofmobility simulation, node movements can be generated from predefined (synthetic)models or from real traces.

Chapter 4Opportunistic Networking:

Applications, Platforms andPerformance Evaluation

The Internet has introduced an enormously accessible andegalitarian platform for creating, sharing and obtaininginformation on a global scale. As a result, we have new waysto allow people to exercise their human and civil rights.

Vinton G. Cerf

The main application area considered in this thesis falls under opportunisticnetworking. In this chapter we introduce the context for research problems thatpertain to the design and evaluation of these specific types of networks, and wemotivate a subset of research questions addressed in our publications. First, weprovide a historical overview of ad hoc communication networks, positioning op-portunistic networks in this landscape. Next, we describe the system setup andelaborate on a particular type of opportunistic networking application in the con-text of a location-aware content distribution scenario. This scenario will be thefocal point of the feasibility study and the performance evaluation in Papers Dand E. In this chapter, we also survey mobile opportunistic system prototypes and(commercially or experimentally) deployed systems in this area. Finally, we intro-duce mathematical tools used for modeling and analyzing the performance of theinvestigated networking application.

4.1 Mobile Ad Hoc Networks and Delay TolerantNetworking

The concept of wireless multi-hop, self-organizing communication networks is asold as the first packet-switching radio networks. An example of the latter is theDARPA’s Packet Radio Network Project (PRNet) [93] (1973). PRNet featured a

27

28Chapter 4. Opportunistic Networking: Applications, Platforms and Performance

Evaluation

distributed architecture of broadcast radios and relied on the combination of Alohaand carrier-sense multiple access (CSMA) protocols for the dynamic sharing of thebroadcast radio channel. The system however revealed scalability, security and en-ergy management issues. To address these issues, in 1983 DARPA launched theSurvivable Radio Networks (SURAN) program [94], aiming to develop techniquesthat would provide survivable communication for thousands of nodes in a dynami-cally changing network, under severe stress conditions. After these initial attempts,in 1997 the U.S. Army started a project called Tactical Internet [95] with the goalto develop the architecture and routing protocols for seamless data communicationacross heterogeneous networks. The project provided an important feedback on theuse of modified commercial Internet networking protocols in the tactical environ-ment, based on extensive modeling, simulations as well as the evaluation of a largescale deployment including several thousands of radio devices and controllers.

The emergence of portable communication devices and the introduction of wire-less technologies such as Wi-Fi and Bluetooth in the early 2000’s, initiated nearlytwo decades of research e�orts on mobile ad hoc networks (MANETs) and theshift from military deployments to the commercial sector and the standardizationcommunity. MANETs were envisioned as general-purpose in the sense that theirdesign should support any legacy TCP/IP application, rather than to adhere toany specific application. Possible applications spanned from tactical operations toemergency and disaster recovery services, to home and enterprise networking. Thecore of MANET research focused on enabling legacy Internet services in infras-tructureless, autonomous networks—in e�ect, replicating the functionality of wirednetworks by replacing wired links with wireless ones. MANETs inherited commoncharacteristics found in wireless networks: varying signal-to-noise ratio, di�erenttransmission ranges of di�erent nodes, interference; and, in addition, exhibitedcharacteristics specific to ad hoc networking: an autonomous and infrastructurelesstopology, the mobility of the nodes, multi-hop communication where each node actsalso as a router, and the uncoordinated sharing of the wireless medium by all nodesin the transmission range. Routing protocols developed for wired networks wereno longer suitable for networks with frequent topology changes; new protocols hadto be invented, but the problem was being solved by using wrong tools, that is,by trying to find end-to-end paths under the assumption that such paths alwaysexist. In practice, however, the existence of end-to-end connectivity turned out tobe feasible, at best, only in dense deployments and at added costs of decentralizednetwork management.

Concurrently with the main thrust of research e�orts in the MANET area, theconcept of delay-tolerant networking (DTN) started to develop. The DTN conceptwas built on the foundations of the Interplanetary Internet [96], aiming to achieveinteroperability between diverse network architectures, especially those that suf-fer from frequent network partitioning and large communication delays (e.g. deepspace communication or underwater acoustic networks), but also the types of net-works that are targeted for extreme environments (sensor-based and military adhoc networks). In 2003, Fall [97] proposed the DTN architecture, which was based

4.2 Data Delivery in Opportunistic Networks 29

on the abstraction of message switching and used the store-and-forward operation.The DTN architecture was built as an “overlay” network on top of the transportand below application layers. This overlay, termed a bundle layer, maps the appli-cation data units into protocol units—message aggregates called bundles. Devicesimplementing the bundle layer, called DTN nodes, are responsible for handlingbundle messages by enabling persistent in-network storage (to combat networkpartitioning) and by forwarding bundles in a hop-by-hop fashion to achieve reliabledelivery. DTN concepts however did not pertain exclusively to disconnected mobilenetworks, but in general to any type of networks that experience challenges suchas intermittent connectivity, large delays or include network-level heterogeneities.

Expanding on the DTN paradigm and motivated by the idea that physical pathsand mobility can be treated as a replacement of network links between the nodes,researchers have embarked on developing the concept of store-carry-forward com-munication. Borrowing conceptual ideas of MANETs that no network infrastructureshould be required to enable communication between mobile nodes, opportunisticnetworks have emerged as a particular type of DTN networks. As the term implies,opportunistic communication exploits intermittent connectivity of devices whichinitiate message forwarding when, due to their mobility, they come into directtransmission range of other devices. Whereas in [98] the authors showed that mo-bility in MANETs can be utilized to improve capacity, in general, mobility in thistype of networks is seen as a communication inhibitor. In opportunistic networks,the situation is the opposite: mobilityenables communication between otherwisedisconnected network partitions.

4.2 Data Delivery in Opportunistic Networks

Data delivery in opportunistic networks takes two main forms: routing and contentdissemination. The former communication mode can be seen as source-destination,or user-centric communication, and the latter one as broadcasting/multicasting orcontent-centric communication.

Routing in Opportunistic NetworksThe aim of routing protocols in opportunistic networks is essentially the same asthose of MANETs—to find a path from the source node to the destination over adynamic network topology. Approaching the routing problem equipped with DTNtools, the problem is reformulated such that the assumption of connectivity is re-laxed, where paths are formed by intermittently connected intermediary nodes,exchanging messages in a pairwise manner. One of the first proposed solutions wasthe epidemic routing protocol [99], which modeled message spreading as an infectiontargeting all nodes in the network, and therefore ensuring the minimum delay andmaximum probability that the message finally reaches the destination. An obviouslimitation of this approach was excessive usage of network resources, which moti-vated other researchers to seek resource-e�cient solutions, while trading o� delay


Evaluation

and delivery probability against resource consumption. The acceptable trade-o� isachieved by considering a utility function to find the best possible next hop, i.e.,the forwarding node. Various utility functions have been defined that concerned,most frequently, contact information (last seen contact, frequency or duration),node characteristics and resources (battery, computing power), or social relations.Notable among the first utility-based protocols are PRoPHET [100], a history-basedprotocol that replicates content items over high likelihood paths towards destina-tion, and Spray-and-Wait [101], which limits the number of copies generated bya�ected nodes. Socially-aware protocols optimize the selection of forwarding nodesby utilizing the information on social relations; this information is derived fromthe analysis of the users’ social network interactions. Two notable examples areSimBet [102] and BUBBLE Rap [103]. While SimBet uses the similarity amongnodes as a metric, BUBBLE Rap bases the decision on the community membershipand node’s betweeness centrality. Devising e�cient routing algorithms attractedmuch attention in the opportunistic networking community, and resulted in a largebody of work, as surveyed in [104]. Nevertheless, these protocols support the same,unicast communication mode, in the sense that they are source-destination oriented.

Content Dissemination and the Publish/Subscribe Paradigm

Rethinking the application domain and the general trends of networking paradigmsshifting from user-centric to content-centric, Karlsson et al. [105, 106] envisioned abroadcast communication mode (as opposed to unicast) for opportunistic networks,introducing podcasting as an application for mobile delay-tolerant networks. Thefirst system tailored for opportunistic content dissemination was developed withinthe Podnet project [107]; the later enhancements of the system were addressedin [108, 109]. In this view, mobile nodes use a publish/subscribe scheme to exchangecontent items based on their subscriptions and in a peer-to-peer manner. The con-tent structure is divided into feeds—hierarchically organized logical containers usedfor grouping similar contents—and the dissemination protocol is receiver-driven,allowing the dissemination only of the content items that the user has requested.Thus, whenever two nodes with a subscription to the same feed meet, they will syn-chronize their views of the feed, and exchange the content items either one of thepeers is missing. This system design [108] proved to be e�cient not just for contentdissemination, but also for mobile social networking [27], collaborative music shar-ing [110], and participative crowd entertainment [111]. Moreover, it conceptuallyinspired convergence of similar solutions [29, 112] towards the publish/subscribeparadigm. This is also the system model that we consider.

Performance Metrics

The objective of opportunistic communication schemes is the delivery of messageseither to a specific destination node or to all nodes interested in that piece of

4.3 Location-Based Opportunistic Networks 31

information. To assess the e�ciency of a communication scheme, several metricshave commonly been used.

• Message delivery delay is measured as the time required for a message toreach a single node or all nodes.

• Delivery probability estimates how likely is that a unicast delivery will besuccessful within a given deadline; alternatively, it can be used to quantifythe fraction of nodes that received the message.

• Communication overhead is defined in terms of the number of hops—interme-diate nodes that relay the message, or the number of transmissions (messagecopies) per relay node.

• Probability of the content persistence represents the likelihood that a contentitem, stored only by the nodes participating in the opportunistic exchange andwithout relying on the external infrastructural storage, will remain availablefor a given (possibly very long) time period.

The goals of an e�cient routing or content spreading algorithm—to minimizethe overhead while maximizing the delivery probability are clearly at odds. Most ofthe proposed algorithms seek for some trade-o� by using additional information andheuristics to decide when, that is, to which next-hop node and what messages toforward, nevertheless still relying on epidemic spreading algorithms as a primitive.We will not further discuss routing algorithms as they are outside of the topic of thisthesis—a thorough examination can be found in [113, 114]—but we acknowledgethat some commonalities with content spreading schemes permeate both perfor-mance modeling directions. In addition to the aforementioned epidemic forwardingprinciple, the assumptions of the mixing among nodes need to be postulated. Wewill shortly return to these modeling assumptions in Section 4.5.

4.3 Location-Based Opportunistic Networks

In diverse networking scenarios, ranging from occasional massive public gatherings(e.g., sport events) to static scenarios such as urban participatory sensing, largeamounts of data can be generated. This data is often relevant only locally, and per-haps even temporally, therefore storing them in some remote server in the networkcore is not well justified. Location-aware opportunistic communication emerges asa potential solution, bringing an array of additional benefits. First, the tempo-ral relevance of the content allows the users to control the content availability bystoring it on their devices, in an entirely decentralized manner. Once the contentis no longer deemed relevant, it will be removed from the user devices and lostirreversibly. The lack of a centralized storage also makes it more di�cult to storethe data in a server where it could be easily susceptible to censorship. Second,the use of location-based services entails that the users disclose their location, at


Evaluation

least with some level of accuracy, leading to privacy breach in the users’ physicallocations. With opportunistic communication, the users’ location privacy can bepreserved through obfuscation [115]. Moreover, for applications such as participa-tory sensing, ensuring the validity of the collected and disseminated information,including the verification of the physical origin, is crucial. Possible solutions for theverification of the user-generated content have been proposed [116]. Having recog-nized the above benefits, researcher have investigated location-based opportunisticnetworking in the context of vehicular [117, 118, 119, 120, 121] and pedestrianscenarios [122], in a purely wireless, infrastructure-less domain [123, 124, 125, 126]and in mixed domains [105, 127].

In this thesis we focus on location-aware opportunistic content spreading wherethe content relevance is geographically limited to a dedicated zone. This typeof applications use geofencing to create a virtual boundary—a geofence—aroundthe zone of interest and to trigger the location awareness of of the users whenthey enter or leave the geofenced location. In particular, we consider urban areas,densely populated with mobile nodes and an opportunistic service for spreadinglocal information, such as news, tourist information, transportation schedule ortra�c alerts. The service can also target opportunistic social networks in transientcommunities [31].

In this scenario, users generate content and assign to it an area of relevance.The dedicated area can be specified either explicitly, by defining the coordinatesof a geofenced polygon, or implicitly e.g., “the city square” and relying on third-party APIs to determine the region of interest. Alternatively, content items canbe generated by non-mobile entities, and injected into the wireless domain eithervia cellular networks or edge nodes to a set of visiting users. An example can be acity transportation agency that wishes to disseminate real-time tra�c updates andchanges to public transportation routes due to an unexpected event. In accordancewith the generic architecture of opportunistic podcasting [105], content items arepublished on channels with descriptive meta-data, where the location also representsan implicit filter.

As shown on Fig. 4.1, the mobility of users is such that they enter the area,traverse through it and then leave. While inside the area, a user occasionally getsin contact with other users and if any of the two users participating in the contactevent does not have the content item they are both interested in, an exchangeoccurs. All users are willing to contribute to content spreading for a certain timeperiod; upon expiry of this period or after leaving the locale, users stop spreadingthe information.

This is not an entirely novel use-case since several earlier studies have tar-geted similar scenarios. Terms such as hovering information [122] and floatingcontent [123, 124, 118] have been used to described the property of content residingin a locale without a dedicated storage. Floating content is envisioned to supportinfrastructure-less, distributed content sharing over a certain area called anchorzone. The necessary conditions for providing such service are established in [125]and [128] for the case of stationary, respective mobile nodes. Focusing on vehicu-

4.4 Opportunistic Communication Systems: Prototypes and Applications 33

Figure 4.1: Location-aware content dissemination.

lar communication, [120] presents a simulation-based study under what conditionsvehicular mesh networks (highway and city-wide cases) can sustain informationaround the region where the information was generated. In [119] the authors pro-posed an overlay for vehicular delay-tolerant networks called Locus. The commonobjective of [119, 122] is to assess the system performance and di�erent cachingalgorithms for the content survival. Ali et. al [126] perform an empirical studyby deploying a floating-content application to investigate the feasibility and per-formance of the service in a campus/o�ce environment. While the above systemsconsider an area of interest marked by an anchor zone with a fixed radius, ourservice domain is somewhat di�erent: we consider a geofenced area. Althoughsuch di�erence is not essential to the system operation, it allows for di�erent—andperhaps easier to human understanding—semantics.

4.4 Opportunistic Communication Systems: Prototypesand Applications

Perhaps the most rewarding recognition for the potential of device-to-device com-munication came from the 3GPP consortium which proposed support for proximityservices (ProSe) in LTE networks (as of Release 12). Meanwhile, for more than adecade, researchers have been developing opportunistic system prototypes, mid-dlewares and proof-of-concept applications, some of these have been transformedinto commercial products. Our main focus is on opportunistic systems for urbandeployments, therefore next we survey most relevant systems in this category andpresent their main, distinct features.

The Haggle [129] project defined the architecture for opportunistic, data-centricnetworking, in which application logic is separated from the underlying network.The idea behind the design of the system architecture is to enable seamless networkconnectivity and application functionality by taking advantage of both infrastruc-


Evaluation

ture and ad hoc networking possibilities. Thus, in the absence of infrastructureconnectivity, mobile devices using the system can dynamically organize ad hoctopologies by forming connections via available wireless technologies, e.g., Blue-tooth or Wi-Fi. Similarly to the Podnet system [108], Haggle uses publish/sub-scribe model to facilitate search and retrieval of data from other users; however,the di�erence between the designs of these systems is that content in Haggle isunstructured and matching is performed by using keywords.

The SCAMPI architecture (Service Platform for Social-Aware Mobile and Per-vasive Computing) [29, 130] defines an opportunistic router called SCAMPI router,based on the protocol and architecture specifications of the DTN Research Group[131], as well as some platform-specific extensions, e.g., for creating applicationlayer units—messages that map to lower layer bundle units. The router uses vari-ous mechanisms to discover peers and nearby services, such as IP multicast/broad-cast beaconing and static IP discovery for direct peer sensing, and transitive peerdiscovery supported by learning remote peers from their direct neighbors. Contentdiscovery is based on message metadata, whereas the content exchange is supportedby the publish/subscribe interface. Liberouter [132] represents an entire system fordo-it-yourself local networking, built around the SCAMPI stack and ported to small,inexpensive computing platforms, such as Raspberry Pis1.

CAMEO [28] is a context-aware middleware for opportunistic mobile social net-working, where the contextual information is derived from the local user, theirdevice and interaction with other users and physical environment. This makesapplications built on top of CAMEO also social-aware, and helps optimizing for-warding decisions.

The above platforms have been accompanied with mobile services and applica-tions, to demonstrate the platforms’ functionalities. Here & Now [31] and SocialPal [133] are two applications built on top of the SCAMPI platform. The formerapplication enables opportunistic social networking and experience sharing, whilethe latter one provides techniques to estimate the social path length between twousers of a social network. SmartCitizen [134] uses the CAMEO dissemination pro-tocol to stimulate citizens’ active participation in collecting environmental data andto facilitate sharing useful contents related to the quality of life in their city.

In addition, there are several commercial applications available from the applica-tion market, relying on or enhancing opportunistic networking mode. Among these,Open Garden’s FireChat [135] is probably the best known mobile social applicationfor o�-the-grid communication between smartphone users, having gained popular-ity during political protests, natural disasters and large festivals. Twimight [136]is an Android-based Twitter client for disseminating messages and sensor data inopportunistic manner, if no other connectivity is viable, for example, in disasterrecovery situations.

Finally, let us mention several notable systems engineered for rural deploy-ments. Experimental DTN deployments in rural areas have been used to enable

1http://www.raspberrypi.org/

http://www.raspberrypi.org/

4.5 Mathematical Tools for Modeling Content Dissemination and of PerformanceEvaluation 35

communication in remote villages with nomadic population [137] or in developingregions [138], as well as for wildlife tracking and environmental monitoring [139,140, 141].

4.5 Mathematical Tools for Modeling ContentDissemination and of Performance Evaluation

A rigorous and credible performance evaluation of opportunistic networking appli-cations often requires both analytic and simulation-based studies. By means ofsimulations, one can achieve higher levels of realism for the investigated scenarios,especially when based on real mobility data. Simulation-based studies, however,entail certain computational and time expenses—due to repetitions through whichstatistical confidence is achieved—and prohibit qualitative, more general assessmentof the solution at hand. Analytical performance evaluation is therefore required tocomplement the simulation studies.

In this section we survey analytical approaches to modeling the spread of in-formation on networks from two perspectives. A general approach, inspired bymathematical modeling of epidemic spreading as well as the specific case of model-ing opportunistic networks are presented. Then, we elaborate on the mathematicaltools that we used to model opportunistic content dissemination in a dynamic net-work of nodes and to evaluate the feasibility and performance of the investigateddissemination schemes. These tools are applied in Papers C–E. In the following,we also provide rationale for the chosen modeling approach.

First, the system model needs to consider contact processes and the underlyingpairwise (or otherwise group) mobility. This is achieved by making assumptionsabout individuals’ mobility, where most often individuals are considered to be sta-tistically equivalent, their movements mutually independent and governed by anergodic and spatially uniform mobility process. As a result, nodes exhibit homo-geneous mixing,where each node is equally likely to meet any other node. For thesake of mathematical tractability, homogeneous mixing is modeled as a Poisson pro-cess representing sequence of contact events, with the same intensity across di�erentnode pairs. While, on the one hand, the exponential property of inter-contact timesis justified by empirical analysis of real-life traces, the assumption of homogeneity,on the other hand, runs the risk of oversimplifying the contact process and, as weshowed as an example in Paper C, may lead to inaccurate assessment of the sys-tem performance. As a remedy to this oversimplification, several works introducedframeworks with di�erent levels of heterogeneity: from moderately heterogeneouspopulations with a small number of mobility classes [142] to extreme cases treatingcontact processes for each pair of nodes separately [143, 144], as well as account-ing for contact dynamics which depart from exponential inter-contact times [145],specifically for Pareto and hyper-exponential contact patterns. Although being ableto capture diversity of interactions among nodes, these schemes have some limita-tions. The focus of these models is on the heterogeneity of contact patterns. Other


Evaluation

research problems may consider the properties of the messages communicated, e.g.the age of information [146] or the version number [147]. Thus, re-purposing theheterogeneous mobility-based models to capture di�erent protocol metrics and theheterogeneity (diversity) of the content states adds another level of complexity andrenders them impracticable. In addition, such models may incur unnecessary com-plexity even for the problems they are aiming to address. As suggested in [148],modeling the spreading of information in a heterogeneous network with a Markovchain where every state represents not only the number of informed nodes, but alsoexactly which nodes are informed, can be simplified under certain assumptions.Specifically, if the pairwise contact rates are derived from the same probability dis-tribution, [148] shows that all the states with the same number of informed nodesare statistically equivalent.

The main questions about epidemic message forwarding concern the number ofinfected (informed) nodes, the potential of content to persist in the network, theage of content and so on. Modeling epidemic processes can be approached fromvarious angles, and below we discuss the two most common directions: the classicalapproach from mathematical epidemiology and the network-theory based approach.

The first approach is inspired by the classical theory from the field of mathemat-ical epidemiology [149, 150], and is based on the assumption that the populationcan be divided into di�erent groups or compartments depending on their infectiousstatus (the stage of disease). Within a group, individuals are considered indistin-guishable. The most basic labeling includes susceptible nodes that have not yetcontracted disease, but in contact with infected ones become infected and eventu-ally, after the incubation period expires, turn to recovered. Additional states can beincluded to denote, for instance exposed individuals—those that have been infectedby the disease but cannot yet spread it to others, or individuals immune to thedisease. By defining the basic, individual level processes that govern the transitionof individuals from one compartment to another, we are able to model the evolu-tion of the number of individuals in each compartment as a function of the time.Depending on the chain of the states that an individual passes through, di�erentmodels can be defined, most common being susceptible-infected-susceptible (SIS) orsusceptible-infected-recovered (SIR). The extensions of the simple compartmentalframework, including age-specific or spatially distinct groups, have been devisedand studied, for example in [151]. This approach has been also adopted in otherfields dealing with mean-field dynamic processes, such as physics and chemistry,where it was rephrased as a stochastic reaction-di�usion process [152, 153]. Giventhe (nearly a century) long history of epidemic modeling, there has been a wealthof results for di�erent configurations of the models, as well as the proofs of themodels’ robustness and their predictive ability [149]. However, there has also beenan evidence that the generalization and treatment of all individuals in the samecompartment under the assumption of the identical mixing patterns, may be in-adequate in some real-world scenarios. This realization triggered e�orts towardsmodeling with higher individual-level resolution.

4.5 Mathematical Tools for Modeling Content Dissemination and of PerformanceEvaluation 37

The second theoretical approach rests on the network theory [154] and aimsto explicitly capture the diverse patterns of nodes’ interaction. In network theory,networks are mathematically described as graphs—structures comprising of pointscalled nodes or vertices, and connections between the nodes called links or edges.Some important results are generally known, such that the degree distribution ofthe nodes (vertices) can determine the evolution of processes on the network (e.g.whether information, or a disease manages to spread) [154], or the e�ects of net-work topology on the rate and patterns of disease spread [155]. The contact networkmodel explicitly represents interactions between any two individuals and these in-teractions mediate the spread of the disease in the network. Since building theexact network graph requires knowledge of every individual and every connection,characterizing the network model requires significant e�ort and may even becomeinfeasible when the network scale increases. Hence, researchers typically have towork with approximate network models obtained through mean-field approxima-tions. Several mean-field approximation methods have been established to thisend.

The individual-based mean-field (IBMF) method [156] characterizes the spreadof epidemic through the spectral radius of the adjacency matrix of the networkgraph. The adjacency matrix specifies all connections in the graph, with its elementstaking value one for an edge connecting two neighboring nodes, and zero otherwise.IBMF models give solutions in the form of expectations of each network node beinginfected, where the state of a node is assumed to be independent of the state ofits neighbors. For static networks, their solutions show reasonable fits to numericalresults but the approximations tends to deteriorate when the independence assump-tion fails. The second, degree-based mean-field (DBMF) approach [157] describesthe system in terms of probability that a node of degree k is infected at time t.It does so by regarding all nodes of the same degree as statistically equivalent,e�ectively reducing the complexity of the system. Compared to the individual-based approach, the DBMF approach does not assume independence of the statesof neighboring nodes. This approach has proven to be, to some degree, suitableeven for networks with time-varying connectivity patterns. The third approachmaps the epidemic outbreak—the case when a large portion of the network getsinfected—into a bond percolation problem in networks. This problem can be con-veniently tackled with generating functions [158]2 and it can be shown that, in thelimit of the large graph size, there exists an exact solution of the average behav-ior of graphs under bond percolation. Percolation models are useful to determinethe probability that an infected node belongs to an infinite cluster and have beenapplied in [125]. In addition to the mean-field approximations they rely on, perco-lation models however have other limitations as well. Such models can only answerwhether the percolation happens, but do not allow for tracking the process overtime.

2The title of this reference is Generatingfunctionology—possibly the best mathematics booktitle ever.


Evaluation

An additional complication in modeling epidemics is posed when the networktopology changes at the rate comparable to (or higher than) the disease spreadingrate. Topological changes could be due to mobility, causing frequent rewiring ofgraph edges, or when the network is constantly evolving, i.e., experiencing churn.A study of Volz [159] suggests that, in such scenarios, a static network graph mayonly serve as an approximate model to study the spread of diseases. This calls forthe treatment of networks with time-varying graphs—a novel modeling approachbased on the dynamic network theory [160] which considers not only the spatial,but also the temporal component of the network graph as its integral part.

The main objective of our studies in Papers D and E is to devise a frameworkfor modeling the content spreading process for location-based applications and,subsequently, to analyze what parameters determine the system viability. Further-more, we wish to assess what are the minimum requirements imposed on the nodesparticipating in content forwarding.

To avoid working with time-varying network graphs, we reconcile compartmen-tal SIR models. However, we depart from the classical SIR models, which arecommonly described by systems of equations in the mean-field regime and extendthe system of equations by adding stochastic components to capture the evolutionof the number of nodes in each compartment. The analytic system model featuresseveral essential parameters: the node arrival rate (to the entire system or a spe-cific compartment), node departure rate representing the rate at which nodes leavethe area, recovery rate, which determines the time that nodes are actively spread-ing content and the rate of infection, derived via node contact rate. Introducingthe stochastic component in our models allows for capturing fine-grained fluctua-tions in the number of nodes per each compartment. This approach can be furtherexploited, as we show, to establish conditions for content survival via stochasticstability analysis. In particular, since we are dealing with SIR systems describedby a set of continuous-time di�erential equations, we utilize Lyapunov theory tostudy the system stability.

Chapter 5Summary of Original Work

In science, if you know what you are doing—youshould not be doing it. In engineering, if you do notknow what you are doing—you should not be doing it.Of course, you seldom, if ever, see either pure state.

Richard Hamming, You and Your Research

This chapter summarizes the original work presented in this thesis. All papersincluded in the thesis appear in the same way they were published; minor formattingchanges may apply. At the end of the chapter, we list other publications that theauthor of the thesis has published during the study period and that are not includedin this thesis.

Paper A: Revisiting the Modeling of User AssociationPatterns in a University Wireless Network

Ljubica PajeviÊ, Viktoria Fodor and Gunnar KarlssonPublished in Proc. IEEE Wireless Communications and Networking Conference(WCNC), 2018.

Summary: This paper presents an analysis of a large trace of user associationsin a university wireless network, which includes around one thousand access pointsover five campuses. The trace is obtained from RADIUS authentication logs and itsmerit is in its recency, scale and duration. We propose a methodology for extract-ing association statistics from these logs, and look at visiting time distributionsand processes of user arrivals to access points. We find that a large fraction ofthe network—around half of all access points—experiences time-varying Poissonarrival process, and association distributions can be modeled by two-stage hyper-exponential distributions at most of the access point. While network associations

39

40 Chapter 5. Summary of Original Work

in campus wireless networks have been extensively studied in the literature, ourstudy reveals changing patterns in user arrival processes and association durations,which seem to be characteristic for networks of predominantly mobile users, andallows the use of tractable network occupancy models.

The author of this thesis performed the work under the supervision of the secondand third authors. The article was written in collaboration with the second author.

Paper B: Predicting the Users’ Next Location from WLANMobility Data

Ljubica PajeviÊ, Viktoria Fodor and Gunnar KarlssonPublished in Proc. IEEE International Symposium on Local and Metropolitan AreaNetworks (LANMAN), 2018.

SummaryAccurate prediction of user mobility allows the e�cient use of resources in ourubiquitously connected environment. In this work we study the predictability ofthe next location, considering a campus scenario with highly mobile users. Weutilize Markov predictors, and estimate the theoretical predictability limits. Basedon the mobility traces of nearly 7400 wireless network users, we estimate that themaximum predictability of the users is on average 82%, and we find that the bestMarkov predictor is accurate in 67% of the time. In addition, we show that mod-erate performance gains can be achieved by leveraging multi-location prediction.

The first two authors formulated the problem addressed in this publication. The au-thor of this thesis implemented the prediction framework and performed the analysis,under the supervision of the second and third authors. The article was written incollaboration with the second author.

Paper C: Epidemic Content Distribution

Ljubica PajeviÊ, Gunnar Karlsson and Ólafur HelgasonPublished in Proc. ACM International Conference on Modeling, Analysis and Sim-ulation of Wireless and Mobile Systems (MSWiM), 2013.

Summary: Epidemic content dissemination has been proposed as an approachto mitigate frequent link disruptions and support content-centric information dis-semination in opportunistic networks. Stochastic modeling is a common methodto evaluate performance of epidemic dissemination schemes. The models introduceassumptions which, on one hand make them analytically tractable, while on theother, ignore attested characteristics of human mobility. In this paper, we inves-tigate the suitability and limitations of an analytical stochastic model for content

41

dissemination by comparison with experimental results obtained from real mobilitytraces. We conclude that, in the extreme cases of heterogeneous contact patterns, ahomogeneous model is unable to capture the performance of content disseminationwith respect to content delivery delays, while showing a reasonable fit in the casesof moderate contact heterogeneity.

The author of this thesis formulated and investigated the problem addressed in thispublication based on the discussion with the third author. The writing was done bythe first author under the supervision of the second and third authors.

Paper D: Modeling Opportunistic Communication withChurn

Ljubica PajeviÊ and Gunnar KarlssonPublished in Elsevier Computer Communications, vol. 96, pp. 123–135, 2016.

Summary: In opportunistic networking, characterizing contact patterns betweenmobile users is essential for assessing feasibility and performance of opportunisticapplications. There has been significant e�orts in deriving this characterization,based on observations and trace analyses; however, most of the previously estab-lished results were obtained by studying contact opportunities at large spatial andtemporal scales. Moreover, the user population is considered to be constant: nouser can join or leave the system. Yet, there are many examples of scenarios whichdo not fully adhere to the previous assumptions and cannot be accurately describedat large scales. Urban environments, such as smaller city districts, are character-ized by highly dynamic user populations. We believe that scenarios with varyingpopulation require further investigation. In this paper, we present a novel model-ing approach to study operation of opportunistic applications in scenarios wherethe population size is subjected to frequent changes, that is, it exhibits churn.We examine two location-based content sharing schemes: a purely opportunisticcase and an infrastructure-supported content sharing scheme, for which we providestochastic models based on stochastic di�erential equations (SDEs). We validateour models in five scenarios: a city area, subway station, conference, campus, anda scenario with a synthetic mobility model and we show that the models providegood representations of the investigated scenarios.

The author of this thesis formulated the problem in collaboration with the secondauthor. The model was proposed and validated by the first author. The article waswritten by the first author under the supervision of the second author.

42 Chapter 5. Summary of Original Work

Paper E: Ensuring Persistent Content in OpportunisticNetworks via Stochastic Stability Analysis

Ljubica PajeviÊ, Viktoria Fodor and Gunnar KarlssonPublished in ACM Transactions on Modeling and Performance Evaluation of Com-puting Systems (ToMPECS). vol. 3, issue 4, article no. 16, 2018.

Summary: The emerging device-to-device communication solutions and the abun-dance of mobile applications and services make opportunistic networking not onlya feasible solution, but also an important component of future wireless networks.Specifically, the distribution of locally relevant content could be based on the com-munity of mobile users visiting an area, if long term content survival can be ensuredthis way. In this paper we establish the conditions of content survival in such op-portunistic networks, considering the user mobility patterns, as well as the timeusers keep forwarding the content, as the controllable system parameter.

We model the content spreading with an epidemic process, and derive a stochas-tic di�erential equations based approximation. By means of stability analysis wedetermine the necessary user contribution to ensure content survival. We showthat the required contribution from the users depends significantly on the size ofthe population, that users need to redistribute content only in a short period withintheir stay, and that they can decrease their contribution significantly in crowdedareas. Hence, with the appropriate control of the system parameters, opportunisticcontent sharing can be both reliable and sustainable.

The author of this thesis formulated the problem in collaboration with the secondand third authors. The theorems for content extinction and content persistence werederived by the first author. The article was written in collaboration with the secondauthor.

43

Publications not included in this thesis

• Viet Trinh Doan, Ljubica PajeviÊ, Vaibhav Bajpai and Jörg Ott, Tracingthe path to YouTube – A quantification of path lengths and latencies towardscontent caches, Accepted for publication in IEEE Communications Magazine.

• Jörg Ott, Ljubica PajeviÊ, Ermias Walelgne, Ari Keränen, Esa Hyytiä andJussi Kangasharju, On the sensitivity of geo-based content sharing to locationerrors, in Proc. IEEE Conference on Wireless On-demand Network Systemsand Services (WONS), 2017.

• Sacha Trifunovic, Sylvia Kouyoumdjieva, Bernhard Distl, Ljubica PajeviÊ,Gunnar Karlsson and Bernhard Plattner, A decade of research in opportunis-tic networks: challenges, relevance, and future directions , in IEEE Commu-nications Magazine, 55(1), 2017.

• Ljubica PajeviÊ and Gunnar Karlsson, Characterizing opportunistic commu-nication with churn for crowd-counting, in Proc. IEEE World of WirelessMobile and Multimedia Networks (WoWMoM), Autonomic and Opportunis-tic Communications (AOC) workshop, 2015.

• Ljubica PajeviÊ and Gunnar Karlsson, A zero dimensional mobility modelfor opportunistic networking, in Proc. IEEE World of Wireless Mobile andMultimedia Networks (WoWMoM), Autonomic and Opportunistic Commu-nications (AOC) workshop, 2011.

• Ljubica PajeviÊ, Emre A. Yavuz, Ólafur Helgason, Gunnar Karlsson andSylvia Kouyoumdjieva, Demo: Opportunistic mobile social networking, inProc. 9th International Conference on Mobile systems, Applications, andServices (MobiSys), 2011.

• Ólafur Helgason, Emre Yavuz, Sylvia Kouyoumdjieva, Ljubica PajeviÊ andGunnar Karlsson, A mobile peer-to-peer system for opportunistic content cen-tric networking, in Proc. 2nd ACM SIGCOMM Workshop on Networking,Systems, and Applications on Mobile Handhelds (MobiHeld), 2010.

• Ólafur Helgason, Emre Yavuz, Sylvia Kouyoumdjieva, Ljubica PajeviÊ andGunnar Karlsson, Demonstrating a mobile peer-to-peer system for oppor-tunistic content-centric networking, demonstration at 2nd ACM SIGCOMMMobiHeld Workshop, 2010.

Chapter 6Conclusion and Future Work

The scientific man does not aim at an immediate result.His work is like that of the planter—for the future.His duty is to lay the foundation for those who are tocome, and point the way.

Nikola Tesla

This thesis investigates the feasibility and performance of content distributionin location-aware opportunistic networks. Carried out by means of either math-ematical analysis or simulations, credible performance evaluation of systems in-corporating mobility as their integral part is impossible without comprehensiveunderstanding and realistic representation of human mobility. Theoretical and em-pirical results presented in this thesis help to fill some research gaps in the spacesof mobility models and of analytic tools for modeling spreading processes. In thischapter, we summarize the main conclusions of this thesis and present some openquestions and guidelines for future work.

Mining, Analyzing and Modeling Human Mobility Patterns

We collected and analyzed WLAN mobility traces of users in a wireless universitynetwork. By considering user mobility from the network perspective, we representedtheir aggregate mobility with queuing models. Specifically, we considered aggregatemobility as arrivals and visits to sections of the network, in particular to singleaccess points; this method maps network access patterns to physical mobility andassociates user movements to areas covered by access points. We found that for asignificant number of cases, relatively simple, but time-varying Markovian modelscan be used to describe access patterns and we also identified the reasons, forexample bursty arrivals of users, when more advanced models would be a better fit.

45

46 Chapter 6. Conclusion and Future Work

Having observed the changing access and mobility patterns of the users in thisscenario as compared to previously established results extracted from other net-works and older traces—a phenomenon that can be attributed to the proliferationof mobile users—we posed a question how this trend would change the understand-ing of individual mobility. We then sought to predict movement of users and foundthat the predictability of the users’ next location has decreased, as compared to theresults of the related, older studies. For this prediction, we used simple, yet e�cientfixed-order Markov predictors which are proven to achieve reasonable accuracy incomparison with the baseline theoretical predictability results. We note that ourfindings have implications for a variety of use-cases, especially the emerging futuremobile networking solutions, where mobility of users, in concert with the mobilityof autonomous devices, will be one of the integral parts of the networked systems.

To study opportunistic applications with a restricted location scope, we havedevised an analytic model for individual mobility and for connectivity patterns ofusers in a population with churn (open population). We parametrized the modelwith the data from real-world datasets as well as from synthetic mobility traces,capturing several scenarios in di�erent contexts of microscopic mobility, and val-idated, by means od simulations, that the underlying (synthetic) mobility modelis able to reproduce the main performance metrics in a close accordance to thetrace-based simulations.

Localized Content Distribution: Feasibility and PerformanceEvaluation

We proposed a framework for modeling and evaluating the performance and thefeasibility of localized epidemic content dissemination schemes. At the core ofthe framework is the open population queueing model, for the user arrivals anddepartures from the system, and the assumption of Poissonain contact processesbetween the nodes in the system. These assumptions allowed us to model the spreadof epidemics (content items) with continuous-time SIR Markov processes, which wefurther approximated and treated with stochastic di�erential equations (SDEs). Bydefining SDE models of two scenarios: an entirely distributed scheme of contentsharing, and the scenario of opportunistic networks with infrastructural nodes, weshowed, via simulations, that the framework is flexible enough to capture epidemicprocesses in these structurally di�erent models. The two models however have acommon denominator relating to the users’ contribution to the system operation:in both it is assumed that users are willing to contribute only a limited amountof their resources for the benefit of the entire system. E�ectively, this is achievedby limiting the time period during which they forward the contents, after theyhave obtained it themselves. Based on the SDE model, we sought to engineer thecontent spreading scheme that would ensure that the content persists in ephemerallocation-aware applications, relaying entirely on the presence and mobility of usersin the area. Given the system model and the specifics of the user population in theconsidered geofenced areas, e.g., the density of users, their sojourn times and contact

47

patterns, the only system parameter that can be engineered is the forwarding time.Hence, we approach the system design problem with the objective to estimate theminimum forwarding time. We addressed the posed problem by means of stochasticstability analysis based on Lyapunov theory and we found that the requirements forforwarding time were, in the worst case of very sparse node densities, comparableto user sojourn times. In dense environments, on the other hand, users can berequired to contribute to the system for only a fraction of the time they spend inthe area.

Future Work

It is clear that the information about people’s locations, as well as their macro-scopic movement patterns, has a great potential and impact for transportation,urban planing and more recently, for activity tracking applications and for noveltransportation sharing-economy platforms. In that sense, vast amounts of dataprimarily being collected for commercial use can also be exploited for research pur-poses. An example of this is the Uber Movement [161] initiative, which promises toallow the usage of traces collected from the cities where the company operates. Aswe have already emphasized, in this thesis we focused mainly on pedestrian mobility,not vehicular, but at larger scales the two cannot be entirely decoupled. Vehiculartraces therefore can be exploited for origin-destination modeling, detecting popularlocations, modeling occupancy and people flows and so on. Even more interestingand fundamental question one may ask is how are the new means of transportationsharing, let alone new emerging technologies such as autonomous vehicles, goingto shape the future of human mobility—the question that must have preoccupiedresearchers at various times, from the invention of the telephone, to the availabilityof a�ordable air transportation to the emergence of social networks.

Considering opportunistic device-to-device communication, it yet remains to seehow the big stakeholders—mobile network operators and smartphone manufacturers—are going to respond to the o�cial release of ProSe services in 3GPP Release 12.While waiting for this conundrum to unravel, and in the light of current a�airsregarding exposures of users’ private data including location, we see an umbrella ofopportunistic applications as particularly interesting. Namely, the location-aware,privacy-preserving data sharing applications and social networks. Even when thedata privacy and protection is not of primary concern, some system deploymentscan benefit from opportunistic communication. For instance, for the deploymentof Internet of Things (IoT) solutions [162], current information-centric based ap-proaches assume wireless coverage through out the environment where the deviceswill operate. On the one hand, to support extreme densification of IoT devices,centralized solutions may become too costly or di�cult to plan and maintain, whileon the other hand, such solutions may be also wasteful of networking (bandwidth)resources since most information that is generated by IoT devices is of local interestand likely to stay where it has originated.

48 Chapter 6. Conclusion and Future Work

The modeling framework we proposed can be used as a starting point towardsperformance evaluation of such deployments, however several questions have tobe addressed first. Since we have studied the feasibility of the content distributiondriven by mobile users, further investigations on the mobility of users immersed into“smart" environments may reveal new mobility and contact patterns, which needto be incorporated in the model. Subsequent modeling e�orts may require morecareful analysis of contact patterns and revisiting the assumption of homogeneousmixing, especially since the assumption of the spatial homogeneity may be violated;consequently, this may require adjusting the model to take into account di�erentlevels of mixing. The follow-up question is how the estimation of scenario-specificsystem parameters and the computation of the engineered parameters could beimplemented in real applications. For this we see two possible solutions. Thefirst solution would entail central entities that would monitor the node density,estimate and compute the desired parameters and would inform other nodes aboutthe requirements. The second solution, in a fully distributed scheme, would haveto be implemented by the nodes themselves. To this end, nodes would use localestimates of the system parameters (arrival and contact rate, sojourn time) and,based on the theoretical results that we derived in Paper E, compute the forwardingtime. UrbanCount [163] could be one solution to estimate the size of the system andthe contact rate in a distributed manner. In addition, practical applications willneed to be able to work with the frequently updated estimation of the fluctuatingnode density and the dynamic adaptation to the changing environment. Thus,future work related to this thesis could be about designing the system with theabove functionalities and implementing those in the content sharing applications.

Bibliography

[1] United Nations Human Rights Council, The promotion, protection and en-joyment of human rights on the internet. 27 June 2016. [Online] Available:http://digitallibrary.un.org/record/845728/files/A_HRC_32_L-20

-EN.pdf. [Accessed: July 2018].[2] G. Fodor, D. Erik, G. Mildh, S. Parkvall, N. Reider, G. Miklós, and Z.

Turányi, “Design aspects of network assisted device-to-device communica-tions”, IEEE Communications Magazine, vol. 50, no. 3, pp. 170–177, 2012.

[3] K. Doppler, M. Rinne, C. Wijting, C. B. Ribeiro, and K. Hugl, “Device-to-device communication as an underlay to LTE-advanced networks”, IEEECommunications Magazine, vol. 47, no. 12, pp. 42–49, 2009.

[4] S. Buchegger and A. Datta, “A case for P2P infrastructure for social net-works – opportunities & challenges”, in Proc. IEEE/IFIP Wireless On-Demand Network Systems and Services (WONS), 2009, pp. 161–168.

[5] Project Loon, [Online] Available: http://loon.co. [Accessed: July 2018].[6] Free Basics Platform, [Online] Available: https://developers.facebook.

com/docs/internet-org. [Accessed: July 2018].[7] R. Sen, S. Ahmad, A. Phokeer, Z. A. Farooq, I. A. Qazi, D. Cho�nes, and

K. P. Gummadi, “Inside the walled garden: Deconstructing Facebook’s FreeBasics program”, SIGCOMM Comput. Commun. Rev., vol. 47, no. 5, pp. 12–24, 2017.

[8] F. Bai, N. Sadagopan, and A. Helmy, “IMPORTANT: A framework to sys-tematically analyze the impact of mobility on performance of routing pro-tocols for adhoc networks”, in IEEE INFOCOM, vol. 2, 2003, pp. 825–835.

[9] G. S. Thakur and A. Helmy, “COBRA: A framework for the analysis ofrealistic mobility models”, in Proc. IEEE INFOCOM, 2013, pp. 3351–3356.

49

http://digitallibrary.un.org/record/845728/files/A_HRC_32_L-20-EN.pdf

http://digitallibrary.un.org/record/845728/files/A_HRC_32_L-20-EN.pdf

http://loon.co

https://developers.facebook.com/docs/internet-org

https://developers.facebook.com/docs/internet-org

50 Bibliography

[10] T. Camp, J. Boleng, and V. Davies, “A survey of mobility models for adhoc network research”, Wireless Communications and Mobile Computing(WCMC): Special Issue on Mobile Ad Hoc Networking: Research, trendsand applications, vol. 2, pp. 483–502, 2002.

[11] S. Kurkowski, T. Camp, and M. Colagrosso, “MANET Scalesimulation stud-ies: The Identifyingncredibles”, SIGMOBILE Mobile Computing and Com-munications Review, vol. 9, no. 4, pp. 50–61, 2005.

[12] K. Pearson, “The problem of the random walk”, Nature, vol. 72, p. 294, 27,1905.

[13] T. Hägerstrand, “What about people in regional science?”, Papers in Re-gional Science, vol. 24, no. 1, pp. 6–21, 1, 1970.

[14] S. Hoogendoorn and P. Bovy, “Pedestrian route-choice and activity schedul-ing theory and models”, Transportation Research Part B: Methodological,vol. 38, no. 2, pp. 169–190, 2004.

[15] A. Hanisch, J. Tolujew, K. Richter, and T. Schulze, “Online simulation ofpedestrian flow in public buildings”, in Proc. of the Winter Simulation Con-ference., vol. 2, 2003, pp. 1635–1641.

[16] O. Helgason, S. Kouyoumdjieva, and G. Karlsson, “Does mobility matter?”,in Proc. IEEE/IFIP Conference on Wireless On-demand Network Systemsand Services (WONS), 2010.

[17] C. Tuduce and T. Gross, “A mobility model based on WLAN traces and itsvalidation”, in Proc. INFOCOM, vol. 1, 2005, pp. 664–674.

[18] C. Bettstetter, “Mobility modeling in wireless networks: Categorization,smooth movement, and border e�ects”, SIGMOBILE Mobile Computing andCommunications Review, vol. 5, no. 3, pp. 55–66, 2001.

[19] M. Musolesi and C. Mascolo, “Mobility models for systems evaluation: Asurvey”, in Middleware for Network Eccentric and Mobile Applications, B.Garbinato, H. Miranda, and L. Rodrigues, Eds. Springer Berlin Heidelberg,2009, pp. 43–62.

[20] N. Aschenbruck, A. Munjal, and T. Camp, “Trace-based mobility modelingfor multi-hop wireless networks”, Elsevier Computer Communications, vol.34, no. 6, pp. 704–714, 2011.

[21] D. Karamshuk, C. Boldrini, M. Conti, and A. Passarella, “Human mobilitymodels for opportunistic networks”, IEEE Communications Magazine, vol.49, no. 12, pp. 157–165, 2011.

[22] J. Treurniet, “A taxonomy and survey of microscopic mobility models fromthe mobile networking domain”, ACM Computing Surveys, vol. 47, no. 1,14:1–14:32, 2014.

Bibliography 51

[23] A. Hess, K. A. Hummel, W. N. Gansterer, and G. Haring, “Data-drivenhuman mobility modeling: A survey and engineering guidance for mobilenetworking”, ACM Computing Surveys, vol. 48, no. 3, 38:1–38:39, 2015.

[24] E. Toch, B. Lerner, E. Ben-Zion, and I. Ben-Gal, “Analyzing large-scale hu-man mobility data: A survey of machine learning methods and applications”,Knowledge and Information Systems, pp. 1–23,

[25] F. Ekman, A. Keränen, J. Karvo, and J. Ott, “Working day movementmodel”, in Proc. ACM SIGMOBILE Workshop on Mobility Models, 2008,pp. 33–40.

[26] C. Boldrini and A. Passarella, “HCMM: Modelling spatial and temporalproperties of human mobility driven by users’ social relationships”, ElsevierComputer Communications, vol. 33, no. 9, pp. 1056–1074, 2010.

[27] S. Kouyoumdjieva, E. A. Yavuz, Ó. Helgason, L. Pajevic, and G. Karls-son, “Opportunistic content-centric networking: The conference case demo”,Demonstration at IEEE INFOCOM, 2011.

[28] V. Arnaboldi, M. Conti, and F. Delmastro, “CAMEO: A novel context-awaremiddleware for opportunistic mobile social networks”, Elsevier Pervasiveand Mobile Computing, vol. 11, pp. 148–167, 2014.

[29] M. Pitkänen, T. Kärkkäinen, J. Ott, M. Conti, A. Passarella, S. Giordano,D. Puccinelli, F. Legendre, S. Trifunovic, K. Hummel, M. May, N. Hegde,and T. Spyropoulos, “SCAMPI: Service Platform for Social Aware Mobileand Pervasive Computing”, in Proceedings of the SIGCOMM Workshop onMobile Cloud Computing (MCC), Helsinki, Finland, 2012, pp. 7–12.

[30] L. Pajevic and G. Karlsson, “Modeling opportunistic communication withchurn”, Elsevier Computer Communications, vol. 96, pp. 123–135, 2016.

[31] T. Kärkkäinen, P. Houghton, L. Valerio, A. Passarella, and J. Ott, “Here &Now: Data-centric local social interactions through opportunistic networks:Demo”, in Proc. of ACM Workshop on Challenged Networks (CHANTS) atMobiCom 2016, 2016.

[32] E. G. Ravenstein, “The laws of migration”, Journal of the Royal StatisticalSociety, vol. 52, no. 2, pp. 241–305, 1885.

[33] S. Isaacman, R. Becker, R. Cáceres, M. Martonosi, J. Rowland, A. Var-shavsky, and W. Willinger, “Human mobility modeling at metropolitanscales”, in Proc. of the ACM International Conference on Mobile Systems,Applications, and Services (MobiSys), New York, NY, USA: ACM, 2012,pp. 239–252.

[34] F. Simini, M. C. González, A. Maritan, and A.-L. Barabási, “A universalmodel for mobility and migration patterns”, Nature, vol. 484, pp. 96–100,26, 2012.

52 Bibliography

[35] Q. Zheng, X. Hong, and J. Liu, “An agenda based mobility model”, in 39thAnnual Simulation Symposium (ANSS’06), 2006.

[36] J. Kim, V. Sridhara, and S. Bohacek, “Realistic mobility simulation of urbanmesh networks”, Elsevier Ad Hoc Networks, vol. 7, no. 2, pp. 411–430, 2009.

[37] X. Liang, J. Zhao, L. Dong, and K. Xu, “Unraveling the origin of exponentiallaw in intra-urban human mobility”, Scientific Reports, vol. 3, 2013.

[38] V. D. Blondel, A. Decuyper, and G. Krings, “A survey of results on mobilephone datasets analysis”, EPJ Data Science, vol. 4, no. 1, p. 10, 5, 2015.

[39] Google location services, [Online] Available: https://developer.android.

com/training/location/index.html. [Accessed: July 2018].[40] N. Kiukkonen, B. J., O. Dousse, D. Gatica-Perez, and J. K. Laurila, “To-

wards rich mobile phone datasets: Lausanne data collection campaign”, inProc. ACM Int. Conf. on Pervasive Services, 2010.

[41] J. Laurila, D. Gatica-Perez, I. Aad, J. Blom, O. Bornet, T.-M.-T. Do, O.Dousse, J. Eberle, and M. Miettinen, “The mobile data challenge: Big datafor mobile computing research”, in Proceedings of the Workshop on the NokiaMobile Data Challenge, 2012, pp. 1–8.

[42] Y. Zheng, X. Xie, and W.-Y. Ma, “Geolife: A collaborative social networkingservice among user, location and trajectory”, IEEE Data(base) EngineeringBulletin, 2010.

[43] A. Stopczynski, V. Sekara, P. Sapiezynski, A. Cuttone, M. M. Madsen, J. E.Larsen, and S. Lehmann, “Measuring large-scale social networks with highresolution”, PLOS ONE, vol. 9, no. 4, pp. 1–24, 2014.

[44] I. Rhee, M. Shin, S. Hong, K. Lee, S. J. Kim, and S. Chong, “On the levy-walk nature of human mobility”, IEEE/ACM Trans. Netw., vol. 19, no. 3,pp. 630–643, 2011.

[45] V. Vukadinovic and S. Mangold, “Opportunistic wireless communication intheme parks: A study of visitors mobility”, in Proceedings of the 6th ACMWorkshop on Challenged Networks (CHANTS), 2011, pp. 3–8.

[46] J. Cranshaw, E. Toch, J. Hong, A. Kittur, and N. Sadeh, “Bridging thegap between physical location and online social networks”, in Proceedingsof the 12th ACM International Conference on Ubiquitous Computing, 2010,pp. 119–128.

[47] J. P. Bagrow, D. Wang, and A.-L. Barabási, “Collective response of humanpopulations to large-scale emergencies”, PLOS ONE, vol. 6, no. 3, pp. 1–8,2011.

[48] X. Lu, E. Wetter, N. Bharti, A. J. Tatem, and L. Bengtsson, “Approachingthe limit of predictability in human mobility”, Scientific Reports, vol. 3,2013.

https://developer.android.com/training/location/index.html

https://developer.android.com/training/location/index.html

Bibliography 53

[49] J. Scott, R. Gass, J. Crowcroft, P. Hui, C. Diot, and A. Chaintreau, CRAW-DAD trace cambridge/haggle/imote/infocom2006 (v. 2009-05-29), 2009.

[50] N. Eagle and A. Pentland, “Reality mining: Sensing complex social systems”,Personal and Ubiquitous Computing, vol. 10, no. 4, pp. 255–268, 1, 2006.

[51] L. Sun, K. W. Axhausen, D.-H. Lee, and X. Huang, “Understanding metropoli-tan patterns of daily encounters”, Proceedings of the National Academy ofSciences, vol. 110, no. 34, pp. 13 774–13 779, 2013.

[52] E. Yoneki, P. Hui, and J. Crowcroft, “Bio-inspired computing and communi-cation”, in, Berlin, Heidelberg: Springer-Verlag, 2008, ch. Wireless EpidemicSpread in Dynamic Human Networks, pp. 116–132.

[53] A.-K. Pietiläinen, E. Oliver, J. LeBrun, G. Varghese, and C. Diot, “Mo-biClique: Middleware for Mobile Social Networking”, in Proceedings of the2Nd ACM Workshop on Online Social Networks (WOSN), 2009, pp. 49–54.

[54] A.-K. Pietiläinen and C. Diot, “Dissemination in opportunistic social net-works: The role of temporal communities”, in Proceedings of the ThirteenthACM International Symposium on Mobile Ad Hoc Networking and Comput-ing (MobiHoc), 2012, pp. 165–174.

[55] C. Petro, CRAWDAD trace hope/amd (v. 2008-08-07), Downloaded fromhttp://crawdad.org/hope/amd/20080807, 2008.

[56] A. Panisson, L. Gauvin, A. Barrat, and C. Cattuto, “Fingerprinting tem-poral networks of close-range human proximity”, in Proc. IEEE Interna-tional Conference on Pervasive Computing and Communications Workshops(PERCOM Workshops), 2013, pp. 261–266.

[57] L. Rossi and M. Musolesi, “It’s the way you check-in: Identifying users inlocation-based social networks”, in Proceedings of the Second ACM Confer-ence on Online Social Networks (COSN), 2014, pp. 215–226.

[58] J.-P. Onnela, J. Saramäki, J. Hyvönen, G. Szabó, D. Lazer, K. Kaski, J.Kertész, and A.-L. Barabási, “Structure and tie strengths in mobile commu-nication networks”, Proceedings of the National Academy of Sciences, vol.104, no. 18, pp. 7332–7336, 2007.

[59] M. González, C. Hidalgo, and A.-L. Barabási, “Understanding individualhuman mobility patterns”, 7196, vol. 453, 2008, pp. 779–782.

[60] I. Barnett, T. Khanna, and J.-P. Onnela, “Social and spatial clustering ofpeople at humanity’s largest gathering”, PLOS ONE, vol. 11, no. 6, pp. 1–12, 2016.

[61] F. Calabrese, L. Ferrari, and V. D. Blondel, “Urban sensing using mobilephone network data: A survey of research”, ACM Computing Surveys, vol.47, no. 2, 25:1–25:20, 2014.

http://crawdad.org/hope/amd/20080807

54 Bibliography

[62] G. Chen, S. Hoteit, A. C. Viana, M. Fiore, and C. Sarraute, “Enrichingsparse mobility information in call detail records”, Computer Communica-tions, vol. 122, pp. 44–58, 2018.

[63] CRAWDAD: A community resource for archiving wireless data at dart-mouth. [Online] Available: http://crawdad.cs.dartmouth.edu. [Accessed:July 2018]. 2013.

[64] G. Barlacchi, M. De Nadai, R. Larcher, A. Casella, C. Chitic, G. Torrisi, F.Antonelli, A. Vespignani, A. Pentland, and B. Lepri, “A multi-source datasetof urban life in the city of Milan and the Province of Trentino”, ScientificData, vol. 2, 2015.

[65] Y. de Montjoye, Z. Smoreda, R. Trinquart, C. Ziemlicki, and V. D. Blondel,“D4D-Senegal: The second mobile phone data for development challenge”,CoRR, 2014. arXiv: 1407.4885.

[66] V. D. Blondel, M. Esch, C. Chan, F. Clérot, P. Deville, E. Huens, F. Morlot,Z. Smoreda, and C. Ziemlicki, “Data for development: The D4D challengeon mobile phone data”, CoRR, 2012. arXiv: 1210.0137.

[67] Y.-A. de Montjoye, C. A. Hidalgo, M. Verleysen, and V. D. Blondel, “Uniquein the crowd: The privacy bounds of human mobility”, Scientific Reports,vol. 3, 2013.

[68] Y.-A. de Montjoye, L. Radaelli, V. K. Singh, and A. ı. Pentland, “Unique inthe shopping mall: On the reidentifiability of credit card metadata”, Science,vol. 347, no. 6221, pp. 536–539, 2015.

[69] B. Gedik and L. Liu, “Protecting location privacy with personalized k-anonymity: Architecture and algorithms”, IEEE Transactions on MobileComputing, vol. 7, no. 1, pp. 1–18, 2008.

[70] P. Wightman, W. Coronell, D. Jabba, M. Jimeno, and M. Labrador, “Eval-uation of location obfuscation techniques for privacy in location based infor-mation systems”, in 2011 IEEE Third Latin-American Conference on Com-munications, 2011, pp. 1–6.

[71] D. Mir, S. Isaacman, R. Caceres, M. Martonosi, and R. Wright, “DP-WHERE:Di�erentially private modeling of human mobility”, in 2013 IEEE Interna-tional Conference on Big Data, 2013, pp. 580–588.

[72] Eduroam, [Online] Available: https://www.eduroam.org.[73] TCPDUMP, [Online] Available: http://www.tcpdump.org.[74] D. Tang and M. Baker, “Analysis of a local-area wireless network”, in Proc.

of the 6th ACM Annual International Conference on Mobile Computing andNetworking (MobiCom), 2000.

http://crawdad.cs.dartmouth.edu

http://arxiv.org/abs/1407.4885

http://arxiv.org/abs/1210.0137

https://www.eduroam.org

http://www.tcpdump.org

Bibliography 55

[75] A. Balachandran, G. M. Voelker, P. Bahl, and P. V. Rangan, “Characterizinguser behavior and network performance in a public wireless LAN”, in Proc.of the ACM SIGMETRICS International Conference on Measurement andModeling of Computer Systems, 2002, pp. 195–205.

[76] F. Chinchilla, M. Lindsey, and M. Papadopouli, “Analysis of wireless in-formation locality and association patterns in a campus”, in Proc. IEEEINFOCOM, vol. 2, 2004.

[77] D. Kotz and K. Essien, “Analysis of a campus-wide wireless network”, Wire-less Networks, vol. 11, no. 1-2, 2005.

[78] T. Henderson, D. Kotz, and I. Abyzov, “The changing usage of a maturecampus-wide wireless network”, Elsevier Computer Networks, vol. 52, no.14, 2008.

[79] M. McNett and G. M. Voelker, “Access and mobility of wireless pda users”,SIGMOBILE Mobile Computing and Communications Review, vol. 9, no. 2,pp. 40–55, 2005.

[80] X. Chen, R. Jin, K. Suh, B. Wang, and W. Wei, “Network performance ofsmart mobile handhelds in a university campus WiFi network”, in Proceed-ings of the ACM Internet Measurement Conference (IMC), 2012, pp. 315–328.

[81] A. Allahdadi, R. Morla, A. Aguiar, and J. S. Cardoso, “Predicting short802.11 sessions from RADIUS usage data”, in IEEE Conference on LocalComputer Networks – Workshops, 2013, pp. 1–8.

[82] M. Zhou, K. Sui, M. Ma, Y. Zhao, D. Pei, and T. Moscibroda, “Mobicamp: Acampus-wide testbed for studying mobile physical activities”, in Proceedingsof the 3rd ACM International on Workshop on Physical Analytics (WPA),2016, pp. 1–6.

[83] M. Afanasyev, T. Chen, G. M. Voelker, and A. C. Snoeren, “Analysis of amixed-use urban WiFi network: When metropolitan becomes neapolitan”,in Proc. 8th ACM SIGCOMM Conference on Internet Measurements (IMC),2008.

[84] ——, “Usage patterns in an urban WiFi network”, IEEE/ACM Transactionson Networking, vol. 18, no. 5, 2010.

[85] M. Lenczner and A. G. Hoen, CRAWDAD dataset ilesansfil/wifidog (v. 2015-11-06), Downloaded from https://crawdad.org/ilesansfil/wifidog/20

151106, 2015.[86] S. Isaacman and M. Martonosi, “Low-infrastructure methods to improve

internet access for mobile users in emerging regions”, in Proceedings of the20th International Conference Companion on World Wide Web (WWW),2011, pp. 473–482.

https://crawdad.org/ilesansfil/wifidog/20151106

https://crawdad.org/ilesansfil/wifidog/20151106

56 Bibliography

[87] W. Bao and B. Liang, “Insensitivity of user distribution in multicell net-works under general mobility and session patterns”, IEEE Transactions onWireless Communications, vol. 12, no. 12, pp. 6244–6254, 2013.

[88] S. Scellato, M. Musolesi, C. Mascolo, V. Latora, and A. T. Campbell, “Nextplace:A spatio-temporal prediction framework for pervasive systems”, in PervasiveComputing, Springer Berlin Heidelberg, 2011, pp. 152–169.

[89] A. Ghosh, R. Jana, V. Ramaswami, J. Rowland, and N. K. Shankara-narayanan, “Modeling and characterization of large-scale Wi-Fi tra�c inpublic hot-spots”, in Proc. IEEE INFOCOM, 2011, pp. 2921–2929.

[90] LEGION Studio, [Online] Available: http://www.legion.com/legion-

studio. [Accessed: June 2018].[91] VADERE simulator, [Online] Available: http : / / www . vadere . org. [Ac-

cessed: July 2018]. 2017.[92] A. Keränen, J. Ott, and T. Kärkkäinen, “The ONE Simulator for DTN

protocol evaluation”, in SIMUTools ’09: Proceedings of the 2nd InternationalConference on Simulation Tools and Techniques, 2009.

[93] R. E. Kahn, S. A. Gronemeyer, J. Burchfiel, and R. C. Kunzelman, “Ad-vances in packet radio technology”, Proceedings of the IEEE Special Issueon Packet Communication Networks, vol. 66, no. 11, pp. 1468–1496, 1978.

[94] J. Westcott and G. Lauer, “Hierarchical routing for very large networks”, inProc. IEEE Military Communications Conference (MILCOM), vol. 2, 1984,pp. 214–218.

[95] D. A. Hall, “Tactical Internet system architecture for Task Force XXI”,in Proceedings of the 1996 Tactical Communications Conference. EnsuringJoint Force Superiority in the Information Age, 1996, pp. 219–230.

[96] V. Cerf, S. Burleigh, A. Hooke, L. Torgerson, R. Durst, K. Scott, E. Travis,and H. Weiss, Interplanetary internet (IPN): Architectural definition, [On-line] Available: www.tools.ietf.org/html/draft-irtf-ipnrg-arch-00,2000.

[97] K. Fall, “A delay-tolerant network architecture for challenged internets”,in Proceedings of the ACM Conference on Applications, technologies, archi-tectures, and protocols for computer communications (SIGCOMM), 2003,pp. 27–34.

[98] M. Grossglauser and D. N. C. Tse, “Mobility increases the capacity of adhoc wireless networks”, IEEE/ACM Transactions on Networking, vol. 10,no. 4, pp. 477–486, 2002.

[99] A. Vahdat and D. Becker, “Epidemic routing for partially-connected ad hocnetworks”, Department of Computer Science, Duke University, Tech. Rep.,2000.

http://www.legion.com/legion-studio

http://www.legion.com/legion-studio

http://www.vadere.org

www.tools.ietf.org/html/draft-irtf-ipnrg-arch-00

Bibliography 57

[100] A. Lindgren, A. Doria, and O. Schelén, “Probabilistic routing in inter-mittently connected networks”, ACM SIGMOBILE Mobile Computing andCommunications Review, vol. 7, no. 3, pp. 19–20, 2003.

[101] T. Spyropoulos, K. Psounis, and C. S. Raghavendra, “Spray and wait: Ane�cient routing scheme for intermittently connected mobile networks”, inProceedings of the ACM SIGCOMM Workshop on Delay-tolerant Networking(WDTN), 2005, pp. 252–259.

[102] E. M. Daly and M. Haahr, “Social network analysis for routing in discon-nected delay-tolerant manets”, in Proceedings of the 8th ACM InternationalSymposium on Mobile Ad Hoc Networking and Computing (MobiHoc), 2007,pp. 32–40.

[103] P. Hui, J. Crowcroft, and E. Yoneki, “Bubble rap: Social-based forwardingin delay-tolerant networks”, IEEE Transactions on Mobile Computing, vol.10, no. 11, pp. 1576–1589, 2011.

[104] Y. Cao and Z. Sun, “Routing in delay/disruption tolerant networks: A tax-onomy, survey and challenges”, IEEE Communications Surveys Tutorials,vol. 15, no. 2, pp. 654–677, 2013.

[105] G. Karlsson, V. Lenders, and M. May, “Delay-tolerant broadcasting”, inProc. of the ACM SIGCOMM Workshop on Challenged Networks (CHANTS),2006, pp. 197–204.

[106] V. Lenders, G. Karlsson, and M. May, “Wireless ad hoc podcasting”, inProc. IEEE Communications Society Conference on Sensor, Mesh and AdHoc Communications and Networks, 2007, pp. 273–283.

[107] PodNet Project: Mobile Distribution of User-generated Content, [Online]Available: http://www.podnet.ethz.ch.

[108] Ó. Helgason, E. Yavuz, S. Kouyoumdjieva, L. Pajevic, and G. Karlsson, “Amobile peer-to-peer system for opportunistic content-centric networking”, inProc. ACM SIGCOMM MobiHeld workshop, New Delhi, India, 2010.

[109] Ó. Helgason, S. Kouyoumdjieva, L. Pajevic, E. A. Yavuz, and G. Karlsson,“A middleware for opportunistic content distribution”, Elsevier ComputerNetworks, vol. 107, pp. 178–193, 2016.

[110] Z. Chen, E. A. Yavuz, and G. Karlsson, “What a juke! a collaborative mu-sic sharing system”, in 2012 IEEE International Symposium on a World ofWireless, Mobile and Multimedia Networks (WoWMoM), 2012, pp. 1–6.

[111] I. Marfisi-Schottman, G. Karlsson, and J. Celander Guss, “Opphos - a partic-ipative light and sound show using mobile phones in crowds”, in ExtremComInternational Conference, Thórsmörk, Iceland, 2013, pp. 47–48.

[112] C. Boldrini, M. Conti, and A. Passarella, “ContentPlace: social-aware datadissemination in opportunistic networks”, in Proceedings of the ACM Inter-national Symposium on Modeling, Analysis and Simulation of Wireless andMobile Systems (MSWiM), 2008, pp. 203–210.

http://www.podnet.ethz.ch

58 Bibliography

[113] T. Spyropoulos, K. Psounis, and C. S. Raghavendra, “E�cient routing in in-termittently connected mobile networks: The single-copy case”, IEEE/ACMTransactions on Networking, vol. 16, no. 1, pp. 63–76, 2008.

[114] ——, “E�cient routing in intermittently connected mobile networks: Themultiple-copy case”, IEEE/ACM Transactions on Networking, vol. 16, no.1, pp. 77–90, 2008.

[115] S. Zakhary, M. Radenkovic, and A. Benslimane, “E�cient location privacy-aware forwarding in opportunistic mobile networks”, IEEE Transactions onVehicular Technology, vol. 63, no. 2, pp. 893–906, Feb. 2014.

[116] S. Gisdakis, T. Giannetsos, and P. Papadimitratos, “SHIELD: A Data Ver-ification Framework for Participatory Sensing Systems”, in Proceedings ofthe 8th ACM Conference on Security & Privacy in Wireless and MobileNetworks (WiSec), 2015, 16:1–16:12.

[117] A. Wegener, H. Hellbruck, S. Fischer, C. Schmidt, and S. Fekete, “Autocast:An adaptive data dissemination protocol for tra�c information systems”, in2007 IEEE 66th Vehicular Technology Conference, 2007, pp. 1947–1951.

[118] J. Ott, L. Kärkkäinen, E. A. Walelgne, A. Keränen, E. Hyytiä, and J. Kan-gasharju, “On the sensitivity of geo-based content sharing to location errors”,in Proc. IEEE/IFIP Conference on Wireless On-demand Network Systemsand Services (WONS), 2017, pp. 25–32.

[119] N. Thompson, R. Crepaldi, and R. Kravets, “Locus: A location-based dataoverlay for disruption-tolerant networks”, in Proceedings of the ACM Work-shop on Challenged Networks (CHANTS), 2010, pp. 47–54.

[120] B. Liu, B. Khorashadi, D. Ghosal, C.-N. Chuah, and M. H. Zhang, “Assess-ing the VANET’s local information storage capability under di�erent tra�cmobility”, in Proc. IEEE INFOCOM, 2010, pp. 1–5.

[121] I. Leontiadis and C. Mascolo, “Opportunistic spatio-temporal dissemina-tion system for vehicular networks”, in Proceedings of the 1st InternationalMobiSys Workshop on Mobile Opportunistic Networking (MobiOpp), 2007,pp. 39–46.

[122] A. A. Villalba Castro, G. D. M. Serugendo, and D. Konstantas, “Hover-ing information – self-organizing information that finds its own storage”, inAutonomic Communication, Springer, 2009, pp. 111–145.

[123] J. Ott, E. Hyytiä, P. Lassila, J. Kangasharju, and S. Santra, “Floating con-tent for probabilistic information sharing”, Pervasive and Mobile Computing,vol. 7, no. 6, pp. 671–689, 2011.

[124] E. Hyytiä, J. Virtamo, P. Lassila, J. Kangasharju, and J. Ott, “When doescontent float? characterizing availability of anchored information in oppor-tunistic content sharing”, in Proc. IEEE INFOCOM, 2011, pp. 3137–3145.

Bibliography 59

[125] E. Hyytiä, P. Lassila, J. Ott, and J. Kangasharju, “Floating informationwith stationary nodes”, in Eighth Workshop on Spatial Stochastic Modelsfor Wireless Networks (SpaSWin), Paderborn, Germany, 2012.

[126] S. Ali, G. Rizzo, V. Mancuso, and M. A. Marsan, “Persistence and availabil-ity of floating content in a campus environment”, in Proc. IEEE INFOCOM,2015, pp. 2326–2334.

[127] L. Pajevic and G. Karlsson, “Characterizing opportunistic communicationwith churn for crowd-counting”, in Proc. IEEE WoWMoM workshop on Au-tonomic and Opportunistic Computing (AOC), 2015, pp. 1–6.

[128] E. Hyytiä and J. Ott, “Criticality of large delay tolerant networks via di-rected continuum percolation in space-time”, in Proc. IEEE INFOCOM,2013, pp. 320–324.

[129] J. Su, J. Scott, P. Hui, J. Crowcroft, E. De Lara, C. Diot, A. Goel, M. H.Lim, and E. Upton, “Haggle: Seamless networking for mobile applications”,UbiComp ’07, pp. 391–408, 2007.

[130] T. Kärkkäinen, M. Pitkänen, P. Houghton, and J. Ott, “SCAMPI applicationplatform”, in Proc. ACM International Workshop on Challenged Networks(CHANTS), New York, NY, USA, 2012, pp. 83–86.

[131] Delay Tolerant Network Research Group (DTNRG), http://irtf.org/

concluded/dtnrgg.[132] T. Kärkkäinen and J. Ott, “Liberouter: Towards autonomous neighborhood

networking”, in IEEE/IFIP Conference on Wireless On-demand NetworkSystems and Services (WONS), 2014, pp. 162–169.

[133] M. Nagy, T. Bui, S. Udar, N. Asokan, and J. Ott, “Spotshare and near-bypeople: Applications of the social pal framework”, in Proceedings of the8th ACM Conference on Security & Privacy in Wireless and Mobile Net-works (WiSec), 2015, 31:1–31:2.

[134] F. Delmastro, V. Arnaboldi, and M. Conti, “People-centric computing andcommunications in smart cities”, IEEE Communications Magazine, vol. 54,no. 7, pp. 122–128, 2016.

[135] O. Garden, Firechat, [Online] Available: https://www.opengarden.com/

firechat.[136] T. Hossmann, F. Legendre, P. Carta, P. Gunningberg, and C. Rohner, “Twit-

ter in disaster mode: Opportunistic communication and distribution of sen-sor data in emergencies”, in Proceedings of the 3rd Extreme Conference onCommunication (ExtremeCom): The Amazon Expedition, 2011, pp. 1–6.

[137] S. Grasic and A. Lindgren, “Revisiting a remote village scenario and its dtnrouting objective”, Elsevier Computer Communications, vol. 48, pp. 133–140, 2014.

http://irtf.org/concluded/dtnrgg

http://irtf.org/concluded/dtnrgg

https://www.opengarden.com/firechat

https://www.opengarden.com/firechat

60 Bibliography

[138] A. Pentland, R. Fletcher, and A. Hasson, “Daknet: Rethinking connectivityin developing nations”, IEEE Computer, vol. 37, no. 1, pp. 78–83, 2004.

[139] Z. J. Haas and T. Small, “A new networking model for biological applicationsof ad hoc sensor networks”, IEEE/ACM Transactions on Networking, vol.14, no. 1, pp. 27–40, 2006.

[140] P. Juang, H. Oki, Y. Wang, M. Martonosi, L. S. Peh, and D. Rubenstein,“Energy-e�cient computing for wildlife tracking: Design tradeo�s and earlyexperiences with zebranet”, SIGARCH Comput. Archit. News, vol. 30, no.5, pp. 96–107, 2002.

[141] G. Werner-Allen, K. Lorincz, M. Ruiz, O. Marcillo, J. Johnson, J. Lees, andM. Welsh, “Deploying a wireless sensor network on an active volcano”, IEEEInternet Computing, vol. 10, no. 2, pp. 18–25, 2006.

[142] T. Spyropoulos, T. Turletti, and K. Obraczka, “Routing in delay-tolerantnetworks comprising heterogeneous node populations”, IEEE Trans. on Mo-bile Computing, vol. 8, no. 8, pp. 1132–1147, 2009.

[143] A. Picu, T. Spyropoulos, and T. Hossmann, “An analysis of the informationspreading delay in heterogeneous mobility DTNs”, in Proc. IEEE WoW-MoM, 2012, pp. 1–10.

[144] Y. Kim, K. Lee, N. B. Shro�, and I. Rhee, “Providing probabilistic guaran-tees on the time of information spread in opportunistic networks”, in Proc.IEEE INFOCOM, 2013, pp. 2067–2075.

[145] C. Boldrini, M. Conti, and A. Passarella, “Performance modelling of oppor-tunistic forwarding under heterogenous mobility”, Elsevier Computer Com-munications, vol. 48, pp. 56–70, 2014.

[146] A. Chaintreau, J.-Y. Le Boudec, and N. Ristanovic, “The age of gossip: Spa-tial mean field regime”, in Proceedings of the Eleventh International JointConference on Measurement and Modeling of Computer Systems (SIGMET-RICS), 2009, pp. 109–120.

[147] T. Kärkkäinen and J. Ott, “Shared content editing in opportunistic net-works”, in Proceedings of the 9th ACM MobiCom Workshop on ChallengedNetworks (CHANTS), 2014, pp. 61–64.

[148] V. Sciancalepore, D. Giustiniano, A. Banchs, and A. Hossmann-Picu, “Of-floading cellular tra�c through opportunistic communications: Analysis andoptimization”, IEEE Journal on Selected Areas in Communications, vol. 34,no. 1, pp. 122–137, 2016.

[149] R. M. Anderson and R. M. May, Infectious Diseases of Humans: Dynamicsand Control, ser. Dynamics and Control. Oxford University Press, Oxford,1992.

[150] M. J. Keeling and P. Rohani, Modeling infectious diseases in humans andanimals. Princeton University Press, 2011.

Bibliography 61

[151] F. Ball, D. Mollison, and G. Scalia-Tomba, “Epidemics with two levels ofmixing”, The Annals of Applied Probability, vol. 7, no. 1, pp. 46–89, 1997.

[152] N. G. van Kampen, Stochastic Processes in Physics and Chemistry (ThirdEdition), First Edition, ser. North-Holland Personal Library. Amsterdam,1981.

[153] D. T. Gillespie, “The chemical Langevin equation”, The Journal of ChemicalPhysics, vol. 113, no. 1, pp. 297–306, 2000.

[154] M. Newman, Networks: An Introduction. New York, NY, USA.: Oxford Uni-versity Press Inc., 2010.

[155] R. Pastor-Satorras and A. Vespignani, “Epidemic dynamics and endemicstates in complex networks”, Physical Review E, vol. 63, no. 6, 2001.

[156] P. Van Mieghem, J. Omic, and R. Kooij, “Virus spread in networks”, IEEE/ACM Transactions on Networking, vol. 17, no. 1, pp. 1–14, 2009.

[157] M. Boguñá and R. Pastor-Satorras, “Epidemic spreading in correlated com-plex networks”, Physical Review E, vol. 66, no. 4, 2002.

[158] H. S. Wilf, Generatingfunctionology. Natick, MA, USA: A. K. Peters, Ltd.,2006.

[159] E. Volz and L. A. Meyers, “Epidemic thresholds in dynamic contact net-works”, Journal of The Royal Society Interface, vol. 6, no. 32, pp. 233–241,2009.

[160] O. Michail and P. G. Spirakis, “Elements of the theory of dynamic networks”,Communications of the ACM, vol. 61, no. 2, pp. 72–81, 2018.

[161] Uber movement, [Online] Available: https://movement.uber.com.[162] E. Borgia, R. Bruno, and A. Passarella, “Making opportunistic networks in

IoT environments CCN-ready: A performance evaluation of the MobCCNprotocol ”, Elsevier Computer Communications, vol. 123, pp. 81–96, 2018.

[163] P. Danielis, S. Kouyoumdjieva, and G. Karlsson, “UrbanCount: Mobile crowdcounting in urban environments”, in Proceedings IEEE Annual InformationTechnology, Electronics and Mobile Communication Conference (IEMCON),2017, pp. 640–648.

https://movement.uber.com

Documents

Performance Analysis of Opportunistic Content Distribution via …1256074/... · 2018. 10. 29. · Introduction Where should I start? Start from the statement of the problem. György