Application platform, routing protocols and behavior models in mobile disruption-tolerant networks (DTNs)

1
Application platform, routing protocols and
behavior models in mobile disruption-tolerant
networks (DTNs)

• Doctoral thesis defense
• Arezu Moghadam
• 13 May 2011

2
Introduction
D
D
3
Internet WiFi or 3G
Communication in mobile DTNs 1 No knowledge
of the routes beyond the immediate
hop 2 Mobility 3 Opportunistic
D2
?
?
D1
DTN Disruption-Tolerant Networks
4
Introduction

• Applications of mobile DTNs
• Covering regions with no infrastructure, e.g.
  natural disasters
• Retrieving data from remote sensor networks
• Sharing music, news, pictures in the subway or
  networks of pedestrians
• Collaborative ad-hoc environments
• Challenges of mobile DTNs
• Networking and connectivity
• No application server or end-to-end communication
  path
• Different routing requirements and models
• Performance of the applications and routing
  algorithms relies on the mobility behavior of
  mobile users

5
Problem scope
Mobile DTNs
Mobility
Routing
Application
A modular app. platform
Popularity-based and interest-aware
communication models
Markov-based mobility model and routing algorithm
6
Problem scope
Mobile DTNs
Applications
Mobility
Routing
Class of disruption- tolerant
Core functional requirements
A modular App platform
7
Motivation
8
Problem
Internet 3G
?
9
Solution

• 7DS platform
• Provides a class of disruption-tolerant
  applications
• Store-carry-forward communication
• Node and service discovery
• Web, email, file-synchronization and
  bulletin-board
• Modular platform for
• application developers

Internet
Suman Srinivasan, Arezu Moghadam, Se Gi Hong,
Henning G Schulzrinne, "7DS - Node Cooperation
and Information Exchange in Mostly Disconnected
Networks", IEEE International Conference on
Communications (ICC), Jun 2007.
10
Email exchange

• Mobile nodes act as mail transport agents (MTA)
• Email client configuration
• SMTP server is set to the 7DS local MTA in the
  email client
• Database
• TTL, relays identities to avoid loops.

11
File synchronization
Pull-based automatic download
12
Bulletin board system

• Push-based data sharing
• Data exchange should be approved by the user
• Metadata in an XML format

Users can generate and share content in the
spirit of Web 2.0
2. Users can search for and read bulletin board
announcements.
1. User publishes announcements on the bulletin
board.
13
Fetches the locally cached web pages.
Emulates a connected communication path in the
absence of Internet
Search the internal cache
Query the local neighbors
BonAHA A thin wrapper around Apples Bonjour
gt rsync
1 - Arezu Moghadam, Suman Srinivasan, Henning
Schulzrinne, "7DS - A Modular Platform to Develop
Mobile Disruption-tolerant Applications", Second
IEEE Conference and Exhibition on Next Generation
Mobile Applications, Services, and Technologies
(NGMAST 2008), Sep 2008. 2 - Suman Srinivasan,
Arezu Moghadam, Henning Schulzrinne, "BonAHA
Service Discovery Framework for Mobile Ad-Hoc
Applications", IEEE Consumer Communications
Networking Conference 2009 (CCNC'09), Jan 2009.
14
Problem scope
Mobile DTNs
Applications
Mobility
Routing
A modular app. platform
Popularity-based and interest-aware
communication models
Markov-based mobility model and routing algorithm
15
Problem scope
Mobile DTNs
Applications
Mobility
Routing
Lack of group communication model
Popularity-based Interest-aware model
16
Routing Problem

• Store-carry-forward
• Storage constraints
• Routing objectives
• Minimize delay
• Maximize throughput
• Per-hop routing vs. source routing
• No end-to-end path
• MANETs routing protocols fail
• Proactive and reactive
• No knowledge of the topology
• Time varying connectivity graph
• Unicast vs. Multicast

Each edge is a contact meaning an opportunity to
transfer data.
u
v
S
w
D
x
gt Routing Models
17
Problem lack of group communication model for
mobile DTNs?

• Any cast communication model
• Emergencies
• Traffic congestion notifications
• Severe weather alerts
• Traditional multicast as a group communication
  model ? Fails!
• No knowledge of the topology
• No infrastructure to track group memberships
• Communication with communities of interest ? Even
  a harder problem!
• Market news, sport events
• Scientific articles
• Advertisement about particular products

Epidemic routing
18
Solution interest-aware communication model

• Our one-to-many communication model with
  communities of users
• Objective transmitting data to users who are
  interested in the content
• Assumptions
• No previous knowledge about the location of the
  recipients
• No knowledge about the mobility behavior of users
• No previous knowledge about interests of users
• Uniform probability of encounter

d
a
X
X
D
e
b
Y
S
Y
f
X
c
X
g
Y
wireless contact data transfer
Arezu Moghadam, Henning Schulzrinne,
"Interest-aware content distribution protocol for
mobile disruption-tolerant networks", 10th IEEE
International Symposium on a World of Wireless,
Mobile and Multimedia Networks, Kos, Greece, Jun
2009.
19
Interest Vector
Monitoring Behavior
Music
Reviewed webpages
Downloaded documents
Restaurant reviews

• User profiling for the Web
• Profiles users based on their downloaded or
  reviewed web content, clicked hyperlinks and
• Music
• The genre of the music user is playing more often
• Topic and category of the documents user has
  downloaded

20
Solution interest-aware communication model
21
LSA

• User profiling for the Web
• Profiles users based on their downloaded or
  reviewed web content, clicked hyperlinks and
• Latent Semantic Analysis
• A low-dimensional topic-based representation of
  web documents is obtained
• Then low-dimensional representations are
  clustered to semantic groups

gt Web recommender
22
Singular Value Decomposition (SVD)
A
U
x
x

m x n
m x r
r x r
r x n
23
Singular Value Decomposition (SVD)
k
k
x
x

m x n
m x r
r x r
r x n
K ltlt r
gt Sim
24
Interest-aware music sharing app.
Rock
P2P Music Bulletin Board
Soul
Vampire weekend
?
Adele
Miles Davis
Jazz
Madonna
Pop
25
Problem with interest-aware Greedy!
h
Y
d
a
X
X
D
e
b
Y
S
Y
f
X
c
X
g
Y
wireless contact data transfer
26
Solution PEEP
T1
T2
T3
T4
T5
T6
T7

• Still interest-aware
• Interest vectors binary
• Learning interests feedback from user, data
  items of each category, play times for music
  files, or LSA
• Transmit-budget
• Amount of data items allowed for transmission at
  each connection
• How to divide the transmit budget?
• Popularity
• Should be estimated

1 0 0 1 1 1 0
1
2
Popular
Arezu Moghadam, Henning Schulzrinne, "PEEP
Popularity-based and Energy Efficient Protocol
for Data Distribution in Mobile DTNs ",
CCNC'2011 - Smart Spaces and Personal Area
Networks, Las Vegas, USA, Jan 2011.
27
Popularity estimation
T1
T2
T3
T4
T5
T6

• Contact window N
• History of the users interests
• Average or weighted average
• Example C6, N8
• Replace the oldest

1 0 1 0 0 1
1 0 0 1 1 1
0 1 0 0 0 0
1 0 0 1 0 0
0 0 1 0 0 0
0 1 0 0 0 0
1 1 0 0 0 0
1 0 1 0 0 0
.62 .37 .37 .25 .12 .25
28
Evaluation of PEEP
gt Simulation details
29
Problem scope
Mobile DTNs
Applications
Routing
Mobility
A modular app. platform
Popularity-based and interest-aware
communication models
Markov-based mobility model and routing algorithm
30
Problem scope
Mobile DTNs
Applications
Mobility
Routing
Markov models to Model users movement
Markov-based Routing algorithm
31
Mobility is a crucial factor!
partition
D
32
Mobility models

• Mobility models usage
• Application provisioning and evaluation of
  routing protocols
• performance analysis
• QoS in cellular networks
• Problem Inadequacy of the current synthetic and
  trace-based mobility models
• Trace-based studies
• Precision and granularity
• Specific population of study
• Our empirical analysis based on a new set of
  traces
• Calculating patterns of human movement and using
  it in designing routing protocols

gt Levy
33
Problem with the current models

• Synthetic models mostly based on RWP
• Simplified assumptions about human movement
• Synthesized or trace-driven models
• Cellular networks
• Handoff predictions for QoS
• Movement of the node is not important within the
  cell
• Mobile DTNs
• No cell-tower or AP
• Impact of the mobility is higher on data
  propagation
• Traces or models extracted for cellular networks
  are not fine-grained enough!
• Traces from a limited number of users from a
  specific class
• Traces from APs with not enough granularity

Arezu Moghadam, Tony Jebara, Henning Schulzrinne,
A Markov Routing Algorithm for Mobile DTNs based
on Spatio-Temporal Modeling of Human Movement
Data ", ACM MSWiM 2011 , Miami Beach, FL, USA,
Oct 2011.
34
Spatial and Temporal Patterns
12am8am Home
35
Sense Networks traces

• GPS traces of a wide-spectrum of mobile users
• Citysense application
• Nightlife discovery
• Friend-finder
• Privacy concerns
• People are owners of their own data
• GPS precision of 20 feet compared to 120miles
  cell-tower coverage
• Population of 10,000 users

36
Data presentation

• Sequence of grids
• G1, G1, G17, G23,, GN
• Learning mechanism
• Ngrams
• A subsequence of N items from a sequence
• Modeling sequences in NLP, gene sequence
  analyzing, speech recognition
• Goal most probable future locations
• Pattern
• Likelihood of traversing a given sequence.

37
Ngrams
Tuples of Grids 5039907665 5038663466 5038414624 5038414623 5060063904 5053345115 . . .
5039907665 1370 10 230 10 0 30 . . .
5038663466 30 130 110 0 0 0 . . .
5038414624 220 110 3420 120 0 60 . . .
5038414623 10 0 50 0 14 0 . . .
5060063904 0 0 12 110 0 0 . . .
5053345115 0 50 0 13 176 343 . . .
. . . . . . . . . . . . . . . . . . . .
Triple of Grids Triple of Grids 5039907665 5038414624 5050607875 5053345115 5038414623 . . .
5039907665 5039907665 1180 110 30 30 10 . . .
5039907665 5038414624 0 230 0 0 0 . . .
5038414624 5038414624 220 2820 10 60 110 . . .
5038414624 5039907665 110 100 0 0 0 . . .
5039907665 5050607875 0 20 10 0 0 . . .
5050607875 5038414624 0 30 133 0 44 . . .
. . . . . . . . . . . . . . . . . . . .

• G1 , G2 , , Gi , , Gn
• Training
• Extract bigram and trigram tables.
• Testing
• Calculating the likelihood of a new observation

38
(No Transcript)
39
Markov chains for users movement

• Set of states
• S S1, s2, , sr
• Transition matrix
• Transitions correspond to consecutive GPS pings
• ? users mobility profiles
• Pattern
• States should be positive recurrent
• Finite hitting times with prob. 1
• Matrix of hitting times

40
Markov-based routing algorithm

• Absorption (hitting) times
• number of transitions until chain arrives
  at state j starting _at_ i
• Select the relay (r) with less absorption time
  than source (s).

41
Monte Carlo simulation
Users locations after each transition
Mobility Generator Engine -------------- Sampling
from the Markov Chains
Routing Algorithm Emulator
42
Performance measure

• Performance objective
• Delay
• Consumed energy
• Family of a-epidemics
• Measure performance curve

a 100
R
a 70
R
R
a 30
R
R
S
R
R
R
R
R
?
43
Evaluation of results
Random Destination
Popular Destination
a 0.1
a 0.2
a 0.3
a 0.7
a 1
44
Conclusion
Mobile DTNs
Applications
Mobility
Routing
45
(No Transcript)
46
Back up slides
47
Rsync Algorithm
48
Current routing models

• Single-source single-destination (no knowledge of
  topology)
• Flooding based protocols
• Epidemic
• Probabilistic routing
• PROPHET 57, RPLM 79, MaxProp 21
• Context or behavior of mobile users
• HiBOp 18, Profile-cast 42, MobySpace 54
• Multicast
• Extends the classical model with group
  memberships to mobile DTNs
• No infrastructure
• No knowledge of the topology (e.g., no multicast
  routers)
• Epidemic based multicast (no knowledge)

49
Current routing models

• Single-source single-destination (no knowledge of
  topology)
• Flooding based protocols
• Epidemic
• Probabilistic routing
• PROPHET 57, RPLM 79, MaxProp 21
• Context or behavior of mobile users
• HiBOp 18, Profile-cast 42, MobySpace 54
• Multicast
• Extends the classical model with group
  memberships to mobile DTNs
• No infrastructure (e.g., no multicast routers)
• No knowledge of the topology
• Epidemic based multicast (no knowledge)

50
Probabilistic routing criteria

• PROPHET
• Delivery predictability calculation.
• Routing with Persistent Link Modeling (RPLM)
• Monitors link connectivity to calculate its cost.
• Dijkstra to find a minimum cost path.
• MaxProp
• Assigning a cost value to each destination based
  on probability.
• Priority queue ? younger messages higher chances.
• MobySpace
• MobyPoint ? each nodes coordinates or mobility
  pattern.
• Distance on each axes probability of contacts or
  presence in a location.

51
Characteristics of the current models
Model objective Delivery ratio Delay Message redundancy Knowledge of topology
Flooding 1-to-1 1-to-many High Low (the least) High ? Buffer congestion Zero
Knowledge based 1-to-1 1-to-many MF the highest (even higher than ER) Moderate Low Provided to the algorithm
Probabilistic 1-to-1 Close to ER with tendency in mobility Close to ER with tendency in mobility Moderate Memory (learning from the past)
Multicast 1-to-many Flooding based is the highest Flooding based is the lowest Flooding based is the highest Required in non-epidemic
52
Interest-aware simulation results

• The ONE simulator for mobile DTNs
• Movement generation based on reality-minings
  mobile traces
• Compared to epidemic multicast with the same
  storage constraints
• The only model with no knowledge about topology
  and group memberships
• Measured relevant and irrelevant documents
  received by mobile users
• Increases received relevant documents by 30
• Decreases received irrelevant documents by 35
• Interest-aware algorithm limits the resource
  usage in terms of the cache and contact duration

The ONE, reality-mining
53
Web recommender systems

• Systems for recommending items (e.g. books,
  movies, CDs, web pages, newsgroup messages) to
  users based on examples of their preferences.
• Many on-line stores provide recommendations (e.g.
  Amazon, CDNow).
• Personalization to the individual needs,
  interests, and preferences of each user.

54
E.g. book recommender
55
Collaborative filtering

• Maintains a database of many users ratings of a
  variety of items
• For a given user, find other similar users whose
  ratings strongly correlate with the current user
• Recommend items rated highly by these similar
  users, but not rated by the current user
• Almost all existing commercial recommenders use
  this approach (e.g. Amazon)

56
Collaborative filtering
57
The ONE Simulator

• A modular simulation environment for mobile DTNs
• Routing package
• Prophet
• Epidemic
• Spray and wait
• Internal and external mobility generation
• RWP
• Map based
• Stationary
• Internal and external message event generation
• Reports of connection and message passing

58
Snapshot of map-based movement
59
The ONE Simulator

• A modular simulation environment for mobile DTNs
• Implements routing packages for one-to-one model
• Prophet
• Epidemic
• Spray and wait
• Internal and external mobility generation
• RWP
• Map based
• Stationary
• Internal and external message event generation
• Reports of contacts and message transmission

60
Interest-aware protocol implementation

• Interest-aware routing as a new module for the
  routing package
• General categories for documents
• Each node randomly assigned with some interest in
  each category
• A sub-population is randomly selected to be in
  the same community of interest
• Documents/messages are generated from nodes
  outside this community
• Coverage, pollution and dropped messages

61
Choice of mobility model for interest-aware

• Synthetic mobility traces
• RWP
• Map-based
• Community-based
• Speed of nodes
• Residence time
• Directions
• More realistic simulation with real-world traces
• Reality-mining traces

62
Users behavior Reality Mining

• Social behavior study
• Users encounters and visited locations
• How predictable is peoples lives?
• How does information flow?
• 100 subjects with Nokia symbian series 6600.
• Logs
• AP, GSM base stations and users encounters, call
  logs.
• Goal learn users behaviors and social network
  studies.

63
Reality-mining database

• MySql database
• Device
• Devicespan
• Person
• person-person contacts
• device-device contacts

Tables in REALMINE
activityspan callspan cellname cellspan celltower coverspan device devicespan person phonenumber
64
Relations we used
person person
PK oid
name password email
device device
PK oid
FK1 macaddr name person_oid
devicespan devicespan
PK oid
FK1 FK2 starttime endtime person_oid device_oid
65
Statistics and simulation set up

• Reality-mining subjects 97
• Total number of encountered devices 20795
• 44 of contacts with duration 0
• 15 of total contacts with devices outside the
  reality-mining
• 66 of these contacts just happened once!
• 40 have been considered in the same community of
  interests
• Fixed number of general categories

66
Optimization criteria for PEEP

• Maximize the number of received items of interest
• Minimize the delay of data distribution
• Not two independent values!
• The more the distribution the less the delay
• has nodes interested
• set of nodes interested in

67
PEEP implementation in The ONE

• PEEP routing as a new module for the routing
  package
• General categories for documents
• Each node is assigned some interest in each
  category based on Zipf distribution
• Distribution of the popular items follows Zipf
  law
• No knowledge of the topology
• Documents/messages are generated uniformly from
  different sources
• Measurements
• Number of received documents of interest over
  time
• Number of received documents of interest over
  contacts
• Speed of the distribution (slope of the graph)

68
Choice of mobility model for PEEP

• Synthetic mobility traces
• RWP
• Map-based
• Community-based
• Speed of nodes
• Residence time
• Directions
• The relative performance of the algorithm should
  be independent from the choice of the mobility
  model
• Our choice RWP
• A constant slope verifies this fact

69
Evaluation of the results

• If storage size is low buffer overflow happens
  too soon
• No chance for the items of interest to survive
• The most important difference with our previous
  work
• Unlimited storage size
• Limited energy (transmit-budget)
• Not far from the reality

70
Low storage size
Epidemic
Interest-aware
71
Medium High storage sizes
72
Levy flight

• Human walk follows a Levy flight distribution
• A random walk for which step size follows a
  power-law distribution
• step size
• Rhee et al.
• GPS traces of 44 users truncated power-law
• Brockmann et al.
• Bank notes is fat tailed power-law
• Gonzalez et al.
• Cell phone traces of 100,000 users truncated
  power-law

Graph from D. Brockmann and F. Theis, Money
Circulation, Trackable Items, and the Emergence
of Universal Human Mobility Patterns, IEEE
Pervasive Computing, Volume 7, Piscataway, NJ,
October 2008.