Social Network Analysis for Routing in Disconnected
Delay-Tolerant MANETs
Elizabeth Daly and Mads Haahr
Distributed Systems Group,
Computer Science Department,
Trinity College Dublin
Dublin, Ireland
{elizabeth.daly,mads.haahr}@cs.tcd.ie
ABSTRACT
Message delivery in sparse Mobile Ad hoc Networks (MANETs)
is difficult due to the fact that the network graph is rarely (if ever)
connected. A k ey challenge is to find a route that can provide good
deliv ery performance and low end-to-end delay in a disconnected
network graph where nodes may move freely. This paper presents
a multidisciplinary solution based on the consideration of t he so-
called small world dynamics which have been proposed for econ-
omy and social studies and have recently revealed to be a successful
approach to be e xploited for characterising information propaga-
tion in wireless networks. To this purpose, some bridge nodes are
identified based on their centrality characteristics, i.e., on t heir ca-
pability to broker information exchange among otherwise discon-
nected nodes. Due to the complexity of the centrality metrics in
populated networks the concept of ego networks is exploited where
nodes are not required to exchange information about the entire net-
work topology, but only locally available information is considered.
Then SimBet Routing is proposed which exploits the exchange of
pre-estimated ‘betweenness’ centrality metrics and locally deter-
mined social ‘similarity’ to the destination node. We present sim-
ulations using real trace data to demonstrate that SimBet Routing
results in delivery performance close to Epidemic Routing but with
significantly reduced overhead. Additionally, we sho w that Sim-
Bet Routing outperforms PRoPHET Routing, particularly when the
sending and receiving nodes have low connectivity.
Categories and Subject Descriptors
C.2.1 [Network and Architecture Design]: Wireless Communi-
cation ; Store and forward networks.
General Terms
Algorithms, Design, Performance.
Keywords
Delay & Disruption Tolerant Networks, MANETs, Sparse Net-
works, Social Network Analysis, Ego Networks.
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
MobiHoc’07, September 9–14, 2007, Montréal, Québec, Canada.
Copyright 2007 ACM 978-1-59593-684-4/07/0009 ...$5.00.
1. INTRODUCTION
A Mobile Ad hoc Network (MANET) is a dynamic wireless net-
work with or without fixed infrastructure. Nodes may move freely
and organise themselves arbitrarily [5]. Sparse Mobile Ad hoc Net-
works are a class of Ad hoc networks where node density is low,
and contacts between the nodes in the network do not occur very
frequently. As a result, the network graph is rarely, if ever, con-
nected and message delivery must be delay-tolerant. Traditional
MANET routing protocols such as A ODV [33], DSR [17], DSD V
[32] and LAR [19] make the assumption that the network graph is
fully connected and fail to route messages if there is not a com-
plete route from source to destination at the time of sending. For
this reason traditional MANET routing protocols cannot be used
in sparse MANETs. To overcome this issue, node mobility is ex-
ploited to physically carry messages between disconnected parts of
the network. These schemes are sometimes referred to as mobility-
assisted routing that employ the store-carry-and-forward model.
Mobility-assisted routing c onsists of each node independently mak-
ing forwarding decisions that take place when two nodes meet. A
message gets forwarded to encountered nodes until it reaches its
destination.
In this paper we propose the use of social network analysis tech-
niques in order to forward data in a disconnected delay-tolerant
MANET. Social networks exhibit the small world phenomenon which
comes from the observation that individuals are often linked by a
short chain of acquaintances. The classic example of a small world
is Milgram’s 1967 experiment, where 60 participants in Nebraska
were asked to forward a letter to be delivered to a stockbroker in
Boston [27]. The letters could only be forwarded to someone whom
the current letter holder knew by first name and who was assumed
to be more likely than the current holder to know the person to
whom the letters were addressed. The results showed that the me-
dian chain length of intermediate letter holders was approximately
6, giving rise to the notion of ‘six degrees of separation’. A more
recent experiment is that conducted by Hsu and Helmy who per-
formed an analysis on real world traces of different university cam-
pus wireless networks [15]. Their analysis found that node encoun-
ters are sufficient to build a connected relationship graph, which is
a small world graph. Therefore social analysis techniques may be
suitable for a number of disconnected delay-tolerant MANETs.
A number of solutions for routing in disconnected networks have
been proposed based on a node’s observed encounters where data
is routed to nodes with the highest ‘probability to deliver’ to a des-
tination node. Such metrics are typically based on either direct or
indirect observed encounters [4, 6, 13, 18, 24].
32
Figure 1: Disconnected Clusters
Some networks may consist of cliques where metrics based on di-
rect or indirect encounters may not find a suitable carrier for the
message. Consider three disconnected clusters in figure 1. Source
node s wishes to send a message to destination node d.However
node s is involved in a highly cliquish cluster in which none of
the nodes have directly or indirectly met destination node d.This
makes the decision of selecting a node to forward data difficult.
The three clusters are linked by bridging ties from i1 to i2 and
from i3 to i4. A path exists between the three clusters using in-
termediate nodes i1, i2, i3 and i4 which form bridges between
the three clusters. Weak acquaintance ties of i1-i2 and i3-i4, il-
lustrated by the dashed lines, become a crucial bridge between the
three tightly connected groups, and these groups would not be con-
nected if not for the existence of these weak ties. We propose to
forward data based on the identification of these ‘bridges’ and the
identification of nodes that reside within the same cluster as the
destination node. Our main contribution of this paper is a new for-
warding metric based on ego network analysis to locally determine
a node’s centrality within the network and the node’s social simi-
larity to the destination node. To the best of our knowledge, this is
the first work to exploit social analysis techniques in this manner.
The remainder of this paper is organized as follows: Section 2 re-
views related work in the area of message delivery in disconnected
networks. Section 3 introduces the concept of node centrality in
a network and shows how it can be used to forward information.
Section 4 discusses social similarity and how it can be used to eval-
uate the strength of a relationship between two nodes. Section 5
presents SimBet Routing. Section 6 compares the performance of
SimBet Routing to Epidemic Routing [38] and the PRoPHET Rout-
ing protocol [24] using real trace data from the MIT Reality Mining
project [1, 7]. Finally, Section 7 concludes the paper.
2. RELATED WORK
A number of projects attempt to enable message delivery by using
a virtual backbone with nodes carrying the data through discon-
nected parts of the network [11, 34]. The Data MULE project uses
mobile nodes to collect data from sensors which is then delivered
to a base station [34]. T he Data MULEs are assumed to have suf-
ficient buffer space to hold all data until they pass a base station.
The approach is similar to the technique used in [2, 11, 12]. These
projects study opportunistic forwarding of information from mo-
bile nodes to a fixed destination. Ho wever, they do not consider
opportunistic forwarding between the mobile nodes.
‘Active’ schemes go further in using nodes to deliver data by as-
suming control or influence over node movements. Li et al. [22]
explore message delivery in disconnected MANETs where nodes
can be instructed to move in order to transmit messages in the
most efficient manner. The message ferrying project [40] proposes
proactively changing the motion of nodes to help deliver data. They
investigate both ‘node initiated’ mobility, where the nodes move to
meet a known message ferry trajectory, or ‘ferry initiated’ mobility,
where the nodes signal communication requests via a long range
radio, and the message ferry moves to meet them. Both assume
control over node movements and in the case of message ferries,
kno wledge of the paths to be taken by these message ferry nodes.
Other work utilises a time-dependent network graph in order to
efficiently route messages. Jain et al. [16] assume knowledge of
connectivity patterns where exact timing information of contacts is
kno wn, and then modifies Dijkstra’s algorithm to compute the cost
edges and routes accordingly. Merugu et al. [26] likewise make
the assumption of detailed knowledge of node future movements.
They route messages over a time-dependent graph with knowledge
of when each node will next be encountered. Handorean et al. [14]
take a similar approach with knowledge of connectivity. However,
they do relax this assumption where only partial information is
known. This information is time-dependent and routes are com-
puted over the time-varying paths available. However, if nodes do
not move in a predictable manner, or are delayed, then the path is
broken. Additionally, if a path to the destination is not available
using the time-dependent graph, the message is flooded.
Epidemic Routing [38] provides message delivery in disconnected
environments where no assumptions are made in regards to con-
trol over node movements or knowledge of the network’s future
topology. Each host maintains a buffer containing messages. Upon
meeting, the two nodes exchange summary vectors to determine
which messages held by the other have not been seen before. They
then initiate a transfer of new messages. In this way, messages are
propagated throughout the network. This method guarantees de-
liv ery if a route is available but is expensive in terms of resources
since the network is essentially flooded. Attempts to reduce the
number of copies of the message are explored in [31] and [36]. Ni
et al. [31] take a simple approach to reduce the overhead of flood-
ing by only forwarding a copy with some probability p<1,which
is essentially randomized flooding. The Spray-and-Wait solution
presented by Spyropoulos et al. [36] assigns a replication number
to a message and distributes message copies to a number carrying
nodes and then waits until a carrying node meets the destination.
A number of solutions employ some form of ‘probability to de-
liver’ metric in order to further reduce the overhead associated with
Epidemic Routing by preferentially routing to nodes deemed most
likely to deliver. These metrics are based on either contact history,
location i nformation or utility metrics.
Burgess et al. [4] transmit messages to encountered nodes in the
order of probability for delivery, which is based on contact infor-
mation. However, if the connection lasts long enough, all messages
are transmitted, thus turning into standard Epidemic Routing. Ac-
knowledgments are sent to all nodes upon delivery, and the deliv-
ered messages are then deleted from the buffers. PRoPHET Rout-
ing [24] is also probability-based, using past encounters to predict
the probability of meeting a node again, nodes that are encountered
frequently have an increased probability whereas older contacts are
degraded over time. Additionally, the transitiv e nature of encoun-
ters is exploited where nodes exchange encounter probabilities and
the probability of indirectly encountering the destination node is
evaluated. Similarly [18] and [37] define probability based on node
encounters in order to calculate the cost of the route. [6] and [13]
use the so-called ‘time elapsed since last encounter’ or the ‘last en-
counter age’ to route messages to destinations. In order to route a
message to a destination, the message is forwarded to the neighbour
who encountered the destination more recently than the source and
other neighbours.
Lebrun et al. [20] propose a location-based delay-tolerant routing
scheme that uses the trajectories of mobile nodes to predict their
future distance to the destination and passes messages to nodes that
33
are moving in the direction of the destination. Leguay et al. [21]
present a virtual coordinate system where the node coordinates are
composed of a set of probabilities, each representing the chance
that a node will be found in a specific location. This information is
then used to compute the best available route.
Musolesi et al. [28] introduce a generic method that uses Kalman
filters to combine and evaluate the multiple dimensions of a node’s
context in order to make routing decisions. The context is created
from measurements that nodes perform periodically, which can be
related to connectivity. The approach only uses a single copy of a
message, which is passed from one node to a node with a higher
‘delivery metric’. The authors propose passing messages for un-
kno wn destinations using a ‘default route’ which is the ‘most mo-
bile’ node available. Spyropoulos et al. [35] use a combination of
random walk and utility-based forwarding. Random walk is used
until a node with a sufficiently high utility metric is found after
which the utility metric is used to route to the destination node.
Our work is distinct in that the SimBet Routing metric is comprised
of both a node’s centrality and its social similarity. Consequently, if
the destination node is unknown to the sending node or its contacts,
the message is routed to a structurally more central node where the
potential of finding a suitable carrier is dramatically increased. We
make no assumptions of control of node movements as in [22, 40]
or knowledge of node future movements as in [14, 16, 26]. Un-
like multi-copy strategies, we assume the existence of a single copy
of each message in the network which reduces resource consump-
tion compared to [31, 36, 38]. We will show that SimBet Routing
improves upon encounter-based strategies where direct or indirect
encounters may not be available [4, 6, 13, 18, 24, 37].
3. CENTRALITY
We estimate a node’s centrality in the network in order to iden-
tify bridges. Centrality in graph theory and network analysis is
a quantification of the relative importance of a vertex within the
graph (e.g., how important a person is within a social network).
The centrality of a node in a network is a measure of the structural
importance of the node. A central node, typically, has a stronger
capability of connecting other network members. There are several
ways to measure centrality. The three most widely used centrality
measures are Freeman’s degree, closeness, and betweenness mea-
sures [9, 10].
‘Degree’ centrality is measured as the number of direct ties that in-
volve a given node [10]. A node with high degree centrality main-
tains contacts with numerous other network nodes. Such nodes
can be seen as popular nodes with large numbers of links to oth-
ers. As such, a central node occupies a structural position (network
location) that may act as a conduit for information exchange. In
contrast, peripheral nodes maintain few or no relations and thus are
located at the margins of the network. Degree centrality for a given
node p
i
is calculated as:
C
D
(p
i
)=
N
k=1
a(p
i
,p
k
) (1)
where a(p
i
,p
k
)=1if a direct link exists between p
i
and p
k
and
i = k.
‘Closeness’ centrality measures the reciprocal of the mean geodesic
distance d(p
i
,p
k
), which is the shortest path between a node p
i
and
all other reachable nodes [10]. Closeness centrality can be regarded
as a measure of how long it will take information to spread from a
given node to other nodes in the network [29]. Closeness centrality
for a given node is calculated as:
C
C
(p
i
)=
N − 1
N
k=1
d(p
i
,p
k
)
(2)
where N is the number of nodes in the network and i = k.
‘Betweenness’ centrality measures the extent to which a node lies
on the paths linking other nodes [9, 10]. Betweenness centrality
can be regarded as a measure of the extent to which a node has
control over information flowing between others [29]. A node with
a high betweenness centrality has a capacity to facilitate interac-
tions between the nodes that it links. In our case it can be regarded
as how well a node can facilitate communication to other nodes in
the network. Betweenness centrality is calculated as:
C
B
(p
i
)=
N
j=1
j−1
k=1
g
jk
(p
i
)
g
jk
(3)
where g
jk
is the total number of geodesic paths linking p
j
and p
k
,
and g
jk
(p
i
) is the number of those geodesic paths that include p
i
.
Freeman’s centrality metrics are based on analysis of a complete
and bounded network which is sometimes referred to as a socio-
centric network. These metrics become difficult to evaluate in net-
works with a large node population because they require complete
knowledge of the network topology. For this reason the concept of
‘ego networks’ has been introduced. Ego networks can be defined
as a network consisting of a single actor (ego) together with the ac-
tors they are connected to (alters) and all the links among those al-
ters. Consequently, ego network analysis can be performed locally
by individual nodes without complete knowledge of the entire net-
work. Marsden introduces centrality measures calculated using ego
networks and compares these to Freeman’s centrality measures of
a sociocentric network [25]. Degree centrality can easily be mea-
sured for an ego network where it is a simple count of the number
of contacts. Closeness centrality is uninformative in an ego net-
work, since by definition an ego network only considers nodes to
which the ego node is directly related and then by definition the
distance from the ego node to all other nodes considered in the ego
network is 1. On the other hand, betweenness centrality in ego
networks has sho wn to be quite a good measure when compared
to that of the sociocentric measure. Marsden calculates the ego-
centric and the sociocentric betweenness centrality measure for the
network shown in figure 2.
Node
Sociocentric
betweenness
Egocentric
betweenness
w1 3.75 0.83
w2 0.25 0.25
w3 3.75 0.83
w4 3.75 0.83
w5 30 4
w6 0 0
w7 28.33 4.33
w8 0.33 0.33
w9 0.33 0.33
s1 1.5 0.25
s2 0 0
s4 0 0
i1 0 0
i3 0 0
Figure 2: Bank Wiring Room network sociocentric and egocen-
tric betweenness [25]
The betweenness centrality C
B
(p
i
) based on the egocentric mea-
sures does not correspond perfectly to the sociocentric measures.
However, it can be seen that the ranking of nodes based on the two
measures of betweenness are identical in this network. This means
that two nodes may compare their own locally calculated between-
34
ness value, and the node with the higher betweenness value may be
determined. In effect, the betweenness value captures ‘how much
a node connects nodes that are themselves not directly connected’.
For example in the network shown i n figure 2, w9 has no connec-
tion with w4. The node with the highest betweenness value con-
nected to w9 is w7, so if a message is forwarded to w7,themes-
sage can then be forwarded to w5 which has a direct connection
with w4. In this way, betweenness centrality may be used to for-
ward messages in a network. Marsden compared sociocentric and
egocentric betweenness for 15 other sample networks and found
that the two values correlate well in all scenarios [25].
4. SIMILARITY
Sociologists have long known that social networks (e.g., networks
of personal acquaintances) display a high degree of transitivity,
showing that there is a heightened probability of two people being
acquainted if they have one or more other acquaintances in com-
mon. In physics literature this phenomenon is called ‘clustering’.
Watts and Strogatz showed that real-world networks exhibit strong
clustering or network transitivity [39]. A network is said to show
‘clustering’ if the probability of two nodes being connected by a
link is higher when the nodes in question have a common neigh-
bour.
The degree of contact between nodes has an important effect in
terms of information dissemination. When the neighbours of nodes
are unlikely to be in contact with each other, diffusion can be ex-
pected to take longer than when the degree of separation is lower.
Consequently, nodes with a lower degree of separation from a given
node are good candidates for information dissemination to that node.
The degree of separation can be measured by the ratio of common
neighbours between individuals in social networks.
Newman analysed the time evolution of scientific collaborations
and observed that the use of examining neighbours, in this case co-
authors of authors, could help predict future collaborations [30].
From this analysis Newman determined that the probability of two
individuals collaborating increases as the number m of their pre-
vious mutual co-authors goes up. A pair of scientists who have
five mutual previous collaborators, for instance, are about twice as
likely to collaborate as a pair with only two, and about 200 times as
likely as a pair with none. Additionally Newman determined that
the probability of collaboration increases with the number of times
one has collaborated before, meaning that past collaborations are a
good indicator of future ones.
Liben-Nowell and Kleinberg explored this theory by using this com-
mon neighbour metric in order to predict future collaborations on
an author database [23]. The probability of a future collaboration
P (x, y) between authors x and y was calculated by:
P (x, y)=|N(x) ∩ N(y)| (4)
Where N(x) and N (y) are the set of neighbours of author x and
y respectively. This probability captures the ‘similarity’ between
nodes x and y, relative to the network topology. The authors anal-
ysed a number of different metrics to capture the similarity between
nodes. The results were promising where links were predicted, us-
ing the common neighbours metric, by a factor of up to 47 improve-
ment compared to that of random prediction.
5. ROUTING BASED ON BETWEENNESS
CENTRALITY AND SIMILARITY
In this section we present SimBet Routing, a forwarding algorithm
based on betweenness centrality and similarity as described in sec-
tions 3 and 4 respectively. The algorithm makes no assumptions
of global knowledge and forwarding decisions are based solely on
local calculations.
5.1 Betweenness Calculation
Betweenness centrality is calculated using an ego network repre-
sentation of the nodes with which the ego node has come into con-
tact. Mathematically, node contacts can be represented by an adja-
cency matrix A, which is an n×n symmetric matrix, where n is the
number of contacts a given node has encountered. The adjacency
matrix has elements:
A
ij
=
1 if there is a contact between i and j
0 otherwise
We consider contacts t o be bidirectional, so if a contact exists be-
tween i and j then there is also a contact between j and i.The
betweenness centrality is calculated by computing the number of
nodes that are indirectly connected through the ego node. The be-
tweenness of the ego node is the sum of the reciprocals of the en-
tries of A
2
[1 − A]
i,j
[8]. The following is a a matrix representa-
tion of the two-hop neighbourhood contacts of node w8 from figure
2.
w8 w6 w7 w9 s4
w8=
w8
w6
w7
w9
s4
⎡
⎢
⎢
⎢
⎣
01111
10110
11011
11101
10110
⎤
⎥
⎥
⎥
⎦
w8 w6 w7 w9 s4
w8
2
[1 − w8] =
w8
w6
w7
w9
s4
⎡
⎢
⎢
⎢
⎣
3
⎤
⎥
⎥
⎥
⎦
Since the matrix is symmetrical, only the non-zero entries abov e
the diagonal need to be considered. In this case the only remain-
ing entry of w8
2
[1 − w8] is 3 and the reciprocal of the value is
0.33 which gives us the egocentric betweenness value for the node.
When a new node is encountered, the new node sends a list of nodes
it has encountered. A new entry is made in the n × n matrix. As an
ego network only considers the contacts between nodes that it has
directly encountered only the entries for the contacts that the newly
encountered node shares in common with the ego node are inserted
into the matrix.
5.2 Similarity Calculation
Node similarity is calculated using the same n×n matrix discussed
in section 5.1. The number of common neighbours between the cur-
rent node i and destination node j is a simple count of the non-zero
equivalent row entries in the matrix. Consider the example matrix
representing the contacts of node w8 in section 5.1. Node w8 has
a similarity with nodes w6, w7, w9 and s4 of 2, 3, 3 and 2 respec-
tively.This example only allows for the calculation of similarity for
nodes that have been met directly, but when nodes exchange a list of
nodes it has encountered as described in section 5.1 we may obtain
useful information in regards to nodes that we have not yet encoun-
tered. As discussed in section 4 the number of common neighbours
may be useful for ranking known contacts but also for predicting
future contacts. It may also represent the possibility of routing to
35
an indirect node through a contact. Hence we maintain a list of
indirect encounters and maintain a separate n × m matrix where
n is the number of nodes that have been met directly and m is the
number of nodes that have not directly been encountered, but may
be indirectly accessible through a direct contact.
w8 w6 w7 w9 s4 w5
w8=
w8
w6
w7
w9
s4
⎡
⎢
⎢
⎢
⎣
01111
10110
11011
11101
10110
⎤
⎥
⎥
⎥
⎦
⎡
⎢
⎢
⎢
⎣
0
0
1
0
0
⎤
⎥
⎥
⎥
⎦
The example abo ve shows the inclusion of an indirect contact node
w5 in the similarity calculation. The fact that node w7 has a contact
with node w5 was learnt during an information exchange between
node w8 and node w7. Since node w8 has no direct contact with
node w5, it is added to the indirect contact matrix. Node w8 and
node w5 share node w7 as a common neighbour and therefore have
a similarity 1.
5.3 SimBet Utility Calculation
The SimBet utility is a value between 0 and 1 and is based on two
components: similarity utility and betweenness utility. Selecting
which node represents the best carrier for the message becomes a
multiple attribute decision problem, where we wish to select the
node that provides the maximum utility for carrying the message.
This is achieved using a pairwise comparison matrix on the nor-
malised relative weights of the attributes. The similarity utility
SimUtil
n
and the betweenness utility BetU til
n
of node n for
deliv ering a message to destination node d compared to node m is
given by:
SimUtil
n
(d)=
Sim
n
(d)
Sim
n
(d)+Sim
m
(d)
(5)
BetU til
n
=
Bet
n
Bet
n
+ Bet
m
(6)
The SimBetUtil
n
(d) is given by combining the normalised rela-
tive weights of the attributes given by:
SimBetUtil
n
(d)=αSimU til
n
(d)+βBetUtil
n
(7)
where α and β are tunable parameters and α + β =1. Conse-
quently these parameters allow for the adjustment of the relative
importance of the two utility values.
5.4 SimBet Routing
This section describes SimBet Routing outlined in algorithm 1. The
algorithm represents the communication between nodes n and m.
Upon reception of a Hello message node n verifies that node m is a
new neighbour. If this is the case, any messages destined for node
m are delivered and an encounter request is sent. Node m replies
with a list of nodes it has encountered. This list of contacts is then
used to update the betweenness v alue on node n and the similarity
v alue as described in sections 5.1 and 5.2 respectively. The two
nodes then exchange a summary vector containing a list of desti-
nation nodes they are currently carrying messages for along with
their own locally determined betweenness value and the similarity
v alue for each destination. For each destination in the summary
vector, node n calculates the SimBet utility of node n and node m
as described in section 5.3. If node n has a higher SimBet utility
Algorithm 1 SimBet Routing Algorithm, pseudo-code of node n
1: upon reception of Hello message h from node m do
2: if newNeighbour(m)==true
3: if msgQueue.hasMsgsForDest(m)==true
4: deliverMsgs(m)
5: requestEncounters(m)
6:
7: upon reception of encounter vector ev from node m do
8: addNodeEncounters(m, ev)
9: updateBetweenness()
10: updateSimilarity()
11: exchangeSummaryVector(m)
12:
13: upon reception of summary vector sv from node m do
14: Vector requestMsgs
15: for all destinations ∈ sv do
16: if m.simBet(d)<simBet(d)
17: requestMsgs.add(d)
18: sendMsgRequest(m, requestMsgs)
19:
20: upon reception of message request vector mrv from node m
do
21: Vector transferMsgs
22: for all messages ∈ mrv do
23: transferM sgs.add(msgQueue.getMsgs(d))
24: sendTransferMsgs(m, transferMsgs
)
25:
26: upon reception of transfer message tm from node m do
27: msgQueue.add(tm)
for a given destination, the destination is added to a vector of desti-
nations for which messages are requested. When all destinations in
the summary vector has been compared, node n sends the message
request list to node m. Node m then removes all messages destined
for the destination node from its queue and forwards them to node
n. Upon receiving a transfer message from node m the message is
added to the message queue of node n.
6. SIMULATION RESULTS
In this section we describe the simulations used to evaluate SimBet
Routing and compare its cost and performance to Epidemic Rout-
ing and PRoPHET Routing. Our first experiment examines inter-
node communication between the entire node population in order
to evaluate the overall performance. In our second experiment we
highlight the conditions where SimBet Routing succeeds in finding
a route while PRoPHET fails, by limiting inter-node communica-
tion to the nodes least connected in the network.
6.1 Simulation Setup
In order to evaluate the premise of routing based on centrality and
similarity we utilised a trace of node contacts from the MIT Re-
ality Mining project [1, 7]. The study consisted of 100 users car-
rying Nokia 6600 smart phones over the course of nine months.
They collected information using call logs, Bluetooth devices in
proximity, cell tower IDs, application usage and phone status. For
our purposes we use Bluetooth sightings in order to identify direct
contacts between nodes where data transfer could have taken place.
The trace file of these sightings was used to generate an event based
simulation. Each time a contact was observed, nodes exchange en-
counters and update their locally calculated ego betweenness and
social similarity values.
36
![Figure 4: Friendship network Eagle and Pentland [7]](/figures/figure-4-friendship-network-eagle-and-pentland-7-31lktgsl.webp)
![Figure 2: Bank Wiring Room network sociocentric and egocentric betweenness [25]](/figures/figure-2-bank-wiring-room-network-sociocentric-and-1kh8q30h.webp)
