∗

Routing in a Cyclic MobiSpace

Cong Liu and Jie Wu

Department of Computer Science and Engineering
Florida Atlantic University
Boca Raton, FL 33431
{cliu8@, jie@cse}.fau.edu

ABSTRACT works with periodic connectivity, and underwater acoustic

A key challenge of routing in delay tolerant networks (DTNs) networks with moderate delays and frequent interruptions.
is to find routes that have high delivery rates and low end- Routing in DTNs is an active research area. The De-
to-end delays. When oracles are not available for future lay Tolerant Network Research Group (DTNRG) [1] has de-
connectivity, opportunistic routing is preferred in DTNs, in signed a complete architecture to support various protocols
which messages are forwarded to nodes with higher delivery in DTNs. A DTN can be described abstractly using a time-
probabilities. We observe that real objects have repetitive space graph, in which each edge contains a set of contacts.
motions, but no prior research work has investigated the A contact is a period of time during which two nodes can
cyclic delivery probability of messages between nodes. In communicate with each other. On the Internet, intermittent
this paper, we propose to use the expected minimum delay connectivity causes loss of data. DTNs, in contrast, support
(EMD) as a new delivery probability metric in DTNs with communication between intermittently-connected nodes us-
repetitive but non-deterministic mobility. Specifically, we ing the store-carry-forward routing mechanism.
model the network as a probabilistic time-space graph with Routing in DTNs poses some unique challenges compared
historical contact information or prior knowledge about the to conventional data networks due to the uncertain and
network. We then translate it into a probabilistic state-space time-varying nature of network connectivity, which signif-
graph in which the time dimension is removed. Finally, we icantly complicates the routing decision. In [2, 3], optimal
apply the Markov decision process to derive the EMDs of delivery paths in a DTN can be discovered by construct-
the messages at particular times. Our proposed EMD-based ing a time-space graph with oracles. In practical situations
routing protocol, called routing in cyclic MobiSpace (RCM), where no oracle is available to reveal future contacts, op-
outperforms several existing opportunistic routing protocols portunistic routing [4, 5] is proposed in which one or more
when simulated using both real and synthetic traces. copies of a message is sent along different paths and each
copy is always forwarded to the node that has a higher de-
livery probability. Metrics for delivery probability can be ei-
Keywords ther short-term metrics (e.g., the time that has elapsed since
Cyclic MobiSpace, delay tolerant networks (DTNs), simula- the last encounter), which have short life-spans and require
tion, trace. frequent updates, or long-term metrics which are relatively
stable over time.
Previous works have proposed a variety of long-term met-
1. INTRODUCTION rics including encounter frequency [5] and social similarity
Delay tolerant networks (DTNs) are occasionally-connected [6]. One advantage of the long-term delivery probability
networks that suffer from frequent network partitioning as metrics is that they are relatively stable once generated from
the result of high mobility, low density, short radio range, historical connectivity information or prior knowledge on the
intermittent power, interference, obstruction, attacks, etc. contact pattern of the nodes, avoiding expensive and fre-
Representative DTNs include sensor networks using sched- quent updates. We notice that most real objects have cyclic
uled connectivity, terrestrial wireless networks that cannot motion patterns, and therefore it is possible and valuable in
ordinarily maintain end-to-end connectivity, satellite net- practice to use some cyclic metric to increase the accuracy
∗This work was supported in part by NSF grants ANI of the estimated delivery probability. In this paper, we pro-
0073736, EIA 0130806, CCR 0329741, CNS 0422762, CNS pose the use of a cyclic long term metric, called the expected
0434533, CNS 0531410, and CNS 0626240. minimum delay (EMD), which is the expected time that an
optimal forwarding scheme takes to deliver a message at a
specific time from a source to a destination, in a network
with cyclic and uncertain connectivity.
Permission to make digital or hard copies of all or part of this work for A MobiSpace is a Euclidean space (or other higher di-
personal or classroom use is granted without fee provided that copies are mensional space) where nodes can be either mobile or static
not made or distributed for profit or commercial advantage and that copies and they can communicate within each other’s transmission
bear this notice and the full citation on the first page. To copy otherwise, to range. We define a cyclic MobiSpace as a MobiSpace where
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
mobility is cyclic. In a cyclic MobiSpace, if two nodes are
MobiHoc’08, May 26–30, 2008, Hong Kong SAR, China. often in contact at a particular time in previous cycles, then
Copyright 2008 ACM X-XXXXX-XX-X/XX/XX ...$5.00.
 1 0.18 0.12

Contact probability
50 0.8 0.16
0.1
0.6 0.14

Contact probability

Contact probability
0.4 0.12 0.08
A B 0.1
p=0.7 0.2
0.06
0 0.08
0 10 20 30 40
0.06 0.04
Time slot
0.04
(a) (b) 0.02
0.02

0 0
0 200 400 600 800 1000 1200 1400 0 200 400 600 800 1000 1200 1400
Time slots (one minute) Time slots (one minute)
Figure 1: Nodes A and B have a contact with p = 0.7
(a) Shifts 01/AM & 03/AM (b) Shifts 32/PM & 21/EVE
during time slot 0 every 50 time slots.

Figure 2: Discretized probabilistic contacts between

different sub-shifts in the UMassDieselNet trace.
the probability that they will be in contact at the same time
in the next cycle is high. A cyclic MobiSpace can be mod-
eled with a probabilistic time-space graph in which an edge
between two nodes contains a set of discretized probabilistic a uniform distribution. Cyclic MobiSpaces are common in
contacts. Each discretized probabilistic contact is associated the real world, since (1) the motion cycle of most objects
with a time slot and a contact probability. The contact prob- are based on human defined or natural cycles of time such
ability is the probability that the two nodes have contacts as hour, day, week, etc; (2) most objects’ motions are repet-
during the time slot in every cycle. To calculate the EMD of itive, time-sensitive, and location-related. Note that cyclic
a message, we translate the probabilistic time-space graph MobiSpace is defined without any assumption on the shapes
into a probabilistic state-space graph, where the time dimen- of the trajectories, nor are the trajectories required to be
sion is removed from the edges. Then, we apply the Markov functions of time as in [7].
decision process to calculate the EMD of the state corre- In this section, we will first model a cyclic MobiSpace as
sponding to the node hosting the message and the current a probabilistic time-space graph, from which we will gener-
time. With EMDs, we propose a routing protocol, called ate a probabilistic state-space graph, which is a probabilistic
routing in cyclic MobiSpace (RCM). graph without time-dependent edges. Our goal is to calcu-
The contributions of this paper are summarized as follows. late the EMD of each message from this probabilistic graph,
for which we provide an extended efficient algorithm.
1. We propose the use of a cyclic, long-term delivery
probability metric, called EMD, to improve the perfor- 2.1 Expected minimum delay (EMD)
mance of opportunistic routing in cyclic MobiSpace.
The EMD is the expected time an optimal opportunistic
2. We model a cyclic MobiSpace as a probabilistic time- forwarding scheme takes to deliver a message given a starting
space graph and to remove the time dimension, we time and a source-destination pair. This optimal scheme
devise a lossless translation of this graph model into a always maximizes the delivery probability of each message
probabilistic state-space graph model. by forwarding a single copy of the message in the network.
As shown in Figure 1(a), assume node A travels a circular
3. We apply the Markov decision process (MDP) to solve trajectory once every 50 time slots, and only in time slot
the EMDs of the states in our probabilistic state-space 1 (Figure 1(a)), it has a contact with static node B with
graph model. We also extend a recent result on solving probability 0.7. The contact probability between A and B in
MDPs and propose an efficient algorithm for calculat- any 50 time slots is shown in Figure 1(b). In the beginning of
ing the EMDs in our state-space graph model. each cycle, i.e., time slot 1, the probability that A can send a
message to B is 0.7, which makes the probability of A having
4. We generate synthetic bus traces mimicking the real
to store the message for 50 time slots before the next possible
traces, which will be available to the research commu-
forwarding opportunity 0.3. Suppose the transmission delay
nity for subsequent research efforts in DTNs.
from A to B is 0.1, then the EMD D0 of a message in time
5. We perform simulations to evaluate and compare RCM slot 0 from A to B can be calculated as follows: D0 = 0.7 ×
with the enhanced versions of some existing proto- 0.1 + 0.3 × (50 + D0 ). Clearly, in any time slot t, the EMD
cols, using the NUS student trace, the UMassDiesel- is Dt = D0 + (50 − t mod 50) mod 50. We introduce EMD
Net trace, and our synthetic traces. Simulation results as a new, accurate delivery probability metric because it
show that RCM outperforms the other protocols. reflects the time-varying delivery probability between each
pair of nodes accurately during each cycle.

2. PROBABILISTIC GRAPH MODELS 2.2 Discretized probabilistic contact

In a cyclic MobiSpace, nodes have cyclic motion and con- We divide the common motion cycle T into small, fixed
tact patterns, and a common motion cycle T exists for all time slots. For each pair of nodes, we introduce a set of
nodes. For example, the common motion cycle T for a bus conceptual discretized probabilistic contacts. A discretized
system is a day. During each day, the positions of a bus probabilistic contact is a tuple (t, p), where t is a time slot
along a regular route at any time can be roughly estimated, in T and p is the contact probability of the two nodes in the
and so can be the contact probability distribution between time slot t.
two buses. Contrary to a cyclic MobiSpace is a network In the following, we show an example of the discretized
with completely random mobility, where the contact distri- probabilistic contact generated from 55 days of contact traces
bution between any pair of nodes over any motion cycle is in the UMassDieselNet trace [8, 9], where we consider each
 1
Tx 0.8

A&B
0.6
C 37/B 0/0
0.4 .5

23
1
0.2

7/

/1
0
0 10 20 30 40 50 B (3
7 ,0
.5 37/C
1
) 30/B 30/1 0/B
0.8
(0,0.7) 33/1

B&C
(30,0.7) C
0.6
0.4 0/0.7 0/0.7
B 0.2
4) 27/1
0 0. 30/A 30/1 0/A
A 0 10 20 30 40 50
1 0, 10/C
1
A (

20

/1
0.8

/1

10
.4

A&C
0.6
0.4
10/A
0/0
0.2
0
0 10 20 30 40 50
(a) G (b) G′
(a) (b)
Figure 4: The time-space graph G and the state-
Figure 3: A physical cyclic MobiSpace (a). The dis- space graph G′ of the cyclic MobiSpace in Fig-
cretized probabilistic contacts between each pair of ure 3(a).
nodes of a common motion cycle T (b).

we generate a probabilistic state-space graph G′ = (V ′ , E ′ ),

sub-shift (see Section 4.3) as a node, one day as the common where V ′ is the set of states and E ′ is the set of links which
motion cycle, and one minute as the unit of a time slot. The are time-independent. G′ is generated as follows. For each
discretized probabilistic contacts between sub-shifts 01/AM node u in G, we create a set of states {ti /u} for each time slot
(the morning sub-shift of shift number 1) and 03/AM are ti when u has one or more discretized probabilistic contacts.
shown in Figure 2(a). Figure 2(b) shows those between an- For example, three states 0/A, 10/A, and 30/A are created
other pair of sub-shifts. in G′ (Figure 4(b)) for node A in G (Figure 4(a)), since A
From these figures, we can see that the discretized proba- has three discretized probabilistic contacts (0, 0.7), (10, 0.4),
bilistic contacts between two nodes in a realistic cyclic Mo- and (30, 0.7) at different time slots. If node u has more than
biSpace gather around only a few consecutive time slots, one contact (with different nodes) in the same time slot,
and the number of discretized probabilistic contacts is much then only one state is created for u corresponding to this
smaller than the total number of time slots. Note that the time slot in G′ .
choice of time slot size is a trade-off between the accuracy1 There are two types of links in G′ : directional link and
of EMDs and the computation time. bidirectional link. The directional link connects the consec-
utive states of a node into a ring. For example, the three
2.3 Probabilistic time-space graph states of node A in G′ are connected to form a ring by three
We model a cyclic MobiSpace as a probabilistic time-space directional links represented by dashed lines as shown in
graph G = (V, E, T ), where V is the set of nodes, E is the set Figure 4(b). The bidirectional link in G′ is created corre-
of edges between the nodes, and T is the common motion sponding to each discretized probabilistic contact in G. For
cycle. An edge between two nodes is a non-empty set of each discretized probabilistic contact between nodes u and
discretized probabilistic contacts between the two nodes. v in time slot t, a bidirectional link is created in G′ between
Figure 3(a) shows a sample cyclic MobiSpace that we will states t/u and t/v (shown as a solid line in Figure 4(b)).
use throughout this paper. In this figure, nodes A and B Each state in G′ is a possible state of a message in the
move in their circular and triangular trajectories respectively network, and each link in G′ is a possible state transition of
with a cycle time of 30 units each, while node C travels along a message. A message is in state 30/A if it is at node A in
its straight-line trajectory with a cycle time of 20 units. Sup- time slot 30. If the message is kept in node A from time slot
pose nodes A, B, and C have contacts only during particular 30 to time slot 0 (which is a time span of 30 slots), then the
time slots in a common motion cycle T = 60 units, and these message transits from states 30/A to 0/A via a directional
contacts are non-deterministic in nature due to uncertainty link. On the other hand, if the message is forwarded in time
of the nodes’ positions, communication failures, etc. The slot 30 from node A to node B, then it transits from states
set of discretized probabilistic contacts for each pair of the 30/A to 30/B via a bidirectional link.
nodes in this cyclic MobiSpace is shown in Figure 3(b). Both types of links are labeled d/pmax , where d is the
The time-space graph G of the network is shown in Fig- transition delay and pmax is the maximal transition proba-
ure 4(a). In G, each edge contains the set of discretized bility. For a directional link, pmax is always equal to 1, which
probabilistic contacts between its nodes. For example, edge means a message can always be kept in a node without be-
(A, B) contains two discretized probabilistic contacts. One ing forwarded. For a bidirectional link, pmax is equal to the
is labeled (0, 0.7), which means its time slot is 0 and contact contact probability of the corresponding discretized proba-
probability is 0.7. bilistic contact. This is because the forwarding probability
of a forwarding node cannot exceed its contact probability
2.4 Probabilistic state-space graph with the receiving node. For simplicity, we let d = 0 for all
In order to remove the time dimension from the edges in G, bidirectional links assuming that message forwarding is al-
ways restricted to a time slot. This assumption is based on
1 the fact that the delay of a message forwarding is generally
Information about the ordering of the physical contacts
within the same time slot is lost. much smaller than the size of the time slot itself.
 0/0 44.7 0.5
1 s1 37/B .5

0
d1,2 /p 1,2

0.
a11

23

1.
1 ax

1
7/
m

5
s1 a12

/1
1 s2

/p
a21 d2,1 /p 2,1
2 0

1
d
/p
d 1,2 1,2 30/B 30/1 0/B Goal 51.7 0.3 66.4

2,g
d 1,
d 2 /p max

d2,g/p 1
2 0.7 0.7
d1,

s s2 0/0.7 0/0.7
g
/p 1
g
/p 1

1 g

0.3

d3
30/A 30/1 0/A 63.1 59.9 0
,

Goal
2 ,g

/1

0
/p

1.
ma

10
3
x

20

/1
g

0
0.

1.
/1

10
.4

6
73.1
s3 20/A 10/A
0/0
49.9
0.4
(a) (b) (c) (d)

Figure 5: (a) A Markov decision process (MDP) model (b) Implicit action (c) The MDP of G′ in Figure 4(b)
with C as destination (d) The values and optimal actions after applying value iteration in Figure 5(c).

2.5 Remarks der the control of the decision maker. MDP is a generalized
In G, a message can be in any node at any time. However Dijkstra’s algorithm for probabilistic graphs.
in G′ , a message can only be at a state in its correspond- We reformulate a variation of MDP as a 5-tuple (S, A, T,
ing time slot. This results in an important difference in D, SG ) as explained below. At any given time, the system
G′ : links are time-independent and they are always avail- can be at only one state s in the set of all states S. Each
able when messages are in the corresponding states. For state s has a set of actions As ⊆ A (the set of all actions).
example, whenever a message is in state 30/A, it always has Only one action a is allowed to take effect at a time. The
a chance of being transited to 30/B through the discretized effect of applying action a is that the system transits from
probabilistic contact available at time slot 30. In the rest of state s to another state. The transition probability function
this paper, we use the EMD of state to refer to the EMDs Ta (s, s′ ), specifies the probability of transiting from state s
of the message in this state. to state s′ when applying action a. Note that for any s and a,
P ′ ′ 2
Our probabilistic graph models differ from the models de- s′ ∈S Ta (s, s ) = 1. The delay function D(s, s ) specifies
fined in previous papers [2, 3] in that a contact in the pre- the delay of transiting from state s to state s′ . A value
vious models is a deterministic connection during a specific function V (s), gives the expected minimum total delay for
period of time, while a discretized probabilistic contact is transiting from s to any goal state g ∈ SG (the set of all goal
conceptual and is drawn from the the cyclic contact history states). V (g) = 0. An MDP is illustrated in Figure 5(a),
or from the prior knowledge of the cyclic contact pattern where s1 , s2 , and g ∈ SG are different states. s1 has two
between a pair of nodes. The purpose of the discretization actions a11 and a21 . If a11 takes effect, the system will transit
is the creation of our graph model G′ with time-independent from s1 to state s2 with delay d1,2 and probability p11,2 .
links where EMDs can be calculated. Note that the transla- The MDP problem is to find the set of expected minimum
tion from G to G′ is lossless since G can be reproduced from total delay values V (s) for each state s to reach any g ∈ SG
G′ by combining the states of the every node. and the corresponding optimal action chosen at each state
The number of states in G′ depends on the number of in order to achieve the minimum values. Value iteration
nodes and the discretized probabilistic contacts of each node [12] solves MDPs by iteratively updating the value functions
which is bounded by the number of time slots in T . Methods (Equation 1) of every state until they converge. In each
to reduce the number of states includes increasing the size of round of the iteration, based on the resulting values in the
time slots and dropping the discretized probabilistic contacts last iteration, the value V (s) of each state s ∈ / SG is updated
whose probabilities are below a certain threshold. by choosing an action a ∈ As such that V (s) is minimized.
For an action a, the value of s is the sum of the delays
D(s, s′ ) + Vt (s′ ) (for each possible next state s′ ) are weighed
3. EXPECTED MINIMUM DELAYS by Ta (s, s′ ) (refer to Figure 5(a)).
IN A STATE-SPACE GRAPH
X
This section reviews and reformulates a variation of the Vt+1 (s) = min {Ta (s, s′ ) × [D(s, s′ ) + Vt (s′ )]}. (1)
Markov decision process (MDP). Using MDP, the values as- a∈A(s)
s′ ∈S
sociated with the states in G′ are updated iteratively and
finally converge to their EMDs. An efficient approach is also The value functions are considered to converge sufficiently
proposed to calculate EMDs by extending a recent work in when the maximum difference between the values of all states
solving MDPs. in two consecutive iterations is less than some threshold
value. When the values of the states are properly initialized
3.1 Markov Decision Process (MDP) (such as iteratively applying Equation 2, which is essentially
State-space searching is a very common problem in AI the Bellman-Ford algorithm), the value iteration is guaran-
planning (or decision-making), which is similar to routing. teed to converge with the minimum values in the states [13].
The Markov decision process (MDP) [10, 11] provides a
mathematical framework for modeling decision-making in 2
In the traditional MDP models, cost function C is used.
situations where outcomes are partly random and partly un- We change it to delay function D in the networking context.
 Goal Goal
Table 1: Comparison of value iteration (VI) and our
C4 C3 extended topological value iteration (E-TVI).
C2
90/1 (a) Resulting value.
0/A 10/A 30/A 120/*
states 0/A 0/B 10/A 30/A 30/B 37/B
C1

/1
VI 59.87 66.41 49.88 63.14 51.70 44.70

83
0/B 30/B 37/B
E-TVI 59.92 66.46 49.91 63.19 51.73 44.73
(b) Computation time.
nodes 100 200 300 400 500
Figure 6: An efficient approach for the MDP in Fig-
VI 0.515 1.256 3.515 7.194 10.25
ure 5(c).
E-TVI 0.137 0.360 0.668 1.225 2.426

V0 (s) = min { min [D(s, s′ ) + V0 (s′ )]} (2) shown in Figure 5(c). Applying value iteration on this MDP,
a∈A(s) s′ :Ta (s,s′ )>0 we get the values (shown inside the states) and the optimal
actions (labeled on the links) of each state in Figure 5(d).
3.2 Deriving EMDs using MDP Note that in Figure 5(d), some of the links are removed such
Our probabilistic time-space graph model G′ differs from as the one from states 0/A to 0/B. This shows the situations
the MDP model in that the states in G′ are not associated where a message is not forwarded from a node with a lower
with actions which tell the actual state transition proba- EMD to a node with a higher EMD.
bilities. However, implicit actions can be determined by
the node’s forwarding preferences. A forwarding preference 3.3 An efficient approach for EMDs
shows the current node’s decision on which node to forward In MDPs, if a state s is a successor state of s′ (there is a
the message to (or to keep the message in the current node) transition from s′ to s), then the value V (s′ ) is dependent
when connected with multiple nodes. Inspired by a research on V (s). This causal relation is transitive. In [13], topology
work on robotics [14], we derive the optimal implicit action value iteration (TVI) is used to reduce the number of rounds
without enumerating all implicit actions. of updates performed to each state in the value iteration. In
Remember that each link in G′ represents a possible state TVI, Kosaraju’s [15] algorithm is used to find the set of
transition and it has a delay d, and a maximal transition strongly-connected components {Ci } and their topological
probability pmax (see Section 2.4). In Figure 5(b), state s order. If component C1 is a successor of C2 (i.e., for any
may transit to three other states for which the corresponding u ∈ C1 and v ∈ C2 , u is a successor of v), then the values of
delays and the maximal probabilities are given. If the pref- the states in C2 are dependent on those in C1 (but not vice
erential order of state transitions is {s1 , s2 , s3 } (which means versa). That is, components C1 and C2 can be calculated
s will transit to s1 whenever possible, and it will transit to separately in a reverse topological order. TVI significantly
s3 only when the transition to neither s1 nor s2 is possi- reduces the number of rounds in value iteration by breaking
ble), the implicit action (which indicates the probabilities the computation of the whole graph into the computation
p1 , p2 , and p3 of transiting from s to s1 , s2 , and s3 respec- of small components.
tively) is calculated as p1 = pmax 1 , p2 = (1 − p1 ) × pmax
2 , TVI cannot be directly applied to our graph model G′
and p3 = (1 − p1 − p2 ) × pmax 3 . Note that the sum of all since in G′ the non-goal states are likely to be connected
transition probabilities of an action calculated in this way as a strongly-connected component. As in Figure 5(c), (1)
is always 1 since for each state s there is at least one bidi- states of the same node are strongly-connected by direc-
rectional link and thus at least one pmax is 1. Also, some tional links, and (2) states of different nodes are connected
transition probabilities might be zero. For example, p3 = 0 by bidirectional links. However, if we group the states of the
if pmax
2 = 1. same time slots into components (not necessarily strongly-
All implicit actions for a state can be obtained from the connected components), we find that these components form
permutation of the preferential order of state transitions a partial order if, given a time slot tb , each link starting at a
from this state. Fortunately, for an optimal forwarding pro- time slot before tb and ending at time slot in or after tb is re-
tocol this order is always an increasing order of the expected moved. For example, if we change the layout of Figure 5(c),
costs of the transitions. For instance, if s can only tran- and remove the corresponding links by setting tb = 0, we
sit to states s1 and s2 , and their costs D(s, s1 ) + V (s1 ) < get the components C4 , . . . , C1 in a partial order as shown
D(s, s2 ) + V (s2 ), then state s1 is preferred over state s2 . in Figure 6. In this figure, the set of removed links are shown
This is because the value of state s is the weight sum p1 × by dashed lines which we denote as Lb .
[D(s, s1 ) + V (s1 )] + p2 × [D(s, s2 ) + V (s2 )], and to minimize Our extended TVI algorithm has two steps. In step one,
this weight sum given p1 + p2 = 1, p1 has to be maximized. the states are updated by TVI after modifying G′ with the
To calculate the EMD of a message in time slot t from following steps: (a) we let tb = 0 and remove Lb from the
node A to node C in the G′ (Figure 4(b)), we modified G′ graph, (b) we add a state s∗ with its time slot being 32 T ,
as follows. We replace all states of C with goal states Goal (c) for each state s which has an outgoing link removed
whose EMDs are 0, and we remove all outgoing links from (in set Lb ), we add a link from s to s∗ . Figure 6 shows
Goal. If state t/A does not exist in G′ , then we add state an example where the state added is 120/∗ and the links
t/A to G′ and add a link from t/A to the consecutive state added are shown by thick lines. The TVI is then applied to
of A. For example, if t = 20, the resulting MDP from G′ is update the component in their reversed topological order,
 0.6 30
i.e., C1 , . . . , C4 . In step two of our extended TVI, we recover Spray&focus* Spray&focus*
Spray&focus Spray&focus
the original graph by removing the added state s∗ and the 0.5 MaxProp*
MaxProp 25
MaxProp*
MaxProp
added links, and adding back the links in Lt . Then, we

Delay (hours)
0.4

Delivery rate
continuously update the components in the same topological 0.3 20

order, i.e., C1 , . . . , C4 , until the values of states in all of these 0.2

15
components converge. Similar to TVI, our extended TVI is 0.1
efficient since it breaks the whole graph into a large number 0 10
of small components and update them separately. 0 5 10 15 20 0 5 10 15 20
Hour Hour
Table 1(a) compares the resulting values of the states in (a) Delivery rate. (b) Delay.
the MDP in Figure 6. We can see that the values computed
by both methods are almost identical. Table 1(b) compares
their computation time in networks with different number Figure 7: Performance comparison of existing pro-
of nodes generated as described in Section 4.2. tocols and their extended versions using the UMass-
DieselNet trace.
3.4 Routing and analysis
We propose our opportunistic routing protocol called rout-
ing in cyclic MobiSpace (RCM), which uses EMD as the We evaluate our protocol, RCM, in the context of other
metric of delivery probability. Its forwarding rule is simple: routing algorithms using a wide variety of traces: NUS stu-
(1) for the single-copy forwarding case, a node u always for- dent contact trace, UMassDieselNet trace, and our synthetic
wards message to node v that it encounters iff the message bus traces. We will describe these network traces and the
has a smaller EMD in v than in u at the current time; (2) our simulation methods using these traces in respective sub-
for the multiple-copy forwarding case, tickets are allocated sections. Simulation results show that RCM improves the
among nodes proportional to the reciprocal of their EMDs. delivery rate and lowers the end-to-end delay in all traces as
RCM uses our long-term metric EMD and it has a small than the compared protocols.
amortized overhead for metadata (routing information), since
it does not require frequent updates of metadata unlike al- 4.1 Protocols in comparison
gorithms that use short-term metrics. For example, for a
We compare RCM against several other approaches. These
DTN based on a bus system that operates for several years,
algorithms are based solely on deciding which messages to
the contacts of the buses in the first few weeks can be gath-
forward during a contact with a given peer. Most of these al-
ered and the generated probabilistic state-space graph can
gorithms approximate delivery probability as the likelihood
be disseminated in the network once and for all. Addition or
of the existence of a delivery path. For our focus on the effi-
removal of bus routes can be reflected in the graph through
ciency of the delivery probability metrics and for fairness in
incremental update.
the comparison, we use the enhanced versions of these pro-
The following theorems show the optimality of our algo-
tocols that make use of the same level of prior knowledge of
rithm. Proofs are simplified due to space limitation.
historical connectivity patterns as RCM does.
Epidemic [16]. A node copies a message to every new
Theorem 1. The value iteration and the extended TVI
node it encounters that has not got a copy already, until its
guarantee that the values of the states converge to EMDs.
copy of the message times out, or it is notified of delivery.
Proof. The algorithms converge to EMDs because (1) Spray&wait [17]. This protocol differs from Epidemic in
the values are non-decreasing in each iteration, and (2) the that it controls the number of copies of each message in the
values are minimum upper bounded by EMDs initially and network. A number L of logical tickets are associated with
in each iteration. each original message. A node i can only copy a message to
another node j it encounters if the message in i owns N > 1
Suppose DAG is the EMD from nodes A to G at the cur- tickets. The new copy in j will have L tickets (L = ⌊N/2⌋),
rent time and DACG is the EMD from nodes A to C then and N − L tickets will remain with the message in i.
to G at the current time, then DAG ≤ DACG since an addi- Spray&focus [18]. This is an extension of Spray&wait
tional constraint, passing C, is placed on the possible mini- where a message with one ticket can still be forwarded from
mum delay paths between A and G. node i to j if the delivery probability of j to the destination
is higher. A node with the later last encounter time with
Theorem 2. The single-copy opportunistic forwarding sch- a destination has a higher delivery probability. Transitivity
eme proposed is the optimal single-copy opportunistic for- [5] is used to improve predictability in their mobility models.
warding scheme in terms of expected delivery latency. We use a enhanced, Spray&f ocus∗ , in which the average
Proof. We need to prove that each forwarding in RCM meeting interval drawn from the contact history is used to
maximizes the EMD of the message. Let the EMD from indicate delivery probability.
nodes A and B to destination G at the current time be MaxProp [8]. A cost is assigned to each node for each
DAG and DBG (DAG > DBG ) respectively. Then A should destination. Each node i keeps track of a probability fji of
forward the message to B in order to maximize the EMD of the next meeting node being j and disseminate it to every
the message if they meet. The assumption that there is a node in the network. The delivery probability from a source
node C through which the message has a larger EMD than to a destination is the total cost on their shortest path, where
B is contradicted by DBG < DAG ≤ DACG . cost of each hop (i, j) is 1 − fji . In our simulation, we use
a copy-controlled version using logical tickets, which differs
from Spray&focus only in the metric of delivery probability.
4. SIMULATION In our enhanced version M axP rop∗ , we let fji /fki = tij /tik
 1 1
Spray&wait 22 24
Spray&focus*
0.8 MaxProp* 0.8 22
RCM 20

Delay (hours)

Delay (hours)
20
Delivery rate

Delivery rate
0.6 0.6 18
18
16
0.4 0.4 16
14
Spray&wait 14
0.2 0.2 Spray&focus* Spray&focus* Spray&focus*
MaxProp* 12 MaxProp* 12 MaxProp*
RCM RCM RCM
0 0 10 10
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Attendance rate Clustering factor Attendance rate Clustering factor

(a) (b) (a) (b)

1 1 22 21
21 20
0.8 0.8 20 19

Delay (hours)

Delay (hours)
19
Delivery rate

Delivery rate

0.6 0.6 18
18
17
17
0.4 0.4 16
16
Spray&wait Spray&wait 15
0.2 Spray&focus* 0.2 Spray&focus* 15 Spray&focus* Spray&focus*
MaxProp* MaxProp* 14 MaxProp* 14 MaxProp*
RCM RCM RCM RCM
0 0 13 13
100 150 200 250 300 350 400 450 500 1 1.5 2 2.5 3 3.5 4 4.5 5 100 150 200 250 300 350 400 450 500 1 1.5 2 2.5 3 3.5 4 4.5 5
# of students # of tickets # of students # of tickets

(c) (d) (c) (d)

Figure 8: Comparison of delivery rate using the Figure 9: Comparison of delay using the NUS trace.
NUS trace.

where tij , tik are the historical encounter times of i with j other iff they are in the same classroom at the same time.
and k respectively. (2) Sessions start on the hour and end on the hour, which
Note that all of the protocols that we implement aim to means that hour is the unit of time for the contact duration.
compare different metrics for delivery probability, and all (3) Only the contacts that take place during class hours are
other optimizations that have orthogonal effects on the per- used. Non-class hours are removed to compress time. For
formance of the protocols are not implemented. The or- example, suppose the last class session on Monday ends at
thogonality means that these optimizations can be added to 9pm, and the first session on Tuesday starts at 8am. If Mon-
all of our implemented algorithms and they are expected to day 8pm to 9pm corresponds to the 10th hour, then Tuesday
provide an equal level of improvement in their routing per- 8am to 9am is the 11th hour. The advantages of the trace
formance. Such optimizations may include buffer manage- synthesized in this model are that they exhibit the same set
ment [8], global estimation of message delivery probability of characteristics to those observed in the real world and
[19] and social centrality of the nodes [6], geometric informa- it provides contact patterns of a large population (several
tion [20], etc. Similarly, we assume that an ideal mechanism orders of magnitude larger than any reported real traces)
serves to clear out buffers in the network of delivered data. over a long period. The schedules of the 4,885 classes and
First, we conduct simulation studies to compare the per- enrollment of 22,341 students for each of the class for each
formance of Spray&focus, MaxProp, and their extended ver- week of 77 class hours are publicly available on [22].
sions using the UMassDieselNet trace. The simulation en- We select a number of students N (100 ≤ N ≤ 500) in
vironment and settings will be the same as specified in the each of experiment due to the memory constraint in the sim-
Section 4.3. As shown in Figure 7, Spray&f ocus∗ is much ulation environment. Contacts related to the non-selected
better than Spray&f ocus in both delivery ratio and delay, students are ignored. We generate non-determinant traces
while M axP rop∗ slightly outperforms M axP rop. Note that by taking absentees into consideration. Each student at-
we use an ideal version of MaxProp where we assume each tends a class with a attendance probability Pattend . Our
node knows the cost function f of the other nodes, which data processing include the following steps. (1) We break
gives an upper performance bound. In practice, MaxProp each class session into several one hour class sessions and
has a much bigger amortized overhead in communication reassigned unique IDs to them. 159 conflicting enrollments
and computation than M axP rop∗ since the latter has a con- are resolved by removing those enrollment with a lower class
stant cost function f . session ID. (2) The selection of N students is tricky. If se-
lected randomly, the network becomes too sparse for mes-
4.2 NUS student contact trace sages to be delivered. On the other hand, when students
Accurate information of human contact patterns is avail- are selected by maximizing their similarities (the number
able in scenarios such as university campuses. As shown of common classes they are enrolled in), the network be-
by the National University of Singapore (NUS) student con- comes over connected. To prevent the above extremes and
tact trace model [21], when the class schedules and student maintain the small-world property in the size-reduced stu-
enrollment for each class on a campus are known, accurate dent networks, we use the following process. We select the
information about contact patterns between students over first student randomly. To select the kth student, we di-
large time scales can be obtained without a long-term con- vide the k − 1th selected students into two groups S1 and
th
tact data collection. Their contact model is simplified in S2 , and
P select the k student
P s as the one with the highest
several ways. (1) Two students are in contact with each score s1 ∈S1 sim(s, s1 ) − s2 ∈S2 sim(s, s1 ) among the stu-
 20 0.3 1 1
Spray&wait Spray&wait
18 Epidemic Epidemic
0.25 0.8 Spray&focus* 0.8 Spray&focus*
16
Number of contacts

MaxProp* MaxProp*
14 0.2 RCM RCM

Delivery rate

Delivery rate
Percentage
0.6 0.6
12
0.15
10
0.4 0.4
8 0.1
6 0.2 0.2
0.05
4
2 0 0 0
6 8 10 12 14 16 18 20 22 24 0 5 10 15 20 25 30 0 5 10 15 20 0 5 10 15 20
Hour Contact duration (seconds) Hour Hour

(a) Contact distribution (b) Duration distribution (a) Delivery rate (1 ticket). (b) Delivery rate (3 tickets).
28 28
Epidemic Epidemic
Figure 10: Statistics in the UMassDieselNet traces. 26 Spray&focus*
MaxProp*
26 Spray&focus*
MaxProp*
24 24
RCM RCM

Delay (hours)

Delay (hours)
22 22
20 20
18 18

dents that are not yet selected, where the similarity func- 16 16
14 14
tion sim is defined as the number of common class ses- 12 12
sions enrolled by two students. We define a clustering factor 10 10
0 5 10 15 20 0 5 10 15 20
C = |S1 |/(|S1 |+|S2 |) which determines the degree of connec- Hour Hour
tivity in the network. (3) If two students enroll in the same (c) Delay (1 ticket). (d) Delay (3 tickets).
class session, they have a discretized probabilistic contact in
2 30 30
the hour of the class session with probability Pabsent . With Epidemic
Spray&focus*
Epidemic
Spray&focus*
25 25
all of the discretized probabilistic contacts, the probabilis- MaxProp*
RCM
MaxProp*
RCM
tic state-space graph can be built. (4) Finally, we generate 20 20

Hop-count

Hop-count
2
traces by generating contacts with probability Pattend for 15 15

each pair of students and each class session they enroll in. 10 10

The default setting in our simulation is Pattend = 0.8, 5 5

C = 0.5, 300 students, one ticket per message. In every 0 0

0 5 10 15 20 0 5 10 15 20
minute, every node sends messages to 20 randomly selected Hour Hour
nodes. As shown in Figures 8 and 9, for every protocols, (e) Forwarding (1 ticket). (f) Forwarding (3 tickets).
the delivery rate increases and the delay decreases as Pattend
increases. All protocols achieve their best performance when
Figure 11: Comparison of different performance
C is approximately 0.6, and the performance of the protocols
metrics using the UMassDieselNet trace.
improves as N decreases and L increases. The delivery rate
of RCM is approximately 25% on average (and up to 50%)
greater than that of the second best protocol M axP rop∗ ,
and the delay of RCM is 15% on average (and up to 60%)
smaller. parsing one of the dispatch records DA all.txt. For exam-
ple, sub-shift 21/AM (the AM sub-shift of shift 21) starts
4.3 UMassDieselNet trace at 6:10AM and ends at 10:30AM. We also obtain a mapping
Before presenting our simulation method and results, we from day and bus to the sub-shifts served by the bus on that
give a brief description of the UMassDieselNet [8, 9] testbed day by parsing DB sheet.txt. For example, on 3/1/2006,
and the traces collected on this testbed. We then describe bus 3007 serves sub-shifts 21/AM, 21/MID, 21/PM, and
how we pre-process these traces. 21/EVE. With the above two mappings, we translate 55
In the UMassDieselNet bus system consisting of 40 buses, days of the bus-to-bus contacts into contacts between sub-
the bus-to-bus contacts (the durations of which are rela- shifts. A virtual contact is created between two sub-shifts
tively short) are logged. Our experiments are performed on if a bus is handed over from one sub-shift to another. Fig-
traces collected over 55 days during the spring 2006 semester ures 10(a) and 10(b) show the distribution of all contacts
with weekends, spring break, and holidays removed due to over a day and the distribution of the contact duration at
reduced schedules. The bus system serves approximately the sub-shift level.
ten routes. There are multiple shifts serving each of these The discretized probabilistic contacts between any pairs of
routes. Shifts are further divided into morning (AM), mid- shifts is then generated from the 55 days of sub-shift based
day (MID), afternoon (PM), and evening (EVE) sub-shifts. contacts. Examples of the discretized probabilistic contacts
Drivers choose buses at random to run the AM sub-shifts. between two pairs are shown in Figures 2(a) and 2(b). We
At the end of the AM sub-shift, the bus is often handed over set the time slot to be one minute. If in the trace of particu-
to another driver to operate the next sub-shift on the same lar day, two sub-shifts have one or more contact during the
route or on another route. Unfortunately, the all-bus-pairs same time slot as a discretized probability contact, then the
contacts provided in the original traces show no discernible contact probability of the discretized probability contact is
1
pattern. Significant effort is needed to obtain the contacts increased by 55 . With the discretized probabilistic contacts,
at a sub-shift level which do exhibit periodic behavior. we can generate the probabilistic time-space graph and the
We obtain the sub-shift level contact by the following probabilistic state-space graph for RCM, the inter-meeting
steps. Each sub-shift has a fixed starting time (TIME AT time for Spray&f ocus∗ , and the next meeting probability f
GARAGE) and a fixed ending time (DRVR CHNG) every- for M axP rop∗ .
day. We obtain a mapping from sub-shifts to these times by Messages are generated from every node (sub-shift) to ev-
 1 2

0.8 1.8

Delay (hours)
Delivery ratio
0.6 1.6

0.4 1.4

Spray&wait
0.2 Spray&focus* 1.2 Spray&focus*
MaxProp* MaxProp*
RCM RCM
0 1
0 50 100 150 200 0 50 100 150 200
Minutes Minutes

(a) Delivery rate (Miami). (b) Delay (Miami).

1 50

0.8 40

Delay (hours)
Delivery ratio
0.6 30

0.4 20

Spray&wait
Figure 12: Our synthetic bus traces are generated 0.2 Spray&focus*
MaxProp*
10 Spray&focus*
MaxProp*
from metro maps: Miami (left) and Madrid (right). 0
RCM
0
RCM
0 50 100 150 200 250 300 350 0 50 100 150 200 250 300 350
Minutes Minutes

(c) Delivery rate (Madrid). (d) Delay (Madrid).

ery other nodes every 10 minutes. The TTL of each message
is 3 days. In different group of simulations, messages are cre-
Figure 13: Comparison of different performance
ated with 1 ticket and 3 tickets respectively. Figures 11(a) to
metrics using our synthetic bus traces.
11(b) show that RCM outperforms all other protocols except
Epidemic in terms of delivery rate by approximately 5-15%.
Similarly, Figure 11(c) and 11(d) shows that RCM outper-
forms all other protocols except Epidemic in terms of delay motion cycle of its route, the bus must stop in its first station
by approximately 5-10%. The insignificant improvement is for the rest of the time. These synthetic bus contact traces
because, the contact between the nodes are highly irregu- and more traces generated from the other maps in the future
lar in these traces. As we see in the simulation performed will be available for the research community.
on our synthetic bus traces, the improvement is much more As shown in Figures 13(a) and 13(b), in the Miami traces,
significant when bus stations are also considered as nodes. RCM and M axP rop∗ deliver all the messages, while
From Figures 11(e) and 11(f), we can see that RCM has Spray&f ocus has a delivery rate of 85%. The delay of RCM
a small overhead compared to Epidemic, but it is on par is 15% less than those of M axP rop∗ and Spray&f ocus.
with all the other protocols. Note that in Figures 11(c) to On the Madrid traces which is much larger, RCM has a
11(f) the delay and hop-count of the messages are averaged significant (more than 40%) of improvement over all the
over the messages that are delivered by all protocols except other protocols, and the delay is only about 40% of that of
Spray&wait. Spray&f ocus. Though the delay is a little larger than that
of M axP rop∗ , considered together with the delivery rate,
4.4 Synthetic bus trace RCM shows a clear superiority over the other protocols.
In this section, we will compare RCM and the other pro-
tocols with two more synthetic bus traces we generated. In- 4.5 Summary of simulation
spired by the NUS student contact trace, we generate two To sum up, RCM outperforms the enhanced versions of
sets of synthetic bus traces from maps of the metros found previous routing protocols with historical connectivity infor-
in the Internet as shown in Figure 12. We develop a tool to mation in terms of both delivery rate and delay. As shown
extract the routes from the maps and then on each route we by the simulation results, RCM has more improvement over
simulate a number of traveling buses. the other protocols when the contact pattern between the
Each bus travels a route starting from one of the station nodes are more regular as in the NUS student trace and
on the route. When several buses travel on the same route, our synthetic bus trace where both bus and bus station are
their starting stations are disperse evenly along the route. nodes. RCM also shows an increasing relative improvement
If the stations of a route arranged as a line instead of a in its performance when the size of the network and the
circle, the motion cycle of the buses in this route is the complexity in the contact pattern increase.
round trip time on this route. The minutes each bus takes
to travel between two consecutive stations is in a Poisson
distribution with mean being 5 minutes. The time that a 5. RELATED WORKS
bus stay at a station is 1 minutes. Buses are in contact In [2], Jain, Fall, and Patra presented a comprehensive
with each other when they are in the same station. Unlike investigation on the DTN routing problem with different
the UMassDieselNet, both stations and buses are nodes in levels of prior knowledge about the network. Specifically,
our synthetic trace. We assign motion cycles to the buses on Dijkstra’s algorithm (with future connectivity information)
each route according to the number of station on that route. or the linear programming approach (when with informa-
For the Miami map, the motion cycles for different routes are tion of future connectivity, traffic demands, etc.) is used
60, 120, and 240 minutes. For the Madrid map, the motion to obtain an optimal path between a source and a destina-
cycles are 30, 120, 180, and 360 minutes. When a bus finish tion. In [3], Merugu, Ammar, and Zegura proposed a DTN
a round trip on its route but it takes a time less than the routing algorithm that is similar in spirit to Dijkstra’s algo-
rithm in [2]. In [7], Liu and Wu propose a model for DTNs [9] X. Zhang, J. F. Kurose, B. Levine, D. Towsley, and
with repetitive mobility. A hierarchical routing is further H. Zhang. Study of a Bus-Based Disruption Tolerant
proposed to make routing in such DTN models scalable. Network: Mobility Modeling and Impact on Routing.
Epidemic routing [16] is the first flooding-based routing In Proc. of ACM MobiCom, 2007.
algorithm. Gossip [4], forwards with probability p. Oppor- [10] R. Bellman. Dynamic Programming. Princeton
tunistic routings, such as [5], forward based on the delivery Univerity Press, 1957.
probability. Different delivery probability metrics are pro- [11] R. A. Howard. Dynamic Programming and Markov
posed including encounter frequency [5], time elapsed since Processes. The MIT Press, 1960.
last encounter [8, 18, 19, 23, 24], social similarity [6], loca- [12] B. Bonet and H. Geffner. Faster Heuristic Search
tion similarity [25], geometric distance [20], etc. Algorithms for Planning with Uncertainty and Full
Traces data [22] available for the research community in- Feedback. In Proc. of IJCAI, 2003.
clude UMassDieselNet trace, the NUS student contact trace, [13] P. Dai and J. Goldsmith. Topological Value Iteration
etc. In [26], several opportunistic routing algorithms are Algorithm for Markov Decision Processes. In Proc. of
simulated in large realistic contact traces. A timely-contact IJCAI, 2007.
probability metric is proposed in this paper which captures
[14] A. J. Briggs, C. Detweiler, D. Scharstein, and
the contact frequency of mobile nodes and is similar to [5]
A. Vandenberg-Rodes. Expected Shortest Paths for
and [8] in spirit.
Landmark-Based Robot Navigation. International
Journal of Robotics Research, 23(7-8):717–728,
6. CONCLUSION July-August 2004.
In this paper we have presented our first research inves- [15] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and
tigating a new cyclic delivery probability metric, expected C. Stein. Introduction to Algorithms, Second Edition.
minimum delay (EMD), and provided methods to achieve it The MIT Press, 2001.
in a cyclic MobiSpace. Our proposed probabilistic routing, [16] A. Vahdate and D. Becker. Epidemic Routing for
routing in cyclic MobiSpace routing (RCM) is evaluated and Partially-connected Ad Hoc Networks. In Technical
compared with the enhanced versions of some existing DTN Report, Duke University, 2002.
routing protocols using the NUS student trace, the UMass- [17] T. Spyropoulos, K. Psounis, and C. Raghavendra.
DieselNet trace, and our synthetic traces. The simulation Spray and Wait: An Efficient Routing Scheme for
results demonstrate that RCM outperforms the compared Intermittently Connected Mobile Networks. In Proc.
protocols in terms of delivery rate and delay. of ACM WDTN, 2005.
Due to space limitation, we left some performance analysis [18] T. Spyropoulos, K. Psounis, and C. Raghavendra.
of our protocol to the future work, which includes the impact Spray and Focus: Efficient Mobility-Assisted Routing
of size of time slots on the performance of RCM. We will also for Heterogeneous and Correlated Mobility. In Proc.
create more synthetic traces with different maps of metro of IEEE PerCom, 2007.
routes and generalize the conditions where RCM can provide
[19] A. Balasubramanian, B. N. Levine, and
most significant improvement.
A. Venkataramani. DTN Routing as a Resource
Allocation Problem. In Proc. ACM SIGCOMM, 2007.
7. REFERENCES [20] U. Acer, S. Kalyanaraman, and A. Abouzeid. Weak
[1] V. Cerf, S. Burleigh, A. Hooke, L. Torgerson, State Routing for Large Scale Dynamic Networks. In
R. Durst, K. Scott, K. Fall, and H. Weiss. Proc. of ACM MobiCom, 2007.
Delay-tolerant Network Architecture. In Internet draft: [21] V. Srinivasan, M. Motani, and W. T. Ooi. Analysis
draft-irrf-dtnrg-arch.txt, DTN Research Group, 2006. and Implications of Student Contact Patterns Derived
[2] S. Jain, K. Fall, and R. Patra. Routing in a Delay from Campus Schedules. In Proc. of ACM MobiCom,
Tolerant Network. In Proc. of ACM SIGCOMM, 2004. 2006.
[3] S. Merugu, M. Ammar, and E. Zegura. Routing in [22] CRAWDAD data set. Downloaded from
Space and Time in Network with Predictable http://crawdad.cs.dartmouth.edu/.
Mobility. In Technical report: GIT-CC-04-07, College [23] H. Dubois-Ferriere, M. Grossglauser, and M. Vetterli.
of Computing, Georgia Tech, 2004. Age Matters: Efficient Route Discovery in Mobile Ad
[4] J. Haas, J. Y. Halpern, and L. Li. Gossip-Based Ad Hoc Networks Using Encounter Ages. In Proc. of
Hoc Routing. In Proc. of IEEE INFOCOM, 2002. ACM MobiHoc, 2003.
[5] A. Lindgren, A. Doria, and O. Schelen. Probabilistic [24] M. Grossglauser and M. Vetterli. Locating Nodes with
Routing in Intermittently Connected Networks. Ease: Last Encounter Routing in Ad Hoc Networks
Lecture Notes in Computer Science, 3126:239–254, through Mobility Diffusion. In Proc. of IEEE
August 2004. INFOCOM, 2003.
[6] D. Elizabeth and H. Mads. Social Network Analysis [25] J. Leguay, T. Friedman, and V. Conan. DTN Routing
for Routing in Disconnected Delay-Tolerant MANETs. in a Mobility Pattern Space. In Proc. of ACM WDTN,
In Proc. of ACM MobiHoc, 2007. 2005.
[7] C. Liu and J. Wu. Scalable Routing in Delay Tolerant [26] L. Song and D. F. Kotz. Evaluating Opportunistic
Networks. In Proc. of ACM MobiHoc, 2007. Routing Protocols with Large Realistic Contact
[8] J. Burgess, B. Gallagher, D. Jensen, and B. N. Levine. Traces. In Proc. of ACM CHANTS, 2007.
MaxProp: Routing for Vehicle-Based
Disruption-Tolerant Networking. In Proc. of IEEE
INFOCOM, 2006.