Contention aware mobility prediction

routing for intermittently connected mobile

networks

Ahmed Elwhishi, Pin-Han Ho & Basem

Shihada

Wireless Networks
The Journal of Mobile Communication,
Computation and Information

ISSN 1022-0038
Volume 19
Number 8

Wireless Netw (2013) 19:2093-2108

DOI 10.1007/s11276-013-0588-7

1 23
Your article is protected by copyright and all
rights are held exclusively by Springer Science
+Business Media New York. This e-offprint is
for personal use only and shall not be self-
archived in electronic repositories. If you wish
to self-archive your article, please use the
accepted manuscript version for posting on
your own website. You may further deposit
the accepted manuscript version in any
repository, provided it is only made publicly
available 12 months after official publication
or later and provided acknowledgement is
given to the original source of publication
and a link is inserted to the published article
on Springer's website. The link must be
accompanied by the following text: "The final
publication is available at link.springer.com”.

1 23
 Author's personal copy
Wireless Netw (2013) 19:2093–2108
DOI 10.1007/s11276-013-0588-7

Contention aware mobility prediction routing for intermittently

connected mobile networks
Ahmed Elwhishi • Pin-Han Ho • Basem Shihada

Published online: 26 April 2013

 Springer Science+Business Media New York 2013

Abstract This paper introduces a novel multi-copy 1 Introduction

routing protocol, called predict and forward (PF), for delay
tolerant networks, which aims to explore the possibility of DTNs [1] are characterized by the lack of end-to-end paths
using mobile nodes as message carriers for end-to-end for a given node pair for extended periods of time,
delivery of the messages. With PF, the message forwarding which demonstrates a completely different design scenario
decision is made by manipulating the probability distribu- from that for the conventional mobile ad-hoc networks
tion of future inter-contact and contact durations based on (MANETs) [2]. Due to the intermittent connections in
the network status, including wireless link condition and DTNs, a node is allowed to buffer a message and wait until it
nodal buffer availability. In particular, PF is based on the finds an available link to the next hop that will be able to
observations that the node mobility behavior is semi- store the message. This process is repeated until the message
deterministic and could be predicted once there is sufficient reaches its destination. This model of routing constitutes a
mobility history information. We implemented the pro- significant departure from that employed in MANETs. This
posed protocol and compared it with a number of existing new aspect of routing is usually referred to as encounter-
encounter-based routing approaches in terms of delivery based, store-carry-forward, or mobility-assisted routing, due
delay, delivery ratio, and the number of transmissions to the fact that nodal mobility serves as a significant factor
required for message delivery. The simulation results show for the forwarding decision of each message.
that PF outperforms all the counterpart multi-copy Depending on the nodal mobility, the routing strategies
encounter-based routing protocols considered in the study. in DTNs can be classified into three categories: precise
(scheduled) contact routing, opportunistic contact routing,
Keywords Utility-based routing  DTNs  Mobility and predicted contact routing.
In scheduled routing strategies, the future network’s
state/topology is predictable, hence allowing message for-
warding to be scheduled ahead of time. The space-time [3],
Electronic supplementary material The online version of this
article (doi:10.1007/s11276-013-0588-7) contains supplementary modified shortest path [4], and interplanetary internet [5]
material, which is available to authorized users. are examples of scheduled routing schemes.
In opportunistic routing, there is no information avail-
A. Elwhishi (&)  P.-H. Ho
able about contact time or place, and the connection is
Department of Electrical and Computer Engineering,
University of Waterloo, Waterloo, Canada simply established at every contact opportunity, such as
e-mail: [email protected] flooding-based schemes [6, 7] that distribute unlimited
P.-H. Ho numbers of copies throughout the network, or controlled
e-mail: [email protected] flooding schemes [8, 29, 30] that distribute only a subset of
message copies. These schemes are considered the main
B. Shihada
techniques for opportunistic routing. This type of routing
Computer, Electrical, Math Sciences and Engineering,
KAUST, Thuwal, Saudi Arabia is considered simple and suitable for the random nodal
e-mail: [email protected] mobility scenarios, at the expense of extensively consumed

123
 Author's personal copy
2094 Wireless Netw (2013) 19:2093–2108

resources as well as the possible performance degradation results and comparisons with the other counterparts. Sec-
when the nodal mobility is not random. tion 8 concludes the paper.
In predicted routing, the prediction of a contact of each
node pair is predicted based on some statistical information
such as the last time of meeting, the frequency of meetings 2 Related work
among nodes, the number of previous contacts, or the
probability of meeting between nodes in some locations. The previously reported encounter-based routing protocols
The techniques that employ this kind of mobility exploi- have focused on the node mobility which is exploited and
tation are also referred to as utility-based routing [9–14, taken as the dominant factor in the message forwarding
31]. The decision of message forwarding to the encoun- decision [28]. Those schemes contributed to the research
tered node is simply based on a developed utility function. community by introducing novel interpretations of the
Although utility based routing schemes have improved observed node mobility in the per-node utility function.
in terms of performance, the previously reported schemes Spyropoulos et al. in [15, 25] developed routing strategies
are subject to respective problems and implementation using different utility routing metrics based on nodal
difficulties. Most importantly, these schemes inevitably mobility statistics, namely Most Mobile First (MMF), Most
take a large amount of transmission bandwidth and nodal Social First (MSF) and Last Seen First (LSF). S. Nelson
memory space, which could easily dominate the network et al. [16] proposed an enhanced version of MSF, where
resource consumption [15]. Further, they suffer from con- the number of message replicas that are transferred during
tention in case of high traffic loads, in which packet drops a contact is proportional to per-node utility function based
could result in a significant degradation of performance and on the evolution of the number of encounters a node has
scalability. Note that the future DTNs may operate on a during a time-window. A. Lindgren et al. in [9] introduced
vast number of miniature and hand-held devices such as a routing technique in DTNs which takes advantage of the
smart phones, tablet computers, personal digital assistants predicted encounter probability between nodes. Jones et al.
(PDAs), and fixed /mobile sensors, which are subject to a in [17] introduced a utility function for DTN routing which
stringent limitation on power consumption and computa- manipulates the minimum expected inter-encounter dura-
tion resources. tion between nodes. Ling et al. in [18] designed a feedback
The paper introduces a novel DTN routing protocol, adaptive routing scheme based on the factors solely
called Predict and Forward (PF), aiming to overcomes the determined by the node mobility, where a node with higher
shortcomings of the previously reported utility schemes. mobility is given a higher factor, and messages are trans-
The main features of the proposed protocol are: (1) mitted through nodes with higher influence factors.
employing an efficient prediction of contact based on the A. Balasubramanian et al. in [10] considered statistics of
probability distribution of meeting between nodes at dif- available bandwidth and the number of message replicas
ferent network partitions, (2) capability to adapt to fluctu- currently in the network in the derivation of the routing
ating network status, traffic patterns /characteristics, and metric to decide which message to replicate first among all
user behaviors, so as to reduce the number of transmis- buffered messages in a node.
sions, message delivery time, and increase delivery ratio. Another scheme is called delegation forwarding [13,
This is achieved by jointly considering node mobility sta- 19], where a custodian node forwards a message copy to an
tistics, congestion, and buffer occupancy, which are sub- encountered node that has a better chance to ‘‘see’’ the
sequently fused in a novel quality-metric function. In destination. The key idea is that a custodian node (source
specific, the link availability and buffer occupancy statis- or relay) forwards a message copy only if the utility
tics are obtained by sampling the channels and buffer space function (represented by the rate of encounters between
during each contact with another node. The developed node pairs) of the encountered node is higher than all the
quality-metric function facilitates decision making for each nodes so far ‘‘seen’’ by a message, and then the current
active data message, resulting in optimized network per- custodian updates its utility value of that message equal to
formance. We will show via extensive simulations that the that of the encountered node.
proposed PF can achieve a significant performance gain The idea of routing using a node mobility profile is
over the previously reported counterparts. introduced by S. Merugu et al. in [3]. It is assumed that
The rest of the paper is organized as follows. Section 2 each node has a set of usually visited locations, referred to
gives a review over the related work. Section 3 describes as the hub list of the node. The introduced protocol routes a
the system model. Section 4 introduces the method of message to one or more locations that have been visited by
estimating the contact probabilities. Section 5 introduces the destination when either the current message custodian
the proposed PF in detail. Section 6 provides the mathe- or the encountered node visits these hubs. The contact
matical analysis. Then, Section 7 provides the simulation probability between two nodes is computed based on all

123
 Author's personal copy
Wireless Netw (2013) 19:2093–2108 2095

nodes’s hub-visited probabilities. By assuming that each any two nodes located in a community within the trans-
node knows its next hub, a node forwards a message to a mission range of each other can establish a contact to
number of its neighbors with a higher probability to visit a exchange messages. Communities are also assigned unique
hub visited by the destination. Each node is assumed to IDs. Each community is attached to a super node, which
know every other node mobility, and can compute the keeps all contact probabilities of the nodes that belong to
contact probabilities with every other node. Such contact the community. It also keeps track of the contact proba-
relation is further abstracted into a weighted graph, wherein bility of any roaming node visiting its community. We
k-shortest path algorithm is applied to find a list of paths to assume a node can be associated with no more than one
every other destination. The method in [3] is an effective community at a time and is aware of which community it is
approach to deal with the considered scenario. However, located.
this mechanism of forwarding is subject to high overhead, In this mobility model, the geographical service area is
and the assumption of knowing the complete topology partitioned into m communities.We consider only in-com-
regarding the nodal mobility is not realistic in most cases. munity meetings among nodes, while on-way meetings are
In the same line of [3], Q. Yuan et al. in [12] introduced negligible to ensure sufficient meeting time for completing
the idea of the contact probability as the main factor of the message forwarding. By assuming the node future
message forwarding. They introduced a single-copy rout- location is independent of its past location given its current
ing scheme called predict and relay (PER) in the scenario location, and each node has an exponential residence time
where the node movement, including the information on for staying in a community or roaming between commu-
when the contact will occur and how long it will last, is nities, the nodal mobility can be modeled as a continuous
semi-deterministic and can be predicted provided with time Markov chain.
sufficiently long mobility history. By using a time-homo- Specifically, the residence time of node A in a com-
geneous semi-Markov process model, PER enables each munity Ci in each visit is an exponential random variable
node to keep the complete network topology, by which an with parameter kliA . When Node A leaves Community Ci, it
optimal performance can be achieved. Obviously, such a may move to any other community Cj, with a probability of
scheme introduces extremely high computation complexity PACij . The roaming time of node A from a community Ci to
and operational overhead.
any other community Cj is exponentially distributed with
To countermeasure the deficiency of the previously
parameter kri;jA . Denote the location state of a mobile node
reported designs, a number of studies have been reported
[13, 15, 20–22]. Nonetheless, they are subject to various A by its current community, then the user mobility model
limitations due to the simplified assumptions related to the can be characterized by a one-dimensional continuous-time
nodal mobility scenarios [12, 23], or the utility function Markov chain, with a location state space given by
presentation [11, 14, 16]. More importantly, the channel C1 ; C2 ; . . .; Cm . The user movement model over the net-
capacity and buffer occupancy states have not completely work coverage area is described by the transition matrix
been considered as factors in the derivation of utility M of the Markov chain, given by PACij , which is the con-
functions. The impact of these two factors is minimal at ditional probability that a mobile node enters community
low encounter frequencies and low traffic loads, but it is Cj given that it is leaving its current community Ci. For any
P
expected to become more significant as the network community Ci, we have j PACij ¼ 1. The transition prob-
resource availability serves as a bottleneck in the message ability matrix depends on the geographical characteristics
forwarding. of the service area and the network environment under
consideration.

3 System model

As addressed in literature [10, 14, 24], many collected real 4 Estimation of contact probabilities
world traces demonstrated that the nodes follow a semi-
deterministic trajectory, or have some patterns in their Define inter-meeting time between a pair of nodes as the
movements. duration of time instant at which two nodes move out of
To characterize the mobility pattern, in this work we each other’s transmission range until the instant that they
consider a partitioned community-based DTN with super move within each other’s transmission range the next time.
node [23]. In this model, a finite number of mobile nodes Define meeting time as the duration of time in which two
roam freely between a set of communities, each repre- nodes remain in contact until they move out of transmis-
senting a MANET. A community is defined as a group of sion range of each other. Define node inter-arrival time for
nodes who can communicate directly with each other, i.e., a community as the duration from the instant that a node

123
 Author's personal copy
2096 Wireless Netw (2013) 19:2093–2108

departs from a community to the instant that the node arrival time of node A to community Ci is exponential with
arrives back at the same community. parameter kiA pAi , and its mean is
Our objective is to analyze the aforementioned node EiA ¼ 1=ðkiA pAi Þ ð4Þ
mobility model so as to obtain an accurate estimate of the
contact probability between each node pair, where the Inter-meeting time (theory): Nodes A and B meeting at
nodal inter-meeting and meeting times are characterized Ci can occur in two scenarios: (1) node A moves to Ci while
and focused. In the following, we derive the distribution of node B already resides in Ci; (2) node B moves to Ci while
the node inter-arrival time for a community and the dis- node A already resides in Ci. Considering scenario (1), the
tribution of the inter-meeting time. number of meetings between the two nodes at Ci is the
Inter-arrival time (theory): Define N(t) as the number of fraction of node A arrivals to community Ci while node B is
all visited communities by time t for node A. Then, N(t) is a residing there. According to the theory of the inter-arrival
Rt time and the fact that node B resides in community Ci with
non-homogeneous Poisson process with mean ¼ ki ðsÞds. probability pBi , the number of effective meetings between
This mean is the sum of the expected residence and A and B at Ci when node A makes the movement follows
roaming times for node A while it roams from one com- Poisson process with mean kiA pAi pBi t. Hence, the inter-
munity to another. Define Ni(t) as the number of visits by meeting time between A and B at community Ci when node
node A to community Ci by time t. Then, the mean arrival A makes the movement is an exponential random variable
Pn with parameter kiA pAi pBi . Similarly, in scenario (2), the
time to the nth community is EAn ¼ i¼1 ðEi ½Tc þ Eij ½TrÞ,
where Ei[Tc] and Eij[Tr] represent the average residence inter-meeting time between A and B at Ci when node
time at community Ci, and the average roaming time when B makes the movement is an exponential random variable
node A move from community Ci to community Cj, with parameter kiB pBi pAi . As a result, the inter-meeting
respectively. In specific, node A stays at Ci for an expo- time between A and B at Ci is a random variable that is the
nential time with parameter kliA ¼ Ei ½Tc
1
and roam with minimum of the two independent exponential random
variables, which follows an exponential distribution with
probability PAci,j to another Cj for an exponential time with parameter (kiA pAi pBi þ kiB pBi pAi ). Considering all network
parameter kri;jA ¼ Ei ½Tr
1
, then the rate that node A makes a partitions, the inter-meeting time between A and B is a
transfer from state i to state j is calculated as random variable with a distribution of the minimum of the
two nodes inter-meeting times at all the network partitions,
1
ki;jA ¼ ð1Þ which is an exponential random variable with parameter
1=ðkliA Pci;jA Þ þ ð1=kri;jA Þ given by
The expected inter-arrival rate of node A at Ci is Xm
calculated as ðk þ kiB ÞpAi pBi
i¼1 iA
ð5Þ
!
X n
1 Noticeably, the above relation is largely based on the
kiA ¼ 1= k
ð2Þ amount of meeting time intervals between nodes, which is
k¼1 ki;jA
in turn affected by the status and the duration of the status
where n is the total number of visited communities by node of the wireless channel and buffer. These contact
A until it come back to Ci. components are time varying and could change
Since the continuous-time Markov chain of node A is dramatically from time to time. The short period of
irreducible, the limiting probabilities exist and can be contact time or the unavailability of the channel or the
P
obtained by solving pAQA = 0, and pAi ¼ 1, where buffer space may cause unsuccessful message exchange.
i Thus, two factors should be considered in (5): (1) the
Q represents the generator matrix, which is given by impact the contact time duration. This factor is introduced
8
>
< P ki;jA if i 6¼ j as the probability that node A will leave the neighborhood
Qi;jA ¼  ki;jA if i ¼ j ð3Þ of community Ci on or before time unit k, denoted as SAi (k).
>
: j Note that SAi (k) also indicates the distribution of the sojourn
0 otherwise
time at Ci for node A regardless of the next community.
pAi is the limiting probability representing the fraction The time parameter(k) is used as a relative time offset. (2)
of time in which node A resides in community Ci. The probability in which the channel is free and the buffer
Thus, the parameter of inhomogeneous Poisson process, at the encountered node B is not full at Ci, denoted as, Pfi .
N(t), can be presented as kiA pAi t. As a result, the inter- This probability represents the probability of performing a

123
 Author's personal copy
Wireless Netw (2013) 19:2093–2108 2097

successful message exchange when node A encounters

node B. By injecting the two parameters, (5) is reformu-
lated as
Xm
ðk þ kiB ÞpAi pBi SAi Pfi
i¼1 iA
ð6Þ

Consider two nodes, A and B. Let PAB denote the

probability that a contact occurs between the two nodes,
given the expected residence time of both nodes in the
network Rm. The probability of a contact based on the
inter-meeting time between the nodes is calculated as
Pm A
PAB ¼ 1  e i¼1 ðkiA þkiB ÞpAi pBi Si Pfi Rm ð7Þ
Fig. 1 Transition matrix of a node for three communities
The contact probability between a super node H, node
A is calculated as node A, PACij is calculated by observing all the movement
kiA pAi SAi Pfi Rm frequency of node A from Ci to Cj, which is calculated as
PAH ¼ 1  e ð8Þ
The aforementioned relations represent the utility numAij
pACij ¼ ð9Þ
function maintained at each node regarding every other numAi
node and every super node in the network. where numAi is the number of all transitions from com-
munity Ci to any community, and numAij is the number of
4.1 Deriving the utility function parameters transitions from community Ci to specific community Cj.
By keeping track of numAi and numAij , each node could
To calculate the probability of contact (the utility function) generate and refine its own PCij matrix over time.
PAB, the proposed algorithm PF needs to compute the
following parameters: (1) the transition probability matrix,
4.1.2 Probability of sojourn time
Pn, (2) the sojourn time probability distribution matrix,
SAi (k), (3) the fraction of time in which node A resides in
The probability distribution SAi (k), represents the proba-
community Ci ; pAi , which represents the probability dis-
bility that node A will move from community Ci to a
tribution at the steady state, (4) the probability of making
community Cj at or before time k. SAi (k) is based on cal-
successful data exchange between nodes Pf, (5) the com-
culating all residual times node A has spend at community
munity inter-arrival rate matrix, kiA , and (6) the average
Ci and followed by transition to a community Cj, which is
residence time of node A in the network, Rm. These
defined as following:
parameters are retrieved from the nodal contact statistics
history. SAi ðkÞ ¼ PðtijA \kÞ ð10Þ

where iAij is the sojourn time at Ci when Cj is the next

4.1.1 Transition probability matrix visited community. The parameter k is employed as a
relative time offset. This distribution is mainly based on the
The illustration of the transition probability matrix, PAC, of mobility history of a node, which becomes more accurate
the embedded Markov chain for node A that visits three when the network reaches steady state. The probabilities
communities is shown in Fig. 1. At any of those commu- P(tm
ij \ k) for any node A can be obtained by measuring all
nities, the node could choose to stay for a while, or move to times tAij whenever node A moves from community Ci to
another community according to its preferred probability. community Cj. For example, assume that node A has eight
For example, if the node is at the C1, it can either (1) measurements for tAij , which are 1, 2, 5, 5, 4, 6, 7,3.
move to the C2 with probability pAC12 , (2) or stay in C1 with P(tm
ij \ 5) is 1/2 by using the following general formula
the probability pAC11 , (3) or go to C3 with probability pAC13. X
k1
Those mobility probabilities constitute the transition PðtijA \kÞ ¼ PðAm
ij ¼ nÞ ð11Þ
probability matrix PACij . Note that each node has its own n¼0

transition probability matrix at each community that This probability is used to estimate the expected contact
reflects its movement history. The transition probability of time to transfer a message from node A to node D.

123
 Author's personal copy
2098 Wireless Netw (2013) 19:2093–2108

4.1.3 Probability of steady state Whenever node A encounters node B at a community,

node A calculates and maintains the Pf value regarding
The fraction of time node A resides in community i; pAi , is node B of that community (i.e, different Pf values at
calculated as different communities). To handle the fluctuation of
P A channel state, when a node roams from one community
t
pAi ¼ Pi iA ð12Þ and enters new one, it broadcasts its latest statistics
j tj
regarding the channel condition for the community it was
where tAi is the duration of time that node A spend at each in. Then the other nodes update their utility function
visit to community Ci, and tAj is the total time node accordingly.
A spends in all communities at each visit.
4.2 Inter-nodal information exchange
4.1.4 Community inter-arrival time
To estimate the contact probability under the proposed PF
The community mean inter-arrival rate, kiA , is obtained by algorithm, every node or super node needs to keep track of
calculating the residual and roaming times taken by node the contact statistics with every other node in the network.
A while moving from one community to another until it Specifically, the contact statistics is defined as a 7-tuple
returns back to the same community. Then, (2) is applied to \nodeID, P, S, k, p, Pf, Rm [, where P is the transition
calculate kiA . probability matrix, S is the sojourn time probability dis-
tribution matrix, k is the inter-arrival rate at each com-
4.1.5 Probability of successful data exchange munity, p is the fraction of time in which node A resides in
community i, and Pf is the probability of the channel
As discussed earlier, the probability of making successful availability in each community during a contact. All the
data exchange during a contact of any node pair, Pf, is aforementioned parameters become more valid when the
mainly based on the channel state and the available buffer network approaches the steady state except for the proba-
space, which is in turn largely based on the available bility of the channel availability, which needs to be fre-
bandwidth and the traffic load in the network. Pf is calcu- quently updated since it is largely affected by the traffic
lated by measuring the fraction of time at which the load. During each contact, the encountered nodes exchange
channel is busy or free during contacts. The statistics of the the information regarding the message destination that they
channel availability is described as follows. The message don’t have in common. With the update, the custodian node
delivery takes place only when the buffer of the encoun- decides whether or not the message should be forwarded to
tered node is not full, and fails otherwise. In case of failure, the encountered node, which is mainly based on the for-
the time in which the buffer is full during a contact is added warding mechanism of PF. This exchange of summary
to the busy time of the channel, i.e., the busy time is vectors is followed by another important update called
composed of the channel busy time and the time in which transitivity update.
the buffer of the encountered node is full during the nodal Transitivity update: The proposed algorithm has transi-
contact. The statistics of the channel and buffer availability tivity property [9], based on the observation that if node
is described as follows. A frequently encounters node B and B frequently encoun-
Let two nodes A and B be in the transmission range of ters node D, then A has a good likelihood to be able to
each other, and each broadcasts a pilot signal per k time forward messages to D through B. Thus, the congestion
units in order to look for its neighbors within its trans- history of buffer and link availability of node B should be
mission range. Let T(A,B), Tfree, and Tbusy represent the total determined accordingly. In order to maximize the contact
contact time, the amount of time the channel is free and the probability, P(A,D) and to make any message destined to
buffer is not full, and the amount of time the channel is node D goes through node B, a proper update using tran-
busy or the buffer is full, respectively, at node A or B. Thus, sitivity property should be made.
the total duration of time in which node A and B can To deal with this maximization problem (maximizing
exchange information is calculated as P(A,D)) using transitivity update, we use the inverse of the
Tfree ¼ TðA;BÞ  Tbusy ð13Þ probability of contact between two nodes. Using this
relation in the transitivity update is simpler and can
Thus, the probability that the link between node A and adaptively update values only when PðAi DÞ \PðB;DÞ in order
node B is available is calculated as following.
to ensure that node A reaches D through B. Otherwise, if
Tfree P(A,B) [ P(B,D), the transitivity property is not useful since
Pf ¼ ð14Þ
TðA;BÞ node A is a better candidate for forwarding messages

123
 Author's personal copy
Wireless Netw (2013) 19:2093–2108 2099

directly to node D rather than forwarding them through 5.1 Forwarding strategy
B. The inverse of probability of contact between nodes
1 The decision of message forwarding in PF is mainly based
A and D, Pr(A,D), is calculated by PðA;DÞ , where P(A,D) is the
average probability of contact between node A and on the utility function value of the encountered node
D. Pr(A,B) and Pr(B,D), are obtained using the similar way. regarding the destination, and the number of message copy
The calculation of the P(A,B) is given by Eq. (7). The new tokens. If more than one message copies are currently
updated parameter, Pr(A,D), is calculated as follows: carried, the weighted copy rule is applied; otherwise the
forwarding rule is applied.
Pnew
rðA;DÞ ¼ wPrðA;DÞ þ ð1  wÞðPrðA;BÞ þ PrðB;DÞ Þ ð15Þ

where w is a weighting factor calculated from 5.1.1 Weighted copy rule

PrðA;BÞ þ PrðB;DÞ The idea of weighted copy rule was first examined in [27]
w¼ ð16Þ
PrðA;DÞ and [11], and has been proved to achieve improved
delivery delay. The proposed PF scheme will also incor-
Note that w must be less than 1; that is Pr(A,D) [ Pr(A,B)
porate with such a mechanism.
? Pr(B,D).
The source of a message initially starts with L copies. In
The new probability of contact is obtained by applying
the event that any node A that has n [ 1 message copy
the following relation
tokens and encounters another node B with no copies with
1 P(B,D) [ P(A,D), node A hands over some of the message
Pnew
ðA;DÞ ¼ ð17Þ
Pnew
rðA;DÞ copy tokens to node B and keeps the rest for itself
according to the following formula:
Pnew
(A,D) represents the new value of Pr(A,D) that is obtained   
from the transitivity update. The introduced transitivity- PðB;DÞ
NB ¼ NA ð18Þ
update rule has great impact on protocol performance. PðB;DÞ þ PðA;DÞ Þ
Note that predicting the future mobility of nodes relies
where NA is the number of message tokens that node A has,
on their trajectory history, which is recorded and dissem-
P(B,D) is the predicted probability of node B encountering
inated throughout the network in an epidemic fashion or
node D, and P(A,D) is the predicted probability of contact
through the supper nodes attached to each community.
between nodes A and D. This formula guarantees that the
largest number of message copies is spread to relay nodes
that have better information about the destination node.
5 Forwarding strategy in PF
After L message copies have been disseminated and carried
by the encountered custodian nodes, each custodian node
PF is a multi-copy DTN routing protocol. At each for-
carrying the message performs message forwarding
warding step, PF selects the next hop with the highest
according to the forwarding rule as described below.
probability of delivery to the destination. If more than one
node is in the transmission range of node A, node A selects
5.1.2 The forwarding rule
the next hope by comparing a probability metric (utility
function) U(x) for all nodes currently in contact with
• If the destination node is one hop away from an
A, denoted as NA, and for itself, x 2 f Ag [ NA . This metric
encountered node, the custodian node hands over the
indicates the delivery performance to the destination if
message to the encountered node and completes the
node A selects node x as the next hop and forwards the
message delivery.
message to x. The current node then selects the next-hop
• If the utility value of the encountered node relative to
h as the node for which the delivery probability metric is
that of the destination node is greater than that of the
maximized
custodian node by a threshold value, Pth, a custodian
h ¼ arg max UðxÞ node hands over the message to the encountered node.
x2fAg[NA
• If A and B are in transmission of super node H, and
If the selected next hop is the current node (h = A), the H has utility value regarding a node C, which is higher
message will not be forwarded. than that of node A or B regarding to D, and C has

123
 Author's personal copy
2100 Wireless Netw (2013) 19:2093–2108

6.1 Delivery probability

In order to calculate the expected message delivery ratio, any

path of message m between S and D is a k - hop simple path,
denoted as l, which is represented by a set of nodes and links
denoted as fS; h1 h2 . . .:hk1 ; Dg, and fe1 ; e2 ; . . .; ek g,
respectively. The cost on each edge, denoted as
{b1, b2,…, bk}, is the inter-meeting rate (or frequency) of
each adjacent node pair along the path. According to the
forwarding policy of PF, the values of inter-meeting rate
should satisfy {b1 \ b2 \ _ \ bk}. The path cost, PRl(t), is
the probability that a message g is successfully forwarded
from S to D along path l within time t, which represents a
cumulative distribution function (CDF). The probability
density function of a path l with k - hop for one message
copy can be calculated as convolution of k probability dis-
tributions [34] which is calculated as:
higher probability of meeting D than A, B, and H. Then
Prl ðtÞ ¼ p1 ðtÞ  p2 ðtÞ  . . .pk ðtÞ ð19Þ
node A gives the message to H.
Theorem 1 Let the probability distribution function
The complete mechanism of the forwarding strategy in PF
(PDF) for the message delivery along a one-hop path
is summarized as shown in Algorithm 1.
i be denoted as pi ðtÞ ¼ bi ebi t . Thus, the PDF for a k - hop
simple path l with an edge cost fb1 ; b2 ; . . .; bk g can be
6 Analytical model of PF
expressed as
In this section a statistical analysis is conducted to evaluate X
kl
ðk Þ
the performance of PF. Without loss of generality, a mobility Prl ðtÞ ¼ Ci l pi ðtÞ ð20Þ
i¼1
model as described in section 7 is employed in the analysis.
The performance measures in the analysis include the where the coefficients are given as follows:
average delivery probability and the message delivery Y
kl
ðkl Þ bj
delay. The analysis is based on the following assumptions. Ci ¼ ð21Þ
b  bi
j¼1;i6¼j j
• Node mobility is independent and heterogeneous,
where nodes have frequent visits to some locations. The proof is provided in ‘‘Appendix of ESM’’.
• Each node in the network maintains at least one
The probability of message delivery on forwarding path
forwarding path to every other node.
l between any source S, and destination D, within expira-
• Each node belongs to a single community at a time.
tion time T is expressed as:
• The inter-meeting time parameter, bAB, between nodes
A and B follows an exponential distribution with proba- ZT
bility distribution function (PDF), PAB ðtÞ ¼ bðA;BÞ :ebðA;BÞ t , Fl ðTÞ ¼ PRl ðTdl \TÞ ¼ Prl ðtÞdt
where t is the time instance. 0

It has been shown that a number of popular user mobility X

kl ZT
models have such exponential tails (e.g., Random Walk, ¼ Cikl Pi ðtÞdt ð22Þ
Random Waypoint, Random Direction, and Community- i¼1
0
based Mobility [32, 33]). In practice, recent studies based X
kl
ðk Þ
on traces collected from real-life mobility examples argued PRl ðTdl \TÞ ¼ Ci l :ð1  ebi T Þ
that the inter-contact time and the contact durations of i¼1

these traces demonstrate exponential tails after a specific If there are L - 1 copies (excluding the message at the
cutoff point [35]. Based on the mobility model of the source) of message m traversing through L - 1 independent
nodes, the distribution of the inter-contact time can be paths in the network, the maximum probability of message
predicted and calculated using (6). Thus, parameter bAB is delivery can be written as
Pm
calculated as bAB ¼ i¼1 ðkiA þ kiB ÞpAi pBi SAi Pfi . PRmax ðTd \TÞ ¼ maxfPRSD ; PR1 ; PR2 ; ::; PRL1 g ð23Þ

123
 Author's personal copy
Wireless Netw (2013) 19:2093–2108 2101

where PRSD and PRl are random variables representing the The above relation gives an upper bound on the delivery
delivery probability in case of direct message delivery delay since it is conditioned to TSD, TSD \1 and can be
between S and D, and through one of L - 1 paths, taken as benchmark.
respectively. The expected delivery probability of message The average delivery delay of message m can be cal-
g with L - 1 copies traversing on L - 1 paths is calculated as: culated intuitively as:
" !#
 1 X
L1
1
Y
L 1
E EDðS;DÞ ¼ TSD þ Tdl : ð30Þ
PRðTd \TÞ ¼ 1  PRSD ðTSD [ T Þ ð1  PRl ðTdl \TÞÞ L l¼1
PRðTd \TÞ
l¼1
ð24Þ TSD is included in (30) only if TSD \1.
! By assuming X totally generated messages in the net-
L1 X
Y kl
ðkl Þ  bi T  work, the average delivery delay can thus be calculated as
PRðTd \TÞ ¼ 1  ebSD T Ci e ð25Þ
l¼1 i¼1
1X X
DR ¼ Dg ð31Þ
By assuming X totally generated messages in the X g¼1
network, the average of the delivery probability in the
network is calculated as 6.3 Validation of analytical model

1X X
In order to evaluate the accuracy of the mathematical
PR ¼ PRg ð26Þ
X g¼1 expressions in this analysis, PF is examined under two
network status scenarios. In the first scenario, the network
6.2 Delivery delay is operating under no congestion, i.e., all the nodes have
infinite buffer space, and the bandwidth is much larger than
Theorem 2 The expected total time required to deliver a the amount of data to be exchanged between any two
message from S to D along an individual path l can be encountered nodes. In the second scenario, the network is
calculated as operating under limited resources, i.e., the forwarding
Z1 X
kl Z1 opportunities can be lost due to high traffic, limited
ðkl Þ
E½Dl  ¼ PRl ðTdl [ tÞ ¼ Ci : ebi t dt bandwidth, limited buffer space, or contention (i.e., more
i¼1 than one node within the transmission range are trying to
0 0 ð27Þ
X
kl access the wireless channel at the same time). For both
ðk Þ 1
E½Dl  ¼ Ci l : scenarios, 50 nodes move in a 1200 m 9 1200 m network
i¼1
bi
divided into 6 communities. The transmission range is set
Let message m have L - 1 copies (excluding the to 30 to enable moderate network connectivity with respect
message at the source) traversing on L - 1 independent to the considered network size. The traffic load is varied
paths. The minimum delivery delay can be written as: from a low traffic load (i.e., 20 messages generated per
DSD ¼ minfTSD; Td1 ; Td2 ; ::TdL1 g ð28Þ node in 30,000 time units) to high traffic load (i.e., 80
messages generated per node in 30,000 time units). A
where TSD and Tdl are a random variables representing the source node randomly chooses a destination and generates
delivery delay through direct path between S and D and messages to it during the simulation time. In this analysis
through one of L - 1 paths, respectively. The expected the message copies are set to 4 (i.e., forming a maximum of
delay of message m, E[DSD], can be calculated as 4 paths).
Examining PF under the two scenarios is very impor-
Z1
tant; in case of no congestion, the best path that is taken by
E½DSD  ¼ PðTd [ tÞ a message is mainly based on the inter-contact and contact
0 time, while under congestion, the message will be buffered
Z1 !
L1 X
Y kl for longer period of time and forced to take longer path to
ðk Þ
¼ ebSD t Ci l :ebi t dt go around the congested area resulting in more dropping
l¼1 i¼1
0 rate and longer delivery delay.
Z1 ! To enable accurate analysis, the simulation program is
1 Y
L1 X
kl
ðk Þ
bSD t
¼ bSD e Ci l :ebi t dt run for a period of time (warm-up period of 10,000 time
bSD
0
l¼1 i¼1 units) such that each node can build and maintain the best
( !) forwarding paths with every other node in the network.
1 Y
L 1 X
kl
ðkl Þ bi TSD
¼ E Ci :e ; TSD \1 ð29Þ These forwarding paths are mainly based on the congestion
bSD l¼1 i¼1 degree (traffic loads values) considered in the analysis. The

123
 Author's personal copy
2102 Wireless Netw (2013) 19:2093–2108

forwarding path is cached by following the trajectories of

the generated messages during the warm-up stage between
every source destination pair in the network. These mes-
sages are forwarded from node to node according to PF
routing mechanism.
In this analysis, we simplified the calculation by limiting
our study to only the best two of forwarding paths among
all other paths and compare the simulation and theoretical
results of delivery ratio and delivery delay. In most cases, a
message takes the best forwarding path that is based on the
inter-contacts history if the network is not congested and
the buffers operate under their capacity limit.
Figures 2 and 3 compare the simulation and theoretical
Fig. 3 The theoretical and simulation results of delivery delay
results in terms of delivery ratio and delivery delay of the
considered scenarios, respectively.
As seen from the figures, when the network resources are 7 Performance evaluation
enough to handle all the traffic loads (Scenario 1), there is no
dramatic change in the obtained delivery ratio and delivery 7.1 Experiment setup
delay for all traffic loads. That is because messages follow
the best forwarding paths that lead to best performance. The To evaluate the performance of the proposed PF scheme,
simulation and analytical plots for PF present close match an adapted DTN simulator based on that in [26] was
and validate the generality of the analytical expressions. implemented. In the simulation, a community based DTN
Additionally, it is evident that (25) and (30) are tight for all model with several predefined communities was estab-
degrees of traffic loads. When the network resources are lished. Nodes usually revolve around these communities.
limited (i.e, scenario 2), the contention and the overhead of That is, the nodes would stay in the neighborhood of a
MAC layer increase, resulting in longer forwarding paths, community, or move to the neighborhood of other com-
higher drop rate, and longer delivery delay. The simulation munities with their own respective probabilities, which is
and analytical plots are still providing a close match with in turn based on their trajectory history. Two nodes can
small divergence in case of high traffic loads. only communicate when they are associated with the same
Although the contention does affect the accuracy of our community. In the simulation, 120 nodes move in a 1200 m x
theoretical expressions, the error introduced for PF is not 1200 m square consisting of 6 identical nonadjacent com-
large (23 %), even for large traffic loads. Therefore, we munities with area of 200 m x 200 m for each community.
believe the analytical expression is useful in assessing the Initially, nodes are uniformly distributed among the com-
performance in more realistic scenarios with contention. As munities, and each node prefers to reside at its home
an evident in these plots, the actual delay obtained by PF community with a longer periods of time than that for any
becomes increasingly worse than what the theory predicts. other community.
This demonstrates the need to add an appropriate conten- Each node can transmit up to a distance K C 0 meters
tion model when it comes to modeling utility-based away, and each message transmission takes one time unit.
schemes. An effort in that direction can be found in [32]. A slotted collision avoidance MAC protocol with Clear-to-
Send (CTS) and Request-to-Send (RTS), was implemented
for contention resolution. A message is acknowledged if it
is received successfully at the encountered node by sending
back a small acknowledgment packet to the sender.
Each node has its own bounds of sojourn time that it
spends at a community, which varies between 5 and 20
time units, with longer periods at its home community. The
bounds of roaming time is defined in the range of 100–300
time unites, which is the time that a node takes to moves
from a community to another. The prediction time window
for the PF algorithm is fixed to 300 time units. The choice
of the time unit mainly depends on the network scenario
and applications under consideration. The PF algorithm has
Fig. 2 The theoretical and simulation results of delivery ratio each node to generate the transition probability matrix

123
 Author's personal copy
Wireless Netw (2013) 19:2093–2108 2103

P, steady state probability p, community inter-arrival rate results of the delivery delay, delivery ratio, and number of
ki, and sojourn time probability distribution Sij(k). Thus, transmissions, respectively. The buffer space was varied
the simulation was run for a ‘‘warm-up period’’ to reach the from 5 (very limited capacity) to 200 (relatively high
steady state with stable values of these parameters. The capacity) messages to reflect the performance of the pro-
simulation lasted for 30,000 time units in each scenario. tocols under the considered traffic load. As expected, due
The final result is the average of twenty runs. to the high traffic volumes, we see a significant impact
The message inter-arrival time is uniformly distributed upon the message forwarding decisions due to the degra-
in such a way that the traffic can be varied from low (20 dation of utility function values caused by buffer overflow.
messages per node in 30,000 time units) to high (70 mes- Note that when the buffer of the encountered node is full,
sages per node in 30,000 time units). The message time to some messages cannot be delivered even though the
live (TTL) is set to 8,000 time units. Each source node encountered node utility value regarding the message
selects a random destination node, and begins generating destination is better than that of the custodian node. This
messages to it during simulation time. situation results in extra queuing delay, especially in the
To examine the performance of PF, it is compared to case when flooding-based schemes are in place. As shown
schemes belonging to the same category (utility-based rout- in the figures, when the buffer size is small (50 messages or
ing). Each one of the considered schemes employs the nodal less) the performance of the protocols is very sensitive to
mobility in a different way. The following schemes reported in the change of buffer capacity.
the literature were implemented in the simulation. It is observed that the PF scheme produced the best
performance in all scenarios, since it considers the situation
• Delegation forwarding (DF) [13]
that a node may have a full buffer, and then degrades the
• Spray and focus (S&F) [15]
corresponding utility metric, producing the best perfor-
• Predict and forward utility-based routing protocol (PF)
mance. In specific, PF yielded a shorter delivery delay than
• Most mobile first (MMF)[25]
that of PROPHET by 170 %, S&F by 59 %, and SARP by
• Self-adaptive routing protocol (SARP) [11]
25 %. PF can achieve a higher delivery ratio than DF by
• PROPHET [9]
76 %, PROPHET by 81 %, S&F by 68 %, and SARP by
For all the protocols, an attempt has been made to tune 19 %. Although PF produced more transmissions than
the parameters in each scenario separately, in order to MMF and DF, it yielded a smaller delivery delay than that
achieve the best transmission-delay performance. The of MMF by 84 %, and DF by 68 %. As the buffer size
utility’s threshold parameter for Spray and Focus, SARP, increased, the performance of all protocols improved
and PF was set to 150, 30, and 0.07, respectively. Simi- especially for MMF and SARP. When the buffer size is
larly, the number of message copies (L) was set 15 for larger than the traffic demand, the SARP scheme has
Spray and Focus, and 10 for SARP and PF. yielded a competitive performance due to the relaxation of
The performance comparison under the considered buffer capacity limitation. PF still yielded the best perfor-
mobility scenarios is in terms of average delivery delay, mance with a smaller number of transmissions than S&F
delivery ratio, and the total number of transmissions per- by 35 %.
formed for all delivered messages.
7.2.2 The effect of traffic load
7.2 Evaluation scenarios
The main goal of this scenario is to observe the perfor-
We analyze the performance implication of the following. mance impact and how PF reacts under different degrees of
First, the performance of the protocols is evaluated with wireless channel contention. The network connectivity is
respect to the low transmission range and varying buffer kept high (i.e., the transmission range is set to as high as
capacity under high traffic load. Second, with respect to the 40 m) under different traffic loads, while channel band-
high-level of connectivity at each community and varying width is set relatively quite small (i.e., one message
traffic load. Third, with respect to moderate-level of con- transfer per unit of time) in order to create congested
nectivity and varying message TTL. environment. We have two scenarios for nodal buffer
capacity: (1) unlimited capacity; and (2) low capacity (15
7.2.1 The effect of buffer size messages). Figures 7, 8, and 9 show the performance of all
the routing algorithms in terms of the average delivery
To examine the effect of buffer sizes upon the performance delay, delivery ratio, and total number of transmissions,
of PF, the transmission range is set to moderate (i.e., K = respectively.
10) and the traffic load is 60 messages generated per node It is observed that PROPHET produced the longest
(relatively high). Figures 4, 5 and 6 show the experiment delivery delay and requires a higher number of

123
 Author's personal copy
2104 Wireless Netw (2013) 19:2093–2108

Fig. 4 The effect of buffer size on delivery delay

Fig. 7 The effect of traffic load under high buffer capacity on
delivery delay

Fig. 5 The effect of buffer size on delivery ratio

Fig. 8 The effect of traffic load under high buffer capacity on

delivery ratio

Fig. 6 The effect of buffer size on number of transmissions

transmissions compared to all the other schemes, thus it is

not included in Fig. 8. PROPHET produced an order of
magnitude more transmissions than that by PF. Fig. 9 The effect of traffic load under high buffer capacity on
As shown in Figs. 7, 8, and 9, when the traffic load is number of transmissions
increased, the available bandwidth is decreased accord-
ingly, which causes performance reduction. When the second best. This is because in MMF, the effect of buffer
traffic load is moderate (i.e., less that 40 messages), it is size is relaxed, which makes nodes buffer an unlimited
clear that the delivery delay is short in all the schemes, number of messages while roaming among communities.
while PF outperforms all other protocols and MMF is the PF can produce delays shorter than that of PROPHET,

123
 Author's personal copy
Wireless Netw (2013) 19:2093–2108 2105

Fig. 12 The effect of traffic load under low buffer capacity on

Fig. 10 The effect of traffic load under low buffer capacity on
number of transmissions
delivery delay

Fig. 11 The effect of traffic load under low buffer capacity on

Fig. 13 The effect of message TTL on delivery delay
delivery ratio

MMF, DF, S&F, and SARP by 360, 54, 430, 270, and achieve as high as 93 %, compared with 90 % by SARP,
59 %, respectively. Regarding the delivery ratio, PF, 87 % by MMF, 77 % by DF, and 88 % by S&F. Even
MMF, S&F, and SARP can achieve excellent performance though DF produced the lowest number of transmissions, it
of 98 %, while the PROPHET routing degrades below is at the expense of the worst delivery delay.
70 % for high traffic loads. DF can achieve delivery ratio As the buffer capacity is low (e.g., 15 messages) and the
above 90 %. traffic load is high, the available bandwidth decreases and
As expected, the performance of all the schemes the buffer occupancy increases accordingly, which
degrades as wireless channel contention is getting higher, degrades the performance of all protocols, especially for
especially when the traffic load exceeds 50 messages per the PROPHET and MMF. It is observed that PROPHET
node during the simulation period. We observed that PF produced the largest delivery delay. It is notable that PF
can achieve significantly better performance compared to outperforms all the multiple-copy routing protocols in
all the other schemes, due to the consideration of busy links terms of delivery delay and delivery ratio under all possible
in its message forwarding mechanism, where the corre- traffic loads. When the traffic load is high, PF yielded
sponding routing-metric is reduced accordingly. This shorter delivery delay than that of SARP by 28 %, MMF
results in the ability of rerouting the contended messages by 53 %, SF by 41 %, DF by 47 %, and PROPHET by
through the areas of low congestion. However, such a re- 233 %. Although PF requires more transmissions com-
routing mechanism makes messages take possibly long pared to the MMF and DF, the number is still smaller than
routes and results in more transmissions than that of MMF that produced by S&F. PF can achieve delivery ratio above
and DF. In summary, the delivery delay obtained by the PF 76 % for high traffic loads, while the SARP, PROPHET,
in this scenario is shorter than that of PROPHET by 330 %, DF, S&F, and MMF degrades by 66, 47, 51, 62, and 55 %,
MMF by 66 %, S&F by 88 %, DF by 233 %, and SARP by respectively. Even though SARP obtains the second best
30 % respectively. Regarding delivery ratio, PF can performance, its mechanism is not capable of dealing with

123
 Author's personal copy
2106 Wireless Netw (2013) 19:2093–2108

increases. This is due to the increased chance of forwarding

messages to more nodes before they get expired or delivered.

8 Conclusion

The paper introduced a novel multi-copy routing scheme,

called PF, for intermittently connected mobile networks that
are possibly formed by densely distributed and hand-held
devices such as smart phones and personal digital assistants.
PF aims to explore the possibility of taking mobile nodes as
message carriers in order for end-to-end delivery of the mes-
sages. The best carrier for a message is determined by the
Fig. 14 The effect of message TTL on delivery ratio prediction result using a novel contact model based on the
probability distribution of future inter-contact and contact
durations, where the network status, including wireless link
condition and nodal buffer availability, are jointly considered.
We compared PF with a number of counterparts via extensive
simulations. It was shown that PF can achieve shorter delivery
delays than all the existing spraying and flooding based
schemes when the network experiences considerable con-
tention on wireless links and/or buffer space. The study pro-
vides a significance that when nodal contact does not solely
serve as the major performance factor, the DTN routing per-
formance can be significantly improved by further consider-
ing other resource limitations in the utility function and
message weighting/forwarding process.

Fig. 15 The effect of message TTL on number of transmissions

References
the uncertainty of the buffer occupancy status and channel
1. DTN Research Group (DTNRG). http://www.dtnrg.org.
condition, resulting in degradation in its performance.
2. Broch, J., Maltz, D. A., Johnson, D. B., Y.-C. Hu, Y. -Y, & Jet-
Figures 10, 11, and 12 show the performance of all tech- cheva, J. (1998). A performance comparison of multi-hop wireless
niques in terms of delivery delay, delivery ratio, and ad hoc network routing protocols. In ACM MobiCom (pp. 85–97).
number of transmissions, respectively, under this scenario. 3. Merugu, S., Ammar, M. H., & Zegura, E. W. (2004). Routing in
space and time in networks with predictable mobility. Technical
Report, Georgia Institute of Technology.
7.2.3 The effect of message TTL 4. Jain, S., Fall, K., & Patra, R. (2004). Routing in a delay tolerant
network. In ACM SIGCOMM (pp. 145–157).
5. Burleigh, S., Hooke, A., Torgerson, L., Fall, K., Cerf, V., Durst,
In this scenario, we study the effect of message TTL, by B., & Scott, K. (2003). Delay-tolerant networking:an approach to
evaluating the performance of PF under different expiration interplanetary internet. IEEE Communications Magazine, 41,
times. The traffic load is set to 50 message per node, and the 128–136.
6. Vahdat, A., & Becker, D. (2000). Epidemic routing for partially-
buffer size to 15 messages. The main conclusion that can be connected ad hoc networks. Duke University, Tech. Rep. CS-200006.
drawn from Figs. 13, 14, and 15 is that; a smaller value of 7. Tseng, Y., Ni, S., Chen, Y., & Sheu, J. (1999). The broadcast
TTL leads to lower delivery ratio. This degradation of storm problem in a mobile ad hoc network. In ACM/IEEE Mo-
message delivery is due to the fact that the TTL of a message biCom (pp. 151–162).
8. Spyropoulous, T., Psounis, K., & Raghavendra, C. S. (2005).
might be expired before the message gets propagated to Spray and wait: An efficient routing scheme for intermittently
other nodes in order to reach its destination. connected mobile networks. In ACM WDTN (pp. 252–259).
As the TTL increases, the delivery latency of all protocols 9. Lindgren, A., Doria A., & Schelen O. (2004). Probabilistic
increases because more long-lived packets are successfully routing in intermittently connected networks. Lecture Nodes in
Computer Science, 3126, 239–254.
delivered. Noticeably, PR scheme dominates with the 10. Balasubramanian, A., Levine, B. N., & Venkataramani, A. (2010).
smallest delivery latency. The total number of transmissions Replication routing in DTNs: A resource allocation approach.
produced by each protocol increases as the TTL of messages IEEE/ACM Transactions on Networking, 18(2), 596–609.

123
 Author's personal copy
Wireless Netw (2013) 19:2093–2108 2107

11. Elwhishi, A. & Ho, P. (2009). SARP—A novel multi-copy 32. Jindal, A., & Psounis, K. (2009). Contention-aware performance
routing protocol for intermittently connected mobile networks. In analysis of mobility-assisted routing. IEEE Transactions on
IEEE GLOBECOM (pp. 4482–4488). Mobile Computing, 8(2), 145–161.
12. Yuan, Q., Cardei I., & Wu, J. (2012). An efficient prediction- 33. Groenevelt, R., Nain, P., & Koole, G. (2005). The message delay
based routing protocol in disruption-tolerant networks. IEEE in mobile ad hoc networks. Performance Evaluation, 62(1–4),
Transactions on Parallel and Distributed Systems, 23(1), 19–31. 210–228.
13. Erramilli, V., Crovella, M., Chaintreau, A. , & Diot, C. (2008). 34. Akkouchi, M. (2008). On the convolution of exponential distri-
Delegation forwarding. In ACM MobiHoc (pp. 251–260). butions. Journal of Chungcheong Mathematical Society, 21(4),
14. Spyropoulos, T., Psounis, K., & Raghavendra, C. S. (2008). 501–510.
Efficient routing in intermittently connected mobile networks: 35. Karagiannis, T., Boudec, J., & Vojnović, M. (2010). Power law and
The single-copy case. IEEE/ACM Transactions on Networking, exponential decay of intercontact times between mobile devices.
16(1), 63–76. IEEE Transactions on Mobile Computing, 9(10), 1377–1390.
15. Spyropoulos, T., Psounis, K., & Raghavendra, C. S. (2008).
Efficient routing in intermittently connected mobile networks:
The multiple-copy case. IEEE/ACM Transactions on Networking,
16(1), 77–90.
16. Nelson, S., Bakht, M., Kravets, R., & Harris, A. (2009).
Encounter: Based routing in DTNs. SIGMOBILE Mobile Com-
Author Biographies
puting and Communications Review, 13(1), 56–59.
17. Jones, E., Lily, L., Schmidke, J. K., & Ward, P. (2007). Practical
routing in delay-tolerant networks. IEEE Transactions on Mobile Ahmed Elwhishi received the
Computing, 6(8), 943–959. B.S. degree in computer Engi-
18. Ling, S. & Wei, W. (2009). Feedback adaptive routing algorithm neering from Tripoli University,
for DTN. In WRI International Conference on Communications Libya, in 1993, the M.Sc. and
and Mobile Computing, CMC, (Vol. 2, pp. 267–271). Ph.D. degrees from the Electrical
19. Erramilli, V., & Crovella, M. (2008). Forwarding in opportunistic and Computer Engineering
networks with resource constraints. In ACM CHANTS (pp. 41–48). Department, University of Water-
20. Li, Z., & Shen, H. (2008). Utility-based distributed routing in loo, ON, Canada, in 2006 and
intermittently connected networks. InICPP (pp. 604–611). 2012, respectively. He is now a
21. Musolesi, M., & Mascolo, C. (2009). CAR: Context-aware research associate in the Depart-
adaptive routing for delay-tolerant mobile netwoks. IEEE ment of Electrical and Computer
Transactions on Mobile Computing, 8(2), 246–260. Engineering at the University of
22. Elwhishi, A., Ho, P., Naik, K., & Shihadda, B. (2010). ARBR: Waterloo. His research interests
Adaptive reinforcement-based routing for DTNs. In IEEE WiMob are in the areas of delay tolerant
(pp. 376–385). networks, P2P system.
23. Samuel, H., Zhuang, W., & Preiss, B. (2009). DTN based domi-
nating set routing for MANET in heterogeneous wireless net- Pin-Han Ho received the B.Sc.
working. ACM Mobile Networks and Applications, 14(2), 154–164. and M.Sc. degrees from the
24. Chaintreau, A., Hui, P., Crowcroft, J., Diot, C., Gass, R., & Scott, Electrical Engineering Depart-
J. (2007). Impact of human mobility on opportunistic forwarding ment, National Taiwan Univer-
algorithms. IEEE Transactions on Mobile Computing, 6(6), sity, in 1993 and 1995,
606–620. respectively, and the Ph.D.
25. Spyropoulos, T., Turletti, T., & Obraczka, K. (2009). Routing in degree from Queen’s University
delay-tolerant networks comprising heterogeneous node popula- at Kingston in 2002. He is now an
tions. IEEE Transactions on Mobile Computing, 8(8), 1132–1147. associate professor in the
26. DTN java simulator, http://people.ee.ethz.ch/spyropot/dtnsim.html. Department of Electrical and
27. Liao, Y., Zhang, Z., Ryu, B., & Gao, L. (2007). Cooperative Computer Engineering, Univer-
robust forwarding scheme in DTNs using erasure coding. In IEEE sity of Waterloo, Canada. He is
MILCOM (pp.1917–1923). the author or coauthor of more
28. Spyropoulos, T., Rais, R., Turletti, T., Obraczka, K., & Vasilakos, than 150 refereed technical
A. (2010). Routing for disruption tolerant networks: Taxonomy papers, several book chapters,
and design. Wireless Networks, 16(8), 2349–2370. and the coauthor of a book on optical networking and survivability. His
29. Dvir, A., & Vasilakos, A. (2010). Backpressure-based routing current research interests cover a wide range of topics in broadband
protocol for DTNs. In SIGCOMM (pp. 405–406). wired and wireless communication networks, including survivable
30. Alresaini, A., Sathiamoorthy, M., Krishnamachari, B., & Neely, network design, wireless Metropolitan Area Networks such as IEEE
M. (2012). Backpressure with Adaptive Redundancy (BWAR). In 802.16 networks, Fiber Wireless (FIWI) network integration, and net-
INFOCOM (pp. 2300–2308). work security. He is the recipient of Distinguished Research Excellent
31. Zeng, Y., Xiang, K., Li, D., & Vasilakos, A. V. (2013). Direc- Award in the ECE Department of University of Waterloo, Early
tional routing and scheduling for green vehicular delay tolerant Researcher Award (Premier Research Excellence Award) in 2005, the
networks. Wireless Networks, 19(2), 161–173. Best Paper Award in SPECTS’02, ICC ’05 Optical Networking

123
 Author's personal copy
2108 Wireless Netw (2013) 19:2093–2108

Symposium, and ICC ’07 Security and Wireless Communications wireless communication networks, including wireless Metropolitan
Symposium, and the Outstanding Paper Award in HPSR ’02. Area Networks such as IEEE 802.16 networks, Fiber-Wireless (FIWI)
network integration, and optical networks, and green communica-
Basem Shihada has been tions. In 1997, he obtained a Bachelor’s degree in Computer Science
appointed Assistant Professor of from the UAE University. In 2001, he obtained a Master’s degree in
Computer Science in the Math- Computer Science from Dalhousie University, Halifax, Canada. In
ematical and Computer Science 2007, he obtained a Ph.D. degree in Computer Science, University of
and Engineering Division at Waterloo, David R. Cheriton School of Computer Science, Waterloo,
KAUST. He joint Stanford Canada.
University as a visiting faculty
in Stanford Computer Science
in 2008. Prior to his appoint-
ment at KAUST, Dr. Shihada
was a research associate in the
Department of Electrical and
Computer Engineering at the
University of Waterloo in
Ontario, Canada. Dr. Shihada’s
current research covers a wide range of topics in broadband wired and

123