Journal of Network and Computer Applications 87 (2017) 169–184

Contents lists available at ScienceDirect

Journal of Network and Computer Applications

journal homepage: www.elsevier.com/locate/jnca

Social-based energy-aware multicasting in delay tolerant networks MARK

⁎
Animesh Roy , Soumit Bose, Tamaghna Acharya, Sipra DasBit
Indian Institute of Engineering Science and Technology, Shibpur, West Bengal, India

A R T I C L E I N F O A BS T RAC T

Keywords: In a post-disaster environment, traditional network infrastructure is mostly unavailable due to incidental
Delay/disruption tolerant networks destruction and damages. In such scenarios, delay/disruption tolerant networking may be considered as an
Relay selection alternative approach to meet the quintessential communication needs between the victims, field workers and
Social-based multicast centralized authority. In such applications, multicast communication is imperative because rescue and relief
Energy awareness
messages may often need to be shared among many volunteers. In this paper, we address the problem of relay
selection for multicasting single and multiple data items in DTNs considering the application in disaster
management. Both the single-data and multiple-data multicast problems are formulated as bi-objective
optimization problems under the constraints of target delivery time, buffer space and residual energy levels
of relay nodes. Both the problems are shown to be NP-hard. Hence, we propose two heuristic relay selection
schemes exploiting social properties in DTNs. Theoretical analysis is performed for energy consumption and
exhaustive simulation is carried out to evaluate performance of our proposed schemes in ONE simulator
considering real-life data in post-disaster scenarios. Results show that our schemes are energy-saving thereby
ensure longer network lifetime compared to other prominent schemes.

1. Introduction pair. On the other hand, frequent network partitions due to unpre-
dictable node mobility result in rare existence of static path(s) between
Delay/Disruption Tolerant Networks (DTNs) are primarily char- an arbitrary source–destination pair in DTNs. A data is typically
acterized by intermittent connectivity and opportunistic contacts delivered in DTNs following “store-carry-and-forward” principle.
among nodes in the networks. In addition, short communication range, However, routing in such networks has been investigated for more
sparseness, node mobility and short communication duration are also than a decade. Consequently, various unicast as well as multicast
some of the key features of such networks. A DTN may also be viewed schemes can be found (Sobin et al., 2016; Zhu et al., 2015). In unicast
as a special class of mobile ad hoc network (MANET), where an end-to- communication, a message is sent from a source to another destination.
end path between any source–destination pair rarely exists. Such Multicast is the term used to describe communication where a message
networks are especially suitable for applications in challenging envir- is sent to a set of destinations. In DTNs, conventional routing schemes
onments, such as battlefield communications, wildlife monitoring, (e.g. PROPHET, Lindgren et al., 2003, data delivery using message
space communications (Fall, 2003). Besides, in post-disaster situa- ferries, Zhao et al., 2004) exploit the mobility of the DTN nodes to
tions, volunteers/relief workers, engaged in any devastated region, face select the same as relays which forward data opportunistically.
a huge problem of connectivity among themselves, as well as with the As node mobility is unpredictable and volatile in nature, there
relief coordination center, due to complete or partial collapse of the arises the need of new routing schemes based on more stable
existing voice/data network infrastructure. In such cases, necessity of a characteristics of the nodes. To make forwarding decisions based on
flexible, fast-deployable low-cost temporary communication network is the knowledge of long-term and less volatile characteristics of the
paramount. DTN is being considered as an important communication nodes, of late, researchers are leveraging social network concepts in the
paradigm in post-disaster rescue and relief management operations, design of efficient DTN routing schemes (Xia et al., 2015; Wei et al.,
because of its potential to meet most of the said requirements 2014). Social-based opportunistic schemes consider dynamic nature of
(Rodrigues, 2014; Martin-Campillo et al., 2013). the nodes’ interactions for having a better representation of social
Popular routing schemes for MANETs are not suitable in DTNs, structures. Moreira et al. (2013) propose dLife, a routing scheme that
because all of them are based on the common assumption of the captures the dynamic social structures by time-evolving social ties
existence of several end-to-end paths between any source–destination between pair of nodes in the network. In Ciobanu et al. (2014), an

⁎
Corresponding author.
E-mail address: animeshroy801@gmail.com (A. Roy).

http://dx.doi.org/10.1016/j.jnca.2017.03.020
Received 27 July 2016; Received in revised form 9 December 2016; Accepted 22 March 2017
Available online 24 March 2017
1084-8045/ © 2017 Elsevier Ltd. All rights reserved.
A. Roy et al. Journal of Network and Computer Applications 87 (2017) 169–184

algorithm is proposed that leverages online social relations (on 2.1. Conventional approach
Facebook or Google+), interests and contact history of a node. In
Ciobanu and Dobre (2012), it is shown that a regular pattern exists in a Wang et al. (2010) show that delegation forwarding is efficient for
node's contact history, and the number of future contacts can be multicast communication. Here, the main idea is to assign a quality
predicted by modeling the time series as a Poisson distribution. In value and a level value to each node for each destination. The quality
Quercia et al. (2010), the authors deal with the problem of recom- value of a node is static, but level value is dynamic. Initially, the level
mending relevant social events from mobile phone location data. It value of each node is equal to its quality value. During the routing
may be noted that social-based schemes (Wei et al., 2014; Moreira process, a node compares its own level value with the quality value of
et al., 2013; Ciobanu et al., 2014; Ciobanu and Dobre, 2012; Quercia the node it meets. A message holder only forwards the message to a
et al., 2010) outperform traditional approaches that are mainly based node with a higher quality than its own level. In addition, the message
on flooding (Vahdat and Becker, 2000; Spyropoulos et al., 2005) and holder also raises its own level to the quality of the higher quality node.
mobility-based predictions (Lindgren et al., 2003; Zhao et al., 2004). So a node will forward a message only if it encounters another node
In some applications, multicast may be more useful for data whose quality metric is greater than any node met by the message so
dissemination than unicast. In post-disaster scenario, messages from far.
the central control center, as well as from the leaders of various relief End-to-end path is not guaranteed in DTNs. Remote communica-
management groups, often need to be disseminated to all or a subset of tion devices can be used to achieve connectivity among the nodes in the
members in the groups. Arguably, in such applications of DTNs, network. It may be noted that remote communication is expensive. So
multicast is more efficient. To the best of our knowledge, the problem imparting an extra remote communication introduces the new challen-
of multicasting in DTNs, especially for applications in disaster manage- ging problem of minimizing its cost. Mongiovi et al. (2012) formulate
ment is not well explored. Further, energy level of a node is another the problem of minimizing the remote communication cost for multi-
critical parameter in DTNs, especially in applications like post-disaster cast in DTNs. The formulation is presented as an Integer Linear
scenario, where the nodes are often deployed in locations without any Programming (ILP) problem which is NP hard. The problem is
scope of recharging and replacing their batteries (Martin-Campillo analyzed in the case of scheduled trajectories and known traffic
et al., 2013). Hence, residual energy of the nodes needs to be demands. The authors propose a heuristic solution based on a graph
considered in almost all the activities of the network including design indexing system and also present an adaptive extension that can
of efficient routing strategies. To the best of our knowledge, energy address the problem even with limited knowledge of node mobility.
aware multicast routing in DTNs has not yet received the due attention Lee at al. (2008) focus on the scalability property of DTN multicast
in open literature. routing and propose RelayCast to improve the throughput bound of
Although Epidemic (Vahdat and Becker, 2000) routing scheme is multicast by exploiting node mobility. They mainly propose a multicast
known to achieve high delivery ratio by flooding the data across the routing scheme that extends the two-hop relay algorithm (Grossglauser
network, it may not be considered suitable in the DTNs operated in and Tse) which is based on mobility of nodes, improving network
post-disaster management because of the huge resource requirement in performance. RelayCast is also extended to make it useful in delay
terms of number of relays and total energy consumption (Sobin et al., tolerant coalition networks where multiple domains exist. In this
2016). Furthermore, in disaster scenarios, node-density is likely to be scenario, nodes communicate with other nodes in different domains
low and energy levels of the nodes also become critical (Martin- via gateways due to security concerns and policy reasons.
Campillo et al., 2013). So in such scenario, there needs to be a balance In Wang and Wu (2012), authors propose a dynamic tree based
between delivery ratio and resource consumption. Thus, necessity of a routing scheme, where each leaf node corresponds to a destination.
data forwarding scheme that achieves moderate delivery ratio consum- This scheme mainly uses small number of relay nodes for multicast. It
ing less resources, is evident. In this paper, we propose two efficient aims to provide a non-replication multicasting scheme in DTNs thereby
relay selection strategies aiming to address the multicasting problem in keeping the number of forwarding low. Here data forwarding occurs
DTNs deployed for post-disaster situation management. Social-based based on the node active level together with contact probability level.
parameters are exploited and energy constraint of the DTN nodes is The contact probability level and active level are mainly estimation of
considered in the proposed schemes. Our proposed schemes could be the number of contacts with the destinations and the number of total
implemented in all practical applications of DTNs where nodes are contacts respectively in a given period.
capable of measuring their own residual energy and exchanging the
same with their neighbors upon contact. It may be noted that the 2.2. Social based approach
energy-aware relay selection scheme for single data multicast has
already been reported in Roy et al. (2015). However, results based Only a few studies are reported on social-based multicast in DTNs.
on theoretical analysis and the scheme for multiple data multicast with In Deng et al. (2013), affiliation and language, two important and
energy awareness are not presented in Roy et al. (2015). In addition, representative social features of nodes, are explored in multicasting. In
simulation results in Roy et al. (2015) are not obtained from ONE affiliation-based multicast scenarios, people with similar affiliations
simulator based on post-disaster scenarios. such as those from the same or related organizations are interested in
The rest of this paper is organized as follows. Section 2 describes a the same information, whereas in language-based scenario, people
brief overview of the related work. Our contributions are clarified in sharing common languages are willing to receive the same messages.
Section 3. Section 4 describes our system model. Proposed energy- Considering these two scenarios, a social profile-based multicast
aware multicast schemes for single and multiple data are presented in routing scheme is proposed. For data delivery, nodes with a small
Sections 5 and 6 respectively. Section 7 illustrates the results of the average affiliation distance or large common language ratio to a
performance evaluations. Finally, Section 8 concludes the paper. multicast group are selected as the relay nodes.
A social network aided multicast delivery (SNAMD) scheme is
presented in Chuah (2009). In this, communities of nodes and
connector nodes that can reach multiple communities are identified.
2. Related work Subsequently, the scheme makes use of such information to efficiently
deliver multicast messages. Specifically, it forwards a copy of the
Several multicast solutions are proposed to deal with the problem of message to connectors and let each node make use of either the
data forwarding in DTNs. We have categorized multicast forwarding delivery predictability or the estimated node contact rate in its
schemes into two groups as follows. forwarding decision.

170
A. Roy et al. Journal of Network and Computer Applications 87 (2017) 169–184

In Galluccio et al., a multicast polymorphic epidemic routing fair delivery ratio consuming less resources in terms of number of
scheme is proposed to deliver data in DTNs. In this work, sociality of relays and battery power.
the nodes is utilized for the data dissemination process and adaptive
recovery strategies are used to reduce network overhead. Nodes are not 4. System model
considered as individual members and can be aggregated into groups
sharing interests. The intrinsic sociality and interests of the nodes are This section provides a description of the system model including
exploited for efficient multicasting. network architecture along with the key assumptions. Social relations
The authors in Gao et al. study single, as well as multiple data among DTN nodes based on node contacts are modeled by the social/
multicasting problems with an aim to minimize the cost of data contact graphs (Hossmann et al., 2010) which help in extracting the
delivery, exploiting social characteristics of the nodes. Delivery cost is social network characteristics of the nodes in the DTNs. Social-based
minimized through minimizing the number of relays used. In this approaches utilize various social metrics (e.g., centrality and commu-
work, the relay selection problem is formulated as a unified knapsack nity) which can be estimated from a social/contact graph. Centrality
problem. To deliver data to the destinations, relays are selected based indicates how much a node is connected to other nodes in the network.
on the centrality of the nodes for single data multicast (SDM). On the There are various ways to represent centrality of a node. Centrality is
other hand, for multiple data multicast (MDM), relays are selected also obtainable from online social networks (OSNs) (Sapountzi and
exploiting social community structures. Psannis, 2016) like Facebook. It is well known that to access OSNs,
continuous connectivity to the Internet is necessary. Accessing OSNs
3. Contributions would indeed be very much problematic in the application of post-
disaster management, where conventional network infrastructure is
In this work, considering post-disaster scenarios, we propose two almost absent. So instead of depending on social data from Facebook,
relay selection schemes to multicast data items to the destinations. Our in our proposed work, centrality is calculated based on the interactions
proposed schemes are based on SDM and MDM (Gao et al.) respec- occurred among the DTN nodes at the disaster site. Hence, we consider
tively, where the similarities with Gao et al. are mentioned as follows. a weighted social network model (Gao et al.). According to the model,
vertices in the graph represent the DTN nodes. An edge of the graph is
• Centrality of a node, represented as a Cumulative Contact considered as a Poisson process and the weight along an edge
Probability Gao et al., is used to select relays for multicasting. represents average contact frequency of the node pair. It has been
• An objective of the proposed problem formulation is to minimize the validated in Gao et al. that the pairwise inter-contact time of the nodes
number of relays used for data delivery achieving a particular is exponentially distributed. The nodes are assumed to be randomly
delivery ratio within a specified time constraint. deployed in the network. It is further assumed that initially all the
nodes have equal energy levels.
In post-disaster scenario, DTN nodes are likely to face the challenge of Fig. 1 shows the network topology at an arbitrary time instance
limited battery power (Martin-Campillo et al., 2013). Therefore, in this along with its corresponding weighted social/contact graph. In
paper, we first consider multicasting a single data item in the network Fig. 1(a), there are 13 nodes which are scattered in the network at a
with limited energy resources. Thereafter, we extend the problem to particular instance. The dotted line between a pair of nodes indicates
multicast of multiple data items in the network. The following are the that they met at least once in the past. On the other hand, if there is no
preview of the key contributions of our work along with the differences dotted line between a pair of nodes, it means that the two nodes did not
over (Gao et al.). come in each other's contact till the current time instance. For example,
nodes I and J have come in contact with each other at least once, but no
• We present constrained optimization framework for both single and contacts took place between the nodes I and K.
multiple data multicast problems and subsequently prove them to be The corresponding social graph which is shown in Fig. 1(b) is built
NP-hard. In Gao et al., there is only one objective of relay based on the observed contacts among the nodes. In general, social/
minimization. In post-disaster scenario, where energy limitation contact graphs are not weighted, but here different contact rates
exists at the DTN nodes, relay minimization is not sufficient for between the node pairs are taken as the weights. For example, in the
efficient multicasting. In addition to relay minimization, energy graph (Fig. 1(b)), weight λij is the contact rate between the node pair I
constraint of the nodes is also considered in our proposed work. and J. As per the properties of the social/contact graph, centrality is the
• We propose energy-aware social-based relay selection schemes to cumulative probability of a node to contact others. Hence, a node with
multicast single data, as well as multiple data items aiming to a large number of possible contacts is considered as a node with high
support as many multicast sessions as possible. Alternatively, the centrality, and also as a good message forwarder for other nodes. If we
objective is to maximize the network lifetime by saving energy of the look at the graph, node D is of the highest centrality among all the
nodes. The proposed schemes in Gao et al. are not application nodes in the network. In this study, centrality metric is defined as the
oriented and do not enhance network lifetime. With a focus on the Cumulative Contact Probability (CCP) following Gao et al.:
application of post-disaster management, our proposed schemes are
1 N
designed. Ci = 1 − ∑ j =1,j ≠i e−λij T
• We consider priority of the messages to multicast multiple data
N−1
items. Justification of considering such priority is found analyzing where Ci is the CCP of the node Ni, N is the total number of nodes in
the WhatsApp messages exchanged among the team members of the the network, λij is the contact rate between Ni and Nj and T is the
NGO “Doctors For You” (DFY) during the relief operations after the specified time. The probability that Ni contacts Nj within time T is
Nepal earthquake occurred in 2015. Consideration of priority helps 1 − e−λij T . Here the CCP is the average probability that a randomly
in the delivery of critical messages in the post-disaster environment. chosen node in the network is contacted by Ni within time T.
No such priority is considered in Gao et al. It may be noted that a warm up period is required for having initial
• We theoretically analyze our schemes and results of simulation values of the centrality. During this warm up period, interactions
study are presented to evaluate our proposed schemes in ONE among the nodes are observed to accumulate the contact rates. A node
simulator using real-life data from two recent disasters, does not generate any message in this period. After this warm up
Uttarakhand flood (2013) and Nepal earthquake (2015). period, centrality of a node changes based on the contacts occurred
with the other nodes in the network. Upon contact with each other,
Comparative performance shows that our proposed schemes achieve nodes exchange a few information which includes centrality, available

171
A. Roy et al. Journal of Network and Computer Applications 87 (2017) 169–184

Fig. 1. (a) DTN topology at a particular time instance. (b) Corresponding social/contact graph.

buffer space and residual energy values. Nodes are selected as relays 5.2. Hardness of the problem
among the contacted neighbors to achieve a particular delivery ratio
which is the ratio of the number of delivered destinations to the total It can be shown that the formulated bi-objective optimization
number of destinations. problem is NP-hard. We begin by considering the following optimiza-
tion problem:
n n
max (∑i =1 ci1 xi , ∑i =1 ci2 xi ) (a)
5. Energy-aware single-data multicast n
s.t. ∑i =1 wi xi ≤W (b) (2)
In this section, problem formulation is presented followed by the
where xi ∈ {0, 1}. This is a 0–1 bi-objective knapsack problem with
description of our proposed scheme for single-data multicast.
real-valued weights wi, which is an NP-hard problem (Lust and
Teghem, 2012; Cormen et al., 2009). Now, by defining

5.1. Problem formulation ci1 = −1, for i = 1, …, n (3)

To multicast a single data item, we formulate our problem as ci2 = Ei |Ei > e0 (4)
follows. n=k (5)
To deliver a data item to a set of destinations, minimizing the
number of relays and maximizing their total residual energy levels, wi = log( pi ) (6)
subject to the constraint of achieving a delivery ratio p within a given W = log(1 − p ) (7)
time T. In addition, each selected relay node must have sufficient
residual energy for relaying the data item. the 0–1 bi-objective knapsack problem (2) is transformed exactly into
The problem for single-data multicast as defined above is an the optimization problem (1). This proves that the energy-aware single
optimization problem which has two objectives. One is for selection data multicast problem (1) is NP-hard. Hence we propose a heuristic to
of minimum number of relays and the other is for maximizing the select relays for single data multicast.
summation of residual energy values of the relays, satisfying the
constraints. So the bi-objective optimization problem is formulated 5.3. Proposed scheme
as follows.
In our proposed energy-aware single-data multicast (EASDM)
k k
min ∑i =1 xi , max(∑i =1 xi Ei |Ei ≥ e0 ) (a) scheme, the data source first divides the set of contacted neighbors
k into high-centrality set (HC set) and low-centrality set (LC set). The
s.t. Π pixi ≤ 1 − p (b) nodes which have pi values greater than 0.5 are in LC set and the other
i =1 (1)
nodes are in HC set. Firstly, relays are selected from HC set. For better
where the set of contacted neighbors is R = {R1, R2 , …., Rk }, binary network performance in terms of delivery ratio, delay as well as for
design variable xi ∈ {1, 0} specifies whether or not Ri is selected as longer network lifetime, the nodes with high centrality must be
relay, Ei is the residual energy of Ri, e0 is the threshold which is the carefully used in the data forwarding process so that these nodes
minimum energy requirement of any relay node to be selected, and pi is remain alive as long as possible. Hence we have to keep the high-
∼
the probability that Ri cannot forward data to a random node Nj ∉ R centrality nodes active as long as possible because of maintaining fair
∼
within time T (R = {S} ∪ R ). Here pi is calculated from the centrality network performance in all the communication sessions in DTNs. So
metric (CCP) of the node as shown in Gao et al. It may be noted that a for selecting relays with high centrality, e0 is set to the value of average
node will have low pi value if it has high centrality value. In Eq. (1), the residual energy of the HC set. This value helps to reserve a few high-
constraint (b) represents the probability that a randomly chosen node centrality nodes with comparatively lower residual energy. Suppose
∼
Nj ∉ R is not contacted by the selected relays should be lower than there are two high-centrality nodes with two different residual energy
1 − p , to ensure that the average delivery ratio achieves a value higher values y1 and y2. If e0 = ( y1 + y2)/2 , the node with lower residual energy
than p (Gao et al.). is not considered for selection, because nodes with residual energy

172
A. Roy et al. Journal of Network and Computer Applications 87 (2017) 169–184

greater than or equal to e0 are considered for relay selection. In value, such that r [i ]. pi < r [ j ]. pi for all 1 ≤ i ≤ h , h + 1 ≤ j ≤ k
addition, to select any node as a relay, its residual energy must be Arrange r [1…h] and r [h + 1…k ] in increasing order of r [i ]. Ei and
greater than or equal to the energy (e1) required to receive and transmit r [ j ]. Ei respectively
the data item. So e0 is greater than or equal to e1 which may be y←h
obtained from the specification of the DTN nodes used in the network. z←0
To select relays from the LC set, e0 is set to the value of e1, because p←1
nodes with low centrality are not so efficient in data forwarding. In the ∑ih=1 r [i] . Ei
HC set, if there are h high-centrality nodes with the residual energy e0 = h
Ei ≥ e1, the e0 is defined as follows. selected ← 0
⎧ ∑ih=1 Ei // relays are selected satisfying constraints
⎪
e0 = ⎨ h , if relays to be selected from HC set while y ≥ z and p > 1 − p do
⎪ e, if y = 0 then
⎩ 1 if relays to be selected from LC set
//to select relays from LC set
The relay selection procedure in EASDM is summarized in y←k
Algorithm 1. The detailed steps are as follows. In the HC set, the node z←h+1
with the highest residual energy is considered first for selection, and it e0 = e1
is selected as a relay if Ei value of the node is equal or more than e0. end
Now if the pi value of the selected node alone does not satisfy the if r [y]. Ei > e0 then
constraint (b) in Eq. (1), the node with the second-highest residual
p ← p*r [y]. pi
energy is also considered for selection from the HC set and again the
selected ← selected + 1
constraint needs to be checked with the pi value of the node satisfying
x [selected] ← r [y]. id
its Ei ≥ e0 . The process of selection continues in the HC set till either
end
the constraints are satisfied or the HC set is exhausted. If the
y←y−1
constraints are still not satisfied, relays are selected from the LC set
end
one by one in decreasing order of the residual energy till the constraints
are satisfied or the LC set is exhausted. In this way, relays are chosen to
meet the performance requirement for data delivery. After initial relay
5.3.1. Algorithmic complexity
selections at the source, data is forwarded to the selected relays. When
The EASDM scheme comprises the following tasks.
a selected relay Ri finds another node Rj whose centrality metric is
higher than that of Ri, data is forwarded to Rj, and Ri then removes its
Task 1: Dividing the contacted neighbors into HC and LC sets.
local data copy. In this way, data is eventually delivered to the
Task 2: Arranging the nodes in increasing order of their residual
destinations.
energy.
Algorithm 1. Algorithm for relay selection in EASDM. Task 3: Executing the while loop.
In case of Task 3, the worst case arises when all the contacted
Input: pi and Ei values of the contacted neighbors neighbors are selected as relays. Now considering n number of
Output: Selected set of nodes as relays contacted neighbors, the worst case complexities of the above tasks
are as follows.
// r[] represents the set of contacted neighbors
// k is the total number of contacted neighbors Task 1: O(n).
// h indicates the number of nodes with high centrality Task 2: O (n log n ).
metric Task 3: O(n).
// x[] holds the selected node-id Hence, O (n log n ) is the worst case time complexity of the proposed
scheme.
Divide r [1…k ] into r [1…h] and r [h + 1…k ] based on a threshold

Fig. 2. Selection of relays following EASDM scheme.

173
A. Roy et al. Journal of Network and Computer Applications 87 (2017) 169–184

5.3.2. Illustrative example 6.2. Hardness of the problem

The relay selection procedure of EASDM is illustrated in Fig. 2. Data
source S is assumed to multicast a data item achieving delivery ratio It can be shown that the formulated bi-objective optimization
p=0.9 within specified time T=22 h. Let us consider e1 = 10% and S has problem is NP-hard. We begin by considering the following optimiza-
seven contacted neighbors (A, B, C, D, E, F and G) with their respective tion problem.
2-tuple parameter values of pi and Ei. Following the proposed EASDM
⎛ m 1 m ⎞
scheme, two sets are formed with high-centrality and low-centrality max ⎜∑ j =1 cj yj , ∑ j =1 cj2 yj ⎟ (a)
nodes respectively. For selection of relays from the HC set, e0 is ⎝ ⎠
n
(40 %+ 30 %+ 50%)/3 = 40%. For the low-centrality set, e0 is e1. Firstly, s.t. ∑i =1 w1i yj ≤ W1 (b)
relays are selected from the HC set. From the HC set, node C will be m
∑ j =1 w2j yj ≤ W2 (c)
selected first as a relay, because it is with the highest residual energy. n
∑i =1 w3i yj ≤ W3 (d) (9)
Since the constraint (b) in Eq. (1) is not satisfied by only taking the node
C as relay, the same selection process is continued on the HC set and where yj ∈ {0, 1}. This is a 0–1 bi-objective multiple-constrained
node A is also selected as a relay. Node B is reserved and not selected, knapsack problem with real-valued weights, which is an NP-hard
because its residual energy falls below e0 (40%). After taking C and A as problem (Lust and Teghem, 2012; Cormen et al., 2009). Now, by
relays, the constraint (b) is still not satisfied. So the process of selecting defining
relays from LC set starts. In the LC set, node E has highest residual
energy and its Ei is greater than e1 (10%). So E is selected as relay in the m ⎧
⎪
n ⎫
⎪

∑ c1j yj = − ⎨Rj | ∑ yj > 0⎬

LC set. As the constraint is satisfied after selecting A, C and E as relays, ⎩ i =1
⎪
⎭
⎪

j =1 (10)
the process of relay selection stops and the algorithm terminates.
n

6. Energy-aware multiple-data multicast cj2 yj = Ej | ∑ yj > 0

i =1 (11)
In this section, we present the problem formulation followed by the yj = xij (12)
description of our proposed scheme for multicasting multiple data
items. w1i = si (13)

W1 = Bj (14)
6.1. Problem formulation
w2j = log(1 − pij ) (15)
To multicast multiple data items, we formulate our problem as
follows. W2 = log(1 − p ) (16)
To deliver a set of data items d1, …, dn with sizes s1, …, sn to
destination sets D1, …, Dn , minimizing the number of relays and also w3i = si e0 (17)
maximizing their total residual energy levels, while meeting a given W3 = Ej (18)
delivery ratio p within the time periods T1, …, Tn respectively.
The aforementioned problem for multiple data multicast may be the 0 − 1 bi-objective multiple-constrained knapsack problem (9) is
formulated as a constrained bi-objective optimization problem. The transformed exactly into the optimization problem (8). This proves that
objective functions are as follows, the energy-aware multiple data multicast problem (8) is NP-hard.
Hence, we propose a heuristic for our relay selection problem in
(i) minimizing the number of relays, and multiple data multicast.
(ii) maximizing the sum of residual energy values of the relays. Relays
are selected from the set of contacted neighbors R = {R1, …, Rm} 6.3. Proposed scheme
with buffer sizes B1, …, Bm . Here xij ∈ {0, 1} indicates whether data
n
item di is placed on relay Rj or not, and ∑i =1 xij si is the total size of 6.3.1. Basic idea
all the data items given to the relay Rj. Also, pij is the probability In DTNs, every relay node needs to be active as long as possible for
that Rj can send data di to its destinations. Here Ej is the residual enhancement of network lifetime. Alternatively, network lifetime is
energy of Rj. The relay selection problem is formulated as follows: dependent on the nodes’ lifetime which is dependent on their residual
⎛ m ⎞ energy. In addition to the residual energy of the nodes, in our proposed
n n
min {Rj | ∑i =1 xij > 0} , max ⎜∑ j =1 Ej | ∑i =1 xij > 0⎟ (a) heuristic, we also consider priority of the data items based on their
⎝ ⎠
n
delivery time constraints. In the existing MDM scheme (Gao et al.), no
s.t. ∑i =1 xij si ≤ Bj for j = 1, …, m (b) such priority is considered. In other words, in Gao et al., each data item
m
Π (1 − pij ) xij ≤1−p (c) is assumed to have the same delivery time constraint. But in many
j =1 applications, all the data items may not necessarily have the identical
n
∑i =1 xij si⋅e0 ≤ Ej for j = 1, …, m (d) delivery time constraints. For example, in a post-disaster scenario,
messages related to situational awareness are likely to have higher
(8)
priority thereby lower delivery time constraint than the sentimental
where e0 is the energy required to transmit and receive data of 1- messages.
byte. The value of e0 can be obtained from the specification of the Justification of considering such priority is found analyzing real
nodes. In Eq. (8), the first constraint (b) signifies that data item di disaster information. In April 2015, Nepal experienced a major severe
of size si is given to the relay node Rj, if its available buffer space Bj earthquake, and from different parts of the world, many organizations
is more than or equal to si. The second constraint (c) guarantees reached there to offer rescue and relief services. Doctors For You
that, for each data item, the average probability of delivery to its (DFY), a pan-India humanitarian organization with international
destinations is higher than p. Here pij can be calculated from the presence, sent a team to Nepal for the same purpose. Members of
centrality metric of Rj. The third constraint (d) ensures that the the team created a WhatsApp group to exchange messages for
relay node Rj has sufficient energy to forward all the data items improving situational awareness during the first few days after the
stored in its buffer space. disaster. The WhatsApp conversations received in smart phones are

174
A. Roy et al. Journal of Network and Computer Applications 87 (2017) 169–184

Table 1
Sample WhatsApp messages exchanged among DFY team members.

Date Time Phone no. Message

28/04/2015 11:18 PM xxxxx84882 I've picked up Shan. Going to hotel

28/04/2015 11:18 PM xxxxx97734 Great
29/04/2015 1:01 AM xxxxx84882 Keep all volunteers on hold as per nepal govt instructions for all foreign medical teams
29/04/2015 1:05 AM xxxxx84882 All foreign medical teams are on hold and frustrated as the ground situation is different from the govt version
29/04/2015 1:09 AM xxxxx84882 All relief material stagnating in the airport as there is no organized distribution even in Kathmandu city
29/04/2015 1:29 AM xxxxx48932 Thanks mridul
29/04/2015 1:33 AM xxxxx03910 Do d best on behalf of us

forwarded to the server as email attachment for giving access to the The process of selection continues in the set with high delivery
messages (Bhattacharjee et al., 2016). We have analyzed WhatsApp probability till either the constraints are satisfied or the nodes in set
messages exchanged among the DFY team members. A sample of the with high delivery probability exhaust. If the constraint (c) is still not
WhatsApp messages is shown in Table 1. It is clear that the third, satisfied, relays are selected from the set with low delivery probability
fourth and fifth messages in Table 1 are situational updates, where the one by one in decreasing order of the residual energy till the constraints
other messages are normal conversational or sentimental. Here situa- are satisfied or the set with low delivery probability is exhausted. Thus,
tional update messages need to have higher priority than the other relays are selected to meet the performance requirement for data
messages. delivery. When relay selections are completed for the data item with
Suppose, a source node has three data items d1, d2 and d3 which are highest priority, the data item with next highest priority is considered
to be delivered to their respective destination sets with the fixed and relays are selected for this data item also in the same way as above.
delivery ratio 0.8 within the specified time intervals of 18 h, 16 h and The heuristic for relay selection terminates when there are no data
22 h respectively. Here the data item d2 is of the highest priority, items remaining for relay selections. After initial relay selections at the
because among the three data items, it needs to be delivered within the data source, data items are forwarded to the selected relays.
least time period (16 h) achieving the required delivery ratio. Afterwards, each relay node acts as the source for the data items
Accordingly, d1 is of next-highest priority and d3 is of the lowest received, and chooses relays in the same way for delivering data items
priority. As the MDM scheme Gao et al., does not have any priority to the destinations that are not in its set of contacted neighbors. In this
consideration, all the data items d1, d2 and d3 are delivered targeting way, the data items are delivered to the destinations.
the same delivery time constraint of 16 h. But this results in usage of
Algorithm 2. Algorithm for relay selection in EAMDM.
more number of relays, because d1 and d3 are also to be delivered
within 16 h that is smaller than their actual delivery time constraints
Input: s (size) and T (time constraint) // Parameter values of the
(18 and 22 h). It may be noted that consideration of such priority helps
data items
in improving the relay minimization objective. Consideration of non-
B (available buffer), qij (average probability that the node cannot
uniform delivery time constraints among multiple data items enables
deliver the data item to its destinations, i.e., 1 − pij ) and E
repetitive use of few relays with high energy levels. On the contrary, if
multiple data items have identical delivery time constraints, a larger (residual energy)
number of relays needs to be used to complete simultaneous delivery of // Parameter values of the contacted neighbors
all data items. So we conjecture that in practical applications, con- Output: Selected set of nodes as relays
sideration of such priority helps in designing efficient relay minimiza-
tion solutions. // d[] represents the set of data items
// r[] represents the set of contacted neighbors
// m is the total number of contacted neighbors
6.3.2. Details of the heuristic // b indicates the number of nodes with sufficient buffer
With the above motivation, we present an energy-aware multiple // h indicates the number of nodes with high average
data multicast (EAMDM) scheme. The relay selection procedure is delivery probability
shown in Algorithm 2. To start with, at the source node, data items are // p is the delivery ratio
prioritized based on their individual delivery time constraints. Data x[] holds the selected node-id
item with lower delivery time is of higher priority. Firstly, relays are
selected for the data item with highest priority. The set of contacted Arrange d [1…n] in decreasing order of d [i ]. T
neighbors is divided into the two sets of nodes with sufficient buffer // repeat for each data item
and insufficient buffer. Nodes which have available buffer space more Divide r [1…m] into r [1…b] and r [b + 1…m] based on s, such that
than the size of the data item with highest priority, are in the set of r [i ]. B ≥ s and r [ j ]. B < s for all 1 ≤ i ≤ b , b + 1 ≤ j ≤ m
nodes with sufficient buffer and others are in the set of nodes with
Divide r [1…b] into r [1…h] and r [h + 1…b] based on a threshold
insufficient buffer. This satisfies the constraint (b) in Eq. (8), consider-
value, such that r [i ]. qij < r [ j ]. qij for all 1 ≤ i ≤ h , h + 1 ≤ j ≤ m
ing the particular data item with highest priority (i=1). Now the set of
nodes with sufficient buffer is again divided into the two sets of nodes Arrange r [1…h] and r [h + 1…b] in increasing order of r [i ]. E and
with high delivery probability and low delivery probability based on a r [ j ]. E respectively
predefined threshold value (0.5). The nodes in both the sets with high y←h
delivery probability and low delivery probability are arranged in z←0
increasing order of residual energy respectively. Next, the node with p←1
highest residual energy satisfying the energy constraint of (d) is selected ← 0
selected as a relay. Now if the pij value of the selected node alone // relays are selected satisfying constraints
does not satisfy the constraint (c), the node with the second-highest while y ≥ z and p > 1 − p do
residual energy is also considered for selection from the set with high if y = 0 then
delivery probability, and again the constraints (d) and (c) need to be y←b
checked with the Ej and pij values of the node. z←h+1

175
A. Roy et al. Journal of Network and Computer Applications 87 (2017) 169–184

end constraint (c) in Eq. (8) is not satisfied. So the node B is also selected as
if r [y]. E > e0 then a relay and the performance constraint (c) is still not satisfied. Now
p ← p*r [y]. qij there are no nodes in the set of high-delivery probability. Hence, node
selected ← selected + 1 C is selected from the set of low-delivery probability. After this, the
x [selected] ← r [y]. id performance constraint is satisfied for the data item d1 and the nodes
end A, B and C are selected as relays.
y←y−1 Next data item d2 is to be relayed. In this case, A, B and E are the
end nodes with sufficient buffer, because each of them has available buffer
space equal to or more than the size of data item d2 after taking d1.
Among the nodes A, B and E, E is the only member in the set of nodes
with low-delivery probability. In the same way, nodes A and B are
6.3.3. Algorithmic complexity
selected as relays. As the performance constraint (c) is still not
The EAMDM scheme comprises the following tasks. Task 1:
satisfied, node E is also selected as a relay. So for the data item d2,
Arranging the data items in decreasing order of their priority. Task
nodes A, B and E are selected as relays.
2: Dividing the contacted neighbors into the set of nodes with sufficient
buffer and insufficient buffer size. Task 3: Dividing the nodes with
sufficient buffer into high and low delivery probability sets. Task 4: 7. Performance evaluations
Arranging the nodes in increasing order of their residual energy. Task
5: Executing the while loop. In this case, the worst case arises when all This section presents performance results of both qualitative and
the contacted neighbors are selected as relays. Now considering n quantitative analyses of our proposed schemes.
number of data items and m number of contacted neighbors, the worst
case complexities of the above tasks are as follows. Task 1: O(n logn). 7.1. Qualitative analysis
Task 2: O(m). Task 3: O(m). Task 4: O(m logm). Task 5: O(m). Here
the Tasks 2–5 are repeated for every data item in the set. So the worst In this subsection, we present the qualitative analysis for the
case time complexity of the proposed scheme is O(nm logm). proposed schemes. We consider smartphone Nokia N95 as DTN node.
This is ARM11 microprocessor (Wang and Zwolinski, 2014) based
6.3.4. Illustrative example smartphone running Symbian OS 9.2. The specification and energy
The relay selection procedure in EAMDM is illustrated in Fig. 3. model for such phone are provided in Perrucci et al. (2011).
Data source S is assumed to multicast two data items d1 and d2 Communication energy is consumed due to transmission and reception
achieving delivery ratio 0.9 within specified time 13 and 22 h respec- of data. In Nokia N95 specification Perrucci et al. (2011), power
tively. Let us consider e0 = 10% and S has seven contacted neighbors consumption for transmitting data at a speed 700 kBps through WiFi
(A, B, C, D, E, F and G) with their respective 3-tuple parameter values IEEE802.11 (ad hoc) is 1629 mW. So energy consumption for trans-
of qij, Ej and Bj, where qij (i.e., 1 − pij ) is the average probability that mitting 1-byte data is as follows.
the node cannot deliver the data item to its destinations within its time
1629
constraint. For simplicity, here we assume, q1j = q2j = qj for all the = 2.272 μJ
potential relay nodes. We further consider that data items d1 and d2 are 700 × 1024
of sizes 6 and 4 bytes respectively. Here d1 is of higher priority than d2, According to Perrucci et al. (2011), power consumption for receiving
since d1 needs to be delivered within smaller time interval than d2. So data at a speed 700 kBps through WiFi IEEE802.11 (ad hoc) is
d1 is to be relayed first. A, B and C are the nodes with required buffer 1375 mW. So energy consumption for receiving 1-byte data is as
space among the contacted neighbors of S, because each of them has follows.
available buffer space more than the size of data item d1. Now among
1375
the nodes A, B and C, A and B are in the set of high-delivery probability = 1.918 μJ
700 × 1024
and C is the only member in the set of nodes with low-delivery
probability. Firstly, node A is selected as a relay, because it has higher Computation energy is consumed due to execution of various
residual energy than that of B. But after selecting this, the performance instructions. Table 2 provides energy requirement to implement

Fig. 3. Relay selection process in EAMDM scheme.

176
A. Roy et al. Journal of Network and Computer Applications 87 (2017) 169–184

Table 2 Let us consider a DTN with 40 nodes and a source node has 5
Power consumption for relevant algorithmic constructs as per instruction set of ARM11 contacted neighbors. The data item is assumed to be 20-byte in size.
microprocessor.
Now for this example, the energy consumption and storage overhead as
Algorithmic constructs Power consumption/instruction (mW) per Table 3 are evaluated for a particular node as follows.
(i) Communication energy
LOAD 225
COMPARE 751
Transmission energy:
STORE 125
ADD 200 SDM – Energy consumption = 2.272 × 20 = 45.44 μJ .
MUL 174 EASDM – Energy consumption = 2.272 × 20 = 45.44 μJ .
SUB 205.2 Reception energy:
SDM – Energy consumption = 1.918 × 20 = 38.36 μJ .
EASDM – Energy consumption = 1.918 × 20 = 38.36 μJ .
different relevant algorithmic constructs (Wang and Zwolinski, 2014).
(ii) Computation energy
As per ARM11 microprocessor (Wang and Zwolinski, 2014), processor
speed is 515 MIPS. We use the following formula to evaluate energy
(7.8924 + 11.1104 × 5) × (25 + 43 × 5)
consumption for computation: SDM – Energy consumption = 515
= 29.56 μJ .

Total no. of instructions (I) EASDM – Energy consumption

Energy = Power(P) × Time (t) = Power (P) × (13.2808 + 6.2334 × 5) × (44 + 23 × 5)
515 × 106 = 515
= 13.72 μJ .
Further, in a node (Nokia N95), the total stored energy is calculated (iii) Storage overhead
using following formula:
SDM – Total number of bytes = 4 × 5 = 20 .
Q×V
E= EASDM – Total number of bytes = 6 × 5 = 30 .
1000 From the above analysis, we clearly observe that the communica-
where E, Q, V are the energy (W h), electric charge (mA h) and voltage tion energy consumption is the same in both SDM and EASDM. But the
(V) respectively. In our case, battery capacity (Q)=1500 mA h, voltage computation energy consumption is lower in EASDM than SDM. On
(V)=3.8 V. Hence, total stored energy (E)=20,520 J. the other hand, storage overhead is a little higher in EASDM than SDM.
Summarily, total energy consumption is lower in EASDM than SDM at
7.1.1. EASDM scheme the cost of a few more storage which is affordable for low-power devices
Table 3 presents the comparative energy consumption (commu- now-a-days.
nication and computation) and storage overhead for the proposed
EASDM with the SDM (Gao et al.). 7.1.2. EAMDM scheme
Communication energy: Assuming size of the data item is s, energy In MDM (Gao et al.), the task of community detection (Hui et al.,
consumptions in transmission and reception are as follows. 2007) requires a significant number of information exchange between
Transmission: Energy consumed in SDM is (2.272 × s )) μJ . the nodes when a contact occurs. The exchanged information between
Energy consumed in EASDM is (2.272 × s ) μJ . these two nodes, includes the following:
Reception: Energy consumed in SDM is (1.918 × s ) μJ.
Energy consumed in EASDM is (1.918 × s ) μJ . (i) Set of encountered nodes (N1) with contact duration.
Computation energy: Let k be the number of contacted neighbors. (ii) Set of nodes (N2) in familiar set.
SDM: The number of instructions to be executed is (25 + 43 × k ). (iii) Set of local community nodes (N3).
The energy consumption for computation is (iv) No. of nodes (N4) in the familiar sets of all the nodes that are in
(7.8924 + 11.1104 × k ) × (25 + 43 × k )
515
μJ . the local community.
EASDM: The number of instructions to be executed is (44 + 23 × k ).
The energy consumption for computation is We assume that the node-id and contact duration are 1-byte and 4-
(13.2808 + 6.2334 × k ) × (44 + 23 × k )
515
μJ . byte in size respectively. In our proposed scheme EAMDM, to reduce
Storage overhead: In the SDM scheme, a source node stores the the energy consumption, we consider centrality. Here the CCP value
centrality values of the contacted neighbors. But in our proposed which is a float value (4-byte) is exchanged between two nodes when a
EASDM scheme, a source node stores the residual energy and also contact occurs.
centrality values of the contacted neighbors. Centrality and residual Table 4 presents the comparative energy consumption (commu-
energy of a node are assumed to be a float (4-byte) and an integer (2- nication and computation) and storage overhead for the proposed
byte) values respectively. Considering k number of contacted neigh- EAMDM with the MDM (Gao et al.).
bors, Communication energy: Energy consumptions in transmission and
reception are as follows.
Storage overhead in SDM is (4 × k ) bytes. Transmission: Energy consumed in MDM is
Storage overhead in EASDM is (6 × k ) bytes. (2.272 × (N1 + (N1 × 4) + N2 + N3 + N4 )) μJ.

Table 3
Comparison of energy consumption in single-data multicast.

Scheme Communication energy (μJ) Computation energy (μJ) Storage overhead (byte)

Transmission Reception

SDM (Gao et al.) 2.272 × s 1.918 × s (7.8924 + 11.1104 × k ) × (25 + 43 × k ) 4×k

515

EASDM 2.272 × s 1.918 × s (13.2808 + 6.2334 × k ) × (44 + 23 × k ) 6×k

515

177
A. Roy et al. Journal of Network and Computer Applications 87 (2017) 169–184

Table 4
Comparison of energy consumption in multiple-data multicast.

Scheme Communication energy (μJ) Computation energy (μJ) Storage overhead (byte)

Transmission Reception

MDM (Gao et al.) 2.272 × (N1 + (N1 × 4) + N2 + N3 + N4 ) 1.918 × (N1 + (N1 × 4) + N2 + N3 + N4 ) Pmdm × Imdm (N1 + (N1 × 4) + N2 + N3 + N4 )
515

EAMDM 2.272 × 4 1.918 × 4 Peamdm × Ieamdm 6×m

515

Energy consumed in EAMDM is (2.272 × 4) μJ. (i) Communication energy: Considering only the exchange of
Reception: Energy consumed in MDM is messages upon node-contact.
(1.918 × (N1 + (N1 × 4) + N2 + N3 + N4 )) μJ. Transmission energy:
Energy consumed in EAMDM is (1.918 × 4) μJ.
Computation energy: It is assumed that, for a particular session, n MDM – Energy consumption
is the total number of data items, m is the number of contacted =2.272 × (4 + 4 × 4 + 3 + 2 + 1) = 59.07 μJ .
neighbors, b is the number of nodes with sufficient buffer space, h is the EAMDM – Energy consumption = 2.272 × 4 = 9.08 μJ .
number of nodes with high centrality among the nodes with sufficient Reception energy:
buffer space. MDM – Energy consumption
MDM: Total no. of instructions =1.918 × (4 + 4 × 4 + 3 + 2 + 1) = 49.86 μJ .
Imdm = m × (60 log m + 18n ) + 60 log m + m × 37, power consumption EAMDM – Energy consumption = 1.918 × 4 = 7.67 μJ .
Pmdm = m × (17.3808 × log m + n × 4.4322) + 17.3808 × log m . (ii) Computation energy
+ m × 9.5646
The energy consumption for computation is MDM – Energy consumption =1.908 mJ .
{(m × (17. 3808 × log m + n × 4. 4322) + 17. 3808 × log m . EAMDM – Energy consumption = 0.082 mJ .
+ m × 9. 5646) × (m × (60 log m + 18n ) + 60 log m + m × 37)}/515 μJ (iii) Storage overhead
EAMDM: Total no. of instructions
Ieamdm = 60 × log(nh ) + 60 × (b − h ) + 33 × log b + 20 × m + 2 , power MDM – Total number of bytes = 4 + 4 × 4 + 3 + 2 + 1 = 26 .
consumption EAMDM – Total number of bytes = 6 × 5 = 30 .
Peamdm = 17.38(log n + log h + b − h ) + 8.1584 × log b + 4.952 × m . From the above analysis, we clearly observe that the communica-
+ 0.35 tion and computation energy consumption are lower in EAMDM than
The energy consumption for computation is MDM. But the storage overhead is slightly higher in EAMDM than
{(17.38(log n + log h + b − h ) + 8.1584 × log b + 4.952 × m + 0.35) × . MDM.
(60 × log(nh ) + 60 × (b − h ) + 33 × log b + 20 × m + 2)}/515 μJ
Storage overhead: In our proposed EAMDM scheme, a source node 7.2. Quantitative analysis
stores the centrality and also residual energy values of the contacted
neighbors. Storage overhead in MDM is (N1 + (N1 × 4) + N2 + N3 + N4 ) This subsection presents the quantitative analysis for both the
bytes. Storage overhead in EAMDM is (6 × m ) bytes. proposed schemes, EASDM and EAMDM.
Let us consider DTN with 40 nodes. Let us also consider in the
process of forming a small community, for a particular node, N1 = 4 , 7.2.1. Simulation set-up
N2 = 3, N3 = 2 , N4 = 1. Further, it is also assumed that n=10, m=5, b=2, In 2013, devastating floods and landslides occurred in the North
h=1. Now for this example, the energy consumption as per Table 4 is Indian state of Uttarakhand. In 2015, Nepal experienced a major
evaluated as follows. disaster by an earthquake of magnitude 7.8. For both the disasters,
Crisis Maps have been created by the Google Crisis Response team as

Fig. 4. Disaster impact map published by Google crisis response team.

178
A. Roy et al. Journal of Network and Computer Applications 87 (2017) 169–184

shown in Fig. 4. The Uttarakhand disaster map (http://google.org/ average of at least 10 runs. We also statistically analyze the simulated
crisismap/2013-uttrakhand-floods?gl=in) covers from Badrinath to results.
Rudraprayag (8.5 km×8.5 km) and the Nepal disaster map (http:// Simulation results: During simulation using Nepal disaster map in
google.org/crisismap/2013-uttrakhand-floods?gl=in) covers from ONE (Keranen et al., 2009) simulator, exchanged WhatsApp messages
Kathmandu to Bharatpur (14 km×14 km). We evaluate the perfor- among the Doctors For You (DFY) team members during relief
mance of our proposed schemes in ONE (Keranen et al., 2009) operations after the Nepal earthquake, are generated by the DTN
simulator taking two such real life post-disaster scenarios. In the nodes.
disaster-affected areas, several NGO groups took major part in rescue EASDM scheme: Figs. 5 and 6 show the simulation results in terms
and relief operations. During simulation based on the Nepal disaster of the four said performance metrics. In Figs. 5(a) and 6(a), variation of
map, WhatsApp messages (a sample shown in Table 1) which were delivery ratio with the delivery time constraint (TTL) is depicted. We
exchanged among the Doctors For You (DFY) team members in the observe that delivery ratio increases with increasing time constraint.
course of relief operations after the Nepal earthquake, are fed into the Proposed EASDM scheme provides almost the same delivery ratio as
ONE (Keranen et al., 2009) simulator. From the exchanged WhatsApp the SDM scheme, but it achieves higher delivery ratio than the
messages among the DFY volunteers, it is observed that in disaster PROPHET. It may be noted that highest delivery ratio is obtained in
relief operations, usually short messages are generated and the size of a Epidemic, because it uses all the nodes as relays for data delivery.
message is not more than 120 bytes. So we set the range of message Precisely, when the time constraint is 6 h, in Uttarakhand data set, the
size as 20–120 bytes. According to the various disaster response values of delivery ratio are 0.44, 0.68, 0.69 and 0.82 in PROPHET,
agencies and NGOs, post-disaster rescue and relief operations are SDM, EASDM and Epidemic respectively, whereas these values are
generally performed during daytime. So we set 12 hours (43,200 s) as 0.37, 0.59, 0.6 and 0.64 respectively for Nepal data set.
the simulation time. The important simulation parameters with their In Figs. 5(b) and 6(b), variation of average cost with time constraint
values are specified in Table 5. For energy consumption of the DTN is shown. Here average cost is in terms of number of relays which
nodes, we use energy model provided in Perrucci et al. (2011). During increases with increasing time constraint. The reason is, increase in
the simulation, data is multicast to 10 randomly selected destinations. TTL results in higher delivery ratio which increases number of relays
used. Significantly, average cost is lower in our proposed EASDM
scheme than the competing schemes. The reason is that EASDM
7.2.2. Performance metrics and candidate schemes
scheme uses the high-centrality nodes judiciously. In SDM, centrality
We consider four simulation metrics viz. delivery ratio, average
is the only consideration for choosing a node as a relay. This results in
cost, average delay and average residual energy. The metrics are
the selection of nodes with high centrality as relays more frequently in
defined as follows.
every multicast session. So the high-centrality nodes become inactive

• Delivery ratio: The ratio of the number of delivered destinations to

soon for over usage. The number of relays increases due to unavail-
ability of high-centrality nodes. In our proposed scheme, in addition to
the total number of target destinations.
• Average cost: The average number of relays used to receive a data
centrality metric, we also consider a node's residual energy for relay
selections. Hence high-centrality nodes with comparatively lower
item at the destinations, which receive the message.
• Average delay: The average value of the delay to each delivered
residual energy are reserved to be used for next subsequent sessions.
Precisely, when the time constraint is 6 h, in Uttarakhand data set,
destination.
• Average residual energy: The average value of the residual energy
values of average cost are 5.69%, 29.69% and 46.04% lower in EASDM
than SDM, PROPHET and Epidemic respectively, whereas these values
of each node after the completion of simulation.
are 15.89%, 39.52% and 52.07% lower in EASDM respectively for
Nepal data set.
Following candidate schemes are considered for performance
In Figs. 5(c) and 6(c), variation of average delay with the time
comparisons.
constraint is shown. Average delay increases with time for achieving

• Proposed EASDM and EAMDM

higher delivery ratio. We observe, average delay is slightly longer in our

• SDM (Gao et al.), MDM (Gao et al.)

proposed EASDM scheme than the SDM scheme. The reason is, as the

• Epidemic (Vahdat and Becker, 2000)

EASDM scheme does not use all the available high-centrality nodes, it

• PROPHET (Lindgren et al., 2003)

incurs slightly longer latency to achieve the required delivery ratio
within the specified time constraint. It may be noted that lowest
delivery ratio is obtained in Epidemic, because it uses the flooding
7.2.3. Results and discussion approach for data delivery. Precisely, when the time constraint is 6 h, in
We conduct two sets of experiments to evaluate performance of Uttarakhand data set, values of average delay are 2.04% and 10.66%
both the proposed schemes for each of the two post-disaster scenarios, higher in EASDM than SDM and Epidemic respectively, whereas these
Uttarakhand and Nepal disaster map, comparing with the other values are 3.08% and 11.56% higher in EASDM respectively for Nepal
candidate schemes. In the presented plots, each data point is the data set. But EASDM achieves 3.83% and 5.72% lower average delay
than PROPHET in Uttarakhand and Nepal map respectively.
Table 5 Figs. 5(d) and 6(d) show the average residual energy of each node
Simulation parameters. after the total simulation time of 12 h. Significantly, average residual
Parameter Value
energy is found to be highest in EASDM among all the candidate
schemes. In EASDM, residual energy of a node is also criteria for relay
Simulation time 12 h selections. Average residual energy of each node is lowest in Epidemic
Warmup time 20% of the simulation time as all the nodes work as relays in this scheme. Precisely, in Uttarakhand
No. of nodes 60
Initial Energy 20,520 J
data set, values of average residual energy in SDM, PROPHET and
Scan Energy 0.223 J/s Epidemic are 94.313 mJ, 137.375 mJ and 221.001 mJ lower than that
Base Energy 0.067 J/s of EASDM respectively. The same values in the respective schemes are
Transmit Energy 0.432 J/s 120.187 mJ, 153.001 mJ and 298.125 mJ lower than that of EASDM
Range of Message size 20–120B
for Nepal data set.
Transmission speed, range 250 kBps, 10 m
Movement model Map-based movement EAMDM scheme: Similar to EASDM scheme, two sets of simulation
results are plotted in Figs. 7 and 8. It may be noted that here all the data

179
A. Roy et al. Journal of Network and Computer Applications 87 (2017) 169–184

Fig. 5. Performance of EASDM on Uttarakhand disaster map.

Fig. 6. Performance of EASDM on Nepal disaster map.

items do not have identical delivery time constraint and the proposed creases with increasing time constraints. In the proposed EAMDM
EAMDM scheme prioritizes the data items based on their individual scheme, a data item with lower delivery time constraint is assigned
time constraint. In Figs. 7(a) and 8(a), variation of delivery ratio with the higher priority. As a result, EAMDM produces higher delivery ratio than
delivery time constraint (TTL) is depicted. We observe that for all the the MDM scheme, when the time constraint is lower. For the data items
schemes (MDM, EAMDM, Epidemic, PROPHET), delivery ratio in- with higher time constraint, because of lower priority, EAMDM achieves

180
A. Roy et al. Journal of Network and Computer Applications 87 (2017) 169–184

Fig. 7. Performance of EAMDM on Uttarakhand disaster map.

Fig. 8. Performance of EAMDM on Nepal disaster map.

slightly lower delivery ratio than the MDM scheme. Precisely, when the In Figs. 7(b) and 8(b), variation of average cost with time constraint
time constraint is 6 h, in Uttarakhand data set, the values of delivery is shown. With the increase in time constraint, average cost in terms of
ratio are 0.38, 0.59, 0.6 and 0.71 in PROPHET, EAMDM, MDM and number of relays increases for achieving higher delivery ratio.
Epidemic respectively, whereas these values are 0.28, 0.47, 0.48 and Interestingly, average cost is found to be lower in our proposed
0.58 respectively for Nepal data set. EAMDM scheme than the competing schemes. Precisely, when the

181
A. Roy et al. Journal of Network and Computer Applications 87 (2017) 169–184

time constraint is 6 h, in Uttarakhand data set, values of average cost imperative, especially in energy-hungry tasks like data forwarding. So
are 8.96%, 24.57% and 45% lower in EAMDM than MDM, PROPHET in such scenario, there should be a balance between delivery ratio gain
and Epidemic respectively, whereas these values are 10.54%, 31.37% and resource consumption. In this case, an acceptable delivery ratio
and 49.82% lower in EAMDM respectively for Nepal data set. needs to be achieved consuming less resources in terms of relay and
In Figs. 7(c) and 8(c), variation of average delay with the time energy. It is clear from the results that our proposed schemes, EASDM
constraint is shown. Average delay increases with the increase in time and EAMDM, meet this requirement thereby enhancing network
constraint. However, average delay is slightly longer in our proposed lifetime. Moreover, message generation rate at a DTN node is not the
EAMDM scheme than the MDM scheme. The reason is, as the EAMDM same for all the times in practical scenario. For example, in post-
scheme uses lesser number of relays, it incurs slightly longer latency to disaster environment, message generation rate is high when the DTN
achieve the required delivery ratio within the specified time constraint. nodes are disseminating situational updates, while this rate is low when
Precisely, when the time constraint is 6 h, in Uttarakhand data set, the nodes are sending sentimental messages. First of all, a DTN node
values of average delay are 3.13% and 9.36% higher in EAMDM than must have sufficient residual energy to forward data when message
MDM and Epidemic respectively, whereas these values are 3.07% and generation rate is high. Results clearly show that average residual
9.18% higher in EAMDM respectively for Nepal data set. But EAMDM energy of the nodes is highest in both EASDM and EAMDM. For
achieves 6.63% and 6.52% lower average delay than PROPHET in instance, in EASDM, a DTN node is capable to transmit 172 KB more
Uttarakhand and Nepal map respectively. than Epidemic at 250 kBps, since EASDM saves 298.125 mJ more than
In Figs. 7(d) and 8(d), the average residual energy of each node is Epidemic on average for every node in Nepal map. In post-disaster
shown after the completion of the simulation time. Significantly, in environment, usually short messages are exchanged. It is found that
EAMDM, average residual energy is higher than all the competing the maximum size of a message is 120B. If maximum message size is
schemes. Residual energy of a node is considered in EAMDM to select considered, the number of messages that can be forwarded by each
relays. Precisely, in Uttarakhand data set, values of average residual node using EASDM is 1433 (172KB/120B) more than Epidemic.
energy in PROPHET, MDM and Epidemic are 97.187 mJ, 124.5 mJ Statistical analysis of simulated results: Now to validate our
and 242.813 mJ lower than that of EAMDM respectively, whereas simulation, we statistically analyze the results obtained from the
these values for the same are 143.087 mJ, 194.4 mJ and 365.713 mJ simulation runs at the time constraint of 6 h. Figs. 9 and 10 are shown
lower than EAMDM respectively in Nepal data set. at 95% confidence interval.
From the above discussion, it is observed that Epidemic which EASDM scheme: Fig. 9(a) shows the delivery ratio of each scheme
floods the data across the network achieves highest delivery ratio at the at 95% confidence interval on Uttarakhand and Nepal disaster maps.
cost of maximum relay usage and battery power dissipation. Because of Apart from PROPHET, the width of the confidence interval is small in
huge resource consumption in terms of relay and battery power, it may each of the schemes. So PROPHET has highest variability in delivery
be considered unsuitable in the DTN application of post-disaster ratio among all the candidate schemes. It may be noted that the
management. Node-density is low in DTNs and especially, in the difference between the means of SDM and EASDM is not statistically
application of post-disaster management, DTN nodes may often significant, because the interval bars of these two schemes overlap.
required to be deployed in locations having almost no scope of In Fig. 9(b), average costs of all the schemes are shown at 95%
recharging and replacing their batteries. In the presence of the said confidence interval on Uttarakhand and Nepal disaster maps. As width
constraints, judicious utilization of any node's battery power is of the confidence interval is higher in Epidemic and PROPHET, they

Fig. 9. Confidence intervals for EASDM.

182
A. Roy et al. Journal of Network and Computer Applications 87 (2017) 169–184

Fig. 10. Confidence intervals for EAMDM.

have higher variability in average cost than SDM and EASDM. It is clear In Fig. 10(d), average residual energy of all the schemes are shown at
that the differences between the means of the schemes are statistically 95% confidence interval on Uttarakhand and Nepal disaster maps. Here
significant, because the interval bars of the schemes do not overlap. the width of the confidence interval is small in every scheme. This means,
Fig. 9(c) shows the average delay of each scheme at 95% confidence all the schemes have small variability in average residual energy. It may be
interval on Uttarakhand and Nepal disaster maps. PROPHET has noted that the differences between the means of the schemes are
highest variability in average delay, because its width of the confidence statistically significant as the interval bars of the schemes do not overlap.
interval is highest among all. It is observed that the differences between
the means of SDM, EASDM and PROPHET are statistically not
significant, because the interval bars of these schemes almost overlap. 8. Conclusion
In Fig. 9(d), average residual energy of all the schemes are shown at
95% confidence interval on Uttarakhand and Nepal disaster maps. All In this paper, we study the necessity of energy-aware data
the schemes have small variability in average residual energy, as the forwarding schemes for multicast in DTNs. Since the existing social-
width of the confidence interval is small in every scheme. It is also clear based multicast schemes do not consider the energy limitations of the
that the differences between the means of the schemes are statistically DTN nodes, these are not suitable for multiple multicast sessions.
significant, because the interval bars of the schemes do not overlap at all. Supporting multiple multicast sessions is imperative in the scenario
EAMDM scheme: Fig. 10(a) shows the delivery ratio of each scheme where recharging facility is rarely available for the nodes in the
at 95% confidence interval on Uttarakhand and Nepal disaster maps. network, e.g. in post-disaster scenario. The two problems for energy-
The width of the confidence interval is small in almost each of the aware relay selection in single and multiple-data multicast are shown to
schemes. It indicates that there is not much variability in delivery ratio be NP-hard. We propose energy-aware relay selection schemes to
of the schemes. As the interval bars of MDM and EAMDM overlap, the multicast single and multiple data items. Relays are selected based
difference between their means is not statistically significant. on the centrality as well as residual energy of the nodes. Complexity of
In Fig. 10(b), average costs of all the schemes are shown at 95% each of the schemes is shown to be polynomial time. We evaluate the
confidence interval on Uttarakhand and Nepal disaster maps. In this performance of our proposed schemes theoretically and also by
case also, not so much variability is observed in average cost of the exhaustive simulation through ONE simulator based on real disaster
schemes. So the width of the confidence interval is small in almost maps and messages exchanged among the DFY (NGO) volunteers
every scheme. Interestingly, the differences between the means of the during the relief operations after the Nepal earthquake. Combining the
schemes are statistically significant, because the interval bars of the results of qualitative and quantitative analyses, we may conclude that
schemes do not overlap. the proposed schemes are able to achieve acceptable delivery ratio
Fig. 10(c) shows the average delay of each scheme at 95% confidence consuming less resources in terms of number of relays and battery
interval on Uttarakhand and Nepal disaster maps. In PROPHET, highest power of the nodes. Interestingly, the nodes using our proposed
width of the confidence interval indicates that variability in average delay schemes can forward significant number of messages more than the
is highest in this scheme among all the schemes. Interval bars of the other prominent schemes without any necessity of recharging them.
schemes do not overlap here. So the differences between the means of Hence, both analyses confirm that the proposed schemes are energy-
the schemes are probably statistically significant. saving thereby enhance network lifetime, while maintaining fair net-
work performance.

183
A. Roy et al. Journal of Network and Computer Applications 87 (2017) 169–184

In future, we intend to validate the existing study by developing delay tolerant networks, In: Proceedings of the ACM/IEEE MobiArch.
Keranen, A., Ott, J., Karkkainen, T., 2009. The ONE simulator for DTN protocol
smartphone based test-bed and perform the necessary experimenta- evaluation, In: Proceedings of the ACM SIMUTOOLS.
tion. Further, future work may also be carried out in the development Lee, U., Oh, S.-Y., Lee, K.-W., Gerla, M., 2008. Scalable multicast routing in delay
of congestion-aware scheme, since repeated selection of few nodes with tolerant networks, In: Proceedings of the ICNP.
Lindgren, A., Doria, A., Schelen, O., 2003. Probabilistic routing in intermittently
high centrality and energy level may result in congestion in the DTNs. connected networks. ACM SIGMOBILE Comput. Commun. Rev. 7 (3), 19–20.
Lust, T., Teghem, J., 2012. The multi-objective multidimensional knapsack problem: a
Acknowledgment survey and a new approach. Int. Trans. Oper. Res. 19 (4), 495–520.
Martin-Campillo, A., Crowcroft, J., Yoneki, E., Marti, R., 2013. Evaluating opportunistic
networks in disaster scenarios. J. Netw. Comput. Appl. 36 (2), 870–880.
This work is partially supported by Information Technology Mongiovi, M., Singh, A.K., Yan, X., Zong, B., Psounis, K., 2012. Efficient multicasting for
Research Academy (ITRA), Government of India, under ITRA-Mobile delay tolerant networks using graph indexing, In: Proceedings of the IEEE
INFOCOM.
grant [ITRA/15(58)/Mobile/DISARM/01]. We also acknowledge
Moreira, W., Mendes, P., Sargento, S., 2013. Opportunistic routing based on users daily
Doctors For You (DFY) team for sharing WhatsApp messages ex- life routine, In: Proceedings of the IEEE International Symposium on a World of
changed during Nepal earthquake. Wireless, Mobile and Multimedia Networks (WoWMoM), pp. 1–6.
Perrucci, G., Fitzek, F., Widmer, J., 2011. Survey on energy consumption entities on the
smartphone platform, In: Proceedings of the IEEE 73rd Vehicular Technology
References Conference (VTC Spring), pp. 1–6.
Quercia, D., Lathia, N., Calabrese, F., Lorenzo, G.D., Crowcroft, J., 2010. Recommending
Bhattacharjee, S., Basu, S., Roy, S., DasBit, S., 2016. Best-effort delivery of emergency social events from mobile phone location data, In: Proceedings of the IEEE
messages in post-disaster scenario with content-based filtering and priority- International Conference on Data Mining, pp. 1–6.
enhanced PRoPHET over DTN, In: Proceedings of the COMSNETS. Rodrigues, J., 2014. Advances in Delay-tolerant Networks (DTNs): Architecture and
Chuah, M.C., 2009. Social network aided multicast delivery scheme for human contact- Enhanced PerformanceWoodhead Publishing Series in Electronic and Optical
based networks, In: Proceedings of the 1st Annual Workshop on Simplifying Materials. Woodhead Publishing, (ISBN 978-0-85709-840-5)https://books.google.
Complex Networks for Practitioners (SIMPLEX). co.in/books?id=hJCuAwAAQBAJ.
Ciobanu, R.I., Dobre, C., 2012. Predicting encounters in opportunistic networks, In: Roy, A., Acharya, T., DasBit, S., 2015. Energy-aware social-based multicast in delay-
Proceedings of the 1st ACM Workshop on High Performance Mobile Opportunistic tolerant networks, In: Proceedings of the 81st IEEE Vehicular Technology
Systems, pp. 1–6. Conference (VTC Spring), pp. 1–5.
Ciobanu, R.-I., Marin, R.-C., Dobre, C., Cristea, V., Mavromoustakis, C.X., 2014. Onside: Sapountzi, A., Psannis, K.E., 2016. Social networking data analysis tools & challenges.
socially-aware and interest-based dissemination in opportunistic networks, In: Future Gener. Comput. Syst.. http://dx.doi.org/10.1016/j.future.2016.10.019.
Proceedings of the IEEE Network Operations and Management Symposium Sobin, C.C., Raychoudhury, V., Marfia, G., Singla, A., 2016. A survey of routing and data
(NOMS), pp. 1–6. dissemination in delay tolerant networks. J. Netw. Comput. Appl. 67, 128–146.
Cormen, T.H., Leiserson, C.E., Rivest, R.L., Stein, C., 2009. Introduction to Algorithms http://dx.doi.org/10.1016/j.jnca.2016.01.002.
3rd edition. MIT Press, (ISBN 9780262258104). Spyropoulos, T., Psounis, K., Raghavendra, C., 2005. Spray and wait: an efficient routing
Deng, X., Chang, L., Tao, J., Pan, J., Wang, J., 2013. Social profile-based multicast scheme for intermittently connected mobile networks, In: Proceedings of the ACM
routing scheme for delay-tolerant networks, In: Proceedings of the IEEE ICC. SIGCOMM Workshop on Delay-tolerant Networking, pp. 252–259.
Doctors for you, URL 〈http://doctorsforyou.org/〉. Vahdat, A., Becker, D., 2000. Epidemic routing for partially connected ad hoc networks.
Fall, K., 2003. A delay-tolerant network architecture for challenged internets, In: Technical Report CS-200006. Duke University.
Proceedings of the ACM Special Interest Group Data Communication Workshop Wang, Y., Wu, J., 2012. A dynamic multicast tree based routing scheme without
(SIGCOMM). replication in delay tolerant networks. Parallel Distrib. Comput. 72 (3), 424–436.
Galluccio, L., Lorenzo, B., Glisic, S., 2016. Sociality-aided new adaptive infection Wang, W., Zwolinski, M., 2014. An improved instruction-level power model for ARM11
recovery schemes for multicast dtns. IEEE Trans. Veh. Technol., 99. http:// microprocessor, In: Proceedings of the High Performance Energy Efficient
dx.doi.org/10.1109/TVT.2015.2450202. Embedded Systems (HIP3ES).
Gao, W., Li, Q., Zhao, B., Cao, G., 2012. Social-aware multicast in disruption-tolerant Wang, Y., Li, X., Wu, J., 2010. Multicasting in delay tolerant networks: delegation
networks. IEEE/ACM Trans. Netw. 20 (5). forwarding, In: Proceedings of the IEEE GLOBECOM.
Google crisis map, Uttarakhand Flooding, http://google.org/crisismap/2013- Wei, K., Liang, X., Xu, K., 2014. A survey of social-aware routing protocols in delay
uttrakhand-floods?Gl=in (2013). tolerant networks: Applications, taxonomy and design-related issues. IEEE
Google crisis map, Nepal Earthquake, 2015, https://google.org/crisismap/2015-nepal- Commun. Surv. Tutor. 16 (1), 556–578.
earthquake. Xia, F., Liu, L., Li, J., Ma, J., Vasilakos, A., 2015. Socially aware networking: A survey.
Grossglauser, M., Tse, D.N.C., 2002. Mobility increases the capacity of ad hoc wireless IEEE Syst. J. 9 (3), 904–921.
networks. IEEE Trans. Netw. 10 (4). Zhao, W., Ammar, M., Zegura, E., 2004. A message ferrying approach for data delivery in
Hossmann, T., Spyropoulos, T., Legendre, F., 2010. Know thy neighbor: towards optimal sparse mobile ad hoc networks, In: Proceedings of the MobiHoc, pp. 187–198.
mapping of contacts to social graphs for DTN routing, In: Proceedings of the IEEE Zhu, K., Li, W., Fu, X., Zhang, L., 2015. Data routing strategies in opportunistic mobile
INFOCOM, pp. 1–9. social networks: taxonomy and open challenges. Comput. Netw. 93, 183–198.
Hui, P., Yoneki, E., Chan, S., Crowcroft, J., 2007. Distributed community detection in http://dx.doi.org/10.1016/j.comnet.2015.10.018.

184