SPECIAL SECTION ON BIG DATA SERVICES AND COMPUTATIONAL INTELLIGENCE

FOR INDUSTRIAL SYSTEMS

Received October 8, 2015, accepted November 14, 2015, date of publication December 9, 2015,
date of current version December 23, 2015.
Digital Object Identifier 10.1109/ACCESS.2015.2507201

Toward Highly Dependable Power-Aware

Mobile Ad Hoc NetworkSurvivability
Evaluation Framework
ZHIPENG YI AND TADASHI DOHI
Hiroshima University, Higashihiroshima 739-8527, Japan
Corresponding author: T. Dohi (dohi@rel.hiroshima-u.ac.jp)

ABSTRACT Network survivability is an attribute that network is continually available even if a

communication failure occurs, and is regarded as one of the most important concepts to design dependable
computer networks. In the existing work, a power-aware mobile ad hoc network (MANET) is described by
a Markov regenerative process, and takes account of the variability in power level, which is caused by the
possible low-battery state in each communication node. However, it implicitly ignores effects by the so-called
border effects, and lacks the reality in modeling. In this paper, we revisit a power-aware MANET model
taking account of border effects and quantify the network survivability more accurately.

INDEX TERMS Network survivability, network connectivity, power-aware MANET, semi-Markov model,
DoS attack, border effects, battery charge.

I. INTRODUCTION survivability based on k-connectivity, which implies that

There exist a number of challenging issues to provide big data
services in ubiquitous circumstance. The drastic improve-
ment of network performance is definitely needed to pro-
cess a large amount of data, especially, in the system level.
On the other hand, it is important to keep the high service
level on the big data stream and evaluate the network depend-
ability in the design phase to develop highly dependable
ubiquitous network systems. Network survivability is defined
as an attribute that network is continually available even
though a communication failure occurs, and is regarded as
the most fundamental issue to design resilient communi-
cation networks. Since unstructured networks, such as P2P
network and mobile ad hoc network (MANET), can change
dynamically their configurations, the survivability require-
ment for unstructured networks is becoming much more pop-
ular than static networks. In the near future, it is expected that
this trend may be accelerated even in the big data service.
Although quantitative network survivability is defined by
various authors [1][3], it is still a challenging issue from
the complex and autonomous properties of unstructured
networks. Xing and Wang [4] perceive the survivabil-
ity of a wireless ad hoc network as the probabilistic
k-connectivity [5][7], and provide a quantitative analysis
on impacts of both node misbehavior and failure. They
approximately derive the lower and upper bounds of network
every node pair in the network can communicate with at
least k neighbors. Unfortunately, the upper and lower bounds
of network survivability in [4] are not always tight to quantify
the network survivability. Yi and Dohi [8] are motivated by
the above fact and extend the seminal model [4] by introduc-
ing the compound distributions of Poisson model.
Okamura et al. [9] evaluate a power-aware MANET by
using a Markov regenerative process (MRGP) and investigate
an effect of variability in power level, which is caused by
the low-battery state in each node, but does not consider the
possible case where the battery in each mobile node can be
re-charged at the lower battery state in their MRGP mod-
eling framework. In addition, Okamura et al. [9] implicitly
assume that the so-called border effects can be ignored in their
modeling. It is well known that the shape of communication
area with border effects strongly depends on the network
properties. Laranjeira and Rodrigues [10] develop a geom-
etry analysis to quantify the network connectivity, and refine
the reliability assurance of wireless sensor networks. More
specifically, it is common to assume in the network analysis
that the communication node is uniformly distributed in an
ideal communication area. Since the border effects in network
communication areas tend to decrease both the communica-
tion coverage and the node degree, they must reflect the whole
network survivability including availability and reliability.

2169-3536
2015 IEEE. Translations and content mining are permitted for academic research only.
VOLUME 3, 2015 Personal use is also permitted, but republication/redistribution requires IEEE permission. 2665
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Laranjeira and Rodrigues [10] show that the relative average
node degree for nodes in borders is independent of the node
transmission range and of the overall network node density in
a square communication area.
Most recently, Yi and Dohi [11] revisit a power-aware
MANET model in Xing and Wang [4] and Yi and Dohi [8]
taking account of both border effects and the possibility
of re-charge, and quantify the network survivability more
accurately. They suppose that each node state is modulated
by a semi-Markov process and that the node density in an
arbitrary communication area is given by a simple Poisson
model, where two types of communication areas are
considered; square area [10] and circular area [6]. In this
paper we further extend the above result for the other stochas-
tic models by including a binomial model and a negative
binomial model [8]. We derive analytically the upper and
lower bounds of network survivability [4] as well as an
approximate form based on the expected number of active
nodes [8] in both square [10] and circular [6] areas, under
a general assumption that the battery life in each node is
non-exponentially distributed. Also, we perform the transient
analysis as well as the steady-state analysis [11] of network
survivability, and complements our early paper [11].
The remaining part of this paper is organized as follows.
In Section II, we define the state of each node in our stochastic
model to describe the behavior of a power-aware MANET.
Based on the familiar semi-Markov analysis, the transition
behavior of the network node is analyzed. Section III is
devoted to the network survivability analysis, where the node
isolation, network connectivity and network survivability are
defined. Here, we present the network survivability formula
with the probability that the active nodes are k-connected
and present three stochastic models; Poisson model. binomial
model and negative binomial model. In Section IV, we intro-
duce the border effects in two kinds of communication areas;
square area and circular area. Based on some geometric ideas
in [6] and [10], we improve the quantitative network surviv-
ability measures taking account of border effects. Numerical
examples are given in Section V, where we compare three
scenarios on battery re-charge with varying node transmis-
sion radius and number of nodes in both of steady-state
network survivability and transient network survivability.
We also compare our refined network survivability models
with the existing ones without border effects. Finally, the
paper is concluded with some remarks in Section VI.
II. MODEL DESCRIPTION
A. NODE CLASSIFICATION
Since nodes in MANETs cooperate with the routing pro-
cesses to maintain network connectivity, each of nodes is
designed as it behaves autonomously, but its discipline to
require, send and receive the route information, is defined
as a strict protocol. At the same time, it is also important
to define the protocol in order to prevent propagation of
the erroneous route information caused by malicious attacks.
Xing and Wang [4] and Yi and Dohi [8] consider a MANET
2666
Z. Yi, T. Dohi: Toward Highly Dependable Power-Aware MANET
that suffers such a malicious attack, whose node states are
defined as follows:
Cooperative state (C): a node complies with all routing
and forwarding rules.
Selfish state (S): a node may not forward control or data
packets for others for the sake of power saving.
Malicious state (M ): a node launches Jellyfish or Black
hole DoS attack.
Jellyfish state (J ): a node being cooperative in the
routing stage reluctant in forwarding data packets.
Blackhole state (B): a node disrupting legiti-
mate path selections by broadcasting fakes route
replies.
Failed state (F): a node is unable to initiate or response
route discoveries.
Moreover, each node may be classified into the following
states in terms of the battery state:
Fully charged battery state (H ): the battery is fully
charged.
Low battery state (L): the battery level is low and may
cause a failure due to out of power.
It is essential to characterize the power state in power-aware
device. The modeling approach by Okamura et al. [9] and
Yi and Dohi [11] can be considered as an incremental one
in technique but significant extension in reality. For common
DoS attacks, the node in Jellyfish attack receives route
requests and route replies. The main mechanism of Jellyfish
state is to delay packets without any reason. On the other
hand, the node in Blackhole attack can respond a node with a
fake message immediately by declaring as it is in the optimal
path or as it is only one-hop away to other nodes.
Suppose that each of states, i = C, S, M , has one of
two sub-states; H and L. For instance, iH means a node in
the state i with high energy level and iL means a node in the
state i with low energy level. The failed state F also has one of
two sub-states; energy exhaustion (EF) and DoS attack
detection (DF).
B. SEMI-MARKOV NODE MODEL
Based on the above node classification, we consider a semi-
Markov model to describe the stochastic behavior of a
node by combining the network state and the battery state.
We define the node behavior as follows:
A Cooperative node (CH or CL) may become a
Malicious node (MH or ML) when it launches DoS
attack, and a low-battery Cooperative node (CL) may
become a Failed node due to energy exhaustion (EF).
A Cooperative node (CH or CL) may become a Selfish
node (SH or SL) for saving the power.
A Malicious (MH or ML) node cannot become a Coop-
erative node (CH or CL) again, but may become a Failed
node by two reasons: energy exhaustion (EF) and DoS
attack detection (DF).
A node in Failed state (DF or EF) may become a Coop-
erative node (CH ) again after it repairs and responds to
routing requests for others.
VOLUME 3, 2015
Z. Yi, T. Dohi: Toward Highly Dependable Power-Aware MANET

Each node may become the low battery state as operating

time passes, but may become fully-charged battery state
from the low battery state again by re-charge.
From above assumptions, we can define the state space
S {CH , CL, SH , SL, MH , ML, EF, DF}, and the time-
dependent transitions rates from state i to state j (i, j S)
by i,j (t).
Similar to the original idea by Xing and Wang [4],
we describe the transition behavior of each node, by a
stochastic process {Z (t), t 0}, associated with the space S.
Let Xn denote the state at transition time tn . Define
Pr(Xn+1 = xn+1 |X0 = x0 ., , , Xn = xn )
= Pr(Xn+1 = xn+1 |Xn = xn ), (1) FIGURE 1. State transition diagram.

where xi S for 0 i n + 1. From Eq. (1), the stochastic

process {Xn , n = 0, 1, 2, ...} constitutes a continuous-time
occur, it is difficult to clarify the connected topology for the
Markov chain with state space S, when all the transition times
MANET. Since the communication availability of a MANET
are exponentially distributed. However, since the transition
depends on the existing paths between two nodes, it is intu-
time from one state to another state is subject to the time-
itively understood that the network survivability strongly
inhomogeneous behavior of a node, it is not realistic to
depends on the connectivity. In general, it is said that the
characterize all the transition times by only exponentially
MANET is k-connected, if there are at least k disjoint com-
distributed random variables. For instance, if a sensor node
munication paths connecting one node to the other node.
is more inclined to fail due to energy consumption as the
When k = 1, it means the probability that there is at least one
operating time passes, and the less residual energy is left, then
communication path connecting one node to the other node,
the more likely a sensor changes its behavior to selfish. This
and is equivalent to the network reliability. Hence, thinking of
implies that the future action of a node may depend on how
higher network survivability is reduced to highly dependable
long it has been in the current state and that the transition time
MANET design.
intervals should obey arbitrary probability distributions.
Given a MANET M, let (M) denote the vertex-
From the above reasons it is common to assume a
connectivity of M. Based on the definition of connectiv-
semi-Markov process (SMP) for {Z (t), t 0} to describe
ity, the network survivability of M, denoted by NSk (M),
the node behavior transitions, which is defined by
is defined as the probability that all active (survived) nodes
Z (t) = Xn , tn t tn+1 . (2) are k-connected [4], i.e.,

Letting Tn = tn+1 tn be the sojourn time between the n-th
and (n + 1)-st transitions, we define the associated
SMP kernel Q = (Qij (t)) by
Qij (t) = Pr(Xn+1 = j, Tn t|Xn = i) = pij Fij (t), (3)
where pij = limt Qij (t) is the transition probability
between state i and j (i, j = ch, cl, sh, sl, mh, ml, ef , df ) cor-
responding to S, and Fij (t) = Pr(Tn < t|Xn+1 = j, Xn = i)
is the transition time distribution from state i to j. Figure 1
illustrates the transition diagram of the homogeneous SMP,
{Z (t), t 0}, under consideration, which is somewhat dif-
ferent from the MRGP in [9]. By using the Laplace-Stieltjes
transform (LST) we can obtain analytically the steady-state
probability of each node (see Appendix).
III. QUANTITATIVE NETWORK SURVIVABILITY
A. NETWORK SURVIVABILITY MEASURES
In a MANET, the transmission of a packet from one node
to another node must go through any path which is made by
its neighbor nodes. Since the topology of a MANET keeps
changing dynamically due to many reasons, such as node
mobility, even when node failures or DoS attacks do not
NSk (M) = Pr((Ma ) = k), (4)
where Ma is a sub-network of M and includes all active
nodes of M. In the above definition, we need to find all the
possible paths between arbitrary node pairs in a MANET.
Unfortunately, it is very difficult to enumerate all the com-
munication paths between arbitrary two nodes especially in
a large-scaled MANET. For this state-explosion problem,
we employ an approximate method to derive the network
survivability. For a geometric graph G with N vertices, the
minimum number of neighbor nodes of one node in G by
(G) is defined as the minimum node degree and (G) is
defined as the vertex-connectivity of G. It turns out that
(G) (G), i.e., the network connectivity is no longer
greater than the minimum number of neighbors of any
node. When N is sufficiently large, the probability that G is
k-connected approximately equals to the probability that
every vertex has at least k neighbors. So, it is immediate to
see that
Pr((G) = k) Pr((G) k). (5)
However, it should be noted that every neighbor does not
always provide effective outgoing paths, because only the

VOLUME 3, 2015 2667

 Z. Yi, T. Dohi: Toward Highly Dependable Power-Aware MANET

cooperative neighbor can transmit a packet for other node. derive approximately the low and upper bounds of network
Hence, a necessary condition for a MANET to be k-connected survivability instead when the number of nodes is sufficiently
is that every node has at least k cooperative degree. Let (M) large by considering the network connectivity of a node in a
denote the minimum of the cooperative degree of all nodes in MANET. The upper and lower bounds of network survivabil-
a MANET M. Then, we have ity are given by
Pr((M) = k) Pr( (M) k). (6) NSk (M)U = (Pr(D(c,u) k))ND , (10)
Remind that the network survivability is defined as the prob- NSk (M)L = max(0, 1 E[Na ](Pr(D(c,u) < k))), (11)
ability that all active nodes are k-connected to M, so that the
quantitative survivability of M can be given by respectively, where u is an arbitrary node index in the active
network Ma . In Eq. (11), E[Na ] = bN (1 Pf )c is the
NSk (M) Pr( (Ma ) k). (7) expected number of active nodes in the network, where bxc
An immediate effect of node misbehaviors and failures in is the maximum integer less than x, Pf is the steady-state
MANETs is the node isolation problem [4]. It is a direct cause probability of a Failed node, and N denotes the total number
for network partitioning, and eventually affects the network of mobile nodes. In Eq. (10), ND is the number of node
survivability. The node isolation problem is caused by four points whose transmission ranges are mutually disjoint over
types of neighbor; Failed, Selfish, Jellyfish and Blackhole the MANET area. Let A and r be the area of MANET and
nodes. If all the neighbors of a node are Failed nodes, Selfish the node transmission radius, respectively. The number of
nodes or Jellyfish nodes, then it can no longer communicate disjoint points is given by ND = bN /(r 2 )c, where
with other nodes. On the other hand, if one of neighbors is = N /A is the node density.
Blackhole, it gives the other node a faked one-hop path, and Next, we give an approximate form of the network sur-
can always shutdown the communication. In this case, it is vivability based on the expected number of active nodes [8].
said that the node is isolated by the Blackhole neighbor. Getting help from the graph theory, the expected network
Furthermore, if there exists a Blackhole node, then the mini- survivability is approximately given by the probability that
mum cooperative degree (Ma ) of network Ma becomes 0, the active node in the network is k-connected:
E[Na ]
NSk (M)E = 1 Pr(D(c,u) < k) .

and the network survivability is always reduced to 0. (12)
To formulate the above isolation problem, we define the
node degree D(u) for node u by the maximum number of By the well-known total probability law, we have
neighbors [5]. Let D(i,u) be the number of node us neighbors N
X
at state i {c, s, j, b, f } corresponding to {C, S, J , B, F}. Pr(D(c,u) < k) = Pr(D(c,u) < k|D(u) = d)
Then the isolation problem in our model can be formulated d=k
as follows: Given node u with degree d, i.e., D(u) = d, Pr(D(u) = d), (13)
if D(s,u) + D(f ,u) + D(j,u) = d or D(b,u) 1, the cooperative
degree is zero, i.e., D(c,u) = 0, and u is isolated from the so that we need to find the explicit forms of Pr(D(c,u) <
network, so it holds that k|D(u) = d) and Pr(D(u) = d). From Eqs. (12) and (13),
it is easy to obtain
Pr(D(c,u) = 0|D(u) = d) = 1 (1 Pb )d + (1 Pc Pb )d ,
Pr(D(c,u) < k|D(u) = d)
(8)
k1  
where Pc is the steady-state probability of a node in a d
X d
= 1 (1 Pb ) + Pcm (1 Pc Pb )dm
Cooperative state and Pb is the steady-state probability of a m
m=0
node launching Blackhole attacks. In Appendix, we give the
k1
steady-state probability in our SMP model. X
= 1 (1 Pb )d + Bm (d, Pc , 1 Pc Pb ), (14)
Hereafter, a node is said to be k-connected to a network if
m=0
its associated cooperative degree is given by k (1). Given
node u with degree d, i.e., D(u) = d, u is said to be where Bm denotes the multinomial probability mass
k-connected to the network if the cooperative degree is k, function.
i.e. D(c,u) = k, which holds only if u has no Blackhole neigh- Since the node distribution Pr(D(u) = d) strongly depends
bor and has exactly k Cooperative neighbors, i.e., D(b,u) = 0 on the model property, we introduce three specific stochastic
and D(c,u) = k, respectively. Then it is straightforward to see models [8] in the following:
that
  1) POISSON MODEL [4]
d
Pr(D(c,u) = k|D(u) = d) = (Pc )k (1 Pc Pb )dk . Suppose that N mobile nodes in a MANET are uniformly
k
distributed over a 2-dimensional square with area A. The
(9)
node transmission radius, denoted by r, is assumed to be
Strictly speaking, it is still difficult to find the probability identical for all nodes. To derive the node degree distribution
distribution of (Ma ) k in Eq.(7). Xing and Wang [4] Pr(D(u) = d), we divide the area into N small grids virtually,

2668 VOLUME 3, 2015

Z. Yi, T. Dohi: Toward Highly Dependable Power-Aware MANET

(E[N ] "
so that the grid size has the same order as the physical Xa
size of a node. Consider the case where the network area is NSk (M)BE = Bd (E[Na ], p) (1 Pb )d
k=0
much larger than the physical node size. Then, the probability #)E[Na ]
k1
that a node occupies a specific grid, denoted by p, is very X
Bm (d, Pc , 1 Pc Pb ) .
small. With large N and small p, the node distribution can be
m=0
modeled by the Poisson distribution: (22)
d If each node is assigned into a communication network area
Pr(D(u) = d) = e , (15)
d! of a node with probability p = r 2 /A, then the corresponding
where = r 2 , and = E[Na ]/A is the node den- binomial model results a different survivability measure.
sity depending on the underlying model. Finally, substituting
3) NEGATIVE BINOMIAL MODEL [8]
Eqs. (13) - (15) into Eqs. (10) - (12) yields
The negative binomial model comes from a mixed Pois-
0(k, Pc ) ND son distribution instead of Poisson distribution. Let f () be
  
NSk (M)PU = ePb 1 , (16) the distribution of parameter in the Poisson model. This
0(k)
implicitly assumes that the parameter includes uncertainty,
0(k, Pc )
   
P
NSk (M)L = 1 E[Na ] 1 e Pb
1 , and that the node distributions for all disjoint areas have
0(k) different Poisson parameters. Then the node distribution can
(17) be represented by the following mixed Poisson distribution:
d
Z
0(k, Pc ) E[Na ]

 
NSk (M)PE = ePb 1 , (18) P(D(u) = d) = e f ()d. (23)
0(k) 0 d!
For the sake of analytical simplicity, let f () be the gamma
where 0(x) = (x 1)! and 0(h, x) = (h 1)!ex h1
l=0 x /l!
l
P
probability density function with mean r 2 N (1 Pf )/A and
are the complete and incomplete gamma functions, coefficient of variation c. Then we have
respectively. 0(a + d)

b a
 
1 d

P(D(u) = d) = = d (a, b),
d!0(a) 1 + b 1+b
2) BINOMIAL MODEL [8]
(24)
It is evident that the Poisson model just focuses on an ideal
situation of mobile nodes. In other words, it is not always where a = b1/c2 c and b = bA/( r 2 N (1 Pf )c2 )c. It should
easy to measure the physical parameters such as r and A be noted that Eq. (24) corresponds to the negative bino-
in practice. Let p denote the probability that each node is mial probability mass function with mean r 2 N (1 Pf )/A,
assigned into a communicate network area of a node. For the and that the variance is greater than that in the Poisson
expected number of activate nodes E[Na ], we describe the model. From Eq. (24), we can obtain alternative representa-
node distribution by the binomial distribution: tions of the network survivability with an additional model
  parameter c.
E[Na ] d
Pr(D(u) = d) = p (1 p)E[Na ]d (E[N ]
Xa
"
d NB
NSk (M)U = d (a, b) (1 Pb )d
= Bd (E[Na ], p), (19)
k=0
k1
#)E[ND ]
where Bd is the binomial probability mass function. X
Substituting Eq. (19) into Eqs. (10) - (12) yields alternative Bm (d, Pc , 1 Pc Pb ) ,
formulas of the network survivability: m=0
(E[N ] " (25)
Xa NSk (M)NB
L = 1 E[Na ]
NSk (M)BU = Bd (E[Na ], p) (1 Pb )d (E[N ] "
Xa
k=0 1 d (a, b) (1 Pb )d
k1
#)E[ND ]
X k=0
Bm (d, Pc , 1 Pc Pb ) , k1
#)!
X
m=0 Bm (d, Pc , 1 Pc Pb ) ,
(20) m=0
(26)
NSk (M)BL = 1 E[Na ] (E[N ] "
Xa
NSk (M)NB d (a, b) (1 Pb )d
(E[N ] "
Xa E =
1 Bd (E[Na ], p) (1 Pb )d k=0
k=0 k1
#)E[Na ]
X
k1 Bm (d, Pc , 1 Pc Pb ) .
#)!
X
Bm (d, Pc , 1 Pc Pb ) , (21) m=0
m=0 (27)

VOLUME 3, 2015 2669

 Z. Yi, T. Dohi: Toward Highly Dependable Power-Aware MANET

IV. BORDER EFFECTS OF NETWORK

COMMUNICATE AREA
The results on network survivability presented in Section III
are based on a non-informative assumption that network
area A has a node density = E[Na ]/A. This means that the
expected number of neighbors of a node in a MANET has the
same value as r 2 . In other words, such an assumption is not
realistic in real world network communication circumstance.
It is well recognized that the border effects tend to decrease
both the communication coverage and the node degree of a
node, which reflect the whole network availability.

FIGURE 3. Border effects in circular area.

Next we consider a circular area for communication

in Fig. 3. For the circular area A with radius r0 , define the
origin O in A and use the coordinates r for a node in this
figure. Nodes in the circular communication network area,
that are located at least r away from the border, are called
the center nodes, which are shown as a node v in Fig. 3.
They have a coverage area equal to r 2 and an expected
node degree E[Na ] r 2 /A. On the other hand, nodes located
closer than r to the border are called the border nodes (node u
in Fig. 3) which have a smaller coverage area, leading to a
FIGURE 2. Border effects in square area.
smaller expected node degree.
Bettstetter [6] calculates the average node degree for nodes
in borders for a circular communication area. The expected
First we consider a square area for communication
node degree of a node c in a circular communicate area can
in Fig. 2. Given the square area with side L in this figure, the
be obtained by
borders correspond to regions B and C. We call the rectangu- s
lar region B the lateral border, the square region C the corner Na

r r 2

border, and the square region I the inner region, respectively. c = 4(1 r ) arcsin + 2r (2r + r ) 1
2 2 3
,
2 2 4
In Fig. 2, for a node v located in the inner region of the
(29)
network area, the expected effective number of neighbors,
EI , is indeed given by r 2 . These results are seen from the where r = r/r0 . The above formula can be further simplified
fact that the effective coverage area of a point in the inner by using Taylor series as
region I is precisely equal to r 2 . However, the effective 
4r

coverage area for a point in the border region B (node u1 ) or c Na r 1
2
. (30)
3
C (node u2 ) is less than r 2 as shown in the shadow areas
of Fig. 2. Consequently, the expected effective number of By replacing the square border effect parameter
neighbors of nodes located in the border areas must be smaller in Eqs. (16)-(18), (20)-(22), and (25)-(27) by s in Eq.(28),
than r 2 . Since the connectivity properties of the network we obtain the improved network survivability measures
depend on the expected effective number of neighbors of taking account of square border effects. Also, using c
nodes, it is needed to obtain the expected effective number in Eq.(30), we derive the network survivability measures in a
of neighbors of nodes in these regions (I, B and C), in order circular communicate area as well.
to understand the connectivity properties in the network.
Laranjeira and Rodrigues [10] show that the relative aver- V. NUMERICAL EXAMPLES
age node degree for nodes in borders is independent of the A. COMPARISON OF NETWORK SURVIVABILITY
node transmission range and of the overall network node In our numerical experiments, we set model parameters as
density in a square communication area. Then, the expected follows:
number of neighbors of a node in a MANET is given by [10]:
CH ,CL (t) = MH ,ML = Gamma(t, 5, 1/600),
r 2 CH ,MH (t) = CL,ML (t) = SH ,MH (t) = Exp(t, 1/6e + 7),
s = , (28)
L2 CH ,SH (t) = Exp(t, 1/720.0),
where = (L 2r)2 + 3.07492r(L 2r) + 2.461344r 2 . CL,EF (t) = Gamma(t, 2, 1/900),

2670 VOLUME 3, 2015

Z. Yi, T. Dohi: Toward Highly Dependable Power-Aware MANET

TABLE 1. Comparison of lower and upper bounds with approximate network survivability.

CL,SL (t) = SH ,CH (t) = Exp(t, 1/180),
MH ,DF (t) = ML,DF (t) = Exp(t, 1/480),
SL,CL (t) = Exp(t, 1/360),
SL,ML (t) = Exp(t, 1/6e + 7),
DF,CH (t) = Uniform(t, 30, 120),
EF,CH (t) = Uniform(t, 30, 90),
pB = 0.1, pJ = 0.9,
where pB and pJ are the Blackhole attack ratio and the
Jellyfish attack ratio of DoS attack. Exp, Gamma and
Uniform are exponential, gamma and uniform p.d.f.s:
Exp(t, x) = xext , t 0, (31)
a a1
b t e bt
Gamma(t, a, b) = , t 0, (32)
0(a)
1 ity in three cases in the approximate model for a given
Uniform(t, min, max) = , min t max .
max min
(33)
To analyze the effect of battery re-charge, we consider three
cases of transition time from low battery states (CL, SL, ML)
to fully charged battery states (CH , SH , MH ):
Case (1) : (iL),(iH ) (t) = Gamma(t, 2, 1/2400),
Case (2) : (iL),(iH ) (t) = Exp(t, 1/4800),
Case (3) : (iL),(iH ) (t) = 0, i {C, S, M },
where (iL),(iH ) (t) = 0 in Case (3) denotes that there is no
battery re-charge in the MANET.
Suppose the following network parameter:
A = 1000 (m) 1000 (m).
To compare several stochastic models with different com-
bination; three node degree models (Poisson, binomial and
negative binomial), the lower and upper bounds versus
an approximate network survivability, existence of border
effects, we consider Case 1, and change the transition radius
from r = 80 to r = 130 and connectivity requirement from
k = 1 to k = 3. The comparative results are shown in Table 1.
From this table, we can see that the difference between three
node degree models of network survivability is very small for
the specific values of r and k. For example, when r = 120 and
k = 1, the difference among three models are less than 0.0003
for the lower and upper bounds and the approximate network
survivability. Then, we attempt to understand the differences
among three node degree models with three battery charge
cases. The results are shown in Table 2. From the table, we can
see that these are the similar results to Table 1. The difference
among three models is small and the battery charge case in
Exp is higher than others when r is small.
Table 3 presents effects of communication range of a
node r on k-connected (k = 1, 2, 3) network survivabil-
N = 500, where No Charge indicates (iL),(iH ) (t) = 0.
We find that the network survivability increases as the com-
munication range of a node r increases, and that the MANET
with battery re-charge is more survivable. More specifically,
when r is small (e.g. r = 80), the network survivability for
Exp is higher than 68%, and the case with No Charge
is less than 16%. On the other hand, even when the mean
transition time for Gamma is equal to that for Exp, there exists
large difference on the network survivability for small r.
However, when r is sufficiently large, the difference among
Exp, Gamma and No Charge is very small. Moreover, as
connectivity requirement k increases, the survivability takes
a lower level when r is small. This result means that the net-
work survivability is more sensitive to battery re-charge with
small r. In Table 4, we investigate the sensitivity of the total
number of nodes N on the network survivability measures,
where the transmission range r is fixed as 100. Note once
again that k = 1 corresponds to the network reliability. From
this result, it can be seen that when the number of nodes is

VOLUME 3, 2015 2671

 Z. Yi, T. Dohi: Toward Highly Dependable Power-Aware MANET

TABLE 2. Comparison of three battery charge cases with approximate network survivability.

TABLE 3. Steady-state network survivability for varying node transmission radius r .

TABLE 4. Steady-state network survivability for varying number of node N.

greater than 500, the network reliability is higher than 90%.
However, once the reliability attains the maximum value with
N = 700, it decreases gradually as the number of nodes
increases. Because of increasing number of nodes, it turns
out that the network connectivity increases. However, from
Table 4 with k = 2, 3, we come to know that the network
survivability does not show the monotone tendency on N ,
similar to the network reliability. This is because the number
of Blackhole nodes increases as the total number of nodes in
the whole network increases.
In Table 5, we focus on the network reliability (k = 1) and
compare the upper and lower bounds of network survivability
in Eqs.(10) and (11) with our approximate formula in Eq.(12),
where the number of nodes is N = 500 and the transmission
range changes from r = 80 to r = 130. In this table, the
values in Square and Circular are calculated based on
Eqs. (28) and (30), respectively. From the result, we can see
that the difference between lower and upper bounds of net-
work reliability is rather remarkable for some specific values
on r. For example, when r = 80, the difference between

2672 VOLUME 3, 2015

Z. Yi, T. Dohi: Toward Highly Dependable Power-Aware MANET

TABLE 5. Steady-state network reliability for node transmission radius r TABLE 6. Steady-state network reliability for varying number of node N
with/without border effects in case(1). with/without border effects in case(1).

the lower and upper bounds with/without border effects are degree models with lower and upper bounds and an approxi-
0.8648 (Ignorance), 0.8016 (Square) and 0.8135 (Circular). mate form. From Table 7, the transient network survivability
On the other hand, the approximate network reliability always has almost the same initial values, and the difference between
takes a value between lower and upper bounds. This result them will be remarkable as time elapses.
tells us that the approximate network reliability in Eq.(12)
is more useful than the bounds for quantification of net-
work reliability. Table 6 presents the dependence of the
number of nodes N on the steady-state network reliability
among three situations with/without border effects. From
these results, it is shown that the network reliability without
border effects (Ignorance) is higher than those with border
effects (Square and Circular).

B. TRANSIENT ANALYSIS OF NETWORK SURVIVABILITY

Next we calculate the transient network survivability with
the limiting probabilities Pch, j (t) ( j {ch, cl, sh, sl, mh, ml,
ef , df }) corresponding to S, by taking the Laplace inver-
sion of Eqs.(56)-(63) in Appendix. We apply the well- FIGURE 4. Transient probability of Cooperative state in three cases.
known Abates algorith [12] for the numerical inversion of
Laplace transforms. Reminding these properties on transi- Figure 4 illustrates the transient probability of cooperate
tion probabilities, we set N = 500 and r = 100, and consider state at time t. We can see that three cases have the sim-
the transient network survivability at time t of three node ilar values in the first 500 seconds and become different

VOLUME 3, 2015 2673


TABLE 7. Transient network survivability with three stochastic models.
FIGURE 5. Transient network survivability of Gamma case.
FIGURE 6. Transient network survivability of EXP case.
after that. Because three node degree models show the similar
tendency, we focus on only the Poisson model to investigate
the impact on transient network survivability here. We set
the total number of nodes N = 500 and transmission radius
r = 100. Then, we plot the transient network survivability
of three battery charge cases; Gamma, Exp and No charge,
with lower/upper bounds and approximate solution based on
the behavior of the transient probabilities at arbitrary time t,
in Fig. 5, Fig. 6 and Fig. 7, respectively. From these figures,
it can be seen clearly that the lower/upper bounds and
2674
Z. Yi, T. Dohi: Toward Highly Dependable Power-Aware MANET
FIGURE 7. Transient network survivability of No Charge case.
FIGURE 8. Comparison of approximate transient netwrok survivability.
approximate solution of network survivability have almost
the same initial values, and the differences among them also
become remarkable as time elapses. All three battery charge
cases have a higher transient network survivability when
connectivity requirement k is lower. When the k becomes
higher (k = 3), the transient network survivability gets closer
to 0.0624 (Gamma)/0.6096 (Exp)/0.026 (No Charge) with
time t elapsing. Finally we compare the approximate solution
of three battery charge cases in terms of the transient network
survivability. Figure 8 depicts the transient network surviv-
ability by varying the connectivity requirement k. It is shown
VOLUME 3, 2015
Z. Yi, T. Dohi: Toward Highly Dependable Power-Aware MANET

Z
that if there is no battery charge, the transient network surviv- qmh,ml (s) = exp{st}F mh,df (t)dFmh,ml (t) (41)
ability drops down as the operation time goes on. However, Z0
the transient solution with battery charge (Exp) still keeps qmh,df (s) = exp{st}F mh,ml (t)dFmh,df (t) (42)
higher levels in the same situation. This fact implies that Z0
the battery charge of node leads to an better performance of qml,mh (s) = exp{st}F ml,df (t)F ml,ef (t)dFml,mh (t)
MANETs. 0
(43)
Z
VI. CONCLUSION qml,df (s) = exp{st}F ml,mh (t)F ml,ef (t)dFml,df (t)
In this paper, we have revisited the network survivabil- 0
(44)
ity models in MANETs by taking account of the battery Z
re-charge and border effects in both square and circular com- qml,ef (s) = exp{st}F ml,mh (t)F ml,df (t)dFml,ef (t)
0
munication areas. Getting idea from the network connectivity, (45)
we have presented the approximate network survivability for-
Z
mulae by calculating the probability that all expected number qsh,ch (s) = exp{st}F sh,mh (t)dFsh,ch (t) (46)
Z0
of active nodes in the MANET is k. In numerical experiments,
we have considered two cases where the transition time from qsh,mh (s) = exp{st}F sh,ch (t)dFsh,mh (t) (47)
Z0
lower battery states to fully charged battery states are given
by the gamma and exponential distributions. We have also qsl,sh (s) = exp{st}F sl,cf (t)F sl,ml (t)dFsl,sh (t) (48)
Z0
compared the steady-state network survivability with/without
battery re-charge. It has been shown numerically that the qsl,cl (s) = exp{st}F sl,sh (t)F sl,ml (t)dFsl,cl (t) (49)
network survivability with battery re-charge was higher than Z0
that with no battery charge, when r was small, and that the qsl,ml (s) = exp{st}F sl,sh (t)F sl,cl (t)dFsl,ml (t) (50)
approximate network reliability always took a middle value Z0
between the lower and upper bounds. In future, we will qdf ,ch (s) = exp{st}dFdf ,ch (t) (51)
develop a comprehensive simulation model and investigate Z0
whether the approximate method for network survivability qef ,ch (s) = exp{st}dFef ,ch (t), (52)
0
itself can work well in several random network environments.
Then it will be needed to develop an efficient algorithm to where in general () = 1(). We also define the recurrent
count all the paths between an arbitrary pair of nodes. time distribution from state CH to state CH and its LST by
Hch,ch (t) and hch,ch (s), respectively. Then, from the one-step
APPENDIX transition probabilities from Eqs.(34)-(52), we have
Z
In this Appendix, we derive the steady-state probability of our
SMP model. Let 1/ij denote the mean transition time from hch,ch (s) = exp{st}dHch,ch (t)
0
state i to state j.
R Define the Laplace-Stieltjes transform (LST) = qch,cl (s)
by qij (s) = 0 exp{st}dQij (t). From the familiar SMP [qcl,ch (s) + qcl,ml (s)(qml,mh (s)qmh,df (s)
analysis technique, it is immediate to see that
qdf ,ch (s) + qml,df (s)qdf ,ch (s) + qml,ef (s)
Z
qef ,ch (s))/k(s)) + qcl,sl (s)
qch,cl (s) = exp{st}F ch,mh (t)F ch,sg (t)dFch,cl (t) (34)
0 [qsl,sh (s)(qsh,ch (s) + qsh,mh (s)(qmh,df (s)
Z
qdf ,ch (s) + qmh,ml (s) (qml,df (s)qdf ,ch (s)
qch,mh (s) = exp{st}F ch,cl (t)F ch,sh (t)dFch,mh (t) (35)
Z0 + qml,ef (s)qef ,ch (s)))/k(s)) + qsl,ml (s)
qch.sh (s) = exp{st}F ch,cl (t)F ch,mh (t)dFch,sh (t) (36) (qml,mh (s)qmh,df (s)qdf ,ch (s) + qml,df (s)
Z0 qdf ,ch (s) + qml,ef (s)qef ,ch (s))/k(s)]
qcl,ch (s) = exp{st}F cl,ml (t)F cl.sl (t)F cl,ef (t) + qcl,ef (s)qef ,ch (s)]/l(s) + qch,mh (s)
0 [qmh,ml (s) (qml,df (s)qdf ,ch (s) + qml,ef (s)
dF (t) (37)
Z cl,ch qef ,ch (s)) + qmh,df (s)qdf ,ch (s)]/k(s)
qcl,ml (s) = exp{st}F cl,ch (t)F cl,sl (t)F cl,ef (t) + qch,sh (s)[qsh,ch (s) + qsh,mh (s)(qmh,ml (s)
0
dF (t) (38) (qml,df (s)qdf ,ch (s) + qml,ef (s)
Z cl,ml qef ,ch (s)) + qmh,df (s)qdf ,ch (s))
qcl,sl (s) = exp{st}F cl,ch (t)F cl.ml (t)F cl,ef (t) /k(s)], (53)
0
dFcl,sl (t) (39) where
Z
qcl,ef (s) = exp{st}F cl,ch (t)F cl,ml (t)F cl,sl (t) l(s) = 1 qcl,sl (s)qsl,cl (s) (54)
0
dFcl,ef (t) (40) k(s) = 1 qmh,ml (s)qml,mh (s). (55)

VOLUME 3, 2015 2675

 Z. Yi, T. Dohi: Toward Highly Dependable Power-Aware MANET

Let Pch,i (t) denote the transition probability from the initial i {ch, cl, sh, sl, mh, ml, ef , df } corresponding to S. Based
state CH to respective states i {ch, cl, sh, sl, mh, ml, on the LSTs, pch,i (s), we can obtain Pi = limt Pch,i (t) =
ef , df } corresponding to S. Then, the LSTs of the transition lims0 pch,i (s) from Eqs.(56)-(63).
R
probability, pch,i = 0 exp{st}dPch,i (t), are given by
n o REFERENCES
pch,ch (s) = qch,mh (s) qch,sh (s) qch,cl (s) /hch,ch (s) [1] R. J. Ellison, D. A. Fisher, R. C. Linger, H. F. Lipson, T. Longstaff,
and N. R. Mead, Survivable network systems: An emerging disci-
(56) pline, Softw. Eng. Inst., Carnegie Mellon Univ., Pittsburgh, PA, USA,
Tech. Rep. CMU/SEI-97-TR-013, 1997.
n
pch,cl (s) = qch,cl (s) qcl,ch (s) qcl,ml (s) qcl,sl (s) [2] J. C. Knight and K. J. Sullivan, On the definition of survivability,
o n o Dept. Comput. Sci., Univ. Virginia, Charlottesville, VA, USA,
qcl,ef (s) / hch,ch (s)k(s) (57) Tech. Rep. CS-TR-33-00, 2000.
n o [3] D. Chen, S. Garg, and K. S. Trivedi, Network survivability performance
pch,sh (s) = qch,sh (s) + qch,cl (s)qcl,sl (s)qsl,sh (s)/l(s) evaluation: A quantitative approach with applications in wireless ad-hoc
n o networks, in Proc. 5th ACM Int. Conf. Model. Anal., Simul. Wireless,
qsh,mh (s) qsh,ch (s) /hch,ch (s) (58) Mobile Syst. (MSWiM), 2002, pp. 6168.
n o [4] F. Xing and W. Wang, On the survivability of wireless ad hoc networks
pch,sl (s) = qch,cl (s)qcl,sl (s)/l(s) with node misbehaviors and failures, IEEE Trans. Dependable Secure
n o Comput., vol. 7, no. 3, pp. 284299, Jul./Sep. 2010.
qsl,cl (s) qsl,sh (s) qsl,ml (s) /hch.ch (s) [5] C. Bettstetter, On the minimum node degree and connectivity of a wire-
less multihop network, in Proc. 3rd ACM Int. Symp. Mobile Ad Hoc Netw.
n (59) Comput. (MobiHoc), 2002, pp. 8091.
[6] C. Bettstetter, On the connectivity of ad hoc networks, Comput. J.,
pch,ml (s) = qch,mh (s)qmh,ml (s) + qch,sh (s)qsh,mh (s) vol. 47, no. 4, pp. 432447, 2004.

qmh,ml (s) + qch,cl (s) qcl,ml (s) + qcl,sl (s) [7] C. Bettstetter, J. Klinglmayr, and S. Lettner, On the degree distribution of
 k-connected random networks, in Proc. IEEE Int. Conf. Commun. (ICC),
[qsl,ml (s) + qsl,sh (s)qsh,mh (s)qmh,ml (s)] May 2010, pp. 16.
on o [8] Z. Yi and T. Dohi, Survivability analysis for a wireless ad hoc network
/l(s) qml,mh (s) qml,df (s) qml,ef (s) based on semi-Markov model, IEICE Trans. Inf. Syst., vol. E95-D, no. 12,
n o pp. 28442851, 2012.
/ hch,ch (s)k(s) (60) [9] H. Okamura, Z. Yi, and T. Dohi, Network survivability modeling and
n analysis for power-aware MANETs by Markov regenerative processes,
Telecommun. Syst., vol. 60, no. 4, pp. 471484, 2015.
pch,mh (s) = qch,mh (s) + qch,sh (s)qsh,mh (s) + qch,cl (s) [10] L. A. Laranjeira and G. N. Rodrigues, Border effect analysis for reliability
 assurance and continuous connectivity of wireless sensor networks in the
qcl,ml (s)qml,mh (s) + qcl,sl (s)[qsl,ml (s) presence of sensor failures, IEEE Trans. Wireless Commun., vol. 13, no. 8,
o
qml,mh (s) + qsl,sh (s)qsh,mh (s)] /l(s) pp. 42324246, Aug. 2014.

n o n o [11] Z. Yi and T. Dohi, Survivability analysis for power-aware mobile ad hoc
network taking account of border effects, in Proc. 12th IEEE Int. Conf.
qmh,ml (s) qmh,df (s) / hch,ch (s)k(s) Adv. Trusted Comput. (ATC), Aug. 2015, pp. 476483.
[12] P. Valk and J. Abate. (Sep. 3, 2003). Numerical Inversion of Laplace
n (61) Transform With Multiple Precision Using the Complex Domain. [Online].

pch,df (s) = qch,cl (s) qcl,ml (s)(qml,mh (s)qmh,df (s) Available: http://library.wolfram.com/infocenter/MathSource/5026/

+ qml,df (s)) + qcl,sl (s)[qsl,sh (s)qsh,mh (s)

[qmh,df (s) + qmh,ml (s)qml,df (s)] + qsl,ml (s)
(qml,mh (s)qmh,df (s) + qml,df (s))] /l(s)

 neering) degree from Hiroshima University, Japan,
+ qch,sh (s)qsh,mh (s) qmh,ml (s)qml,df (s)
 
+ qmh,df (s) + qch,mh (s) qmh,ml (s)qml,df (s)
o n o
+ qmh,df (s) qdf ,ch (s)/ hch,ch (s)k(s) (62)
n
pch,ef (s) = qch,mh (s)qmh,ml (s)qml,ef (s)/k(s) + qch,sh (s)
qsh,mh (s)qmh,ml (s)qml,ef (s)/k(s) + qch,cl (s)
 Ph.D. (Engineering) degrees from Hiroshima
qcl,ml (s)qml,ef (s)/k(s) + qcl,sl (s)[qsl,ml (s)
qml,ef (s) + qsl,sh (s)qsh,mh (s)qmh,ml (s)
o ment of Industrial and Systems Engineering,
qml,ef (s)]/k(s) + qcl,ef (s) /l(s) qef ,ch (s)

/hch,ch (s). (63)
From Eqs.(56)-(63), the transient solutions, Pch,i (t),
i {ch, cl, sh, sl, mh, ml, ef , df }, which mean the proba-
bility that the state travels in another state i at time t, can
be derived numerically, by means of the Laplace inversion
technique (e.g. see [12]). As a special case, it is easy to
derive the steady-state probability Pi = limt Pch,i (t),
ZHIPENG YI received the B.S.E. degree from the
University of Electronic Science and Technology
of China, China, in 2007, and the M.Sc. (Engi-
in 2012, where he is currently pursuing the
Ph.D. degree with the Department of Information
Engineering, Graduate School of Engineering. His
research area is network survivability analysis.
TADASHI DOHI received the B.Sc., M.Sc., and
University, Japan, in 1989, 1991, and 1995,
respectively. In 1992, he joined the Depart-
Hiroshima University, as an Assistant Professor.
Since 2002, he has been a Full Professor with the
Department of Information Engineering, Gradu-
ate School of Engineering, Hiroshima University.
In 1992 and 2000, he was a Visiting Research
Scholar with the University of British Columbia, Canada, and Duke Uni-
versity, USA, respectively, on leave of absence from Hiroshima University.
His research areas include software reliability engineering, dependable com-
puting, and performance evaluation. He is a Regular Member of ORSJ, IPSJ,
IEICE, REAJ, and IEEE.

2676 VOLUME 3, 2015