IEEE BE-BTECH NS2 PROJECT@ DREAMWEB TECHNO SOLUTION

Joint Optimal Data Rate and Power Allocation
in Lossy Mobile Ad Hoc Networks
with Delay-Constrained Traffics
Songtao Guo, Member, IEEE, Changyin Dang, Senior Member, IEEE, and
Yuanyuan Yang, Fellow, IEEE
Abstract—In this paper, we consider lossy mobile ad hoc networks where the data rate of a given flow becomes lower and lower along
its routing path. One of the main challenges in lossy mobile ad hoc networks is how to achieve the conflicting goal of increased network
utility and reduced power consumption, while without following the instantaneous state of a fading channel. To address this problem,
we propose a cross-layer rate-effective network utility maximization (RENUM) framework by taking into account the lossy nature of
wireless links and the constraints of rate outage probability and average delay. In the proposed framework, the utility is associated with
the effective rate received at the destination node of each flow instead of the injection rate at the source of the flow. We then present a
distributed joint transmission rate, link power and average delay control algorithm, in which explicit broadcast message passing is
required for power allocation algorithm. Motivated by the desire of power control devoid of message passing, we give a near-optimal
power-allocation scheme that makes use of autonomous SINR measurements at each link and enjoys a fast convergence rate. The
proposed algorithm is shown through numerical simulations to outperform other network utility maximization algorithms without rate
outage probability/average delay constraints, leading to a higher effective rate, lower power consumption and delay. Furthermore, we
conduct extensive network-wide simulations in NS-2 simulator to evaluate the performance of the algorithm in terms of throughput,
delay, packet delivery ratio and fairness.
Index Terms—Mobile computing, mobile ad hoc networks, cross-layer optimization, network utility maximization, congestion control, power
control, outage probability
Ç
1 INTRODUCTION
MOBILE nodes in ad hoc networks are usually powered
by battery with energy constrained [1], [2], [3]. Hence,
wise allocation of power is critical in wireless mobile ad hoc
networks (MANETs) for both prolonging battery life of the
mobile devices and increasing utilization of the scarce wire-
less spectrum. The aim of power control is to intelligently
adjust the transmission power to the least that is required to
send the data packet to meet the required signal-to-interfer-
ence-plus-noise ratio (SINR) threshold and achieve the qual-
ity of service (QoS) goals in wireless channels.
Existing power-control schemes, either centralized, such
as [4], [5], [6], or distributed, such as [7], [8], [9], [10]), always
assume quasi-stationarity of the fading wireless channels
and are based on the observed SINR at the receiver or the
knowledge of the gains of all the links. Thus, the implicit
assumption made is that the power-control updates are
made every time the fading state of the channel changes.
However, in wireless communication channels, which
exhibit fast fading where the fades can change within milli-
seconds, this might not always be practical. Very frequent
power-control updates can also consume a lot of energy.
Therefore, it is desirable that the power-control does not need
to be updated when the channel meanders from a fading state
to another. This is one of our motivations of this work.
Another fundamental task performed frequently by ad
hoc networks is to regulate the allowed source rates to
maximize the total utility of the users subject to resource
constraints. One common approach to achieving this
objective is network utility maximization (NUM) through
the cross-layer design of joint power and rate control. The
cross-layer design takes users’ service requirements and
interference-limited channel into account to jointly opti-
mize network performance. In fact, the link capacities
determined by the power allocation at the physical layer
affect the arrival rates at the transport layer, and vice
versa. With this coupling, resource optimization within
layers is not competent.
Therefore, motivated by the benefits, a cross-lay design
of NUM has extensively been studied [11], [12], [13], [14],
[15], [16]. In [11], Chiang analyzed such a joint congestion
and power control problem by using Lagrangian primal-
dual approach. In [12], the authors considered joint routing
and rate allocation that maximizes the lifetime of an ad hoc
network supporting variable-rate transmissions (link adap-
tation). In [13], a joint opportunistic power scheduling and
end-to-end rate control (JOPRC) scheme was presented for
S. Guo is with the College of Electronic and Information Engineering,
Southwest University, Chongqing, 400715, P.R. China.
E-mail: stguocqu@gmail.com.
Y. Yang is with the Department of Electrical and Computer Engineering,
Stony Brook University, Stony Brook NY 11794, USA.
E-mail: yuanyuan.yang@stonybrook.edu.
C. Dang is with the Department of Manufacturing Engineering and Engi-
neering Management, City University of Hong Kong, 83 Tat Chee
Avenue, Kowloon, Hong Kong. E-mail: mecdang@cityu.edu.hk.
Manuscript received 9 Sept. 2012; revised 17 Aug. 2013; accepted 11 Dec.
2013. Date of publication 15 Jan. 2014; date of current version 11 Feb. 2015.
Recommended for acceptance by M.Y. Eltoweissy.
For information on obtaining reprints of this article, please send e-mail to:
reprints@ieee.org, and reference the Digital Object Identifier below.
Digital Object Identifier no. 10.1109/TC.2013.2296043
IEEE TRANSACTIONS ON COMPUTERS, VOL. 64, NO. 3, MARCH 2015 747
0018-9340 ß 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

wireless ad hoc networks by modeling the time-varying
wireless channel as a stochastic process. In [16], the authors
modeled an efficient spatial-TDMA based adaptive power
and rate cross-layer scheduling problem as a mixed integer
linear program (MILP). A cross-layer design of power- and
rate-allocation was studied in [17] for MANETs with lossy
links. All of these works similarly assume the channel to be
quasi-stationary and track the instantaneous channel state;
meanwhile, the data rate of a given flow is treated
unchanged from hop to hop along its route. In wireless net-
works, however, wireless links are typically lossy and the
packet loss rate is often high due to channel impairment
such as fading and interference. As a result, the data rate of
a flow decreases along its route because of the lossy nature
of wireless links. Clearly, these previous studies that follow
fast fades directly are not feasible or desirable in this situa-
tion due to mobility of the nodes and/or environment. This
fact also motivates our work in this paper.
More recently, there have been many studies on the
problem of power control with constraints on outage proba-
bility in wireless networks with lossy links, in which outage
probability is referred to as the probability that the smallest
maximum achievable rate among all of users is smaller than
a specified transmission rate. In [18], Kandukuri and Boyd
proposed a new method of optimal power control in inter-
ference-limited fading wireless channels with outage-proba-
bility specifications, where interior point methods are
employed to find the solution. In [19], the authors addressed
a joint power control and multiuser detection problem with
outage-probability constraints in a Rayleigh fast-fading
environment. Furthermore, in [20], they proposed a general-
ized framework for solving the problem under modest
assumption on the underlying channel fading. Unlike previ-
ous work, which dealt with Rayleigh fast-fading model,
each user is allowed to have a different fading distribution.
In [21], the outage probability of multiple antenna multicast
channels was investigated and its upper and lower-bounds
were derived by using extreme value theory. A link outage
probability based resource allocation scheme for multi-
radio multi-channel wireless mesh networks was proposed
in [22]. However, the objective of these works is to minimize
the total power of all users while meeting their outage tar-
gets or minimize the outage probability, whereas they do
not consider the maximization problem of network utility.
An emerging challenging issue to be addressed for
dynamic power and rate allocation in next generation
MANETs is how to meet with user’s heterogeneous trans-
mission QoS requirements. Among others, the demand for
wireless high-speed connectivity for both delay-tolerant
“packet” data and delay-sensitive “circuit” data is expected
to rise significantly in the next decade. Accordingly, how to
optimally allocate both power and rate to support delay-
constrained data traffic while sufficiently considering the
lossy nature of links becomes an important research issue.
Therefore, this paper aims to provide concrete answers to
the aforementioned problems.
In this paper, we first propose a novel rate-effective net-
work utility maximization (RENUM) framework, which
takes account of the lossy nature of links in the objective
function and the constraints of rate outage and average
delay. In particular, the minimization of the overall power
and delay of all links is used as the optimization objective
as well, which is the distinct difference between effective
network utility maximization (ENUM) with outage con-
straints in [23] and our work. In our framework, the trans-
mission rate at the source node of each flow is called the
injection rate, and the data rate correctly received at the des-
tination node is called the effective rate. Since the effective
rate is typically lower than the injection rate, it is natural to
examine the utility corresponding to the effective rate. Then
the original non-convex RENUM is converted into a convex
one by employing some logarithmic transformation. The
convex RENUM problem is further decomposed into three
separate maximization sub-problems of rate control, power
and delay allocation. Three corresponding distributed sub-
algorithms are developed, in which some forms of broad-
cast message passing is required for power-allocation. In
practice, it may be desirable to avoid such overhead and
thus we include a near-optimal scheme that makes use of
autonomous SINR measurements at each link for power-
allocation. To the best of our knowledge, this paper is the
first work to address the limits of dynamic power and rate
allocation over the lossy links under various average delay
and rate-outage probability constraints.
The contributions and findings of the paper can be briefly
summarized below.
The problem of the frequent rate and power updates
induced by fast-fading is addressed, which is imper-
ative especially in a MANET with limited energy, by
taking account of the average SINR and rate-outage
probability, where flow rate changes per hop.
A rate-effective network utility maximization frame-
work is proposed by considering the lossy nature of
wireless links and the rate outage constraints with a
small tolerable value, which is necessary because the
loss of wireless links and packets occurs frequently
due to channel impairment and mobility of nodes in
MANETs.
The minimization of the overall power and average
delay of all links is used as the optimization objective
while the effect of the loss and the rate outage on
optimal joint power, rate and delay control is studied
to support delay-constrained data traffic.
The distributed cross-layer power and rate control
algorithms with/without explicit broadcast mes-
sage passing are proposed by solving the non-con-
vex RENUM problem. Furthermore, our extensive
numerical and simulation results demonstrate that
the RENUM outperforms other NUM schemes
without rate outage probability/average delay
constraints in terms of effective rates, link delay
and power consumption.
The reminder of this paper is organized as follows. In
Section 2, we introduce the system model. Section 3 out-
lines the RENUM framework. In Section 4, we propose a
distributed joint rate, power and average delay control
algorithm. In Section 5, we further give the near-optimal
rate and power control algorithm making use of noise
measurements instead. Illustrative numerical and simula-
tion results are provided in Section 6. Finally, Section 7
concludes the paper.
748 IEEE TRANSACTIONS ON COMPUTERS, VOL. 64, NO. 3, MARCH 2015

2 SYSTEM MODEL
2.1 Network Model
In this paper, we consider a MANET that consists of a set
L ¼ f1; 2; . . . ; KLg of logical links, shared by a set
S ¼ f1; 2; . . . ; KSg of sources, where KL and KS denote the
number of links and sources, respectively. Each source
s 2 S emits one flow to its destination node using a set of
links LðsÞ L for its route. Due to the unreliability of the
network, routing may be adapted and optimized with the
improved routes communicated to the transport layer. In
practice, our formulated optimization model in Section 3
for the RENUM problem is suitable for the scenario of
varying routes. This is because our model is based on
hop-by-hop rate control and the proposed RENUM algo-
rithm in Section 4 is a distributed hop-by-hop algorithm
by considering the lossy nature of each link. Each source s
injects data into the network at an average rate of
xs 2 ms; Ms½ Š where ms and Ms denote the minimum and
maximum rates, respectively. For each link l, the set
SðlÞ ¼ s j l 2 LðsÞf g indicates the set of sources that use
link l. Each source s attains a utility function UsðxsÞ, which
is a function of the effective rate at the destination of
source s. A utility function can reflect the satisfaction of a
source with resource allocation. Different shapes of utility
functions can lead to different types of fairness among
users [24]. Take the following utility function, parameter-
ized by a 0, as an example
UsðxsÞ ¼
ð1 À aÞÀ1
ðxsÞ1Àa
; if a 6¼ 1
log xs; otherwise:

If a ¼ 0, system throughput maximization is achieved; if
a ¼ 1, proportional fairness among users is attained; and if
a ¼ 2, then harmonic mean fairness is achieved. In previ-
ous work, the utility function is usually assumed to be
strictly concave, strictly increasing, and twice differentia-
ble [17]. Concavity of UsðÁÞ means that the marginal utility
(i.e., the increase in utility obtained for a fixed increase in
the resource) decreases as the amount of resource
increases. It is reasonable to assume that UsðÁÞ is increas-
ing since a source always prefers to have more resource,
when the amount of all other resources stays the same. In
general, all the nodes in the network are assigned the
same utility function which aims at providing a uniform
measurement of network performance.
The wireless channel is a shared medium and interfer-
ence-limited. In the physical layer, code-division-multiple-
access (CDMA) is used so that each link may simulta-
neously communicate within the same spectrum allocation,
at the expense of multiple-access interference. We have an
underlying assumption that nodes can transmit and receive
simultaneously. It is indeed possible since simultaneous
transceivers have been implemented in [25]. The main nota-
tions and their corresponding definitions are summarized
in Table 1.
2.2 Fast-Fading and Average SINR
We let pl denote the power used by the transmitter of link l,
which is subject to a simple power limit pmin
l pl pmax
l ,
where pmin
l and pmax
l denote the minimum and maximum
transmission powers respectively, and gl be the signal to
interference and noise ratio of link l. The signal power at the
receiver of link l is given by GllFllpl and the total interfer-
ence power at the receiver of link l is given byP
k6¼l GlkFlkpk. Thus, the SINR for link l is
glðPPÞ ¼
GllFllpl
P
k6¼l GlkFlkpk þ s2
; (1)
where s2
is the thermal noise power at each receiver;
PP ¼ ðp1; p2; . . . ; pKL
ÞT
denotes a vector of transmission
powers; Glk is the effective positive slow-varying path gain
between the transmitter of link k and the receiver of link l.
In the analysis below, we assume that it is constant; Flk
models Rayleigh fast-fading, which is assumed to be inde-
pendent exponentially distributed random variables with
unit mean [26]. In other words, the received signal power at
receiver l from transmitter k is an exponentially distributed
random variable with mean EðGlkFlkpkÞ ¼ Glkpk.
Furthermore, in Rayleigh fast-fading environment, the
instantaneous capacity of each link l is defined on the
Shannon capacity as
clðPPÞ ¼ W logð1 þ ÁglðPPÞÞ; (2)
where is “SINR-gap” that reflects a particular modulation
and coding scheme. Without loss of generality, we assume
that ¼ 1; W is the base-band bandwidth, which is normal-
ized by a fixed packet size.
For fast-fading channel, very frequent power updates
will induce the fast change of the instantaneous capacity
clðPPÞ according to (2). The high fading rate means that itera-
tion update-rate would need to be increased to keep track of
the dynamic fading state, which makes the instantaneous
state vary too quickly to track properly. Indeed, we
TABLE 1
List of Notations
Notation Definition
L Set of logical links
S Set of sources
KL Number of links
KS Number of sources
xs Data rate of source s
ms Minimum rate at which source s transmits data
Ms Maximum rate at which source s transmits data
Us(·) Utility function of source s
γl SINR for link l
γl Average SINR for link l
pl Transmission power in link l
pmin
l Minimum transmission power
pmax
l Maximum transmission power
P Vector of transmission powers
cl Instantaneous capacity of link l
Pr(Rl) Outage probability at link l with transmission rate Rl
Glk Path gain with slow-varying
Flk Path gain with Rayleigh fast-fading
Ravg
l Average rate correctly received over link l
ys Average rate correctly received at destination node
ζmax
l Maximum value of outage probability on link l
φl(P) Probability that the data is correctly received
Rth
l Threshold of data rate Rl
τl Sum of transmission delay and queuing delay on link l
E(τl) Average packet delay on link l
ϑl Specific delay-tolerant amount
Hs Number of hops for flow s
GUO ET AL.: JOINT OPTIMAL DATA RATE AND POWER ALLOCATION IN LOSSY MOBILE AD HOC NETWORKS WITH... 749

explicitly take into account the average SINR and the notion
of a rate-outage probability, whose definition can be found
in Section 2.4 to allow the network to experience a limited
amount of fading-induced congestion so as to avoid the
fast-update problem.
Therefore, instead of using instantaneous SINR, we will
use statistical variation of SINR, i.e., average SINR, in our
scheme.
glðPPÞ ¼
EðGllFllplÞ
E
P
k6¼l GlkFlkpk

þ s2
¼
Gllpl
P
k6¼l Glkpk þ s2
: (3)
2.3 Capacity and Data Rate with Outage
It is clear from (2) that the capacity of link depends on chan-
nel state information (CSI) such as channel gain between
transmitter and receiver. In this paper, similar to [23], we
assume that there is receiver CSI and no transmitter CSI at
each link, i.e., the receiver knows the value of CSI, and both
the transmitter and the receiver know the distribution of
CSI, since transmitter CSI results in extra overhead and con-
siderable additional complexity. In this case there are two
capacity definitions: Shannon capacity and capacity with
outage. It was shown in [19] that the capacity of a fading
channel with receiver CSI (i.e., capacity with outage) is less
than the Shannon capacity of an additive white Gaussian
noise (AWGN) channel with the same average SINR. This
indicates that fading and poor channel states can reduce
Shannon capacity when only the receiver has CSI, because
the transmission strategy must incorporate the effect of
these poor states.
The capacity with outage for fading channel with receiver
CSI is defined as the maximum rate that can be transmitted
over a channel with an outage probability that the transmis-
sion cannot be decoded with negligible error probability.
The data is correctly received if cðPPÞ is greater than or equal
to the ingress date rate R. Otherwise, the data received over
the transmission burst cannot be decoded correctly with
probability approaching 1, and the receiver declares outage.
Here, the outage probability is thus given by PrðRlÞ ¼
PrðRl clðPPÞÞ at link l with transmission rate Rl.
The outage probability characterizes the probability of
data loss or, equivalently, the probability of deep fading. At
a link l with transmission rate Rl, the probability that data is
received correctly is 1 À PrðRlÞ. The rate-effective correctly
received over link l is thus
Rl;eff ¼ Rl Á 1 À PrðRlÞð Þ: (4)
For a flow s traversing multiple hops, the effective data
rate correctly received at the destination node ys satisfies
ys xs
Y
l2LðsÞ
1 À PrðRlÞð Þ; (5)
where xs is the source rate of flow s, and LðsÞ is the set of
link along the route. It can be seen that the data rate of a
flow decreases every hop along its route and the data rate
correctly received at the destination node decreases as the
link or hop count along the route of a flow increases.
2.4 Rate-Outage Probability
As the fading rate increases, more timely feedback of the
end-to-end congestion state, which must be feedback to the
source to adapt its transmission data rate explicitly through
message passing, becomes increasingly difficult. To allevi-
ate this problem, the rate-outage probability should be con-
strained to a small tolerable value instead of zero-level
tolerance on fading-induced congestion.
Similar to [17], the rate-outage probability is defined as
the probability of experiencing fading-induced congestion,
i.e., the probability that the ingress rate Rl to link l exceeds
the instantaneous channel capacity cl. Therefore, the rate
outage constraint can be expressed as
PrðRl clðPPÞÞ zmax
l ; 8l 2 L; (6)
where zmax
l 2 ð0; 1Þ is the maximum value of the rate outage
probability on each link l. Accordingly, the z-outage capac-
ity on link l is defined as the highest rate that the outage
probability PrðRlÞ is less than zmax
l . It is noted that there is a
one-to-one mapping between the outage probability and the
corresponding outage capacity. For instance, the z-outage
capacity on each link l over Rayleigh fading is given by [27]
clðPP; zÞ ¼ W log 1 þ KlglðPPÞ
Â Ã
; (7)
where Kl ¼ Àlogð1 À zmax
l Þ.
For the Rayleigh fading model, we can write the rate-out-
age probability in closed form by using (3) as [17]
Pr Rl clð Þ ¼ Pr W log 1 þ KlglðPPÞ
Â Ã
Rl
À Á
¼ 1 À flðPPÞ; (8)
where
flðPPÞ ¼ exp À
Rth
l
glðPPÞ

(9)
is regarded as the probability that the data is correctly
received, and Rth
l ¼ expðRl=WÞ À 1 is defined as the data
rate threshold.
2.5 Average Delay
We now consider the delay property to be included in
constraints and objective function. The average delay a
packet experiences when traversing a network is an
important design consideration in some real-time applica-
tions. The delay includes both queuing and transmission
delays. The packet traffic entering the multihop network
at the transmitter of link is assumed to be a Poisson distri-
bution and to have an exponentially distributed length
with a mean of K bits. In this study, we model each link
as an M/M/1 queue, which is a good approximation for
the behavior of individual links in densely connected net-
works satisfying the above assumptions [28]. Under the
feed-forward routing and Poisson input assumptions, by
applying Burke’s theorem [28], the average packet delay
on link l is thus given by
EðtlÞ ¼
K
clðPP; zÞ À
P
s2SðlÞ xs
;

where tl is the sum of the transmission delay and queuing
delay on link l. For real-time applications, generally the
average delay EðtlÞ is required to be no more than a specific
delay-tolerant amount #l ! 0, i.e., EðtlÞ #l. Equivalently,P
s2SðlÞ xs clðPP; zÞ À K
#l
.
3 RENUM WITH RATE OUTAGE CONSTRAINT
In this section, a novel rate-effective network utility maximi-
zation framework is proposed by taking into account the
lossy nature of links and the rate outage constraints with a
small tolerate value. At the ith link of flow s, the data rate is
denoted by xlðs;iÞ
s where lðs; iÞ denotes the unique link num-
ber of the ith link of flow s. For clearer presentation, notation
xi
s will substitute for xlðs;iÞ
s . At the destination of flow s, the
correctly received data rate can be represented as xHsþ1
s ,
where Hs ¼ jLðsÞj denotes the number of hops for flow s. It
can be considered as the data rate over a link which connects
the receiver and itself [23]. For example, in Fig. 1, the effective
rate of flow 1 is given by x5
1. Thus similar to [23], the utility
function of flow s is associated with the effective rate xHsþ1
s .
There are two main objectives to be achieved in this
problem formulation, i.e., to maximize the overall rate-
effective network utility across all flows and minimize
the overall power and average delay of all links, which is
necessary to reduce interference and prolong the life of
nodes powered by batteries. The first objective can be
expressed mathematically by
max
x
X
s2S
Us xHsþ1
s
À Á
and the second objective can be characterized by
min
#;p!0
w1
X
l2L
pl þ w2
X
l2L
#l:
Based on the fact that minfðÁÞ is equivalent to max À fðÁÞ, by
combining the above two objectives and considering the
constraints in Section 2, the RENUM problem can be formu-
lated as
max
x;#;p!0
X
s2S
Us xHsþ1
s
À Á
À w1
X
l2L
pl À w2
X
l2L
#l (10)
Subject to
xiþ1
s xi
s 1 À Prðlðs; iÞÞð Þ; i ¼ 1; 2; . . . ; Hs; 8s; (11)
Pr
X
s2SðlÞ
xl
s clðPPÞ
0
@
1
A zmax
l ; 8l; (12)
X
s2SðlÞ
xs clðPP;zÞ À
K
#l
; 8l; (13)
ms xs Ms; 8s; (14)
pmin
l pl pmax
l ; 8l; (15)
where Prðlðs; iÞÞ denotes the outage probability of the ith
link of flow s, and w1 and w2 are fixed weights to trade these
two conflicting objectives. The constraints can be explained
as follows:
Flow conservation constraint. Constraint (11) shows
that for each flow, the effective data rate received at
the receiver is no more than the injection rate at the
transmitter for route links by considering the lossy
link property.
Rate-outage probability constraint. Constraint (12)
describes that at each link, there exists an upper
bound (a small tolerable value) on the rate-outage
probability.
Average delay constraint. Constraint (13) ensures that
the average delay on link l is guaranteed not to
exceed a specific delay-tolerant amount #l.
Rate fairness constraint. Constraint (14) demonstrates
that the transmitting rate of each source is bounded
by the maximum and minimum transmission rates,
Ms and ms; in order to keep the fairness and improve
the efficiency of wireless links.
Energy constraint. Constraint (15) describes that the
transmission power at each transmitter on link l is
constrained by the maximum and minimum trans-
mission powers, pmax
l and pmin
l , in order to save the
finite energy of each transmitter and satisfy the
diverse quality of service requirement, respectively.
Note that the outage parameter zmax
l reflects the trade-
off between rate and outage. With the increase of zmax
l ,
higher data rate can be transmitted, but higher data loss
may occur.
Remark 1. Compared with other NUM problems, the
RENUM framework has two unique features: (i) the
network utility is based on the achieved rates at
the destinations instead of the injected rates at the sour-
ces; (ii) the effect of rate-outage probability and average
delay on the optimal rate and power allocation is con-
sidered accurately.
For wireless links experiencing Rayleigh fading, the
inequality constraint (11) becomes
xiþ1
s ¼ xi
sflðs;iÞðPPÞ; i ¼ 1; 2; . . . ; Hs; 8s: (16)
In general, the RENUM problem (10) is a non-convex
problem due to the utility function associated with the
effective rate xHsþ1
s and the existence of constraint (11).
Therefore, the problem needs to be converted into a con-
vex one under some regularity conditions by employing
some auxiliary variables and appropriate transformation.
Define ~xl
s ¼ logðxl
sÞ, ~pl ¼ logðplÞ, ~#l ¼ logð#lÞ, ~Ls ¼ logðmsÞ;½
x 1
1
1
12
3
1
5
1
2
2
2
2
3
2
4
Flow 1 : Effective rate
Destination 1
Destination 2
: Effective rate
Source 1
Source 2
Flow 2
x
x
x
x
x
x
xx 4
1
Fig. 1. An example network with lossy links.

logðMsÞŠ, ~Il ¼ logðpmin
l Þ; logðpmax
l Þ
Â Ã
, then the optimization
problem (10) can be transformed as follows:
max
~xs; ~#l;~pl
X
s2S
Us exp
À
~xHsþ1
s
ÁÀ Á
À w1
X
l2L
expð~plÞ À w2
X
l2L
expð ~#lÞ
(17)
Subject to
~xiþ1
s ~xi
s À clðs;iÞ
ð~PPÞ; i ¼ 1; . . . ; Hs; 8s; (18)
cl
ð~PPÞ Kl; 8l; (19)
X
s2SðlÞ
exp
À
~xl
s
Á
clðexpð~PPÞ; zÞ À
K
expð ~#lÞ
; 8l; (20)
where ~xs 2 ~L, ~pl 2 ~Il and clðs;iÞ
ð~PPÞ ¼ Àlogðflðs;iÞð~PPÞÞ:
Next, we will explore the convexity of the RENUM
problem (17)-(20) in preparation for the development of
an algorithm attaining a global solution. At first, we will
show the convexity of the constraint set in (18)-(20).
Theorem 1. The constraint set comprised by (18)-(20) is convex.
Proof. The source rates xs are transformed logarithmically
to ensure that the left-hand side of the constraints is
convex as sum-exp is convex. Since clðs;iÞ
ð~PPÞ is twice
differential with respect to ~PP and its Hessian matrix is
positive semidefinite, the right-hand side of constraint
(18) is concave and the left-hand side of constraint (19)
is also convex. It is easy to verify that the Hessian
matrix of the right-hand side of constraint (20) is nega-
tive semidefinite in ~#l and ~pl, which means that it is
concave. We therefore conclude that the constraint set
in (18)-(20) is convex. tu
The next issue to be addressed is to show the convexity of
the objective of the transformed problem (17).
Theorem 2. The objective of the transformed problem (17) is con-
cave over ~xHsþ1
s , ~pl and ~#l:
Proof. Since the objective is separable in each variable, it is
clearly concave in ~pl and ~#l (sum of negative exponen-
tials). The concavity in ~xHsþ1
s follows from the concavity
Usðexpð~xHsþ1
s ÞÞ of in ~xHsþ1
s . Since the utility function
UsðxsÞ is twice-differentiable, increasing and strictly-con-
cave in xs, it is also (log, xs)–concave. We further have
that the objective function in (17) is concave about ~xHsþ1
s .
Problem (17) is thus a convex problem with a unique
global optimal point. tu
It is not difficult to see that the Slater condition [29] is sat-
isfied, i.e., there exist feasible solutions of ~xHsþ1
s , ~pl and ~#l
such that the constraints hold with strict inequality, and
strong duality holds, which implies that the optimal values
of the primal and dual problems are equal. Thus, the opti-
mization problem (17) can be efficiently tackled by using
modern convex programming schemes.
4 DISTRIBUTED ALGORITHM FOR RENUM
PROBLEM
In this section, we use Lagrange-duality method for solving
RENUM problem to develop a distributed algorithm. The
first step for the Lagrange-duality method is to introduce
Lagrangian Multipliers associated with some constraints of
the original problem. For RENUM problem that has multi-
ple constraints, there are also various ways to introduce
Lagrangian Multipliers that might result in different dual
problems. For the following proposed solution, Lagrangian
Multipliers are chosen aiming to facilitate implementing it
in the real time. We refers to the following development as
Algorithm A.
As a first step, the Lagrangian multipliers i
s, ml, yl,
i ¼ 1; 2; . . . ; Hs; s 2 S; l 2 L are introduced for users and
links with respect to constraints (18), (19) and (20), respec-
tively. Here, i
s, ml and yl can be interpreted as flow conser-
vation price for the ith lossy link of flow s, i.e., congestion
price, the price for keeping the rate-outage probability at
each link l within a small tolerable value, and the price for
the average delay no more than a specific delay-tolerant
amount, i.e., queueing delay price, respectively. The objec-
tive function of the Lagrangian dual problem of (17) can be
then expressed as
Dð; m; yÞ ¼ max
f~x2~L; ~#;~p2~Ig
Lð~x; ~#; ~p; ; m; yÞ
¼ max
f~x2~L; ~#;~p2~Ig
L ~x ð~x; ; m; yÞ þ L~pð~p; ; m; yÞ
þ L~#ð ~#; yÞ;
(21)
where L~xð~x; ; m; yÞ is the Lagrangian function with respect
to source rate ~x and Lagrangian multipliers ; m and y;
L~pð~p; ; m; yÞ represents the Lagrangian function with
respect to transmission power ~p and Lagrangian multipliers
; m and y; L~#ð ~#; yÞ denotes the Lagrangian function with
respect to transmission delay ~# and Lagrangian multiplier
y. The expressions of these Lagrangian functions can be
obtained by multiplying the Lagrangian multipliers by the
corresponding constraints (18), (19), (20) and combining the
similar terms.
By the linearity of the differentiation operator, the objec-
tive (21) can be decomposed into three separate maximiza-
tion subproblems due to the separable nature of the
function
P1 : D~xð; m; yÞ ¼ max
~x2~L
L~xð~x; ; m; yÞ;
P2 : D~pð; m; yÞ ¼ max
~p2~I
L~pð~p; ; m; yÞ;
P3 : D~#ð; m; yÞ ¼ max
~#
L~#ð ~#; yÞ;
s:t: ~# 2 R, where R denotes the real number.
Subproblem P1 is the rate control subproblem in terms of
rate variable ~x, P2 represents the power control subproblem
to achieve the minimum power consumption, and P3
denotes the average delay control subproblem to give an
optimal delay-tolerant amount ~#. Note that indices s, i and l
are dropped in P1, P2 and P3 since they are applicable for
all s, i and l: The maximization involved in the calculation
of the dual function has been decomposed into three simple

maximization subproblems, which implies that the optimal
resource allocation problems can be separated and solved
in a distributed manner.
Accordingly, the dual problem of the original (primal)
problem is given by
min
;m;y
Dð; m; yÞ: (22)
Because the primal problem is convex and also satis-
fies the Slater’s condition [29], the duality gap between
the optimal value of the primal problem and that of the
dual problem becomes zero. This suggests that we can
design primal-dual algorithms to find the optimal
powers, rates, delay-tolerant amount and prices. More
specifically, they can be updated iteratively in such a
way that the Lagrangian function Lð~x; ~#; ~p; ; m; yÞ is max-
imized with respect to ~x, ~p and ~# for the primal, and the
dual function Dð; m; yÞ is minimized with respect to ,
m, and y for the dual.
Next, we will take an iterative approach to solving the
dual problem (22): at iteration t, the maximization problems
P1, P2 and P3 are solved for fixed , m, and y, then a subgra-
dient method is employed to update the dual variables.
Subsequent iterations t þ 1, t þ 2, . . . would repeat this pro-
cess until convergence to the globally optimum resource
allocation.
4.1 Source Rate Subalgorithm
The source rates are updated by solving the maximiza-
tion problem P1. Since problem P1 is strictly concave for
fixed , m, and y, we can use a subgradient projection
with a sufficiently small fixed step-size to find a maxi-
mizer. For convenience, we define 0
s ¼ 0, 8s. Taking par-
tial derivative of L~xð~x; ; m; yÞ with respect to ~xi
s, we have
the subgradient on the data rate ~xi
s as follows:
for i ¼ 1; 2; . . . ; Hs,
rD1
À
~xi
s
Á
¼ i
s À iÀ1
s À ylðs;iÞexp
À
~xi
s
Á
; (23)
and for i ¼ Hs þ 1,
rD1ð~xi
sÞ ¼ U0
s exp
À
~xHsþ1
s
ÁÀ Á
exp
À
~xHsþ1
s
Á
À Hs
s ; (24)
where U0
sðÁÞ is the first derivative of the utility function.
Due to the strict concavity of the Lagrangian function
L~xð~x; ; m; yÞ by Theorem 1 and Theorem 2, the dual func-
tion Dð; m; yÞ is differential everywhere by [30], Prop.
6.1.1]. Taking partial derivative of with respect to i
s, the
subgradient on the price i
s is given by
rD
À
i
s
Á
¼ ~xi
s À ~xiÀ1
s À clðs;iÞ
ð~PPÞ: (25)
Based on (24) and (25) and by utilizing the complementary
slackness condition, we can obtain the destination of flow s,
the flow conservation price and the effective rate as follows:
Hs
s ¼ U0
s exp
À
~xHsÀ1
s
ÁÀ Á
exp
À
~xHsÀ1
s
Á
; (26)
~xHsþ1
s ¼ ~xHs
s À clðs;HsÞ
ð~PPÞ: (27)
Remark 2. clðs;iÞ
ð~PPÞ is referred to as the data loss rate at the
ith link of flow s. Since clðs;iÞ
ð~PPÞ is a decreasing function
on ~PP, which means that the larger the power is, the
smaller the loss rate. We can observe from (26) and (27)
that at the destination of flow s, the effective rate
becomes larger as the power increases. Since Hs
s is a
non-increasing function of the effective rate ~xHsþ1
s , it can
be viewed as the effective flow weight, which guarantees
the fairness among destination nodes.
Next, we use the subgradient projection method to find
the optimal solution for ~xi
s and i
s. We have the following
theorem.
Theorem 3. At the ith link of flow s, the flow rate ~xi
s can be
updated as follows:
~xi
sðt þ 1Þ ¼ ~xi
sðtÞ þ rD1ð~xi
sÞ
Â Ã
~Ls
; i ¼ 1; 2; . . . ; Hs; (28)
where is a sufficiently small fixed step-size.
It is noted that the rate update depends on the link con-
gestion and the QoS price induced by tolerant delay and the
difference of available buffer spaces between adjacent nodes
i
s À iÀ1
s .
At the receiver of the ith link of flow s, by using the sub-
gradient projection method to solve the dual problem (22),
the price i
s can be updated as follows:
i
sðt þ 1Þ ¼ i
sðtÞ À rD
À
i
s
ÁÂ Ãþ
; i ¼ 1; 2; . . . ; Hs: (29)
Equation (29) reveals that i
s is intimately related to the
available buffer space at the receiver of the link, which
determines in turn the difference between the incoming rate
~xi
s À clðs;iÞ
ð~PPÞ and the outgoing rate ~xiþ1
s . For instance, the
more the available buffer space the receiver has, the smaller
the difference between the incoming rate and the outgoing
one, which results in an increasing price i
s. In fact, this is a
back-pressure algorithm where the flow with greater queue
length difference is granted higher priority. If each node is
assumed to have the same total buffer size, then i
s À iÀ1
s is
exactly the queue length between the transmitter and
receiver at the ith link of flow s. The detailed rate control
subalgorithm for subproblem P1 is summarized in Table 2.
It is clear that the complexity of the subalgorithm is OðH2
s Þ.
4.2 Link Power and Delay Subalgorithm
Since the twice order derivatives of L~pð~p; ; m; yÞ and
L~#ð ~#; yÞ with regard to ~p and ~# respectively are negative
semidefinite, i.e., they are concave, there exist optimal solu-
tions of ~p and ~#. Taking partial derivative of L~pð~p; ; m; yÞ
with respect to ~pl, and coming back to the solution space PP
instead of ~PP, it is easy to verify that the subgradient on the
transmission power pl is
rD2ðplÞ ¼ VlðPPÞGll þ
X
n6¼l
GnlMnðPPÞ À w1; (30)
where
VlðPPÞ ¼
ylWKl
1 þ KlglðPPÞ
À ðl
s þ mlÞ
Rth
l
ðglðPPÞÞ2
; (31)

MnðPPÞ ¼
VnðPPÞðgnðPPÞÞ2
Gnnpn
: (32)
Therefore, we have the following theorem.
Theorem 4. The transmission power is updated for each link
l 2 LðsÞ with message passing by
plðt þ 1Þ ¼ plðtÞ þ rD2ðplÞ½ Š
pmax
l
pmin
l
¼ plðtÞ þ GllVlðtÞ À w1 þ
X
n6¼l
GnlMnðtÞ
!pmax
l
pmin
l
;
(33)
where each MnðtÞ is referred to as the messages comprising
local information at each link n, which is a scaled measurement
of the receiver SINR. A distributed power-update algorithm
can be implemented through message passing: each receiver on
link n broadcasts its message MnðtÞ and estimates MnðtÞ
through training sequence.
Similarly, the link average delay is obtained by solving
L~#ð ~#; yÞ and updated by
#lðt þ 1Þ ¼ ðylðtÞK=w2Þ1=2
: (34)
The link power and delay subalgorithm for subproblems
P2 and P3 is summarized in Table 3. It is clear that the com-
plexity of the subalgorithm is OðK2
LÞ.
4.3 Lagrangian Multiplier Update Subalgorithm
The updates in (33) and (34) are shown after transformation
back to the original co-ordinate space. Furthermore, by
using the subgradient projection, prices ml and yl are
updated by
mlðt þ 1Þ ¼ mlðtÞ À Kl À cl
ð~PPÞ
À ÁÂ Ãþ
; (35)
and
ylðt þ 1Þ ¼ ylðtÞ À rDðylÞ½ Šþ
; (36)
respectively, where
rDðylÞ ¼ À
X
s2SðlÞ
exp
À
~xl
s
Á
þ clðexpð~PPÞÞ;zÞ À
K
expð ~#lÞ
:
It is worth noting that in principle, mobile ad hoc networks
are independent of any established infrastructure or central-
ized mechanism. Each node operates in a distributed peer-
to-peer mode, acting as an independent router and an end
host. It can move independently in any direction, thus
change its links to other devices frequently. In such an infra-
structure-less network, there does not exist any central coor-
dination or control. Thus, centralized solutions are
impractical. The proposed RENUM algorithm has to be
implemented in a distributed manner in the absence of cen-
tral coordination, through the cooperation of nodes. To elab-
orate, on one hand, each node must give the feedback such
as message Mn, based on the local information such as SINR
gl and prices i
s, ml and yl of other links in the interference
set, rather than the individual link, and broadcast the mes-
sage to its neighbor nodes. On the other hand, the distributed
nature of the RENUM algorithm can cope with the absence of
the central coordination in ad hoc networks: source s can
solve Equation (28) separately without need to coordinate
with other sources because i
s summarizes all the congestion
information source s needs to know; Equations (33) and (34)
can be implemented by an individual link l using only the
accumulated information such as message Mn from its inter-
ference set and its own information such as ingress rate Rl,
average SINR gl and prices l
s, ml and yl of link l.
The complexity of the algorithm can be derived as fol-
lows. Assuming that the ellipsoid method is used to itera-
tively update dual variables, the required number of
iterations for convergence is Oðm2
Þ, where m denotes the
size of the problem. Let KL and KS denote the number of
links and sources, respectively. The ellipsoid method will
need OðK2
SK2
LÞ iterations to obtain iÃ
s
È É
and ~xiÃ
s
È É
and
OðK2
LÞ iterations to obtain mÃ
l
È É
, yÃ
l
È É
, pÃ
l
È É
and #Ã
l
È É
.
Hence, the total complexity of the algorithm is OðK2
SK4
LÞ.
4.4 Relationship between RENUM Algorithm
and TCP
In this section, we discuss the relationship between the pro-
posed RENUM algorithm and transmission control protocol
(TCP).
Our proposed source rate control subalgorithm is closely
related to the congestion control in TCP. It is well known
that a TCP sender maintains two windows, the congestion
window and the TCP maximum window, for congestion
TABLE 3
Link Power and Delay Subalgorithm for Flow s
TABLE 2
Distributed Rate Control Subalgorithm for Flow s

control. The congestion window is the flow control imposed
by the sender, while the TCP maximum window, i.e., the
window advertised by the receiver, is the flow control
imposed by the receiver. A source/sender s keeps a variable
called congestion window size Ws that determines the num-
ber of packets it can transmit in a round-trip time (RTT) ti
over link i, the time from sending a packet to receiving its
acknowledgment from the destination. This implies that the
source rate xi
s in RENUM is approximately equal to the ratio
of window size to RTT, i.e., xi
s ¼ Ws=ti, in packets/second.
Compared to the wired networks, for multihop MANETs,
the traditional TCP cannot distinguish effectively between
congestion loss induced by buffer overflow and packet loss
caused by mobility and shared channel contention among
wireless nodes. This is because that buffer overflow-
induced packet losses are rare and packet drops due to
wireless link contention dominate.
To make the congestion control mechanism of TCP adapt
to multihop MANETs, we employ a cross-layer method,
which adaptively adjusts the TCP maximum window size
according to the number of RTS (request to send) retry
counts of the IEEE 802.11 DCF (distributed coordination
function) protocol of the MAC layer at the TCP sender, to
control the number of TCP packets in the network and thus
reduce the channel contention. In our method, not only the
TCP receiver but also the TCP sender participate in adapt-
ing the maximum window size. More specifically, the
receiver still determines the maximum window size accord-
ing to its buffer space, and the sender also adjusts the maxi-
mum window size according to the average RTS retry
counts of its MAC layer.
Equation (28) in Theorem 3 reflects the TCP congestion
window update while Equation (29) describes the link con-
gestion price update. The source rate control subalgorithm
shows that if the data loss rate clðs;iÞ
ð~PPÞ at the ith link of
flow s becomes higher, which implies that congestion
occurs at the ith link of flow s, then congestion price i
s will
increase. In this case, source s decreases its window size Ws
and transmission rate. Otherwise, if the demand for the
weighted incoming flow rate ylðs;iÞ expð~xi
sÞ at the ith link of
flow s is less than the available buffer spaces between adja-
cent nodes of flow s or the allowed maximum queue length
between the transmitter and receiver at the ith link of flow
s, i
s À iÀ1
s , then source s increases its window size and
transmission rate. Herein, i
s À iÀ1
s can be regarded as the
TCP receiver’s advertised window, i.e., the TCP maximum
window. Thus, the proper reduction of the advertised win-
dow by intermediate transmitters/receivers can control the
number of packets sent from a TCP source.
To achieve the adaptive adjustment for the TCP maxi-
mum window size i
s À iÀ1
s according to the number of
RTS retry counts, in the RTS/CTS (request to send/clear
to send) scheme of the IEEE 802.11 MAC protocol, each
node has a variable, ssrc (station short retry count), to
record the retry count of the current RTS frame. Each
node determines its current channel access according to
the value of ssrc. To make the node know its channel
state, we introduce a new variable, rrc, to record the retry
count of the last RTS frame. The bigger the rrc is, the
more RTS collisions the node experiences and thus the
worse the channel state is. According to the value of rrc
from the MAC layer, the TCP sender can know its local
channel contention state and estimate the channel conten-
tion state of the forwarding nodes. Therefore, the TCP
sender can reduce shared channel collisions in the net-
work by dynamically adjusting the TCP maximum win-
dow size i
s À iÀ1
s based on its local RTS retry counts.
Such a method improves the TCP performance by con-
trolling the packets injected by TCP senders and thus
avoids the congestion caused by packet losses due to low
link quality, which is owing to that we tackle TCP degra-
dation in the wireless domain through the cooperation
between the transport layer and the MAC layer.
In addition, in RENUM algorithm, some information is
required for the subalgorithm updates such as the local
information of SINR glðPPÞ, prices yl; l
s; ml, ingress rate Rl
and message Mn from the neighbor set. The local informa-
tion can be computed by each node itself. Message Mn can
be included in a distributed protocol by making use of
explicit message passing. To achieve low communication
overhead, we employ a type of piggybacking similar to the
method in [31], i.e., the information of Mn is piggybacked
onto acknowledgment (ACK) packets. A field is reserved in
an ACK packet header that is sent from the receiver back to
the transmitter. Along the path, each intermediate link accu-
mulates the information of message Mn into this field. When
the ACK reaches the transmitter, this reserved field forms
the required information for those updates. This way, the
network can also provide explicit feedback information to
the TCP sources to dynamically adjust their window size
when congestion is detected.
To reach an inference and update for data rate and
transmission power, each node maintains a local broad-
cast message pool. The local broadcast message pool
includes n entities (where n is the number of down-
stream links of the node). Each entity includes the fol-
lowing fields: (1) ID of sender and ID of receiver, which
are used to uniquely identify the link, (2) sequence no.,
which is used to judge the freshness of the entity, (3)
message type, which is set as “1” to identify that the
information is received from other nodes, (4) fields to
record message Mn. At the beginning, the sender of the
link generates an entity of the broadcast message packet
for a link. In this entity, the ID of sender, ID of receiver
and message Mn are filled in the corresponding field.
When a data packet is received successfully, the ACK
packet piggybacking a broadcast message is transmitted.
The broadcast message packet contains all the entities in
the local broadcast message pool that have been broad-
cast. When a broadcast message packet is received, each
entity is handled as follows: According to ID of sender
and ID of receiver, if an entity did not exist earlier, this
entity is added in the local broadcast message pool. If an
entity already exists, and the newer entity has a larger
sequence number, the newer entity is also added in the
pool, and the older entity is discarded. If the newer
entity has a smaller sequence number, this newer entity
is discarded.
4.5 Packet Routing and Flow Contention
In this section, we extend our rate control subalgorithm to
handle the route outages by jointly considering the packet

routing at the network layer and the flow contention at the
MAC layer.
The routing of data packets is determined by congestion
price i
s and flow distribution over the links they traverse.
Intuitively, a node prefers to forward the flow to a neighbor
with lower congestion price or on the route with a lower
flow rate. Such a scheme guarantees that each flow travels
through links in a decreasing order of congestion prices.
This indicates that the packet routing is in accordance with
the rate control since we can observe from (25) and (28) that
the increase of data loss rate clðs;iÞ
ð~PPÞ would lead to the
increase of congestion price i
s, and in turn the decrease of
flow rate ~xi
s. Accordingly, the routing of data packets will
inevitably make the adjustment based on the variations of
congestion price and flow rate. Therefore, the routing imple-
mentation requires a supplementary mechanism which
coordinates the necessary message exchange. Node i needs
to collect local measures such as average SINR gl and mes-
sage Mn as well as reports of congestion prices j
s from its
neighbor j to which it forwards flow s. Moreover, node i is
responsible for calculating its own congestion price by (28),
and then providing i
s to its neighbors from which it
receives traffic of flow s.
Remark 3. It is known that the ad hoc on-demand distance
vector (AODV) routing protocol uses the messages of
route requests (RREQs), route replies (RREPs) and route
errors (RERRs) to discover and maintain routes. As long
as the endpoints of a communication connection have
valid routes to each other, AODV does not play any role.
Therefore, there exists a flow between the source and the
destination for the AODV routing protocol as long as the
communication links connecting them are valid. Clearly,
our proposed RENUM algorithm is effective for the case
that AODV keeps the discovered routes. As for the case
when AODV needs to discover a new route, the RENUM
algorithm will employ the messages of RREQs, RREPs
and RERRs to show that a new route is found or a link
break has occurred, and accordingly update the set of
links LðsÞ connecting source s and its destination. During
discovery, the broken link hinders the data transmission,
which in turn causes the increase of congestion price i
s.
As a result, source s will have to choose the set of links
with lower congestion prices to transmit data. This adap-
tive routing shows the proposed hop-by-one source rate
subalgorithm can also work well for the case of discover-
ing a new route.
In the meanwhile, both the interflow contention and the
intraflow contention at the MAC layer could cause the route
outage of data transmission. The interflow contention is
referred to as the contention that each multiple flow encoun-
ters from other flows which pass by the neighborhood. The
intraflow contention is the MAC layer contention for the
shared channel among the nodes of the same flow, which
are in each other’s interference range. These two types of
flow contentions could result in severe collisions and con-
gestion. A good solution to the flow and congestion control
problem in ad hoc networks must consider the MAC layer
characteristics and respond quickly to the congestion. The
essence of addressing the intraflow contention problem is to
prevent the first few nodes on the path from injecting more
packets than that next nodes can forward. Therefore, in this
paper, one way to prevent the intraflow contention is to
assign high channel access priority to each node when it
only receives a packet, and those packets which have tra-
versed more hops. That is to say, the source node tries to
hold the current packets until the preceding packet is trans-
mitted out of its interference range. This mechanism only
considers the interference in a single flow. If the next hop of
the current receiver is busy or interfered by other transmis-
sion, the receiver cannot seize the channel even with the
highest priority.
Hence, we introduce the backward-pressure scheduling
mechanism to deal with the interflow contention. Its basic
idea is to keep nodes from transmitting to their already con-
gested downstream nodes, hence, yield the channel access
to the congested nodes to clear congestion as well as to
avoid severe medium contention. The mechanism includes
a transmission blocking procedure and a transmission
resuming procedure. It requires that each node monitors
the number of individual flow in the shared outgoing
queue. Let ns denote the number of packets of flow s. If ns
reaches a backward-pressure threshold, the transmission of
flow s from its upstream node will be blocked. When the
node successfully forwards some packets to its downstream
node so that ns is less than the backward-pressure thresh-
old, it initiates the transmission resuming procedure to
allow the upstream node to transmit packets of flow s. The
transmission blocking procedure takes advantage of the
RTS/CTS exchange in the IEEE 802.11 MAC protocol to
restrict the transmission from the upstream nodes. The
transmission resuming procedure adopts the receiver-initi-
ated transmission and uses the three-way handshake CTS/
DATA/ACK instead of the normal four-way handshake
RTS/CTS/DATA/ACK, because the downstream node
already knows the restricted upstream node has packets
designated to it.
5 DISPOSING OF POWER ALLOCATION OVERHEAD
In pursuit of the globally optimum solution to the RENUM
problem (10)-(13), explicit message passing between links is
required for power allocation (33). While broadcast is a via-
ble way to realize such inter-link communication, the ideal
scheme is to use some type of indirect measurement to
achieve a similar goal.
In this section, we provide an algorithm with the follow-
ing property: each link’s power-allocation is based on the
locally measured interference level caused by other links.
No explicit message passing is required. These messages
are embedded in a header field within ACK packets that tra-
verse the source’s reverse-path and incur little overhead.
The algorithm is referred to as Algorithm B.
5.1 Link Power Subalgorithm
The following development concentrates on the fact that at
the optimum solution, the ingress rate to a link is matched
to its capacity (all constraints are active). Considering a
particular link l, we can rearrange the constraint in terms
of the link transmission power, resulting in the following
power update

plðt þ 1Þ ¼
plðtÞ
glðPPðtÞÞ
’lðxs; #lÞ; 8l; (37)
where ’lðxs; #lÞ ¼ expð 1
W ð
P
s2SðlÞ xs þ K
#l
ÞÞ À 1.
It is observed from (37) that gl is a scaled noise measure-
ment, while the term ’lðxs; #lÞ makes use of an estimate of
the ingress rate and the link average delay. When both the
rate and the link delay are fixed, we can interpret ’lðxs; #lÞ
as a SINR-threshold, and convergence of the update follows
from [32]. Clearly, the measurement of noise and SINR can
be carried out by the interference level caused by local other
links. It is also found that the update rate is a decreasing
function of the average delay #l. In other words, the longer
the average delay is, the lower the rate of power update
becomes. It is also consistent with the power update in prac-
tical wireless networks.
5.2 Source Rate and Delay Subalgorithm
We now obtain a rate allocation under the assumption that
link powers are fixed. Plugging (37) into (10) results in the
following optimization:
max
x;#;p!0
X
s2S
Us xHsþ1
s
À Á
À w2
X
l2L
#l À w1
X
l2L
pl
glðPPÞ
’lðxs; #lÞ
(38)
s:t: xiþ1
s xi
sflðs;iÞ; i ¼ 1; 2; . . . ; Hs; 8s; (39)
Pr
X
s2SðlÞ
xl
s clðPPÞ
!
zmax
l ; 8l: (40)
Remark 4. The above method does not guarantee the opti-
mality of the joint rate and power allocation of the origi-
nal problem (10) since the optimization with the fixed
link powers PP replaces function PPÃ
ððxÞ, which denotes
the simultaneous minimum power solution to (37) for a
specified xx. However, we do not have a closed-form
expression for this function.
Similar to (17), by employing some auxiliary variables
and logarithmic transformation, the optimization problem
(38) can be transformed to the unconstrained optimization
DDð; m; yÞ ¼ max
f~x2~L; ~#;~p2~Ig
L1ð~x; ; m; yÞ þ L2ð~p; ; m; yÞ
þ L3ð ~#; yÞ; (41)
where L1ð~x; ; m; yÞ is the Lagrangian function with respect
to source rate ~x and Lagrangian multipliers ; m; y;
L2ð~p; ; m; yÞ represents the Lagrangian function with respect
to transmission power ~p and Lagrangian multipliers ; m; y;
L3ð ~#; yÞ denotes the Lagrangian function with respect to
transmission delay ~# and Lagrangian multipliers y.
For (41) with ~PP fixed, we observe that the objective is
strictly concave in ~x. We have the subgradient on the data
rate ~xi
s as follows for i ¼ 1; 2; . . . ; Hs,
rDDð~xi
sÞ ¼ i
s À iÀ1
s À
w1exp
À
~xi
s
Á
Ms
W
; (42)
where Ms ¼ ~pl
glð~PPÞ
expð 1
W ð
P
s2SðlÞ expð~xsÞ þ K
expð ~#lÞ
ÞÞ and for
i ¼ Hs þ 1; the subgradient is the same as that in (24).
Theorem 5. At the ith link of flow s, the flow rate ~xi
s without
explicit message passing can be given by
~xi
sðt þ 1Þ ¼ ~xi
sðtÞ þ rDDð~xi
sÞ
Â Ã
~Ls
; i ¼ 1; 2; . . . ; Hs; (43)
where is a sufficiently small fixed step-size.
Similarly, the objective in (41) is also strictly concave in ~#l
for ~PP fixed. We can therefore make use of the delay update
~#lðt þ 1Þ ¼ ~#lðtÞ þ rDDð ~#lðtÞÞ
Â Ãþ
; 8l; (44)
where is a sufficiently small fixed step-size and
rDDð ~#lðtÞÞ ¼
w1KMs
Wexpð ~#lðtÞÞ
: (45)
The update of i
s and ml is the same as that in (29) and (35),
respectively.
Ms are messages that are accumulated from each link
along the route of a source s. Each of these terms involves a
noise measurement that is scaled by the total ingress rate
and the delay a per-link, a locally measurable quantity. As
in Section 4, these terms may be progressively accumulated
into an ACK packet header as it traverses from the receiver
back to the source.
This source rate update is guaranteed to converge if (43)
and (44) can be iterated until convergence for a single
source, keeping other sources fixed, and before proceeding
to the next source. In practice, we have observed such con-
vergence even when these messages are used for continu-
ous asynchronous updates.
6 PERFORMANCE EVALUATION
In this section, we evaluate the performance of the proposed
RENUM algorithm by numerical results and network-wide
simulation.
6.1 Numerical Results
In this section, we implement the proposed RENUM algo-
rithm in MATLAB and provide numerical examples to illus-
trate the advantages of RENUM over other network utility
maximization (NUM) algorithms without link outage/aver-
age delay constraints in terms of convergence, effective
rates, link delay and power consumption.
Consider a MANET with k links, where the links are
numbered from 1 to k. The link outage constraint is set to be
zmax
l ¼ 0:1, l 2 1; 2; . . . ; k. An example network with four
sources having fixed routes is shown in Fig. 2. Flows 2 and
3 traverse all the links from link 2 to link 4 and from link 1
to link 5, respectively, and flow 1 is three hops to its destina-
tion over links 3 and 4. The receiver of link 1 is interfered by
the transmitters of flows 1 and 4. A transmission bandwidth
is 1 MHz, giving a W ¼ 125 KHz baseband on each link. We
let ms ¼ 0:1 kbps, Ms ¼ 0:5 Mbps, K ¼ 1 Mbit, Glk ¼ d4
lk,
pmin
l ¼ 1 mw and pmax
l ¼ 100 mw. We average over 50 trials
to obtain a small relative error (within 10 percent of the
average value).
For clearer illustration, we use a-fairness utility function
UsðxsÞ ¼ ð1 À aÞÀ1
x1Àa
s with a ¼ 5. We compare the perfor-
mance of RENUM with JOPRC in [13], both of which

propose joint power scheduling and rate control schemes in
wireless ad hoc networks. However, the latter employs
opportunistic power scheduling and a stochastic subgra-
dient algorithm. Figs. 3a and 3b show the comparison of the
injection rates and the effective rates on flows 1 and 3,
respectively. Clearly, the effective rates received at destina-
tions are smaller than the injection one at transmitters. As
shown in Fig. 3b, both RENUM and JOPRC can converge
fast, whereas RENUM yields higher effective rates for flows
1 and 3 than JOPRC does on their destination nodes. It is
seen from Fig. 3a that the injection rates for both flows 1
and 3 in RENUM is larger than those in JOPRC while the
total power consumption of the two for all flows are almost
equivalent, as shown in Fig. 3c. This is because that RENUM
explicitly takes into account the data loss in its objective
function, whereas JOPRC is concerned with the injection
rates only. The data loss can induce the increase of source
rate to maximize the network utility.
Fig. 4a compares the convergence of Algorithms A
and B proposed in this paper and their power consump-
tion. It is clear that Algorithm B converges much faster
than Algorithm A, whereas compared with Algorithm A,
Algorithm B is not optimal and it consumes more power
to achieve the convergence. It is justified since no explicit
message is passed to the transmitters; moreover, it needs
more power to measure interference level caused by
other links. We now consider the channels with fast fad-
ing. Only algorithm B is considered due to its attractive
convergence rate and fully-distributed power-updates.
At each system period t, each link fading gain Glk is
computed with Clarke’s model [33, Ch. 5] at a carrier
frequency of 2.4 GHz and we use v ¼ 10 km/hr to calcu-
late the maximum Doppler frequency for the worst-case
situation of node moving in opposite directions. Fig. 4b
shows the effect of node mobility velocity on the
effective rate of flow 3 by Algorithm B. It is clear that as
the node mobility velocity increases, the effective rate of
flow 3 decreases rapidly. The RENUM with Algorithm B
can converge fast in low mobility velocity, whereas the
effective rate of flow 3 oscillates severely in high mobil-
ity velocity, in this case, it is controlled by a rate-outage
target 20 percent.
Figs. 5a and 5b show the effect of maximum rate outage
probability on efficient rate and flow delay when RENUM
converges to optimum. We can observe from Fig. 5a that
for a given maximum rate outage probability, the flow
effective rate on destination decreases as the number of
hops increases. Especially, due to the competition for
wireless links among multiple flows, the decrease
becomes sharper. We can also find that as the maximum
rate outage probability increases, there exits a bound on
the effective rate increase. Once crossing this bound, the
effective rate begins to decrease. As shown in Fig. 5b, its
effect on flow delay is opposite to that. This is reasonable
since at the beginning, the larger rate outage probability
means more tolerance on fading-induced congestion and
more packets forwarded, which further induces the
increase of the effective rate and the decrease of flow
delay. However, once the tolerance is beyond a bound,
more packets injected into the wireless links will make
the congestion more serious. Furthermore, a lot of packets
are dropped and flow delay increases as well.
We also compare the performance of RENUM with
ENUM proposed in [23]. Fig. 6 depicts the effective rate,
power consumption and delay obtained by RENUM and
ENUM for flow 3. We can find that both can achieve the
similar effective rate, whereas flow 3 by RENUM con-
sumes less transmission power and has smaller average
delay in comparison to ENUM. It is justified since
ENUM only takes into account the effect of the lossy fea-
ture of wireless links on rate control to maximize the
network utility, however, does not consider the effect of
the lossy feature on transmission power and link delay.
Especially, the objective of ENUM is not to minimize the
total power consumption and flow delay.
6.2 Network-Wide Performance
To evaluate the network-wide performance of the pro-
posed RENUM algorithm, we also implement it in NS-2
simulator [34].
Link 1
Flow 2
Flow 3
Link 2 Link 3 Link 4 Link 5
Flow 1
Flow 4
Interference
Flow
Wireless Link
Fig. 2. MANET topology.
Fig. 3. Performance comparison between RENUM and JOPRC. (a) Injection rates of flows 1 and 3 on their own sources. (b) Effective rates of flows 1
and 3 on their own destinations. (c) Total power consumption comparison between ERNUM and JOPRC on all links.

To elaborate, the source rate control subalgorithm is
implemented in the transport layer while the power alloca-
tion subalgorithm is in the physical layer. Unless otherwise
stated, the default parameter settings are as follow: the two-
ray ground reflection model is used as the radio propaga-
tion model, IEEE 802.11 DCF (distributed coordination
function) as the MAC protocol, Drop Tail as the interface
queue type, Omniantenna as the antenna model and AODV
(ad hoc on-demand distance vector routing) as the routing
protocol. An interface queue at the MAC layer could hold
50 packets before they are sent out to the physical link. Link
breakage is detected from MAC layer feedbacks. A routing
buffer at the network layer could store up to 64 data pack-
ets. This buffer keeps data packets waiting for a route, such
as packets for which route discovery has started but no
reply arrived yet.
In the simulation, 60 nodes are randomly placed in a
1,000 m Â 1,000 m square area. The source of each TCP flow
randomly selects one node as the destination, which is at
least m hops away from the source. In our studies, we
choose m ¼ 1; 3; 5. There are a total of 20 flows with the
same CBR/UDP traffic in the network. In these topologies,
each node is 200 m away from its closest neighbors. The
transmission range and the interference range are 250 and
550 m, respectively. The data transmission rate is 2 Mbps.
The packets generated are of size 1,000 bytes in the simula-
tion. The mobility model used for topology generation is
the random waypoint model [35]. We conduct simulation
using five different moving speeds, 0, 1, 5, 10 and 20 m/s.
The speed of 0 m/s means all nodes are static. The pause
time of each node is equal to 0, and therefore each node
moves continuously in the network if its maximum speed is
greater than 0. All results are averaged over 30 random sim-
ulations with 80 s simulated time each.
In the simulation, the following several important perfor-
mance metrics are evaluated.
Aggregate end-to-end throughput. The amount of data
delivered to the destinations per second.
End-to-end delay. The accumulative delay in data
packet delivery due to buffering of packets, new
route discoveries, queuing delay, MAC-layer
retransmission, and transmission and propagation.
delays.
Packet delivery ratio. Percentage of data packets
received at the destinations out of the total data
packets generated by CBR traffic sources.
Fairness index. The commonly used fairness index for
all flows xið1 i nÞ, i.e.,
f ¼
ð
Pn
i¼1 xiÞ2
ðn Á
Pn
i¼1 x2
i Þ
; (46)
where xi denotes the end-to-end throughput of the
ith flow.
We first verify the performance of RENUM in terms of
the aggregate end-to-end throughput and end-to-end delay.
Fig. 7a depicts the aggregate end-to-end throughput of
RENUM under different number of hops. We observe that,
when the number of hops for each flow increases, the aggre-
gated end-to-end throughput of RENUM decreases. This is
reasonable because multihop flows with longer paths have
to pass more lossy links, thus lose more data packets for the
same injection traffic, which results in that the effect rates
received at destinations decrease. In addition, we also find
that a larger number of hops can induce unstable aggregate
Fig. 4. (a) Comparison of convergence and power consumption between
Algorithms A and B. (b) Effect of the node mobility on the effective rate
by Algorithm B.
Fig. 5. (a) Effect of maximum rate outage probability on efficient rate.
(b) Effect of maximum rate outage probability on flow delay.
Fig. 6. Performance comparison between RENUM and ENUM. (a) Comparison of effective rate for flow 3 between the two. (b) Comparison of power
consumption for flow 3 between the two. (c) Comparison of delay for flow 3 between the two.

end-to-end throughput. The underlying reason is that the
flows with longer paths will encounter not only the inter-
flow contention, but also the intraflow contention, which
could result in collisions and congestion, and significantly
limit the performance of ad hoc networks.
Fig. 7b shows end-to-end delay of RENUM with different
total injected traffic when the speed of node mobility is
1 m/s. We can observe that the end-to-end delay is low
when the network is in light traffic load, then rapidly
increases with traffic load and finally trends to be stable in
spite of high traffic load. This is because as the traffic load
becomes heavy, more flows take part in the contention for
shared channels, and the network begins to get congested,
which results in the longer round-trip time (RTT) of TCP
packets. In the meanwhile, we can see that a large number
of hops leads to long end-to-end delay. The reason is that a
large number of hops can prolong the time of ACK feed-
back, data retransmission and new route discovery, and
lead to a lot of accumulated packets in the outgoing queue
at each node, which can greatly increase the queueing delay.
We now explore the impact of node mobility on packet
delivery ratio, throughput and fairness.
Fig. 8 specifies data packet delivery ratio with different
node moving speeds. The purpose of considering mobil-
ity is only to illustrate that our RENUM framework can
work well in mobile scenarios with an on-demand rout-
ing scheme. We have the following observations. First,
the data packet delivery ratio of nodes with lower moving
speeds is larger than that with higher moving speeds.
Second, the larger number of hops, the smaller packet
delivery ratio. The reason behind these observations is
that if nodes move rapidly in the network, the mobility
would cause packet losses and make TCP connection
unreliable. Moreover, when the path length of TCP flows
increases, the probability that TCP packets collide with
their ACKs would become larger, and the TCP sender has
fewer possibilities to accurately evaluate the channel state
of forwarding nodes according to its local information.
Fig. 8b indicates fairness index with different speeds of
node mobility. We can see that RENUM algorithm achieves
comparatively stable fairness among all flows and allocates
bandwidth fairly even for the nodes with higher moving
speed. Moreover, we also observe that the number of hops
becomes the main factor that influences the fairness of
resource allocation among flows, i.e., the longer the length of
TCP flows is, the worse the fairness among the flows is. This
is because that in the RENUM framework, we consider rate
fairness constraint, which makes it possible that the flows
with poor channel conditions have an opportunity to trans-
mit data, and employ the rate-effective based congestion
control mechanism with an explicit congestion notification
including the link congestion prices and the channel states.
We now study the effect of variability and unexpected
events such as sudden flow changes and cyberattacks on
throughput in the random topology.
Fig. 9a shows the effect of sudden flow changes on
throughput in the random topology. In this simulation,
the number of TCP flows is 50 at time 0.0 s. At the 40th
second, 15 flows are terminated. At the 80th second, addi-
tional 10 flows are added. The whole simulation lasts
120 s. We can observe that RENUM algorithm is sensitive
to these sudden flow changes and rapidly becomes stable
after these disturbances. As the flows depart suddenly,
the throughput becomes small, and the change is opposite
for sudden arrivals. This experiment verifies that RENUM
algorithm is much robust for sudden flow changes. The
reason behind these observations is that we explicitly
take into account the average SINR and take advantage of
rate-outage probability to allow the network to experience
a limited amount of congestion.
Fig. 9b depicts the evolution of aggregate throughput
when 10 nodes are attacked at 40th second in the random
topology. This simulation aims to verify whether RENUM
algorithm can resist and tolerate the cyberattacks, and can
hold on the normal performance when the network is
attacked. In this simulation, we assume that at the 40th
second, the locations of 10 nodes are disclosed and the
hacker captures a part of these nodes so that in the network
there forms a “black hole” in which nodes can neither trans-
mit nor receive, while the attacker controls the rest of these
nodes to wirelessly interfere with their neighbors by using a
Fig. 7. Performance evaluation of RENUM in the random topology
(speed of node mobility = 1 m/s). (a) Aggregate end-to-end throughput
of RENUM under different number of hops. (b) End-to-end delay of
RENUM under different total injected traffic.
Fig. 8. Impact of node mobility on performance in the random topology.
(a) Data packet delivery ratio with different node moving speeds. (b)
Fairness index with different node moving speeds.
Fig. 9. Impact of sudden flow changes and cyberattacks on throughput in
the random topology (speed of node mobility = 1 m/s). (a) Effect of sud-
den flow changes on throughput (number of hops ¼ 3). (b) Evolution of
aggregate throughput when 10 nodes are attacked at 40th second (num-
ber of hops ¼ 1).

huge amount of power or sending a huge amount of rub-
bish messages to their neighbors so that in the network traf-
fic jamming occurs and the neighbor nodes are not able to
transmit and receive normal data packets. We can observe
from Fig. 9b that at the 40th second, the aggregate through-
put decreases sharply and then oscillates slightly at a low
level; After a short time, the throughput gradually increases,
then recovers up to the level before the attack and finally
becomes stable. The underlying reason is that both connec-
tions and communications between nodes in a MANET do
not rely on a preexisting infrastructure, and each node is an
equally privileged participant in routing by forwarding data
for other nodes. Even though some nodes are disabled due
to cyberattacks, the sources could rapidly adjust their for-
warding nodes and search for new routes to the destinations
by cooperation among nodes so as to avoid the degradation
of network performance. On the other hand, the fact that
data transmissions are decentralized among nodes and the
forwarding nodes are hidden among plenty of network com-
ponents reduces the probability that private information is
eavesdropped.
7 CONCLUSIONS
This paper has revisited a cross-layer design problem for
MANETs involving rate- and power-allocation. By consid-
ering the lossy feature of wireless links and a small degree
of toleration for fading-induced congestion, we formulate
a RENUM problem with the constraints of flow conserva-
tion, rate-outage probability and average delay, whose
objective is to maximize the effective rate related network
utility while minimizing the total power consumption and
average delay of all links. The non-convex RENUM prob-
lem is decomposed into three subproblems of the rate con-
trol at the transport layer and the power and average
delay control at the physical layer. Moreover, we propose
three corresponding distributed subalgorithms with
explicit message passing. In order to alleviate the explicit
message passing, we also develop three distributed subal-
gorithms having near-optimal performance with a conver-
gence rate order of magnitude faster than previous
schemes. Numerical examples illustrate that the proposed
RENUM outperforms other NUM such as JOPRC and
ENUM in terms of convergence, effective rates, link delay
and power consumption. Our NS-2-based network-wide
simulation has also confirmed the effectiveness of RENUM
algorithm and demonstrated that RENUM algorithm can
achieve efficient and fair resource allocation even for
nodes with high moving speed in ad hoc networks.
As a future work, we will implement and evaluate the
proposed RENUM algorithm in a real network testbed/sys-
tem instead of simulations, and further study how to inte-
grate the proposed RENUM framework with TCP-SACK so
as to improve the network performance since TCP-SACK
allows ACKs to carry information about which packet they
are acknowledging.
ACKNOWLEDGMENTS
The work in this paper was supported in part by the grants
from the Research Grants Council of Hong Kong, China
(No. CityU 112910), the National Natural Science Foun-
dation of China (61170248, 61373179, 61373178), New
Century Excellent Talents in University (No. NCET-10-
0877), Natural Science Foundation of Chongqing (CSTC,
cstcjjA40003), Science and Technology Leading Talent
Promotion Project of Chongqing (cstc2013kjrc-ljrccj40001),
Fundamental Research Funds for the Central Universities
(XDJK2013A018), and US National Science Foundation
(NSF) under grant number ECCS-1307576 and US Army
Research Office under grant number W911NF-09-1-0154.
REFERENCES
[1] C. Ma and Y. Yang, “A Battery-Aware Scheme for Routing in
Wireless Ad Hoc Networks,” IEEE Trans. Vehicular Technol.,
vol. 60, no. 8, pp. 3919-3932, 2011.
[2] C. Ma and Y. Yang, “Battery-Aware Routing for Streaming Data
Transmissions in Wireless Sensor Networks,” Mobile Netw. Appl.,
vol. 11, no. 5, pp. 757-767, 2006.
[3] X. Xiang, X. Wang and Y. Yang, “Stateless Multicasting in Mobile
Ad Hoc Networks,” IEEE Trans. Comput., vol. 59, no. 8, pp. 1076-
1090, 2010.
[4] J. Lee, R. Mazumdar, and N. Shroff, “Opportunistic Power Sched-
uling for Dynamic Multi-Server Wireless Systems,” IEEE Trans.
Wireless Comm., vol. 5, no. 6, June 2006.
[5] L. Zhang, Y.-C. Liang, and Y. Xin, “Joint Beamforming and
Power Allocation for Multiple Access Channels in Cognitive
Radio networks,” IEEE J. Selected Areas in Comm., vol. 26,
no. 1, Jan. 2008.
[6] J.T. Wang, “SINR Feedback-Based Integrated Base-Station Assign-
ment, Diversity, and Power Control for Wireless Networks,” IEEE
Trans. Vehicular Technol., vol. 59, no. 1, pp. 473-484, Jan. 2010.
[7] J. Gomez and A.T. Campbell, “Variable-Range Transmission
Power Control in Wireless Ad Hoc Networks,” IEEE Trans. Mobile
Comput., vol. 6, no. 1, pp. 87-99, Jan. 2007.
[8] B. Alawieh, C.M. Assi, and W. Ajib, “Distributed Correlative
Power Control Schemes for Mobile Ad Hoc Networks Using
Directional Antennas,” IEEE Trans. Vehicular Technol., vol. 57,
no. 3, pp. 1733-1744, May 2008.
[9] Q. Qu, L.B. Milstein, and D.R. Vaman, “Cross-Layer Distributed
Joint Power Control and Scheduling for Delay-Constrained Appli-
cations over CDMA-Based Wireless Ad-Hoc Networks,” IEEE
Trans. Comm., vol. 58, no. 2, pp. 669-680, Feb. 2010.
[10] D. Wang, X. Wang, and X. Cai, “Optimal Power Control for Multi-
User Relay Networks over Fading Channels,” IEEE Trans. Wireless
Comm., vol. 10, no. 1, pp. 199-207, Jan. 2011.
[11] M. Chiang, “Balancing Transport and Physical Layers in Wireless
Multihop Networks: Jointly Optimal Congestion Control and
Power Control,” IEEE J. Selected Areas in Comm., vol. 23, no. 1,
pp. 104-116, Jan. 2005.
[12] R. Jantti and S. Kim, “Joint Data Rate and Power Allocation for
Lifetime Maximization in Interference Limited Ad Hoc
Networks,” IEEE Trans. Wireless Comm., vol. 5, no. 5, pp. 1086-1094,
May 2006.
[13] J.-W. Lee, R.R. Mazumdar, and N.B. Shroff, “Joint Opportunistic
Power Scheduling and End-to-End Rate Control for Wireless Ad
Hoc Networks,” IEEE Trans. Vehicular Technol., vol. 56, no. 2,
pp. 801-809, Mar. 2007.
[14] P. Soldati, B. Johansson, and M. Johansson, “Distributed Cross-
Layer Coordination of Congestion Control and Resource Alloca-
tion in S-TDMA Wireless Networks,” Wireless Netw., vol. 14, no. 6,
pp. 949-965, Dec. 2008.
[15] S. Guo, C. Dang, and X. Liao, “Distributed Algorithms For
Resource Allocation of Physical and Transport Layers in Wireless
Cognitive Ad Hoc Networks,” Wireless Netw., vol. 17, no. 2,
pp. 33-356, Feb. 2011.
[16] K. Hedayati, I. Rubin, and A. Behzad, “Integrated Power Con-
trolled Rate Adaptation and Medium Access Control in Wireless
Mesh Networks,” IEEE Trans. Wireless Comm., vol. 9, no. 7,
pp. 2362-2370, July 2010.
[17] J. Papandriopoulos, S. Dey, and J. Evans, “Optimal and Distrib-
uted Protocols for Cross-Layer Design of Physical and Transport
Layers in MANETs,” IEEE-ACM Trans. Netw., vol. 16, no. 6,
pp. 1392-1405, Dec. 2008.

[18] S. Kandukuri and S. Boyd, “Optimal Power Control in Interfer-
ence-Limited Fading Wireless Channels with Outage-Probability
Specifications,” IEEE Trans. Wireless Comm., vol. 1, no. 1, pp. 46-
55, Jan. 2002.
[19] J. Papandriopoulos, J. Evans, and S. Dey, “Optimal Power Control
for Rayleigh-Faded Multiuser Systems with Outage Constraints,”
IEEE Trans. Wireless Comm., vol. 4, no. 6, pp. 2705-2715, Nov. 2005.
[20] J. Papandriopoulos, J. Evans, and S. Dey, “Outage-Based Optimal
Power control for Generalized Multiuser Fading Channels,” IEEE
Trans. Comm., vol. 54, no. 4, pp. 693-703, Apr. 2006.
[21] S.Y. Park and D.J. Love, “Outage Performance of Multi-Antenna
Multicasting for Wireless Networks,” IEEE Trans. Wireless Comm.,
vol. 8, no. 4, pp. 1996-2005, Apr 2009.
[22] C. Shao, X. Hua, and A. Huang, “Outage Probability Based
Resource Allocation in Wireless Mesh Networks,” Proc. IEEE
Global Telecomm. Conf., pp. 1-5, 2010.
[23] Q. Gao, J. Zhang, and S.V. Hanly, “Cross-Layer Rate Control in
Wireless Networks with Lossy Links: Leaky-Pipe Flow, Effective
Network Utility Maximization and Hop-by-Hop Algorithms,”
IEEE Trans. Wireless Comm., vol. 8, no. 6, pp. 3068-3076, June 2009.
[24] J.-W. Lee, M. Chiang, and A.R. Calderbank, “Jointly Optimal Con-
gestion and Contention Control Based on Network Utility Maxi-
mization,” IEEE Comm. Letters, vol. 10, no. 3, pp. 216-218, 2006.
[25] S. Chen, M. Beach, and J. McGeehan, “Division-Free Duplex for
Wireless Applications,” Electronics Letters, vol. 34, no. 2, pp. 147-
148, Jan 1998.
[26] G. Stuber, Principles of Mobile Communication. Kluwer, 1997.
[27] A. Goldsmith, Wireless Communication. Cambridge Univ.
Press, 2005.
[28] D.P. Bertsekas and R.G. Gallager, Data Networks. Prentice Hall,
1987.
[29] S. Boyd and L. Vandenbergh, Convex Optimization. Cambridge
Univ. Press, 2004.
[30] D.P. Bertsekas, Nonlinear Programming. second ed. ed., Athena
Scientific, 1999.
[31] W. Xu, Y. Wang, J. Chen, G. Baciu, and Y. Sun, “Dual Decomposi-
tion Method for Optimal and Fair Congestion Control in Ad Hoc
Networks: Algorithm, Implementation and Evaluation,” J. Parallel
and Distributed Computing, vol. 68, no. 7, pp. 997-1007, 2008.
[32] R. Yates, “A Framework for Uplink Power-Control in Cellular
Radio Systems,” IEEE J. Selected Areas in Comm., vol. 13, no. 7,
pp. 1341-1347, Sept 1995.
[33] T.S. Rappaport, Wireless Communication: Principles and Practice.
second ed. ed., Prentice-Hall, 2002.
[34] NS-2, Network Simulator, http://www.isi.edu/nsnam/ns/, 2012.
[35] D. Johnson and D. Maltz, Ad Hoc Networking. Addison-
Wesley, 2001.
Songtao Guo received the BS, MS, and PhD
degrees in computer software and theory from
Chongqing University, Chongqing, China, in
1999, 2003, and 2008, respectively. He was a
professor from 2011 to 2012 at Chongqing Uni-
versity. He is currently a full professor at South-
west University, China. He was a senior research
associate at the City University of Hong Kong
from 2010 to 2011, and a visiting scholar at Stony
Brook University, New York, from May 2011 to
May 2012. His research interests include wire-
less sensor networks, wireless ad hoc networks and parallel and distrib-
uted computing. He has published more than 30 scientific papers in
leading refereed journals and conferences. He has received many
research grants as a principal investigator from the National Science
Foundation of China and Chongqing and the Postdoctoral Science
Foundation of China. He is a member of the IEEE.
Changyin Dang received the PhD degree (cum
laude) in operations research/economics from Til-
burg University, The Netherlands, in 1991, the
MSc degree in applied mathematics from Xidian
University, China, in 1986, and the BSc degree in
computational mathematics from Shanxi Univer-
sity, China, in 1983. His research interests include
operations research/optimization and related
fields. Because of his significant contributions, he
has been named an excellent researcher by Til-
burg University, The Netherlands, in 1990 and
invited to give talks at universities such as Cornell University, Stanford
University, University of California, University of Michigan, University of
Minnesota, etc., and to work as a research fellow at Yale University in
1994. He has published more than 100 papers in journals including SIAM
Journal on Discrete Mathematics, SIAM Journal on Optimization, Artificial
Intelligence, IEEE Transactions on Automatic Control, IEEE Transactions
on Evolutionary Computation, etc. He is a senior member of the IEEE.
Yuanyuan Yang received the BEng and MS
degrees in computer science and engineering
from Tsinghua University, Beijing, China, and
the MSE and PhD degrees in computer sci-
ence from Johns Hopkins University, Balti-
more, Maryland. She is currently a professor
of computer engineering and computer sci-
ence at Stony Brook University, New York,
and the director of the Communications and
Devices Division at the New York State Center
of Excellence in Wireless and Information
Technology (CEWIT). Her research interests include wireless net-
works, data center networks, optical networks and high-speed net-
works. She has published 280 papers in major journals and refereed
conference proceedings and holds seven US patents in these areas.
She is currently an associate editor-in-chief for the IEEE Transac-
tions on Computers and an associate editor for the Journal of Paral-
lel and Distributed Computing. She has served as an associate
editor for the IEEE Transactions on Computers and IEEE Transac-
tions on Parallel and Distributed Systems. She has served as a gen-
eral chair, program chair, or vice chair for several major
conferences and a program committee member for numerous con-
ferences. She is a fellow of the IEEE.
For more information on this or any other computing topic,
please visit our Digital Library at www.computer.org/publications/dlib.

IEEE BE-BTECH NS2 PROJECT@ DREAMWEB TECHNO SOLUTION

More Related Content

What's hot (20)

Similar to IEEE BE-BTECH NS2 PROJECT@ DREAMWEB TECHNO SOLUTION (20)

Recently uploaded (20)

IEEE BE-BTECH NS2 PROJECT@ DREAMWEB TECHNO SOLUTION