186 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 13, NO.

1, JANUARY 2014
Joint Routing and Resource Allocation for
Delay Minimization in Cognitive Radio Based
Mesh Networks
Amr A. El-Sherif, Member, IEEE, and Amr Mohamed, Member, IEEE
AbstractThis paper studies the joint design of routing and
resource allocation algorithms in cognitive radio based wireless
mesh networks. The mesh nodes utilize cognitive overlay mode to
share the spectrum with primary users. Prior to each transmis-
sion, mesh nodes sense the wireless medium to identify available
spectrum resources. Depending on the primary user activities
and trafc characteristics, the available spectrum resources will
vary between mesh transmission attempts, posing a challenge
that the routing and resource allocation algorithms have to
deal with to guarantee timely delivery of the network trafc.
To capture the channel availability dynamics, the system is
analyzed from a queuing theory perspective, and the joint routing
and resource allocation problem is formulated as a non-linear
integer programming problem. The objective is to minimize the
aggregate end-to-end delay of all the network ows. A distributed
solution scheme is developed based on the Lagrangian dual
problem. Numerical results demonstrate the convergence of the
distributed solution procedure to the optimal solution, as well as
the performance gains compared to other design methods. It is
shown that the joint design scheme can accommodate double the
trafc load, or achieve half the delay compared to the disjoint
methods.
Index TermsDistributed resource optimization, cognitive
mesh networks, joint routing and channel allocation, delay
minimization.
I. INTRODUCTION
C
OGNITIVE radio is a promising technology aiming
at better utilization of available channel resources by
prescribing the coexistence of licensed (or primary) and un-
licensed (secondary or cognitive) radio nodes on the same
bandwidth [1]. One of the key challenges in the design of
cognitive radio networks is the design of dynamic spec-
trum allocation algorithms, which enable the cognitive nodes
to opportunistically access the available wireless spectrum,
without interfering with existing primary nodes. Therefore,
dynamic spectrum access techniques have received signicant
attention. In [2] and [3] the cognitive radio problem was
investigated from an information theoretic standpoint. The
cognitive transmitter is assumed to transmit at the same time
Manuscript received December 29, 2012; revised May 10 and August 18,
2013; accepted October 1, 2013. The associate editor coordinating the review
of this paper and approving it for publication was S. Bahk.
A. A. El-Sherif is with the Department of Electrical Engineering, Alexan-
dria University, Alexandria 21544, Egypt. This work was done while he was
with the Computer Science and Engineering Department, Qatar University,
Doha, Qatar (e-mail: amr.elsherif@ieee.org).
A. Mohamed is with the Computer Science and Engineering Department,
Qatar University, P.O. Box 2713, Doha, Qatar (e-mail: amrm@ieee.org).
Digital Object Identier 10.1109/TWC.2013.112513.122082
and on the same bandwidth of the primary link. Interference
is mitigated through the use of complex precoding techniques
that require perfect prior information about the primary signal.
The concept of a time-spectrum block was introduced in [4]
and protocols to allocate such blocks were proposed. In [5] the
authors derived optimal and suboptimal distributed strategies
for the secondary users to decide which channels to sense and
access under a Partially Observable Markov Decision Process
(POMDP) framework.
The cognitive radio concept is desirable for a wireless
mesh network (WMN) in which a large volume of trafc is
expected to be delivered since it is able to utilize spectrum
resources more efciently. Therefore, it improves network ca-
pacity signicantly. However, the dynamic nature of the radio
spectrum calls for the development of novel spectrum-aware
routing algorithms. In fact, spectrum occupancy is location-
dependent, therefore in a multi-hop path available spectrum
bands may be different at each node. Hence, controlling the
interaction between the routing and the spectrum management
functionalities is of fundamental importance. While cross-
layer design principles have been extensively studied by the
wireless networking research community, the availability of
cognitive and frequency agile devices motivates research on
new algorithms and models to study cross-layer interactions
that involve spectrum management-related functionalities.
Leveraging interactions between routing and spectrum al-
location has been recently investigated in the literature. In
[6], each source node nds candidate paths based on Dy-
namic Source Routing (DSR) and collects information on
link connectivity and quality. For each candidate route, the
algorithm nds all feasible spectrum assignment combinations
and estimates the end-to-end throughput performance for
each combination. Based on this computation, it selects the
route and spectrum assignment with maximal throughput and
schedules a conict-free channel for this route. The authors
in [7] follow a similar path by rst constructing a set of
possible routes and then nd the spectrum allocation policy
that maximizes the throughput. The work in [7] differs from
[6] in the way the routes are constructed. In [7] the routes are
constructed based on what the authors call route robustness,
which is a representation of how often a route would be
disrupted due to primary users activity. A routing and spectrum
selection algorithm for cognitive radio networks was proposed
in [8]. The algorithm chooses the path that has the highest
probability to satisfy the demands of secondary users in terms
1536-1276/14$31.00 c 2014 IEEE
EL-SHERIF and MOHAMED: JOINT ROUTING AND RESOURCE ALLOCATION FOR DELAY MINIMIZATION IN COGNITIVE RADIO BASED MESH . . . 187
2
n
1
n
3
n
9 . 0
9 . 0
6 0 6 . 0
Fig. 1. A 3 node network with primary channel idle probability shown on
the links.
of capacity. However, it does not cover the issue of scheduling.
In [9], a cross-layer optimization problem for a network with
cognitive radios is formulated. The objective is to minimize
the required network-wide radio spectrum resources needed to
support trafc for a given set of user sessions. The problem is
formulated as a mixed integer non-linear problem (MINLP),
and a sequential xing (SF) algorithm is developed where
the integer variables are determined iteratively via a sequence
of linear programs, which provide a near-optimal solution to
the original problem. Distributed implementations that build
over the AODV routing protocol and enhance it for cognitive
scenarios are proposed in [10] and [11]. In [10], a routing
metric aiming at achieving lower cumulative delay caused
by channel switching, queueing and collisions is proposed.
However, the algorithm assumes that each node has global
knowledge of the spectrum choices of the other nodes. The
target of [11] is to maximize the throughput through the design
of cross layer routing, spectrum access, and power control.
However, the data packets are routed along channels whose
qualities have not been assessed. Therefore, the robustness of
the routes is not guaranteed.
In this paper, we study the cross-layer joint design of
routing and resource allocation protocols in cognitive radio
based WMNs. Optimizing these two layers independently can
lead to sub-optimal solutions at best. For example, consider
the 3 node network depicted in Fig. 1, with n
1
as source
and n
3
as destination, and we need to choose between the
direct and indirect paths for routing packets. The numbers
on the links are the probability with which that link sees an
idle primary channel (i.e., a transmission opportunity). If the
routing protocol takes these probabilities as a routing metric,
then the indirect path will be chosen as it is the more reliable
path. However, in a half duplex network, node n
2
can either
be receiving from n
1
or transmitting to n
3
. Therefore, each
link on the indirect path can be active for 50% of the time.
And since the channel is available for 90% of the time, then
each link can only be active for 45% of the time. This means
that it is preferable to forward the packets through the direct
link, where it will be active for 60% of the time and resulting
in a higher throughput.
Many popular multimedia applications, e.g., voice over IP,
IPTV, and online gaming, have strict delay requirements. In
this paper, we aim at designing routing and resource allocation
protocols to minimize the end-to-end delay. Another unique
characteristic of our approach is the queuing theoretic analysis,
which is used to estimate the expected end-to-end delay across
different data streams. Each cognitive mesh node is modeled
as an innite length queue. A queuing theoretic analysis of the
network characterizes the arrival and service rates for each one
of these queues given the routes that pass through a node and
Primary coverage range
Secondary node
Primary node
Fig. 2. Network model.
the resources allocated to that node. The joint routing and
resource allocation design problem is then formulated as an
optimization problem having as objective the minimization of
the end-to-end delay, and having integer valued decision vari-
ables. As formulated, the optimization problem is a non-linear
integer programming (NIP) problem, which has combinatorial
complexity. It is shown that by relaxing the integer constraints
imposed on the decision variables, one can efciently nd a
solution to the relaxed optimization problem. Finally, using
the Lagrangian dual function, a distributed solution to the
optimization problem is presented. The performance of the
proposed protocol is thoroughly studied and compared to
the performance of a disjoint protocol. The disjoint protocol
solves the routing problem rst and then allocates resources
along the constructed routes. The routing metric used favors
links with higher primary idle probability while penalizing
the total number of hops. The resource allocation part aims at
minimizing the end-to-end delay along the preselected routes.
The rest of the paper is organized as follows. The different
aspects of the system model are presented in section II. In
section III the joint routing and resource allocation problem
is formulated as an integer programming optimization problem
and suboptimal approximate solution is provided. Section IV
presents an efcient distributed solution algorithm for the
routing and resource allocation problem. Numerical results are
presented and discussed in V. Finally, the paper is concluded
in section VI.
II. SYSTEM MODEL
A. Network Model
The cognitive radio based wireless mesh network model
used is depicted in Fig. 2. The mesh network consists of
cognitive nodes that opportunistically access one of N non-
overlapping primary channels of equal bandwidth. We assume
that the primary network follows a time-slotted transmission
structure. This assumption will simplify the analysis of our
cognitive network, however, our model and analysis could
be extended to incorporate different primary transmission
schemes.
Each cognitive mesh node will opportunistically access
idle primary channels to transmit its packets. Local chan-
nel availability can be detected using one of the different
spectrum sensing techniques available [12], [13], [14], [15].
The cognitive mesh network employs hybrid TDMA/FDMA
188 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 13, NO. 1, JANUARY 2014
for channel access. Therefore, time is divided into time slots
of xed duration, which are further grouped into frames of
T time slots each. In each time slot, a node selects one
of the available frequency channels to transmit over. Since
the primary network transmissions are also slotted, then it is
customary that the cognitive network adjusts the boundaries
of its time slots to match those of the primary network [16],
[17].
A time slot and a channel pair (t, c), is considered as the
minimum unit for resource allocation, we will call it resource
element. As dened, a resource element is similar to the
resource block concept in LTE systems. At the beginning
of each time slot t, a cognitive mesh node senses channel
c assigned to it in that time slot. If the channel is detected as
idle, the node transmits the packet at the head of its queue
to the next node along the route to its destination, otherwise
it remains silent and keeps sensing the channel in subsequent
time slots. For simplicity, we will assume that cognitive nodes
have access to perfect spectrum sensing information. After any
successful transmission, the receiving node acknowledges the
reception of the packet by transmitting an ACK packet back
to the transmitter.
The cognitive mesh network is modeled as a directed graph
G(V, E), where each vertex v V corresponds to a cognitive
mesh node. During any given time slot, there will be a set
C
v
of channels available to node v. An edge e E exists
between nodes u and v if there exists a channel c C
u

C
v
and the nodes are within transmission range of each other,
i.e., u v R, where u v is the Euclidean distance
between nodes u and v, and R is the transmission range. Since
the graph is directed, there will be two edges between nodes
u and v in this case, the rst has node u as a transmitter,
while the second has node v as transmitter. It is assumed that
all cognitive nodes use the same xed transmission power,
therefore, all nodes have the same transmission range R.
Due to primary nodes activity, channel availability will vary
with time. Therefore, network connectivity is time varying,
which poses a challenge to the routing and resource allocation
protocol design.
In this work, the effect of the wireless interference between
different nodes is modeled based on the protocol model [18],
i.e., simultaneous packet transmissions from interfering nodes
result in the loss of all involved packets. We say that two links
e
1
and e
2
interfere with each other if
1) there is a shared node between e
1
and e
2
(because of
half duplexing, unicast communications, or collisions)
or
2) any node from e
1
is within interference range I > R
from any node from e
2
, and they are using the same
channel.
Because of the ACK packets sent back to the transmitter,
both the links transmitter and receiver need to be free of
interference.
B. Channel Model
The wireless channel between a node and its destination is
modeled as a Rayleigh at fading channel with additive white
Gaussian noise [19]. Success and failure of packet reception is
Idle Busy
c
1 c

Fig. 3. Primary nodes Markov chain model.

characterized by outage events and outage probabilities. De-
tails of the channel model and outage probability calculation
can be found in [19] and [17].
C. Queuing Model
Each node in the cognitive mesh network has an innite
buffer for storing packets of xed length. The nite buffers
case could also be accommodated into our model with slight
modications to the optimization problem formulated in the
next section. The duration of a time slot is enough for the
transmission of a single packet (in addition to the sensing
time and ACK/NACK feedback). Multiple data connections or
streams are present in the network. For data stream f having
node u as source, packet arrivals at the source are modeled as
a stationary Bernoulli process with i.i.d arrivals from slot to
slot and mean
f
u
[20]. In other words, the probability that a
new packet arrives at any given time slot t is
f
u
. Moreover, the
packet arrival processes are assumed to be independent from
one data stream to another. Furthermore, each node maintains
a separate queue for each trafc ow that is routed through it.
The state (idle or busy) of any of the N primary channels is
modeled using a two state Markov chain as shown in Fig. 3.
Using the stationary distribution of the Markov chain, at any
given time slot channel c will be idle (Markov chain in the
off state) with probability
c
. The evolution of any channel is
independent from all other channels.
The main model assumptions can be summarized as follows:
Primary and cognitive networks both employ a time
slotted transmission structure
Cognitive nodes have access to perfect spectrum sensing
information
All cognitive nodes use the same xed transmission
power
In a given time slot, a cognitive node can use at most
one channel for packet transmission
III. JOINT ROUTING AND RESOURCE ALLOCATION
STRATEGY
The main objective in this work is to nd the best routing
and resource allocation strategies in order to minimize the
average end-to-end delay of multiple data connections in the
cognitive radio based wireless mesh network. Because of
the primary nodes activity, the spectrum resources available
to the cognitive mesh nodes are varying in both space and
time. Therefore, any successful routing strategy will have to
work closely with the resource allocation strategy in order to
make sure that any selected route will have enough resources
available to guarantee the required QoS. Because of this strong
interdependence between the routing and resource allocation
strategies, we propose to deal with the routing and resource
EL-SHERIF and MOHAMED: JOINT ROUTING AND RESOURCE ALLOCATION FOR DELAY MINIMIZATION IN COGNITIVE RADIO BASED MESH . . . 189
allocation strategies in a joint fashion rather than separating
the two problems.
Before presenting joint design strategy we need rst to
analyze the effect of the routing and resource allocation
decisions on the network performance. This is achieved by
relying on queueing theory to model the different aspects of
the cognitive mesh network and to form a basis for our routing
and resource allocation protocol design.
A. Queueing Analysis
We start by calculating the average arrival and service rates
at the different nodes in the cognitive mesh network. For this
calculation we need to rst introduce the decision variables
that dene the different routes in the network as well as the
resources allocated to the different links along these routes.
We dene two sets of decision variables:
x
f,e
=
_
1, if link e is routing packets for stream f
0, otherwise
y
t,c
f,e
=
_
_
_
1, if resource element (t, c) is allocated to link e
and stream f
0, otherwise
It is worth mentioning that the same resource allocation pattern
is repeated every TDMA frame, and that the system is assumed
to be stationary.
To dene the average arrival rate at any node v along the
route of data stream f we start by identifying the events
necessary for a packet arrival to take place. A packet from
data stream f enters the queue of node v in a given time slot
t if:
1) there is a resource element (t, c) allocated to one of vs
incoming edges,
2) the primary node owning channel c is either idle during
that time slot or the cognitive node v is out of the
primary nodes interference range,
3) the preceding cognitive node in the route has at least
one packet in its queue to transmit to node v.
Therefore, the probability of a packet arrival at node v
along the route of the data stream f is the joint probability
of these three events. Since these events are independent, this
probability can be written as
a
t
f,v
=

eE
in
v
N

c=1
y
t,c
f,e

f
e(s)

f
e(s)
[I
c
e

c
+ (1 I
c
e
)] (1 P
out
e
),
(1)
where E
in
v
is the set of incoming edges to node v, e(s) is
the source node for edge (link) e,
f
e(s)
is its arrival rate and

f
e(s)
its service rate. By modeling each queue as discrete time
Markov chain, it can be shown that the fraction
f
es
/
f
es
is
the probability that the queue has at least one packet [21], and
hence will transmit a packet to the following node on the route
whenever it has a chance.
c
is the probability that channel c is
idle, P
out
e
is the outage probability between the transmitter and
receiver of link e. Finally, I
c
e
= 1 if the primary node owning
channel c interferes with transmissions over link e, otherwise
I
c
e
= 0. Note that for simplicity we assume perfect sensing at
all cognitive mesh nodes. Interference from other cognitive
nodes is not reected in (1), since the resource allocation
scheme discussed below makes sure that no two interfering
links can share the same resources.
It is clear that the packet arrival probability can vary from
one time slot to another within the same frame since assigned
channels can vary between time slots. This dependence is
emphasized in (1) by the use of the superscript t (which is
the index of the time slot in a TDMA frame). Since the same
resource allocation pattern is repeated each TDMA frame, a
given time slot within any frame will always have the same
packet arrival probability. For the purpose of mathematical
tractability, instead of using different packet arrival probabili-
ties for the different time slots in a TDMA frame, we calculate
the average packet arrival probability per time slot and use it
for all the time slots. The average packet arrival probability
can be calculated as,

f
v
=
1
T
T

t=1
a
t
f,v
, (2)
which is interpreted as the probability that a packet from data
stream f arrives at node v in any time slot. It should be noted
that, using the average arrival rate for all time slots instead
of the time slot dependent arrival rates does not affect the
average behavior of the queues (which is what the following
analysis is based on). In other words, the rate at which queues
are lled (or emptied) as well as the probability that a queue
becomes empty are the same in the two cases. Given this
denition, packet arrivals can be approximated as Bernoulli
trials at each time slot with average success probability
f
v
.
Therefore, the packet arrival process can be approximated as
a Bernoulli process with average arrival rate
f
v
.
Similarly, to calculate the average service rate at any node
v along the route of data stream f, we start by identifying the
events necessary for a successful packet transmission. This
will take place if in a given time slot t
1) there is a resource element (t, c) assigned to one of vs
outgoing edges,
2) the primary node owning channel c is either idle during
that time slot or cognitive node v is out of its interfer-
ence range.
Therefore, the probability of a packet being serviced from
node v along the route of data stream f is dened as the joint
probability of these two events, which is given by
d
t
f,v
=

eE
out
v
N

c=1
y
t,c
f,e
[I
c
e

c
+ (1 I
c
e
)] (1 P
out
e
), (3)
where E
out
v
is the set of outgoing edges to node v.
Similar to the arrival probabilities, the service probabilities
vary between time slots. Here also we resort to using the
average service probability per time slot, calculated as

f
v
=
1
T
T

t=1
d
t
f,v
, (4)
which is the probability that a packet from data stream f leaves
the queue of node v in any time slot. Similar to the arrival
events, the service events can be seen as Bernoulli trials at
each time slot with average success probability
f
v
. Therefore,
190 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 13, NO. 1, JANUARY 2014
the packet service process can be approximated as a Bernoulli
process with average service rate
f
v
.
To further simplify the analysis and protocol implementa-
tion in the next sections, we resort to the following queueing
theory result. Given a stationary and stable queue (its arrival
rate is less than its service rate), the average number of
customers leaving the queue per unit time is equal to the
average number of customers entering the queue per unit time
[22]. In other words, the average departure rate is equal to
the average arrival rate. It is noted that the average departure
rate is not to be confused with the average service rate (the
service rate is the reciprocal of the time a customer spends at
the server before departing). This result can be easily proved
by analyzing the underlying Markov chain model of the queue.
Applying this result to our system, it is noted that the rate
at which packets depart a given queue is equal to that queues
arrival rate. Therefore, in a multi-hop system with single path
routing, all the nodes along that path will have the same arrival
rate as the source node, since the departures from a node are
the arrivals to the next node. This observation will allow us
to deal with the arrival rates at different nodes as constant
values.
B. End-to-End Delay
Having calculated the average arrival and service rates for
each node in the network, we can now quantify the delay
incurred by routing data through any given node.
As discussed above, the arrival and services processes to
and from any node are Bernoulli processes with average rates

f
v
and
f
v
, respectively. A queueing system with such arrival
and departure processes is a discrete time M/M/1 queue [22].
Therefore, given the average arrival and service rates, the
average delay incurred by data stream f packets when passing
through cognitive node v is given by [22],
D
f
v
=
1
f
v

f
v

f
v
. (5)
The end-to-end delay is the accumulation of the incurred
delay by a packet when it goes through the different nodes
along its route from source to destination. Therefore, the
average end-to-end delay for the packets of data ow f is
given by D
f
=

vV
f
D
f
v
, where V
f
is the set of nodes
forming the route for the packets of data stream f.
Using the fact that the arrival rates can be considered as
constants as observed above, it can be easily shown that the
delay is a convex function of the service rates
f
v
.
To validate our analytical model, we simulate a 4 node
network with two trafc streams. Fig. 4 shows a comparison
between the average delay based on our analytical model and
that based on Monte Carlo simulation. The results show that
the delay calculated based on the analytical model is within
10% to 15% from that of the simulation.
C. Optimization Problem Formulation
Here the joint routing and resource allocation design
problem is formulated as a non-linear integer programming
problem. It can be noted from the arrival and service rates
calculations above, and the problem constraints presented next,
0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0
2
4
6
8
10
12
14
16
Avg. sec. arrival rate
A
v
g
.

d
e
l
a
y

(
t
i
m
e

s
l
o
t
s
)


Stream 1 Analytic
Stream 2 Analytic
Stream 1 Simulation
Stream 2 Simulation
Fig. 4. Comparison between analytical and simulation models.
that the routing and resource allocation decisions are based on
the channel idle probabilities, the channel outage probabilities,
and the possible interference between nodes.
The optimization problem is formulated as follows,
min

fF

vV
f
D
f
v
(6)
subject to:

eE
out
f(s)
x
f,e
= 1, f F; (7)

eE
out
v
x
f,e
=

eE
in
v
x
f,e
, v V \{f(s), f(d)}, f F;
(8)
y
t,c
f,e
x
f,e
, e E, t [1, T], c [1, N]; (9)

fF

Le
y
t,c
f,e
1, e E, t [1, T], c [1, N]; (10)
N

c=1
y
t,c
f,e
1, e E, t [1, T], f F; (11)

fF

eE
in
v
N

c=1
y
t,c
f,e
+

fF

eE
out
v
N

c=1
y
t,c
f,e
1,
t [1, T], v V ; (12)

f
v
<
f
v
, v V, f F; (13)
x
f,e
{0, 1}, e E, f F; (14)
y
t,c
f,e
{0, 1}, e E, f F, t [1, T], c [1, N]. (15)
The objective function in (6) is set to minimize the total
average end-to-end delay for all the data streams, where F
is set of all data streams. In this formulation all data streams
are treated equally. However, service differentiation can be
incorporated through the minimization of a weighted sum
of the average end-to-end delays. Constraints (7) and (8)
guarantee that for each data stream a single path from its
source to its destination is selected. Where f(s) and f(d)
denotes the source and destinations nodes for data stream f,
respectively. The single path routing is ensured by allowing
only one active link out from the streams source node in (7),
and at any intermediate node along the route the number of
EL-SHERIF and MOHAMED: JOINT ROUTING AND RESOURCE ALLOCATION FOR DELAY MINIMIZATION IN COGNITIVE RADIO BASED MESH . . . 191
active incoming links equals the number of active outgoing
links, in constraint (8). Since the source node has a single
active outgoing link, then intermediate nodes will have a single
active incoming link and a single active outgoing link. Note
that a given node can be used to route packets for different data
streams over different links. Constraint (9) establishes the link
between the two sets of decision variables, and restricts the
allocation of resources to the links taking part in the selected
routes.
In (10) L
e
denotes the set of links interfering with link e as
per the interference conditions discussed in the previous sec-
tion. This constraint ensures that interfering links are allocated
distinct resource elements, which avoids interference between
cognitive mesh nodes during packet transmissions. Constraint
(11) ensures that a node is assigned no more than one channel
per time slot. This is assuming the available radios can only
access a single channel at any given time. This constraint can
be modied to accommodate multi-channel radios, as well as
different capabilities for different nodes. Constraint (12) is a
half duplex constraint, making sure a node cannot transit and
receive simultaneously. Constraint (13) guarantees the stability
of all the queues in the network. Unstable queues result in
unbounded delays. Finally, constraints (14) and (15) are binary
value constraints ensuring that the decision variable can only
take a value of 0 or 1.
It is noted that the average service rates at each node are
linear in the decision variables as given by (3) and (4), as well
as all the constraints. The objective function is convex in
f
v
,
thus convex in the decision variables as well. Therefore, the
optimization problem is a convex problem.
Despite the problem having a convex objective and linear
constraints, the integer-valued constraints on the decision
variables renders the problem extremely complex to solve. In
order to reduce the complexity of the problem, a suboptimal
relaxation will be presented in the next section that will allow
the problem to be efciently solved.
D. Suboptimal Relaxation
To reduce the high complexity of the nonlinear integer
programming problem presented in the previous section we
propose to relax the binary value constraints (14) and (15) and
allow the decision variables to take any value in the interval
[0, 1]. The resulting optimization problem has the same form
as in (6) to (13), with (14) and (15) rewritten as follows
x
f,e
[0, 1], e E, f F; (16)
y
t,c
f,e
[0, 1], e E, f F, t [1, T], c [1, N].
(17)
This relaxation transforms the nonlinear integer program
into a convex optimization problem with real valued variables
for which efcient solution algorithms exist [23]. However,
the resulting optimal solution for the relaxed problem is
not guaranteed to be optimal for the original integer-valued
problem.
In the relaxed problem, the routing variables x
f,e
take
values in the range [0, 1]. Therefore, the constraints in (7) and
(8) can no longer guarantee single path routes between the
source and destination of the different data streams. Multiple
1
n
4
n
2
n
3
n
6 . 0
1
,

e f
x
4 . 0
2
,

e f
x
3 . 0
4
,

e f
x
3 . 0
3
,

e f
x
7 . 0
5
,

e f
x
Fig. 5. A 4 node network illustrating the different routes resulting from
solving the relaxed optimization problem.
paths between the source and destination of a given data
stream will be constructed, and resource elements will be
assigned to the links along these paths.
One can interpret the resulting multi-path routes as if they
result from the solution of a stochastic routing problem.
Given this interpretation, the fractional values of the decision
variables x
f,e
dene the probability with which a packet
belonging to data stream f is routed through link e. For
example, assume the source node for data stream f has 3
outgoing links, e
1
, e
2
, and e
3
, and the corresponding decision
variables take the values x
f,e1
= 0.3, x
f,e2
= 0.7, and
x
f,e3
= 0.0. This means that 30% of the packets are routed
through link e
1
, 70% of the packets are routed through link
e
2
, and link e
3
is not used to route packets for data stream f.
This interpretation also applies to all intermediate links along
the different paths between the source and the destination.
Similarly, the resource allocation variables y
t,c
f,e
take values
in the range [0, 1]. Solving this optimization problem results
in fractional allocation of resource elements to the different
links in the network. In other words, a resource element can
no longer be seen as the minimum unit for resource allocation
as discussed earlier. Because of the real-valued nature of y
t,c
f,e
,
fractions of a given resource element could be allocated to
different links in the network.
Since the initial design goals were for single path routing
and deterministic resource allocation along those routes, we
have to nd an algorithm to transform the resulting stochastic
routing and resource allocation solutions into the required
deterministic ones. For this transformation to take place, the
decision variables x
f,e
and y
t,c
f,e
have to be converted back
into binary valued variables. In the course of this conversion
process, it is crucial not to violate any of the original problem
constraints.
Algorithm 1 describes our proposed scheme for real-valued
to binary-valued (RV-BV) solution conversion while obeying
the constraints imposed by the optimization problem.
Fig. 5 illustrates a mesh network with 4 nodes (n
1
though
n
4
) and 5 edges (e
1
through e
5
). The network has a single data
stream having node 1 as source and node 4 as destination.
Assume that after solving the relaxed optimization problem
the resulting routing variables x
f,e
are as shown in the gure.
We note that although edge e
5
is the only edge coming out
of node n
3
, x
f,e5
= 0.7 and not 1.0 because constraint (8)
forces x
f,e5
= x
f,e2
+x
f,e4
.
192 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 13, NO. 1, JANUARY 2014
Algorithm 1 Real-Valued to Binary-Valued (RV-BV) conver-
sion
1: for f F do
2: Using the routing variables x
f,e
, construct all possible
paths for data stream f
3: For the paths identied in step 2, calculate the proba-
bility of routing packets through each path
4: Set x
f,e
= 1, e K
e
, where K
e
is the set of links in
the path with highest probability
5: Set x
f,e
= 0, e K
e
6: Set y
t,c
f,e
= 0, e K
e
, t [1, T], c [1, N]
7: end for
8: Dene the set Z =
_
(f, e, t, c) : y
t,c
f,e
> 0
_
9: (f

, e

, t

, c

) = arg max
(f,e,t,c)Z
y
t,c
f,e
.
10: Set y
t

,c

,e
= 1
11: Dene the set Q(f

, e

, t

, c

) = {(t, c, f, e) Z :
given y
t

,c

,e

= 1, and any of the constraints (10), (11),

(12) is violated }
12: Set y
t,c
f,e
= 0, (f, e, t, c) Q
13: Update Z = Z\ {Q

(f

, e

, t

, c

)}
14: if Z = then
15: Goto step 9
16: else
17: End
18: end if
The rst step in our conversion scheme is to use the routing
variables x
f,e
to construct all the possible paths through which
the packets are routed. From the values x
f,e
, it can be seen
that there exists 3 possible paths; Path
1
= e
1
e
3
, Path
2
=
e
2
e
5
, and Path
3
= e
1
e
4
e
5
. If for example x
f,e4
was equal to 0 instead of 0.4, then Path
3
would not have
existed because now one of its links is never used.
Once all the possible paths for the given data stream are
identied, the next step is to calculate the probability of rout-
ing packets through each of these paths. Let the probability for
using Path
i
equals p
i
. At node n
1
packets are routed though
links e
1
and e
2
with probabilities 0.6 and 0.4, respectively.
Since link e
1
is the rst link in both Path
1
and Path
3
, we
set the probabilities p
1
= p
3
= 0.6. Link e
2
is the rst link
in Path
2
, so we set p
2
= 0.4. At node n
2
, packets are routed
through links e
3
and e
4
. The probability of using each of
these links is calculated by normalizing the corresponding
x
f,e
so that the sum over all outgoing edges adds up to 1,
thus, representing a probability. Therefore, the probability of
routing packets through e
3
is calculated as
x
f,e
3
x
f,e
3
+x
f,e
4
. Since
e
3
and e
4
have the same weights, packets are divided equally
between them. So we update Path
1
and Path
3
probabilities
to p
1
= p
3
= 0.6 0.5 = 0.3. Finally, since node n
3
has a
single outgoing link e
5
, all packets will go through this link
with probability 1 despite the fact that x
f,e5
= 0.7. The nal
paths probabilities are p
1
= 0.3, p
2
= 0.4, and p
3
= 0.3,
note that the sum of the probabilities should sum up to 1
for each data stream in the network. Since Path
2
has the
highest probability, it is selected for packet routing and all
other paths are deleted. So we set x
f,e2
= x
f,e5
= 1, and
x
f,e1
= x
f,e3
= x
f,e4
= 0.
After selecting a single path for routing packets we release
all the resources that were allocated to any link not on that
path by setting y
t,c
f,e
= 0 for those links. For the links on
the selected path we need to convert the stochastic resource
allocation into a deterministic one. First we dene the set Z of
all resource elements that are assigned to any of the network
links. Then we nd the highest assignment probability and
set the corresponding decision variable to y
t

,c

,e

= 1. Given
the constraints (10), (11), and (12) we dene a new set Q
of all the resource elements with assignments that conict
with the above assignment. We then release all the conicting
assignments, and update the set Z by removing the element
(f

, e

, t

, c

) and all elements in Q. These steps are repeated

till all the elements are removed from the set Z.
IV. DISTRIBUTED SOLUTION
The nonlinear integer programming problem along with its
relaxed approximation and the conversion scheme described
above are all centralized in nature. They require global infor-
mation about the network to be present at a central point to
be able to nd a solution. In a wireless mesh network, each
node has local information about its environment. These local
information need to be wirelessly communicated to the central
point from all the nodes in the network. In many cases such
communication overhead is not practical, especially in net-
works with a large number of nodes. Therefore, a distributed
and scalable solution scheme in which calculations are done
locally at each node, or at local central points or cluster heads
is desirable. In this section, we propose a decomposition of
the original problem into smaller subproblems that can be
efciently solved in a distributed fashion.
The distributed solution approach is based on dual decom-
position. The rst step is to form the Lagrangian by relaxing
the constraint in (9) which forms the link between the routing
decision variables x
f,e
and the resource allocation decision
variables y
t,c
f,e
,
L =

fF

vV
f
D
f
v
+

fF

eE
T

t=1
N

c=1
p
t,c
f,e
(y
t,c
f,e
x
f,e
), (18)
where p
t,c
f,e
are the Lagrange multipliers.
We note that in the Lagrangian (18) the routing (x
f,e
)
and resource allocation (y
t,c
f,e
) decision variables are separa-
ble. Therefore, the the Lagrangian minimization problem can
be decoupled into two disjoint subproblems, each of which
depends only on a single set of decision variables, x
f,e
or
y
t,c
f,e
. The two disjoint subproblems are:
1) The routing subproblem:
max

fF

eE
x
f,e
T

t=1
N

c=1
p
t,c
f,e
, (19)
subject to constraints (7), (8), and (14).
2) The resource allocation subproblem:
min

fF

vV
f
D
f
v
+

fF

eE
T

t=1
N

c=1
p
t,c
f,e
y
t,c
f,e
, (20)
subject to constraints (10), (11), (12), (13), and (15).
EL-SHERIF and MOHAMED: JOINT ROUTING AND RESOURCE ALLOCATION FOR DELAY MINIMIZATION IN COGNITIVE RADIO BASED MESH . . . 193
Thus, the global end-to-end delay minimization problem de-
composes into a routing subproblem and a resource allocation
subproblem. The resource allocation subproblem distributes
the available channel resources among the different data
streams and network links such the end-to-end delay for each
data stream is minimized, and the stability of the queues of the
network nodes taking part in packet forwarding is guaranteed.
Solution of the resource allocation subproblem may result in
multiple paths for packet routing between a source node and
its destination. The routing subproblem then ensures that the
optimal single path is selected for packet routing between each
source and its destination.
The dual variables p
t,c
f,e
play a key role in coordinating
between the solution of the routing and resource allocation
subproblems. For instance, from the resource allocation point
of view the value p
t,c
f,e
can be interpreted as the price associated
with using a given resource element (t, c) to forward packets
for trafc stream f over link e. It is worth mentioning here
that the same resource element can have different prices when
assigned to different links, or when assigned for different
trafc streams. Therefore, the resource allocation subproblem
tries to minimize the end-to-end delay while utilizing a set
of resources that have a low price. From the routing point of
view, the value

T
t=1

N
c=1
p
t,c
f,e
which appears in (19) can
be interpreted as routing metric associated with link e when it
is used to forward packets for trafc stream f. Similar to the
resource allocation subproblem, each link may have different
metrics when used for forwarding different trafc stream
packets. Therefore, the routing subproblem as formulated in
(19) selects the single path with the largest overall metric.
Because of the different ways the two subproblems use
to deal with the Lagrange multipliers, a mismatch between
the selected path by the routing subproblem and where the
resources are allocated occur. This mismatch necessitates an
update in the value of the Lagrange multipliers as discussed in
the next section. During this update, the price of the resources
along the selected path will decrease, while the other resources
will encounter a price increase. Given the new values for the
prices/routing metrics, more resources may be allocated to
the path selected by the routing subproblem, or a switch to a
new path to which enough resources are allocated may take
place. The dynamics of the Lagrange multipliers updates will
be discussed in details in the results section.
A. Primal-Dual Solution Framework
The convexity of the original optimization problem in (6)
enables us to use the duality theory [24] to solve the problem
via its dual. Algorithm 2 presents the primal-dual algorithm
to solve the optimization problem.
Convexity of the constraints guarantees that the update in
(21) is a subgradient for the dual variables. Thus, as long
as the step sizes are chosen appropriately (e.g., as a square
summable but not summable sequence), the dual update even-
tually converges [25]. A square summable but not summable
sequence satises
(i) 0,
inf

i=1
(i)
2
< inf,
inf

i=1
(i) = inf .
Algorithm 2 Joint Cognitive Routing and Resource Allocation
(JCRA)
1: Set i = 0. Initialize p
t,c
f,e
(i).
2: In the primal domain, solve subproblems (19) and (20).
3: In the dual domain, update the dual variables
p
t,c
f,e
(i) =
_
p
t,c
f,e
(i) +
t,c
f,e
(i)(y
t,c
f,e
x
f,e
)
_
+
, (21)
where []
+
denotes max(0,), and
t,c
f,e
(i) is the step size.
4: Set i = i + 1. Return to step 2 until convergence.
5: Use The RV-BV algorithm to convert to a binary-valued
solution.
One typical example is (i) = a/(b + i), where a > 0 and
b > 0 [25].
The primal-dual algorithm described above allows each sub-
problem to be solved independently. In the next section we will
show how the routing and resource allocation subproblems can
be solved efciently. The described solutions require that each
node identies all its neighboring nodes. This can be achieved
through the use of a simple HELLO protocol for neighbors
discovery [26]. It is noted that, since the network model
studied is static with no mobility or changing topology, this
neighbor discovery need not be performed at regular intervals.
Therefore, the associated overhead is negligible.
B. Routing Subproblem
In the routing subproblem, each edge in the graph modeling
the mesh network has an associated metric or weight. A
given edge may have different metrics or weights for different
data streams. As formulated in (19) the routing subproblems
searches for a single path between each source and its des-
tination such that the sum of the metrics of the edges along
each path is maximized. Formulation in (19) and its associated
constraints is on the form of a linear program which have
efcient solution methods.
In the form presented above, this problem is a classical
problem in graph theory for which there exist several ef-
cient centralized (e.g., Dijkstra and Bellman-Ford algorithms)
and decentralized (e.g., Distributed Bellman-Ford algorithm)
solution algorithms [21]. Given the weights p
t,c
f,e
associated
with each link in the network any of these algorithms can be
employed to solve the routing subproblem in a decentralized
fashion. We note that to start constructing the routes each
node needs only to know the weights p
t,c
f,e
associated with its
outgoing links. Furthermore, since these solution methods are
designed to return a single path routes, the solution provided is
already an integer valued solution, which eliminates the need
for the conversion step.
C. Resource Allocation Subproblem
A decentralized solution to the resource allocation subprob-
lem can be obtained using a primal-dual framework similar
to the one we used to decouple the original problem. First,
we start by forming the Lagrangian for the optimization
problem (20). We note that the rst constraint is responsible
194 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 13, NO. 1, JANUARY 2014
for preventing interfering edges from sharing the same re-
source element. Therefore, this constraint affects the resource
allocation at different nodes simultaneously. The effect of
the remaining constraints is localized at each node, and does
not affect the resource allocation at other nodes. Using this
remark, we form the Lagrangian by combining the objective
function and the rst constraint as follows,
L =

fF

vV
f
D
f
v
+

fF

eE
T

t=1
N

c=1
p
t,c
f,e
y
t,c
f,e
+

eE
T

t=1
N

c=1
l
t,c
e
_
_

fF

Le
y
t,c
f,e

1
_
_
, (22)
where l
t,c
e
are the new Lagrange multipliers. If we change
the order of the summations in the rst and second terms,
and noting that the sum over all the edges in the second and
third terms can be written as a nested summation rst over
the nodes and then over the outgoing edges from each node,
we can then rewrite (22) as
L =

vV
_
_

fF
D
f
v
+

eE
out
v

fF
T

t=1
N

c=1
p
t,c
f,e
y
t,c
f,e
+

eE
out
v
T

t=1
N

c=1
l
t,c
e
_
_

fF

Le
y
t,c
f,e

1
_
_
_
_
. (23)
From (23), it can be seen that the resource allocation
subproblem can be divided into |V | smaller subproblems,
which are solved locally at each node. In order to solve these
subproblems, each node needs to know the resource allocation
solutions for all nodes within its interference range. In other
words, nodes within the same contention domain [26] need
to share their solutions at each iteration. Moreover, each node
needs to know the values of the Lagrange multipliers associ-
ated with its resources, which are updated locally according
to,
l
t,c
e
(i) =
_
_
l
t,c
e
(i) +
t,c
e
(i)
_
_

fF

Le
y
t,c
f,e

1
_
_
_
_
+
, (24)
where []
+
denotes max(0,), and
t,c
e
(i) is the step size.
Here also, the convexity of the constraints guarantees that
the update in (24) is a subgradient for the dual variables. Thus,
as long as the step sizes are chosen appropriately (e.g., as a
square summable but not summable sequence), the dual update
eventually converges.
V. RESULTS AND DISCUSSIONS
We consider a mesh network with 6 cognitive nodes uni-
formly distributed in a 200m x 200m square region. The
transmission range of any node is set to R = 100m, and
the interference range I = 2R. A TDMA frame has T = 12
time slots. Channel parameters used are: transmission power
P = 100mW, SNR threshold = 20dB, path loss exponent
= 3.7, and noise power spectral density N
0
= 10
11
W/Hz. To average out the effect of random channels, random
deployment of primary nodes, and random selection of the
source and destination of the different trafc streams, each
0 10 20 30 40 50
0
5
10
15
20
25
30
35
40
45
Iterations
D
e
l
a
y

(
t
i
m
e

s
l
o
t
s
)
Fig. 6. Average end-to-end delay convergence for a single data stream.
0 50 100 150
0
2
4
6
8
10
12
14
16
18
20
Iterations
D
e
l
a
y

(
t
i
m
e

s
l
o
t
s
)
Fig. 7. Effect of dynamic load on end-to-end delay per data stream.
scenario is repeated 25 times and the results averaged over
these runs. It is also noted that, when there is multiple trafc
streams in the network, these streams are between different
source-destination pairs. Results of the proposed optimized
algorithm are compared with a simple disjoint protocol that
solves the routing problem rst and then allocates resources
along the constructed routes. The routing metric used favors
links with higher primary idle probability while penalizing
the total number of hops. The resource allocation part aims
at minimizing the end-to-end delay along the preselected
routes. Moreover, the results are compared with the Delay
motivated On-demand Routing Protocol (DORP) presented in
[10]. DORP builds over the well-known on-demand routing
protocol (AODV) [27] and denes a routing metric based on
the channel switching delay and the service time of the packet
at the head of each nodes queue. However, it does not take
the effect of queueing delay into consideration.
First, we study the convergence behavior of our primal-dual
distributed solution framework. The mesh network considered
has a fully connected topology, there is a single trafc stream
and a single primary node, hence a single primary channel.
Fig. 6 depicts the evolution of the average end-to-end delay
with time. It is noted that the system converges to the same
solution as the centralized framework in less than 25 iterations.
We then study the impact of varying loads on the network
EL-SHERIF and MOHAMED: JOINT ROUTING AND RESOURCE ALLOCATION FOR DELAY MINIMIZATION IN COGNITIVE RADIO BASED MESH . . . 195
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35
0
10
20
30
40
50
60
70
80
Sec. packet arrival rate
A
v
g
.

E
n
d

t
o

E
n
d

D
e
l
a
y

(
t
i
m
e

s
l
o
t
s
)


1 stream Opt.
3 streams Opt.
1 stream disjoint
3 streams disjoint
1 stream DORP
3 streams DORP
Fig. 8. Average end-to-end delay as a function of arrival rate.
1 1.5 2 2.5 3 3.5 4 4.5 5
10
15
20
25
30
35
40
No. of channels
A
v
g
.

E
n
d

t
o

e
n
d

D
e
l
a
y

(
t
i
m
e

s
l
o
t
s
)


1 stream Opt.
3 streams Opt.
1 stream disjoint
3 streams disjoint
1 stream DORP
3 streams DORP
Fig. 9. Average end-to-end delay as a function of the number of primary
channels.
performance in Fig. 7. The network starts with a single data
stream. After 50 iterations, a second data stream is admitted
into the network. For this second stream the algorithm con-
verges in 25 iterations. We note that the end-to-end delay per
data stream has increased because the available resources are
now shared between the two data streams. After another 50
iterations, one of the data streams leaves the network. In this
case, convergence is achieved in less than 5 iterations, where
the freed up resources are allocated to the remaining stream
and we see a decrease in its delay.
Fig. 8 depicts the average end-to-end delay per data stream
as a function of the arrival rate. We consider N = 2 channels
and primary idle probability
c
= 0.8 for each channel.
The performance gain of the proposed protocol is clear. For
example, with a single data stream and for a secondary packet
arrival rate = 0.25, we can see that the proposed design
strategies can achieve 67 % decrease in delay compared to
the disjoint design and 31 % decrease compared to the DORP
protocol. Furthermore, when the number of trafc streams in
the network is increased to 3, our proposed protocol achieves
a 23 % reduction in delay compared to the DORP protocol,
where the disjoint protocol can only support 3 trafc streams
up to a secondary packet arrival rate less than 0.15 packets
per time slot.
The effect of the number of primary channels is shown in
0.6 0.65 0.7 0.75 0.8 0.85 0.9
0
20
40
60
80
100
120
140
160
180
200
Pri. idle prob.
A
v
g
.

E
n
d

t
o

E
n
d

D
e
l
a
y

(
t
i
m
e

s
l
o
t
s
)


1 stream Opt.
3 streams Opt.
1 stream disjoint
3 streams disjoint
1 stream DORP
3 streams DORP
Fig. 10. Average end-to-end delay as a function of the primary idle
probability.
Fig. 9. We set the primary idle probability
c
= 0.8 for each
channel, and the secondary packet arrival rate = 0.3 for
all trafc streams. It is seen that with a single channel the
proposed protocol can support a single trafc stream only,
adding more streams results in innite delay for at least one
of the streams. While the DORP and disjoint protocol cannot
support any trafc streams in this case. Given our setup,
a single trafc stream cannot benet from more than one
channel. That is because of the single channel per time slot
constraint. Once all the links are using the maximum number
of time slots possible, they cannot benet from any additional
channel resources. The same conclusion applies for 2 and 3
trafc streams. It is noted that for the disjoint case only 1
trafc stream can be supported even with 5 available channels.
Fig. 10 shows the effect of the primary idle probability on
the average end-to-end delay when the network has N = 2
primary channels and secondary packet arrival rate is = 0.3
for each data stream available. As expected, increasing the
primary idle probability provides the cognitive nodes with
more channel access opportunities which results in a decrease
in the delay. Here again we note that the disjoint design
cannot accommodate more than a single trafc streams into
the network, while our proposed design scheme as well as the
DORP scheme are able to accommodate all 3 streams. It is
also clear that the proposed protocol outperforms the DORP
protocol for all values of primary idle probability and for any
number of trafc streams.
Finally, we characterize the maximum admissible secondary
packet arrival rate as a function of the primary idle probability
for the case of 1, 2, and 3 trafc streams in gures 11, 12, and
13, respectively. The maximum admissible secondary packet
arrival rate is measured such that all the queues in the network
are stable. That is, the service rate of any given queue is
larger than its packet arrival rate. It is noted that the maximum
admissible arrival rate of the proposed protocol and that of the
DORP protocol are very close to each other, hence we will
sufce by comparing our protocol with the disjoint scheme.
For the case of a single stream, we note that the proposed
design scheme does not benet from increasing the number
of channels. On the other hand, the disjoint design scheme
benets from the increase in the number of channels and
196 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 13, NO. 1, JANUARY 2014
0 0.2 0.4 0.6 0.8 1
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
Pri. idle prob.
S
e
c
.

p
a
c
k
e
t

a
r
r
i
v
a
l

r
a
t
e

(
p
a
c
k
e
t
s
/
s
e
c
)


1 channel
1 channel disjoint
2 channels disjoint
Fig. 11. Maximum admissible arrival rate vs primary idle probability when
S = 1 stream.
0 0.2 0.4 0.6 0.8 1
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
Pri. idle prob.
S
e
c
.

p
a
c
k
e
t

a
r
r
i
v
a
l

r
a
t
e

(
p
a
c
k
e
t
s
/
s
e
c
)


1 channel
2 channels
1 channel disjoint
2 channels disjoint
Fig. 12. Maximum admissible arrival rate vs primary idle probability when
S = 2 streams.
is able to accommodate a stream with higher arrival rates.
However, the performance of our scheme with only a single
channel coincides with that of the disjoint scheme when using
2 channels.
In the case of 2 streams, both the joint and disjoint designs
benet from the increase in the number of channels. However,
our joint scheme can always accommodate trafc streams
with higher arrival rates. For example, at a primary idle
probability of 0.6 and with a single channel, the joint scheme
can accommodate trafc streams with 75% higher arrival rates
compared to the disjoint scheme. With two channels, the joint
design can admit streams with double the arrival rates of the
disjoint design. The same conclusions can be drawn for the
case of 3 trafc streams. It is noted in this case that the
performance of the joint scheme with 1 channel coincides with
that of the disjoint scheme with 3 channels.
VI. CONCLUSIONS
In this paper we have considered the joint design of routing
and resource allocation schemes in cognitive radio based
WMNs. For this class of networks, cross-layer design schemes
are crucial since disjoint design strategies lead to lower perfor-
mance (in terms of delay, or the number of admissible trafc
streams) or infeasible solutions in many cases. Cross-layer
0 0.2 0.4 0.6 0.8 1
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
Pri. idle prob.
S
e
c
.

p
a
c
k
e
t

a
r
r
i
v
a
l

r
a
t
e

(
p
a
c
k
e
t
s
/
s
e
c
)


1 channel
2 channels
1 channel disjoint
2 channels disjoint
Fig. 13. Maximum admissible arrival rate vs primary idle probability when
S = 3 streams.
minimization of the average end-to-end delay is formulated
as non-linear integer programming problem. To reduce the
complexity of the problem, sub-optimal relaxation of the
binary-valued variables is proposed. Moreover, an efcient
decentralized solution algorithm based on the Lagrangian dual
decomposition. Results demonstrate the efciency of the pro-
posed decentralized solution scheme, and its ability to adapt
to varying network loads. Moreover, bounds on the admissible
number of trafc streams as well as the maximum admissible
arrival rates are characterized as functions of primary users
idle probabilities. Performance gains of the cross-layer design
in comparison with disjoint designs and other design methods
are also demonstrated. It was shown that the proposed design
scheme can accommodate higher trafc load, and achieve
lower delay compared to the disjoint design and the DORP
protocol.
ACKNOWLEDGMENT
This work was made possible by NPRP grant No. 08-374-2-
144 and No. 4-1034-2-385 from the Qatar National Research
Fund (a member of Qatar Foundation). The statements made
herein are solely the responsibility of the authors.
REFERENCES
[1] S. Haykin, Cognitive radio: brain-empowered wireless communica-
tions, IEEE J. Sel. Areas Commun., vol. 23, no. 2, pp. 201220, Feb.
2005.
[2] N. Devroye, P. Mitran, and V. Tarokh, Achievable rates in cognitive
radio, IEEE Trans. Inf. Theory, vol. 52, no. 5, pp. 18131827, May
2006.
[3] A. Jovicic and P. Viswanath, Cognitive radio: an information-theoretic
perspective, in Proc. 2006 IEEE Intl. Symp. Inf. Theory, pp. 24132417.
[4] Y. Yuan, P. Bahl, R. Chandra, T. Moscibroda, and Y. Wu, Allocating
dynamic time-spectrum blocks in cognitive radio networks, in Proc.
2007 ACM MobiHoc, pp. 130139.
[5] Q. Zhao, L. Tong, A. Swami, and Y. Chen, Decentralized cognitive
MAC for opportunistic spectrum access in ad hoc networks: a POMDP
framework, IEEE J. Sel. Areas Commun., vol. 25, no. 3, pp. 589600,
Apr. 2007.
[6] Q. Wang and H. Zheng, Route and spectrum selection in dynamic
spectrum networks, in 2006 IEEE Consumer Commun. Netw. Conf.
[7] C.-F. Shih, W. Liao, and H.-L. Chao, Joint routing and spectrum
allocation for multi-hop cognitive radio networks with route robustness
consideration, IEEE Trans. Wireless Commun., vol. 10, no. 9, pp. 2940
2949, 2011.
EL-SHERIF and MOHAMED: JOINT ROUTING AND RESOURCE ALLOCATION FOR DELAY MINIMIZATION IN COGNITIVE RADIO BASED MESH . . . 197
[8] H. Khalife, S. Ahuja, N. Malouch, and M. Krunz, Probabilistic path
selection in opportunistic cognitive radio networks, in 2008 IEEE
GLOBECOM.
[9] Y. T. Hou, Y. Shi, and H. D. Sherali, Optimal spectrum sharing for
multi-hop software dened radio networks, in Proc. 2007 IEEE Intl.
Conf. Comput. Commun., pp. 19.
[10] G. C., W. Liu, Y. Li, and W. Cheng, Joint on-demand routing and
spectrum assignment in cognitive radio networks, in Proc. 2007 IEEE
International Conf. Commun., pp. 64996503.
[11] L. Ding, T. Melodia, S. Batalama, and M. J. Medley, Rosa: distributed
joint routing and dynamic spectrum allocation in cognitive radio ad hoc
networks, in Proc. 2009 ACM International Conf. Modeling, Analysis
Simulation Wireless Mobile Syst., pp. 1320.
[12] D. Cabric, S. M. Mishra, and R. W. Brodersen, Implementation issues
in spectrum sensing for cognitive radio, in Proc. 2004 Asilomar Conf.
Signals, Syst., Comput., pp. 772776.
[13] S. Enserink and D. Cochran, A cyclostationary feature detector, in
Proc. 1994 Asilomar Conf. Signals, Syst. Comput.
[14] A. Ghasemi and E. Sousa, Collaborative spectrum sensing for oppor-
tunistic access in fading environments, in Proc. 2005 IEEE Symp. New
Frontiers Dynamic Spectrum Access Netw., p. 131-136.
[15] S. M. Mishra, A. Sahai, and R. W. Brodersen, Cooperative sensing
among cognitive radio, in Proc. 2006 IEEE ICC, pp. 16581663.
[16] A. K. Sadek, K. J. R. Liu, and A. Epheremides, Cognitive multiple
access via cooperation: protocol design and performance analysis, IEEE
Trans. Inf. Theory, vol. 53, no. 10, pp. 36773696, Oct. 2007.
[17] A. A. El-Sherif, A. K. Sadek, and K. J. R. Liu, Opportunistic multiple
access for cognitive radio networks, IEEE J. Sel. Areas Commun.,
vol. 29, no. 4, pp. 704715, Apr. 2011.
[18] P. Gupta and P. R. Kumar, The capacity of wireless networks, IEEE
Trans. Inf. Theory, vol. 46, no. 2, pp. 338404, Mar. 2000.
[19] J. G. Proakis, Digital Communications. McGraw-Hill Inc., 1994.
[20] G. R. Grimmett and D. R. Stirzaker, Probability and Random Processes.
Oxford University Press, 2001.
[21] D. Bertsekas and R. Gallager, Data Networks, 2nd ed. Prentice Hall,
1992.
[22] R. W. Wolff, Stochastic Modeling and The Theory of Queues. Prentice
Hall, 1989.
[23] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge Uni-
versity Press, 2004.
[24] D. P. Bertsekas, Nonlinear Programming. Athena Scientic, 2003.
[25] S. Boyd, L. Xiao, and A. Mutapcic, Subgradient methods, Lecture
Notes EE392o, Stanford University, Autumn Quarter, 2004 (2003).
[26] A. Mohamed and H. Alnuweiri, Utility-based optimal rate allocation
for heterogeneous wireless multicast, in Proc. 2007 IEEE Intl. Conf.
Commun., pp. 34633470.
[27] C. E. Perkins and E. M. Royer, Ad hoc on-demand distance vector
routing, in 1999 IEEE Workshop Mobile Comput. Syst. Applications.
Amr El-Sherif (S 00, M 08) received his B.Sc.
(with highest Honors) and M.Sc. degrees in electri-
cal engineering form Alexandria University, Alexan-
dria, Egypt in 2002 and 2005, respectively. He
received his Ph.D. degree in electrical engineering
from the University of Maryland, College Park,
in 2009. He is currently an Assistant Professor in
the Electrical Engineering Department at Alexandria
University, Egypt. His research interests include
cooperative communications and networking, cross-
layer design for wireless networks, multiple access
technologies for wireless and sensor networks, and spectrum sharing and
cognitive radio systems.
Amr Mohamed (S 00, M 06) received his M.S.
and Ph.D. in electrical and computer engineering
from the University of British Columbia, Vancouver,
Canada, in 2001, and 2006 respectively. He has
worked as an advisory IT specialist in IBM Innova-
tion Centre in Vancouver from 1998 to 2007, taking
a leadership role in systems development for vertical
industries. He is currently an assistant professor in
the college of engineering at Qatar University and
the director of the Cisco Regional Academy. He has
over 20 years of experience in wireless networking
research and industrial systems development. He holds three awards from IBM
Canada for his achievements and leadership, and three best paper awards. His
research interests include networking and MAC layer techniques mainly in
wireless networks. Dr. Mohamed has authored or co-authored over 40 refereed
journal and conference papers and one textbook.