Physical Communication 14 (2015) 1–13

Contents lists available at ScienceDirect

Physical Communication
journal homepage: www.elsevier.com/locate/phycom

Full length article

Resource allocation for LTE-based cognitive radio network

with queue stability and interference constraints
Alia Asheralieva ∗ , Kaushik Mahata
School of Electrical Engineering and Computer Science, University of Newcastle, Callaghan NSW, Australia

article info abstract

Article history: We consider the problem of interference management and resource allocation in a cog-
Received 12 February 2014 nitive radio network (CRNs) where the licensed spectrum holders (primary users) share
Received in revised form 20 October 2014 their spare capacity with the non-licensed spectrum holders (secondary users). Under such
Accepted 9 November 2014
shared spectrum usage the transmissions of the secondary users should have a minimal
Available online 18 November 2014
impact on the quality of service (QoS) and the operating conditions of the primary users.
Therefore, it is important to distinguish the two types of users, and formulate the prob-
Keywords:
3GPP LTE
lem of resource allocation considering hard restrictions on the user-perceived QoS (such
CRN as packet end-to-end delay and loss) and physical-layer channel characteristics (such as
Interference management noise and interference) of the primary users. To achieve this goal, we propose to assign the
Resource allocation bandwidth and transmission power to minimize the total buffer occupancy in the system
subject to capacity constraints, queue stability constraints, and interference requirements
of the primary users. We apply this approach for resource allocation in a CRN built upon a
Third Generation Partnership Project (3GPP) long-term evolution (LTE) standard platform.
Performance of the algorithm is evaluated using simulations in OPNET environment. The
algorithm shows consistent performance improvement when compared with other rele-
vant resource allocation techniques.
© 2014 Elsevier B.V. All rights reserved.

1. Introduction in the network. The corresponding resource allocation

problem should be formulated as optimal bandwidth and
Advent of the cognitive radio networks (CRNs) has en- transmission power control with hard restrictions on the
abled efficient spectrum sharing and has brought new user-perceived QoS (such as packet end-to-end delay and
challenges to the classical problems of interference man- loss) and physical-layer channel characteristics (such as
agement and resource allocation in wireless networks. noise and interference) of the primary users.
In CRNs the licensed spectrum holders (also called pri- The related literature on resource allocation and inter-
mary users) can share their bandwidth resources with ference management in CRNs can be divided in three large
non-licensed holders (secondary users). Under such shared groups. The first group [2–4] focuses on devising various
usage of primary licensed spectrum the transmission of the power control policies where satisfying the QoS require-
secondary users should have a minimal impact on the qual- ments of the primary users is a premium. In such a frame-
ity of service (QoS) and the operating conditions of the pri- work the objectives are: (i) to keep the interference caused
mary users [1]. In this context, it is essential to consider by the secondary users to the primary users below a given
the presence of two types of users (primary and secondary) threshold, and (ii) to preserve the admissibility and oppor-
tunism of the secondary users. Through theoretical anal-
ysis and simulations it is shown that all of the proposed
∗ Corresponding author. algorithms [2–4] achieve the objectives, and show con-
E-mail address: [email protected] (A. Asheralieva). sistent performance in terms of transmission power and
http://dx.doi.org/10.1016/j.phycom.2014.11.001
1874-4907/© 2014 Elsevier B.V. All rights reserved.
2 A. Asheralieva, K. Mahata / Physical Communication 14 (2015) 1–13

signal-to-interference-and-noise ratio (SINR). These meth- So far there is no known method that allows us to
ods do not account for important higher-layer QoS met- address all critical issues of interference management and
rics as packet end-to-end delay, loss, or throughput in the resource allocation for a CRN user in a unified framework.
proposed optimization criteria. Therefore, these results are Some authors have focused only on achieving the target
applicable only in some rare cases where the service per- interference caused by the secondary users to the primary
formance of the end-to-end user applications is not a vital users [2–4]. Some other works concentrate only on higher-
issue (e.g., for best-effort traffic streams). layer QoS requirements without considering SINR [5–7].
The second group of papers examine user-perceived The remaining available results do not distinguish between
QoS using theoretical analysis of user behavior in CRNs. the QoS requirements of the primary and secondary
For instance, in [5] a statistical traffic control scheme is users [8,9]. To fill this void, we propose a novel approach
proposed to satisfy the timing constraints of the packets for resource allocation in CRNs. Our method is designed
belonging to different streams with diverse QoS char- to provide best effort service to the secondary users while
acteristics and requirements. Here the authors deploy actively restricting the interference to the primary users. In
admission control and coordinated transmission of the addition, the proposed approach can guarantee acceptable
constant-bit-rate (CBR) and variable-bit-rate (VBR) traf- QoS level for the primary users. Thus, unlike previous
fic. A dynamic channel-selection strategy for autonomous works we simultaneously consider both physical-layer
wireless users transmitting delay-sensitive multimedia requirements and higher-layer QoS guarantees of primary
applications over CRN has been proposed in [6]. Unlike users, and control the bandwidth and the power of the
prior works, the authors of [6] have paid attention to the users to maximize the QoS for both primary and secondary
rate and delay requirements of heterogeneous multime- users.
dia users. To efficiently manage the available spectrum We apply the idea for allocating bandwidth and power
resources in a decentralized manner, the authors use a vir- in a CRN built on a Third Generation Partnership Project
tual queue interface to evaluate the average delays experi- (3GPP) long-term evolution (LTE) standard. LTE is chosen
enced by the traffic with various priority levels. Each user as the implementation platform because of its appealing
utilizes this delay information in a dynamic strategy learn- features, such as spectrum flexibility, fast adaptation to
ing (DSL) algorithm for selecting the channels in order to time-varying channel conditions, high spectral efficiency
maximize some utility function. In [7] the authors propose and robustness against interference [10]. A detailed de-
a spectrum allocation framework that jointly considers the scription of the LTE radio interface can be found, for in-
QoS provisioning for heterogeneous secondary real-time stance, in [11]. In short, LTE is based on the universal
(RT) and non-real time (NRT) users, the spectrum sensing, terrestrial radio access (UTRA) and high-speed down-link
spectrum access decision, channel allocation, and call ad- packet access (HSDPA). In the downlink, LTE uses Orthog-
mission control in a distributed cooperative CRN. Within onal Frequency-Division Multiple Access (OFDMA) which
this framework, the admissible number of the RT users is offers high spectral efficiency and robustness against in-
calculated based on their QoS requirements (measured in terference. Single Carrier Frequency Division Multiple Ac-
terms of channel dropping and packet blocking probabil- cess (SC-FDMA) is used in the uplink of LTE due to its
ities). Then, the available channels are distributed among lower (compared to OFDM) Peak-to-Average Power Ratio
the admitted RT users, whereas the remaining channels are (PAPR) [12]. The numerology of LTE includes a subcarrier
allocated to the NRT users. The set of QoS measures ana- spacing of 15 kHz, support for bandwidths up to 20 MHz,
lyzed in [6,7] does not include physical-layer characteris- and a resource allocation granularity of 180 kHz × 1 ms
tics, such as SINR. However, in CRNs it is essential to control (so-called resource block or RB). The available resource
the interference from the secondary users to the primary blocks are distributed among the users by the medium ac-
users. Even if the allocated capacity (number of channels) cess control (MAC) schedulers in enhanced NodeBs (eNBs).
is large enough for a given traffic stream, the interference Depending on the implementation, the scheduling can be
from the secondary users may result in poor channel qual- done based on the queuing delay, instantaneous channel
ity causing very low service rate, and consequently vary conditions, fairness, etc. [13,14].
unacceptably high delays for the primary users. The rest of the paper is organized as follows. In Section 2
The last group of papers [8,9] focus on the trade-off we describe the system model and radio model of the
between transmission delay and transmission power in network, and formulate the resource allocation problem.
wireless networks. In [8] a delay-power control (DPC) In Section 3 we discuss algorithm implementation. In
scheme is proposed to balance the delay against trans- particular, we present the solution methodology and the
mission power in each wireless link. It is shown that DPC possible settings of algorithm. In Section 4 we carry out
converges to a unique equilibrium. The equilibrium point a performance analysis based on results of simulations in
OPNET environment [15].
satisfies key properties related to the nature of bandwidth
sharing achieved by the links. Distributed resource alloca-
tion based on queue balancing in a multi-hop CRN has been 2. Resource allocation algorithm
investigated in [9]. Here the network resources (power,
channel and data rate) are allocated to the users based on 2.1. The CRN architecture
the assumption of dynamic link capacity. The papers [8,9]
assume all users having equal priorities in accessing the 2.1.1. Network topology
spectrum, and do not pay attention in presenting the QoS We consider a CRN architecture implemented over a
for the primary users. standard infrastructure-based LTE network with frequency
 A. Asheralieva, K. Mahata / Physical Communication 14 (2015) 1–13 3

CRM

Fig. 1. A typical CRN architecture.

division duplex (FDD) as shown in Fig. 1. The network the necessary information from the control mechanisms
consists of n primary (licensed) eNBs (PBs) numbered built in LTE.
PB1 , . . . , PBn . In addition, there are m secondary (unli- An LTE system operates on a slotted time basis. The
censed) eNBs (SBs) numbered SB1 , . . . , SBm . In the sequel time axis is partitioned into a sequence of non-overlapping
we frequently use the notation I = {1, . . . , n}, J = time intervals (slots) of the length Ts each. In the sequel
{1, . . . , m} to denote the indices associated with the PBs the integer-valued index of a time slot is denoted by
and the SBs, respectively. All the eNBs are connected to the t. In LTE, each eNB serves a number of wireless user
backbone server via a central resource manager (CRM). The equipments (UE) located within its service area (cell).
communication between the eNBs, the CRM and the server Resources (counted in terms of RBs) are allocated to the
is realized using the high-speed internet protocol (IP) links users by the eNBs using the standard packet scheduling
to facilitate fast data transmission. procedure described in detail in [16]. In this procedure
each UE is required to transmit its buffer status information
2.1.2. Role of the CRM and cognitive functionality at every time slot. In this way the eNB gets to ‘‘know’’
The primary eNBs operate on their fixed licensed the exact amount of uplink data bits enqueued in the
spectrum bands (primary channels). In this work we buffers of UEs. On the other hand, the eNB readily finds
consider the most general scenario when the frequency out the size of the downlink buffer for each UE. The uplink
bands of the PBs do overlap. The fixed licensed spectrum and downlink buffer status information is used by the
band of PBi is denoted by Bi . Note that Bi is an integer, which scheduling algorithm run by the eNB for allocating the
resources to individual UEs on a slot-by-slot basis.
represents the number of RBs that belong to PBi .
In the framework used in this paper both primary
In the network, a CRM retains control over the net-
and secondary eNBs adopt the standard packet scheduling
work resources (bandwidth and the transmission powers)
procedure described above. The CRM uses the buffer
of the individual PBs and SBs. In real network deployments,
status information (transmitted by the users to the eNBs
the PBs often have some spare capacities. In such a situa-
during the scheduling process) to allocate the resources
tion the CRM may allow one or more SBs to make use of
to different eNBs. At every time slot each eNB is required
this spare capacity. So, when PBi has some spare capac-
to report its aggregated traffic demands to the CRM. The
ity, then CRM might allocate that to one or more SBs. The
justification behind using the aggregated traffic demand
resource allocation policy used by CRM should guarantee
lies in the fact that in an LTE system the queuing and
that the transmission of SBs will have minimal impact on
processing delays of the packets at the buffers of the UEs
the quality of service (QoS) and operating conditions of
are practically negligible compared to that at the buffers of
PBs [1].
the eNB [17]. This is mainly because the amount of traffic
arriving at the buffer of a UE is much less than that arriving
2.1.3. Traffic information exchange for resource allocation at the buffer of a eNB.
In order to make effective resource allocation the CRM After the resources are allocated to the eNBs, the eNBs
needs to monitor the traffic load at different eNBs in both distribute them among their users. This second stage of
uplink and downlink directions. In fact, the CRM can gather allotments utilizes the buffer status information of the
4 A. Asheralieva, K. Mahata / Physical Communication 14 (2015) 1–13

eNB (to be more precise, it is the aggregated size of the

buffers of the UEs, which a eNB reports to the CRM during
the packet scheduling procedure).
In the downlink direction the arrival rate at the queue
of a eNB corresponds to the total traffic arriving from the
CRM to its physical buffer. The CRM can readily keep track
of this information. The service rate of each queue depends
on the bandwidth and transmission power allocated to the
respective eNBs. Consequently, at any slot t the size of the
Fig. 2. A typical CRN architecture.
queue (in bits) and the arrival rate at the queue (in bits
per slot or pbs) can be observed in respective eNBs. This
individual UEs, and can be implemented according to any information (queue size and arrival rate) is sent by the
standard scheduling algorithm. In this paper, however, eNBs to the CRM. CRM uses this information to make the
we concentrate on the first stage only. Many papers have decision on resource allocation.
addressed the second stage allocation problem, see [18,19]. Let qPi (t ) denote the queue size at the buffer of PBi at
the beginning of the tth time slot, and let aPi (t ) denote the
2.1.4. Information exchange and control channels number of bits arrived at the buffer of PBi during tth time
In LTE, the scheduling information (and some other con- slot. Then Lindley’s equation [21] gives
trol information) is carried by the physical uplink control +
qPi (t + 1) = qPi (t ) + aPi (t ) − riP (t ) ,

channel (PUCCH) and the physical downlink control chan- ∀i ∈ I (1a)
nels (PDCCH) in uplink and downlink directions, respec-
tively. According to the standard, a pre-defined number of where ⌈x⌉ = max(0, x); (t ) denotes the number of
+
riP
RBs are reserved for PUCCH transmissions. PDCCH occupies bits served from the buffer of PBi during tth time slot.
the first 1–3 OFDM symbols in a time slot extending over The value of riP (t ) depends on the power and bandwidth
the entire licensed spectrum band of the eNB [20]. allocated to PBi during tth time slot (see (4a) below). Later
Both primary and secondary eNBs in the CRN use above in Section 2.2.2 we shall derive the expression for riP (t ).
control channels to transmit their scheduling information. We can write an equation similar to (1a) for each
Since the secondary eNBs do not have licensed access, the secondary eNB:
CRM reserves a small number of frequency bands (2 RBs +
qSj (t + 1) = qSj (t ) + aSj (t ) − rjS (t ) ,

in our model) for PUCCH and PDCCH transmissions by all
∀j ∈ J (1b)
the SBs. We assume that the frequency band reserved for where qSj (t ) denotes the number of bits waiting to be
control messages of the SBs is non-overlapping with the served at the buffer of SBj at the beginning of tth time slot;
licensed spectrum bands of the PBs. Hence, the control
aSj (t ) denotes the number of bits arrived at the buffer of
signals of the SBs do not interfere with the transmissions
of the PBs. In real network deployments, this can be SBj during tth time slot; rjS (t ) denotes the number of bits
implemented by reserving some bandwidth for control served from the buffer of SBj during tth time slot. The value
signaling between the SBs and their users. This bandwidth of rjS (t ) depends on the power and bandwidth allocated to
should not overlap with the licensed spectrum bands of SBj during tth time slot (see (4b)). An expression for rjS (t )
the PBs. All the SBs (and their users) can use this reserved will be given shortly in Section 2.2.2.
bandwidth for passing their control messages. In this paper we use future queue size of the eNBs
Note, that alternatively SB could use the physical con- given by (1) as the optimization target. The reason behind
trol channels of the PB(s) sharing the bandwidth with this this approach is that the queue size is directly connected
SB. However, in this case if some of the SBs are not allo- to the main QoS metrics, such as round-trip latency and
cated the bandwidth, we will not be able to obtain their loss. Unfortunately, the direct estimation of delay and
scheduling information. loss is rather complex in an LTE system. For instance,
the packet end-to-end delay in LTE consists of various
2.2. Network model delay components, including transmission and queuing
delay, propagation and processing delay, the uplink delay
2.2.1. MAC layer model due to scheduling and delay due to hybrid automatic
The eNB level MAC layer model of the above CRN sys- repeat request (HARQ) [22]. The accurate analysis of
tem is illustrated in Fig. 2. It consists of n primary and m these delay components requires knowledge of many
secondary channels belonging to the PBs and SBs, respec- system parameters which are not available during resource
tively. Each eNB (PB or SB) is represented by the queue (pri- allocation, whereas the queue size can be calculated easily
mary or secondary). The model in Fig. 2 can be applied to using (1).
both uplink and downlink directions. Similarly, the discus-
sion in the rest of the paper is applicable to either direction. 2.2.2. Radio model
Note that the queue associated with the eNB in the Both SC-FDMA (applied in the uplink of LTE system) and
model relates to the physical buffer of this eNB (and not OFDMA (applied in the downlink of LTE system) ensure
to the buffers of the UEs served by the eNBs). In the uplink mutual orthogonality of the RBs (consisting of the sets of
direction the arrival rate to each queue conforms with the OFDMA subcarriers) assigned to different users. This re-
aggregated traffic demand of the users at the respective sults in a minimal level of co-channel interference between
 A. Asheralieva, K. Mahata / Physical Communication 14 (2015) 1–13 5

different transmitter–receiver pairs located within one PBi at tth time slot. Given the channel gains GPS SS
ij , Gjk , i ∈
cell. In absence of frequency reuse, the co-channel interfer- I, k ∈ J, j ∈ J, SINRSij (t ) is estimated using [24]:
ence should be close to zero (assuming the UEs are properly
synchronized to the eNB). Therefore, the signal transmitted pSij (t )GPS
ij
SINRSij (t ) = ,
(received) by a user to (from) its serving eNB (PB or SB) will

ηijS + pPk (t )GPS pSkl (t )GSS
 
be distorted only by the users located in different eNBs (PBs ik + kl
and SBs) and transmitting at the same frequency [23]. k∈I/{i} l∈J

Now, consider a wireless channel between a eNB and a ∀ i ∈ I, j ∈ J (2b)

UE. This wireless link is characterized by some link coeffi-
where ηijS (t ) is the thermal noise power (in Watts) in the
cient, expressing the gain (attenuation) experienced by the
channel between PBi and SBj .
signal transmitted using this wireless link. In LTE the link
In practice, the PBs and SBs are told by the CRM what
coefficients for each wireless channel can be obtained from
power they can use for transmission. Note, that in general
the channel state information (CSI) through the use of spe-
the transmission power allocated to the eNBs cannot
cial reference signals (RSs) [23]. The reference signals (or
exceed their maximal transmission power. Therefore,
pilot signals) are broadcast by all users (and the eNBs) with
some constant transmit power (so-called reference power 0 ≤ pPi (t ) ≤ PiP , ∀i ∈ I (3a)
known to the users and the eNBs). On the receiving side
where PiPis the maximal transmission power capacity of
the power of the received signal is measured and compared S
PBi . The total transmission power needed by SBj is P1j (t ) +
with the reference power. The ratio between the reference
power and received power levels of the reference signal is
S
· · · + PNj (t ), as PijS (t ) denotes the power used by SBj in the
associated with the link gain (attenuation) in the wireless RB(s) of PBi . Hence,
channel between the corresponding transmitter–receiver

0≤ pSij (t ) ≤ PjS , ∀j ∈ J (3b)
pair [20,23]. Consequently, for each eNB in the network we i∈I
can always find the total (aggregated) link gain for all wire-
less channels between the users and the eNB. where PjS is the maximal transmission power capacities of
In passing, we note that in LTE the CSI information is SBj . The values of PiP and PjS are known to CRM.
usually available both for the channels within the eNB If we know the power allocations and channel gains, we
(i.e. between a eNB and its local users) and for the channels can calculate the SINR values using (2). Furthermore, if we
outside the eNB (between the eNB and the other users know the bandwidth allocations, then we can also obtain
connected to other eNBs, and between the local users and the values of riP (t ) and rijS (t ). Suppose that the number of
other eNBs) [20,23]. RBs allocated to PBi during tth time slot is bPi (t ). Then, using
We denote the aggregate gain (attenuation) for the Shannon–Hartley theorem [25] we get
wireless channels between the users of PBi and the users
riP (t ) = ω · bPi (t ) log2 (1 + SINRPi (t )), ∀i ∈ I (4a)
ik , i ∈ I, k ∈ I. Similarly, Gij , i ∈ I, j ∈ J denotes
PS
of PBk via GPP
the aggregate gain (attenuation) for the wireless channels where ω = 180 kHz is bandwidth comprising one RB.
between the users of PBi and the users of the SBj . Likewise, Now, suppose bSij (t ) is the number of RBs that PBi
shares with SBj during tth time slot. Recall, that SBj uses
jk j ∈ J, k ∈ J denotes the aggregated gain (attenuation)
GSS
for the wireless channels between the users of SBj and the a power PijS (t ) to transmit in the RB(s) of PBi . Hence, from
users of the SBk . Shannon–Hartley theorem we note that the number of bits
Let SINRPi (t ) denote the signal-to-interference-and- that SBj can transmit in these resource blocks is given by
noise ratio (SINR) experienced by PBi at tth time slot. Then, rijS (t ) = ω · bSij (t ) log2 (1 + SINRSij (t )), ∀i ∈ I, j ∈ J. (4b)
ik , Gij , i ∈ I, k ∈ I, j ∈ J, the
PS
given the channel gains GPP
By accounting for all the PBs sharing their RBs with SBj we
SINRPi (t ) is estimated using [24]:
get
pPi (t )GPP

rjS (t ) = rijS , ∀j ∈ J.
SINR Pi (t ) =  ii
,
i∈I
ηiP + pPk (t )GPP pSkj (t )GPS
 
ik + kj
k∈I\{i} j∈J
2.3. Resource allocation problem
∀i ∈ I (2a)
We now consider the problem of joint power and
where pPi (t ) is the power used by PBi to transmit during tth bandwidth allocation in described CRN model. First, let us
time slot; pSkj (t ) is the power used by SBj to transmit in the define the unknown n(m + 1) dimensional bandwidth and
RB(s) that it borrows from PBk ; ηiP (t ) is the thermal noise transmission power vectors as
power (in Watts) in the channel belonging to PBi . Note, that
pP1 (t ), . . . , pPn (t ), pS11 (t ), . . . , pSn1 (t ), . . . ,

p :=
the set of values that the summation index k takes in the T
denominator in (2a) excludes i. This accounts for zero co- pS1m (t ), . . . , pSnm (t ) , (5a)
channel interference in OFDMA.
bP1 (t ), . . . , ( ),
bPn t (t ), . . . ,
bS11 bSn1 (t ), . . . ,

b :=
Similarly, let SINRSij (t ) denote the SINR experienced by
T
bS1m (t ), . . . , bSnm (t ) .

SBj while transmitting in the resource block borrowed from (5b)
6 A. Asheralieva, K. Mahata / Physical Communication 14 (2015) 1–13

Note, that at any slot t the number of RBs allocated for minimize the sum of the maximal queues of the PBs and
data transmission in the licensed bandwidth of PBi cannot SBs subject to the constraint (10). To simplify the notation,
exceed Bi RBs. Thus, we skip the index t below
 +
bPi (t ) + bSij (t ) ≤ Bi , ∀ i ∈ I. minimize max qPi + aPi − riP (b, p)

(6a)
i∈I
j∈J
 +
Let us define 
  + max qSj + aSj − rijS (b, p) (11a)
 j∈J
 i∈I
B := b bPi (t ) + bSij (t ) ≤ Bi , bPi (t ),

 j∈J subject to: riP (b, p) ≥ qPi + aPi − Qitar , ∀i ∈ I, (11b)

riP (b, p) ≤ qPi + aPi , ∀i ∈ I,


(11c)
bSij (t ) ∈ Z+ , ∀i ∈ I, j ∈ J (6b)
p ∈ P, b ∈ B. (11d)

which is the set of all admissible values that a vector b can In (11) we minimize the maximum queue size not only
take. Note, that Z+ represents the set of all non-negative for SBs, but also for PBs (this might appear unnecessary
integers. because the queue sizes for PBs are already constrained to
The CRN should allocate the bandwidth and transmis- be bounded). To see the reason, suppose that we do not
sion power to guarantee some QoS levels to the PBs. There- have the first term in (11a), i.e. we minimize the maximum
fore, for each PB we set some minimal guaranteed QoS queue size only for SBs. If we solve this problem, then the
metrics. These are the target queue sizes denoted as Qitar ≥ optimal solution will be such that for all PBs the queue size
0, i ∈ I, expressed in bits, and target interference levels will be equal to the target queue size, i.e.
from SBs denoted as Iitar ≥ 0, i ∈ I, expressed in Watts.1 qPi (t + 1) = Qitar , ∀i ∈ I.
Consequently, we constrain the queue size of the PBs to
be below the specified target queue size, i.e.: Now, if the traffic demands in SBs are very low, then
there might be a situation when the optimal solution for
qPi (t + 1) ≤ Qitar , ∀ i ∈ I, (7) SBs will be such that
which is equivalent to (see (1a)) max qSj (t + 1) ≤ Qitar = qPi (t + 1), ∀i ∈ I.
j∈J
riP (b, p) ≥ qPi (t ) + aPi (t ) − Qitar , ∀ i ∈ I, (8a)
However this is clearly unfair, as the PBs will have larger
riP (b, p) ≤ qPi (t ) + aPi (t ), ∀i ∈ I. (8b) queue sizes than SBs, and therefore will experience longer
delays than SBs.
Note, that in (8) we express the service rate at the PBs given
Note, that the formulation of the resource allocation
by (4) as a function of the bandwidth and transmission
problem given by (11) can be applied only in case if all
power.
the constraints (11b)–(11d) are feasible. However, it is easy
We also constrain the interference from SBs to PBs to be
to check that the constraint (11b) may not be feasible for
below the specified target interference levels. Hence,
some PBs. This is discussed in the next subsection, where
we also suggest a modified version of the primary opti-
 
pSkj (t )GPS
kj ≤ Ii ,
tar
∀ i ∈ I. (9)
k∈I\{i} j∈J
mization problem (11) which can be used freely without
worrying about the feasibility of (11b).
The constraint (9) is used here to protect the users of the
PBs from the potentially heavy interference from the users
2.4. Detecting infeasibility and associated remedy
of the SBs.
Using (3) and (9), the set of all admissible values that p
In some cases the constraint (11b) may be not feasible.
can take is given as:
  This happens when the service capacity of the eNB is not
  enough to guarantee that the queue size of a PB will not
P := p  pS (t )GPS
kj ≤ Ii ,
tar

k∈I\{i} j∈J kj exceed the target queue size. Let us define
 qPi + aPi − Qitar
0 ≤ pPi (t ) ≤ PiP , 0 ≤ pSij (t ) ≤ PjS , Ωi = , ∀ i ∈ I. (12)
i∈I ω · Bi
If SINRPi (p) < 2Ω i − 1, then PBi ’s queue size will exceed

∀ i ∈ I, j ∈ J . (10) the Qitar , and (11b) will be infeasible. Given some p ∈ P we
can construct the set Î = i|i ∈ I, SINR Pi (p) < 2Ωi − 1 .
 

Now, we can formulate the optimization problem using This set consists of all indices associated with the primary
the queue length of the eNBs as an optimization target. eNBs which are unable to meet the target queue size for
To maximize the QoS for the users in the eNBs we will the given p.
When i ∈ Î then PBi is not in a position to share any re-
sources with the SBs. All of PBi ’s licensed bandwidth is used
1 The settings of Q max and I max will be discussed in Section 3.
i i by itself, i.e. pSij = 0 and bSij = 0, ∀j ∈ J. No target queue size
 A. Asheralieva, K. Mahata / Physical Communication 14 (2015) 1–13 7

requirements can be applied to PBi . Consequently, the con- Consequently, the functions fi (p) can be approximated
straints (6a), (7) and (11b) should be changed accordingly by the shifted logistic functions
as ρ
fˆi (p) = gi (SINRPi (p) − 2Ωi + 1)
≤ Bi , ∀i ∈ Î,
bPi (13a) 1
= ,
  1 + exp[−2ρ(SINRPi (p) − Ωi + 1)]
kj ≤ 0,
pSkj GPS ∀i ∈ Î, (13b)
k∈I\{i} j∈J ∀i ∈ I. (16)

riP (b, p) ≥ 0, ∀i ∈ Î. (13c) Now we can use the approximation fˆi (p) instead of the
original indicator function fi (p) to construct the modified
Note that the set of indices Î depends on the exact values constraints given by (15):
of p (and SINRPi (p)). Therefore, before resource allocation

bPi + fˆi (p) · bSij ≤ Bi , ∀ i ∈ I, (17a)
we are unable to perform the feasibility test for PBs (i.e., de- j∈J
tect the PBs which are unable to meet the target queue size  
kj ≤ fi (p) · Ii , ∀ i ∈ I,
tar
requirements). Instead, we should find some simple mech- pSkj GPS ˆ (17b)
anism to deal with infeasibility during resource allocation. k∈I\{i} j∈J

Let us construct the set of feasibility indicators as

riP (b, p) ≥ fˆi (p) · (qPi + aPi − Qitar ), ∀ i ∈ I. (17c)
Ω
fi (p) = H SINRPi (p) − 2 + 1 ,
 
∀i ∈ I (14) Correspondingly, the sets of all admissible values that
the vectors p and b can take are modified as
where H (x) is a well-known Heaviside step-function [26]  
defined as
 
B̃(p) := b bPi + fˆi (p) · bSij ≤ Bi , bPi , bSij ∈ Z+ ,

1,

if x ≥ 0
 j∈J
H ( x) =
0,

otherwise.
∀i ∈ I, j ∈ J , (18a)
Note, that
 
1, i ̸∈ Î(p)
  
fi (p) = P̃ := p  pS GPS ≤ fˆi (p) · Iitar ,

0, i ∈ Î(p). k∈I\{i} j∈J kj kj

0 ≤ pPi ≤ PiP , 0 ≤ pSij ≤ PjS ,
For a given p, the function fi (p) indicates whether PBi
i∈I
is able to meet the constraint (11b). Using the feasibility 
indicator functions fi (p), we can modify the constraints ∀ i ∈ I, j ∈ J . (18b)
(6a), (7) and (11b) as follows:

Note, that the set B̃(p) defined in (18a) is p dependent.


bPi + fi (p) · bSij ≤ Bi , ∀ i ∈ I, (15a)
j∈J With new constraints (17), and feasibility sets (18a),
  (18b), the final optimization problem will take the form
pSkj GPS
kj ≤ fi (p) · Iitar , ∀i ∈ I, (15b) +
minimize max qPi + aPi − riP (b, p)

k∈I\{i} j∈J
i∈I

riP (b, p) ≥ fi (p) · (qPi + aPi − Qitar ), ∀ i ∈ I. (15c)

 +

+ max qSj + aSj − rijS (b, p) (19a)
Hence, if (12) holds, then the expressions (15) will be j∈J
i∈I
equivalent to the constraints (13). Otherwise, (15) is equiv-
alent to the primary constraints (6a), (7) and (11b). subject to: riP (b, p) ≥ fˆi (p) · (qPi + aPi − Qitar ),
Unfortunately, as fi (p) are non-smooth functions, it is ∀ i ∈ I, (19b)
hard to incorporate them in a numerical algorithm to solve
riP (b, p) ≤ qPi + aPi , ∀ i ∈ I, (19c)
an optimization problem. A more convenient approach is
to derive the smooth approximations for fi (p). For this, p ∈ P̃, b ∈ B̃(p). (19d)
note that a well-known smooth approximation to the step
The advantage of the formulation (19) is that there is
function H (x) is the logistic function [27]
no need to detect the feasibility of (11b) before resource
1 allocation.
g ρ (x) =
1 + e−2ρ x
3. Algorithm implementation
where ρ > 0. Larger ρ corresponds to a closer approxima-
tion to H (x), i.e.: 3.1. Solution methodology

1 In this subsection we provide the solution methodology

H (x) = lim .
ρ→∞ 1 + e−2ρ x for the problem (19). Firstly, note that (19) can be
8 A. Asheralieva, K. Mahata / Physical Communication 14 (2015) 1–13

equivalently written as yk be the value of the objective function at (bk , pk ). Note

that yk gives a lower bound on the cost function at node k.
P S
minimize Qmax + Qmax (20a) We will also use the notation (bMIP , pMIP ) to denote the best
subject to: riP (b, p) ≥ fˆi (p) · (qPi + aPi − Qitar ), obtained MIP solution of the primary MIP problem and yMIP
to denote the best obtained cost at (bMIP , pMIP ) which gives
∀ i ∈ I, (20b) an upper bound on the primary MIP problem.
riP (b, p) ≤ qPi + aPi , ∀ i ∈ I, (20c) At each node k the branching is performed if bk is non-
integer and better than the best obtained integer solution
qPi + aPi − riP (b, p) ≤ Qmax
P
, ∀ i ∈ I, (20d) (bMIP , pMIP ) (i.e. when yk < yMIP ). The node k has no
 branches in the following cases:
qSj + aSj − rijS (b, p) ≤ S
Qmax , (20e)
i∈I – RP k has no feasible solutions;
 – bk is integer;
rijS (b, p) ≤ qSj + aSj , (20f) – bk is non-integer and worse than the best obtained
i∈I integer solution (bMIP , pMIP ) (yk > yMIP for minimization
p ∈ P̃, b ∈ B̃(p). (20g) problem).
In all cases above everything is known about the sub-
The above problem (21) is a mixed-integer programming
problem, and the subproblem is called ‘‘fathomed’’. A de-
(MIP) problem due to the integer restrictions on b. This
tailed illustration of the solution process in B&B algorithm
problem can be effectively solved by applying any known
is shown in Fig. 3 [28].
MIP method (in our work, we used branch-and-bound
algorithm [28]).
In branch-and-bound (B&B) algorithm the MIP solution 3.2. Target levels of interference and queue size of PBs
can be found solving the relaxed problem (RP) given by
In this subsection we discuss some possible settings of
P
minimize Qmax S
+ Qmax (21a) the parameters Iitar and Qitar . The acceptable interference
levels at the PBs can be set based on the PBs’ SINR re-
subject to: riP (b, p) ≥ fˆi (p) · (qPi + aPi − Qitar ), quirements. Note that some PBs may operate on very noisy
∀i ∈ I, (21b) channels with high interference from other PBs. Appar-
ently, further reduction of SINRs in these channels will be
riP (b, p) ≤ qPi + aPi , ∀ i ∈ I, (21c)
disastrous, and for the sake of keeping reliable data trans-
qPi + (b, p) ≤
aPi − riP , ∀ i ∈ I, P
Qmax (21d) mission of the PBs it is important to refrain from allocating
 these channels to the SBs (we will set zero transmission
qSj + aSj − rijS (b, p) ≤ Qmax
S
, (21e) power to the SBs in these channels). Other PBs may oper-
i∈I ate on rather noiseless channels with zero or very light in-
 terference from other PBs. These channels may be used by
rijS (b, p) ≤ qSj + aSj , (21f)
the SBs for their data transmissions.
i∈I
Let SINRtar
i be the target SINR of PBi . Then, at any slot t
p ∈ P̃, b ∈ B̃(p)
˜ (21g) the following condition must hold

where: pPi GPP

ii
  SINR Pi =  
ηiP + pPk GPP pSkj GPS
  
˜B̃(p) := b bP + fˆ (p) ·  bS ≤ B , bP , bS ≥ 0, ik + kj
 i i ij i i ij k∈I\{i} j∈J
 j∈J

i ,
≥ SINRtar ∀i ∈ I. (22)

∀ i ∈ I, j ∈ J . (21h) If (22) is feasible, then
  pPi GPP 
ik − ηi ,
ii
One can easily see that (21) is a non-convex problem pSkj GPS
kj ≤ − pPk GPP P
SINRtar
i
which can be solved using any standard interior-point k∈I\{i} j∈J k∈I\{i}

method [29] modified for non-convex problems as it has ∀i ∈ I.

been done, for instance, in [30]. In particular, we apply a
Otherwise (if (22) is not feasible) we should keep zero the
primal–dual interior point algorithm for non-convex prob-
interference from the SBs.
lems described in [31]. If there are no MIP solutions to the
However, at the beginning of slot t (i.e. before alloca-
relaxation, we divide the problem into a set of smaller sub-
tion) we do not know the exact values of pPi (t ) and pSij (t ) for
problems (this is the ‘‘branching’’ part of the B&B algo-
all i ∈ I, j ∈ J. Therefore, to set the target interference level
rithm).
from the SBs to the PBs, we propose to deploy past (avail-
In effect, the solution process in B&B algorithm can be
able) observations of the transmission power, and set:
described by the tree consisting of the nodes (associated
with separate subproblems) and branches (associated  +
pPi GPP
with the problem division) [28]. Let RP k be the relaxed

Iitar = ii
− pPk GPP
ik −η P
i , ∀i ∈ I (23a)
subproblem at node k, (bk , pk ) be the solution of RP k and SINRtar
i k∈I\{i}
 A. Asheralieva, K. Mahata / Physical Communication 14 (2015) 1–13 9

Fig. 3. B&B algorithm.

where enough to keep their queues stable, i.e. for each PB we set

pPi (t − τ + 1) qPi (t + 1) ≤ Qitar = qPi (t ), ∀i ∈ I.


(24)
τ ∈T
pPi = , ∀i ∈ I, T = {1, . . . , T }, (23b) A more sophisticated approach can be used to find the
T
lower QoS bound if the information about the past queue
T is the number of past observations. By doing so, we learn size is available at any slot t. In this case one can set the tar-
about SINR conditions in each channel (from the past ob- get buffer size based on past s ≥ 1 queue size observations
servations), and set the target interference level from the as
SBs to the PBs based on these conditions.
The parameter Qitar represents the target queue size of qPi (t + 1) ≤ Qitar = qPi , ∀i ∈ I, (25a)
PBs, which indicates that the QoS of PBs is satisfied. If we
where
set Qitar to be equal to zero, then after each subsequent
data arrival the queues of PBs will be fully cleared. This qPi (t − τ + 1)

approach represents a ‘‘hard’’ QoS strategy which ensures τ ∈T
qPi = , ∀i ∈ I, T = {1, . . . , T }. (25b)
that PBs are always served with best possible QoS (with T
minimal delay), but does not guarantee that some spare
capacity will be left to serve SBs.
A ‘‘more soft’’ QoS strategy will be to establish some 4. Algorithm performance
lower bound on QoS for PBs which will provide an oppor-
tunity for SBs to be served. To establish this lower bound, 4.1. Simulation model
we will use the main principle of congestion control [32,22]
stating that in an uncongested node the queue size q(t ) is In this section we present the performance analysis
stable, i.e. q(t + 1)/q(t ) ≈ 1, whereas in a congested node of the proposed algorithm using the simulation model of
the queue size is growing very fast, so that q(t + 1)/q(t ) ≫ the network shown in Fig. 1. The model has been built
1. Therefore, to protect the PBs from congestion, it is upon the LTE FDD platform using the OPNET simulation
10 A. Asheralieva, K. Mahata / Physical Communication 14 (2015) 1–13

Table 1
Simulation parameters of the model.
Parameter Value

Radio network model: Pass loss L = 40 log10 R + 30 log10 f + 49, R—distance (km), f —carrier frequency (Hz)
Shadow fading Log-normal shadow fading with a standard deviation of 10/12 dB for
outdoor/indoor users
Penetration loss The average building penetration loss is 12 dB with a standard deviation of 8 dB
Multipath fading Spatial Channel Model (SCM), Suburban macro
UE velocity 0 km/s
Transmitter/Receiver antenna 10 dBi (pedestrian), 2 dBi (indoor)
gain
Receiver antenna gain 10 dBi (pedestrian), 2 dBi (indoor)
Receiver noise figure 5 dB
Thermal noise density –174 dB m/Hz
Cable/connector/combiner 2 dB
losses
Physical profile: Operation mode FDD
Cyclic prefix type Normal (7 symbols per slot)
EPC bearer definitions 348 kbit/s (Non-GBR)
Subcarrier spacing 15 kHz
Admission control PDCCH symbols per subframe 3
parameters:
UL loading factor 1
DL loading factor 1
Inactive bearer timeout 20 s
Buffer status report Periodic timer 5 subframes
parameters:
Retransmission timer 2560 subframes
L1/L2 control parameters: Reserved size 2 RBs
Cyclic shifts 6
Starting RBP for format 1 0
messages
Allocation periodicity 5 subframes
Random access Number of preambles 64
parameters:
Preamble format Format 0 (1-subframe long)
Number of RA resources per 4
frame
Preamble retransmission limit 5 subframes
RA response timer 5 subframes
Contention resolution timer 40 subframes
HARQ parameters: Maximal number of 3 (uplink and downlink)
retransmissions
HARQ retransmission timer 8 subframes (uplink and downlink)
Maximal number of HARQ 8 per UE (uplink and downlink)
processes

and development package [15]. The CRM functionalities maintain admissibility and opportunism of SBs. The second
have been implemented in the evolved packet core (EPC) scheme, presented in [8], is used to balance transmission
connected to the eNBs and external server via 1 MBits/s IP delay against transmission power in the wireless channels
links. The primary eNBs operate on fixed licensed spectrum of the eNBs.
bands of the length B1 = 25 RBs, B2 = 50 RBs, B3 = 100 To differentiate the performance of different algorithms
RBs (spanning 5 MHz, 10 MHz and 20 MHz, respectively). in simulations we use the following notation in the sequel:
The radio model of the network has been developed
according to the ITU-T Recommendation M.1225. The – DPCPC (Duo Priority Class Power Control) for the
maximal transmission power of all eNBs (primary and sec- scheme described in [2];
ondary) is equal to 46 dB m. Other simulation parame- – DPC (Delay Power Control) for the scheme derived
ters are set in accordance with the requirements of the LTE in [8];
specifications [10] (the simulation parameters of the net- – SQC (Soft Quality of service based Control) for the
work model are listed in Table 1). scheme proposed in this paper with Iitar and Qitar
In this paper we observe and compare the performance calculated from (22) and (24), respectively, with T = 1;
of the proposed algorithm with the performance of – SPIC (Soft Past Information based Control) for the
two most relevant schemes applicable to the considered scheme proposed in this paper with Iitar and Qitar
system model. In the first scheme described in [2] the calculated from (22) and (24), respectively, with T =
power is allocated to the eNBs to protect QoS of PBs and 10.
 A. Asheralieva, K. Mahata / Physical Communication 14 (2015) 1–13 11

Fig. 4. Average number of iterations and solution time in the proposed algorithm.

Fig. 5. Average packet end-to-end delay in the PBs and the SBs.

All the algorithms are simulated with identical LTE consists of one object, whereas one image consists of
parameters and under identical network deployment five objects. The object size is constant and equal to
scenarios (such as channel quality, traffic load, and user 1000 bytes [33].
behavior). The target SINR is equal SINRtar
i = 10 dB in all
simulations. 4.2. Simulation results
The user traffic consists of three most frequently used
network applications: VoIP, video and HTTP. The number
First, we observe the complexity of the proposed algo-
of users of each type is distributed in proportion 2:3:5
rithm in terms of average number of iterations and the so-
for voice, video and data users, respectively. The following
lution time. Fig. 4 shows performance of SQC with different
models are used to simulate voice, video and web users:
number of PBs and SBs, n and m. Each eNB (PB or SB) serves
– The VoIP services model is ON–OFF model with ex- 50 users. Results show that the proposed algorithm has
ponentially distributed ON–OFF periods. The mean relatively low complexity (less than 70 iterations or 115
duration of ON and OFF periods is 0.65 s and 0.352 s, mcs for relatively large network comprising 15 × 4 eNBs).
respectively. The VoIP traffic is generated by using the These results are rather expected, since B&B was always
G.723.1 (12.2 kbps) codec with a voice payload size 40 described as one of the most effective and less complex al-
bytes and a voice payload interval 30 ms [33]. gorithms (see, for instance, reports on complexity of B&B
– Video services are simulated using a high resolu- provided in [28]).
tion video model with a constant frame size equal to We also evaluate performance of the different algo-
6250 bytes, and exponentially distributed frame inter- rithms in terms of average packet end-to-end delay, aver-
arrival intervals (with mean equal to 0.5 s) [33]. age transmission power and average SINR of the users in
– Web users in simulations are HTTP1.1 users generating PBs and SBs. Figs. 5–7 demonstrate simulation results in
pages or images with exponential page inter-arrival in- the network with n = 3 PBs with varying load, and m = 4
tervals (mean equal 60 s). It is assumed that one page SBs with constant load in SBs (each SB serves exactly 50
12 A. Asheralieva, K. Mahata / Physical Communication 14 (2015) 1–13

Fig. 6. Average transmission power in the PBs and the SBs.

Fig. 7. Average SINRs in the PBs and the SBs.

users). Results show that the duo-priority power control there is no protection of PBs in terms of interference from
scheme proposed in [2] provides consistent performance SBs, as well as no differentiation between the QoS require-
in terms of the average transmitted power and SINR for pri- ments of PBs and SBs [8]. Hence, performance of DPC com-
mary eNBs, but does not improve higher-layer QoS since pared to other schemes is relatively good when it comes to
the packet end-to-end delay in PBs is the largest when SBs, but not acceptable to PBs.
compared to other schemes. Such results can be explained Both schemes, SQC and SPIC proposed in this paper
as follows. Recall, that the objective of power allocation show very similar performance; they work better than
in DPCPC is to minimize the total transmission power DPCPC and DPC in reducing the packet end-to-end delay
of all eNBs (primary and secondary) subject to SNR re- and maintaining the target SINR levels for PBs. Such
quirements of PBs [2]. Consequently, we observe reduced performance can be easily reasoned by the fact that in
consumption of transmission power in the terminals be- our algorithm we take into account both SINR and QoS
longing to the users of PBs and SBs, and improved SINR requirements of PBs. We note in passing, that SPIC shows
conditions in the wireless channels of PBs. Unfortunately, slightly better performance that SQC in terms of packet
the delay and loss requirements of PBs are not considered end-to-end delay because it uses more long history for
in this scheme [2], and as a result, the users of all eNBs get resource allocation, and therefore is more effective in
significant delays during their packet transmissions. maintaining queue stability in PBs.
Delay-power control scheme proposed in [8] demon-
strates better performance than DPCPC in terms of packet 5. Conclusion
end-to-end delay for primary eNBs, but very ineffective
when it comes to SINR which is much lower than that in In this paper we consider a problem of transmission
other schemes. Results also show that the users of SBs and power and bandwidth allocation in a cognitive LTE based
PBs in average have similar consumption of transmission network consisting of primary and secondary eNBs with
power, SINR and delays. To understand such performance, different requirements on SINR and application-layer QoS.
note that in this scheme all eNBs are treated equally, and We propose a novel strategy for resource allocation in
 A. Asheralieva, K. Mahata / Physical Communication 14 (2015) 1–13 13

which the total queue size in eNBs (primary and sec- [19] F. Capozzi, et al., Downlink packet scheduling in LTE cellular
ondary) is minimized subject to interference requirements networks: key design issues and a survey, IEEE Commun. Surv. Tutor.
15 (2) (2013) 678–700.
and queue stability constraints of the primary eNBs. Per- [20] Physical Channels and Modulation. 3GPP TS 36.211. (Release 8).
formance of the algorithm has been evaluated using simu- [21] D.V. Lindley, The theory of queues with a single server, Math. Proc.
lations in different network deployment scenarios. Results Cambridge Philos. Soc. 48 (2) (1952) 277–289.
[22] J. Mo, J. Walrand, Fair end-to-end window-based congestion control,
show that the proposed algorithm outperforms previously IEEE/ACM Trans. Netw. 8 (5) (2000) 556–567.
proposed schemes. [23] I.F. Akyildiz, D.M. Gutierrez-Estevez, E.C. Reyes, The evolution to 4G
cellular systems: LTE-advanced, Phys. Commun. 3 (2010) 217–244.
[24] F. Baccelli, B. Blaszczyszyn, Stochastic Geometry and Wireless
References Networks, Volume I—Theory, Volume 3, in: Foundations and Trends
in Networking, vol. 3–4, NoW Publishers, 2009.
[1] M. Buddhikot, Understanding dynamic spectrum access: mod- [25] R.V.L. Hartley, Transmission of information, Bell Syst. Tech. J. (1928).
els, taxonomy, and challenges, in: Proc. IEEE DySPAN, 2007, [26] R. Bracewell, Heaviside’s unit step function, H (x), in: The Fourier
pp. 649–663. Transform and Its Applications, third ed., McGraw-Hill, New York,
2000, pp. 61–65.
[2] S. Sorooshyari, C.W. Tan, M. Chiang, Power control for cognitive radio
[27] P.-F. Verhulst, Notice sur la loi que la population poursuit dans son
networks: axioms, algorithms, and analysis, IEEE/ACM Trans. Netw.
accroissement, Correspondance mathématique et physique, Vol. 10,
20 (3) (2012) 878–891.
pp. 113–121, Retrieved 2009.
[3] S. Parsaeefard, A.R. Sharafat, Robust distributed power control in
[28] A.H. Land, A.G. Doig, An automatic method of solving discrete
cognitive radio networks, IEEE Trans. Mob. Comput. 12 (4) (2013)
programming problems, Econometrica 28 (3) (1960) 497–520.
609–620. [29] S.P. Boyd, L. Vandenberghe, Convex Optimization, Cambridge
[4] M. Zhou, X. Zhao, A power allocation algorithm based on target
University Press, 2004.
capacity, J. Inf. Comput. Sci. 10 (6) (2013) 1839–1845.
[5] S.Y. Lien, K.C. Chen, Statistical traffic control for cognitive radio [30] R.J. Vanderbei, D.F. Shanno, An interior-point algorithm for noncon-
vex nonlinear programming, Comput. Optim. Appl. 13 (1–3) (1999)
empowered LTE-advanced with network MIMO, in: IEEE INFOCOM,
231–252.
2011, pp. 80–84. [31] D.M. Gay, et al., A primal–dual interior method for nonconvex
[6] H.P. Shiang, M. van der Schaar, Queuing-based dynamic channel nonlinear programming, in: Advances in Nonlinear Programming,
selection for heterogeneous multimedia applications over cognitive in: Applied Optimization, vol. 14, 1998, pp. 31–56.
radio networks, IEEE Trans. Multimedia 10 (5) (2008) 896–909. [32] V. Jacobson, Congestion avoidance and control, in: Proc. SIGCOMM,
[7] A. Alshamrani, X. Shen, L.-L. Xie, QoS provisioning for heterogeneous Vol. 18 (4), 1988, pp. 314–329.
services in cooperative cognitive radio networks, IEEE J. Sel. Areas [33] IPOGUE, Internet study 2007 and 2008/2009, Research Report.
Commun. 29 (4) (2011) 819–830.
[8] F. Baccelli, N. Bambos, N. Gast, Distributed delay-power control
algorithms for bandwidth sharing in wireless networks, IEEE/ACM
Trans. Netw. 19 (5) (2011) 1458–1471. Alia Asheralieva has received her Master’s De-
[9] W. Wang, K.G. Shin, W. Wang, Distributed resource allocation based gree in Asian Institute of Technology in Bangkok,
on queue balancing in multihop cognitive radio networks, IEEE/ACM Thailand. She is currently pursuing her Ph.D. at
Trans. Netw. 20 (3) (2011) 837–850. University of Newcastle, Australia. Her research
[10] 3GPP TR 25.913. Requirements for Evolved UTRA (E-UTRA) and interests include Wireless Networks, Cognitive
Evolved UTRAN (E- UTRAN), Release 8. Networks, Performance Evaluation and Teletraf-
[11] E. Dahlman, S. Parkvall, J. Skod, 3G Evolution: HSPA and LTE for fic.
Mobile Broadband, Academic Press, Oxford, UK, 2007.
[12] G. Berardinelli, L.A. Ruiz de Temino, S. Frattasi, M.I. Rahman, P.
Mogensen, OFDMA vs. SC-FDMA: Performance comparison in local
area IMT-A scenarios, IEEE Wirel. Commun. 15 (5) (2008) 64–72.
[13] C.Y. Wong, R.S. Cheng, K.B. Lataief, R.D. Murch, Multiuser OFDM Kaushik Mahata received his Ph.D. from Upp-
with adaptive subcarrier, bit and power allocation, IEEE J. Sel. Areas sala University, Sweden in 2003. Since then he
Commun. 17 (10) (1999) 1747–1758. has been with the University of Newcastle, Aus-
[14] Anna Larmo, et al., The LTE link-layer design, IEEE Commun. Mag. 47 tralia. His research interests include Signal Pro-
(4) (2009) 52–59. cessing, System Identification, Optimization and
[15] OPNET Website: www.opnet.com. their applications.
[16] E-UTRA; MAC protocol specification. 3GPP TS 36.321. (Release 8).
[17] H. Holma, A. Toskala, LTE for UMTS: Evolution to LTE-Advanced, John
Wiley and Sons, 2011.
[18] S.-B. Lee, et al., Proportional fair frequency-domain packet schedul-
ing for 3GPP LTE Uplink, in: Proc. INFOCOM, 2009, pp. 2611–2615.