Upload
sahathermal6633
View
215
Download
2
Embed Size (px)
DESCRIPTION
Scheduling in a Cognitive Radio Networks
Citation preview
Scheduling in a Cognitive Radio Network
Dong Huang, Cyril Leung, Zhiqi Shen and Chunyan Miao
School of Computer Engineering, Nanyang Technological University, Singapore
Email: {dhuang, ASCYMiao}@ntu.edu.sgDepartment of Electrical and Computer Engineering, University of British Columbia
Vancouver, Canada, Email: [email protected] of Electrical and Electronic Engineering, Nanyang Technological University, Singapore
Email: [email protected]
AbstractIn this paper, we consider the channel access andpower allocation problem for secondary users in a time divisionmultiple access (TDMA) cognitive radio network. An algorithmis proposed to minimize the total transmit power while ensuringthat SU queue lengths are stable. Simulation results show thatthe proposed algorithm can achieve a good tradeoff between thepower consumption and the queue length.
I. INTRODUCTION
Dynamic power allocation is an effective method to transmit
data efciently over wireless channels and is important in
wireless communications due to the limited energy in wire-
less devices. Algorithms based on the concept of congestion
control have been successfully applied to the resource allo-
cation problem in wireless networks [1], [2], [3], [4]. These
algorithms optimize the network performance by fully utilizing
the network capacity and avoiding excessive congestion inside
the network. In addition, the network capacity region (i.e.,
the set of all feasible data rates) has been shown to be a
powerful technique for the resource allocation problem in
wireless networks [5], [6].
Cognitive radio (CR) is a promising technology for improv-
ing the spectrum utilization by allowing secondary users (SUs)
to access unoccupied licensed channels opportunistically. The
resource management in CR has been studied widely [7], [8],
[9], [10]. A decentralized cognitive MAC protocol based on
the theory of Partially Observable Markov Decision Process
(POMDP) is proposed in [7]. The goal of the proposed
cognitive MAC protocol is to optimize the performance of
secondary users while limiting the interference perceived by
primary users (PUs). The impacts of the transmission power
of SUs on the occurrence of spectrum opportunities and the
reliability of opportunity detection are investigated in [8]. In
order to maximize the throughput of SUs, an opportunistic
scheduling policy for cognitive radio networks is developed in
[9]. In [10], the rate adaptation problem in a cognitive radio
network is formulated as a general sum constrained switching
control Markovian dynamic game. Most of these works are
based on Markov Decision Process (MDP). Finding the opti-
mal policy may be computationally prohibitive. Suboptimal
algorithms with low complexity may be more suitable for
online implementation in wireless networks.
In this paper, we consider a Time Division Multiple Access
(TDMA) cognitive radio system model [10], where multiple
SUs compete to access an unlicensed channel at each time
slot. To provide a distinct tradeoff in power consumption and
queue length, we develop an algorithm for the access rule and
rate adaptation problem by introducing a control parameter. By
varying the control parameter value, we can achieve a good
tradeoff between the power consumption and the queue length.
The rest of the paper is structured as follows: The network
model and problem denition are presented in Section II.
In Section III, the achievable rates for a transmission policy
in which the SUs transmit with xed equal powers are rst
discussed. An algorithm with variable transmit power is then
proposed. Simulation results are presented in Section IV, and
conclusions are provided in Section V.
II. NETWORK MODEL AND PROBLEM DEFINITION
Consider a TDMA cognitive radio network with N SUlink pairs; the SUs share a common channel with PUs. SUs
are allowed to access the channel when they sense that the
channel is not being used by PUs. We assume that at most
one SU accesses the channel in each time slot, i.e. there are no
collisions among SUs. SUs can obtain historical PU channel
usage information from the primary user base station (PUBS).
SUs are able to exchange information, such as individual
SU link pair channel gains and individual SU queue lengths,
through a control channel. For brevity, we dene the set
N = {1, 2, . . . , N}. For SU n, the number of bits arrivingin time slot t is dn(t). We assume dn(t) are stationary andergodic with rate n, i.e., limT
1T
T1t=0 dn(t) = n. We
also assume that the number of arriving bits for all SUs
at every time slot is bounded and there exists a constant
dmax > 0 such that dn(t) dmax, n N. Let hn(t) bethe channel quality state of user n and it is assumed thathn(t) H = {H1, . . . , HM}, n N.
Let S(t) be the channel indicator variable. When the chan-nel is available, i.e., PU does not transmit, S(t) = 1; S(t) = 0otherwise. The channel availability is modeled as a two state
Markov chain, as shown in Fig. 1. Let q denote the probabilitythat the channel is not available, then q = /( + ). Themaximum allowable probability of collision with an SU given
that the PU transmits is .
Let Pn(t) [0, Pmax] be the transmit power of SU n in ___________________________________ 978-1-61284-307-0/11/$26.00 2011 IEEE
Fig. 1: Channel availability model.
time slot t. The corresponding bit rate is given by [11]
(Pn(t), hn(t)) = log2
(1 +
hn(t)Pn(t)
N0
), (1)
where N0 denotes the noise power and hn(t) is the channelgain. Let Xn(t) be the number of bits in the queue of SU nat the beginning of time slot t. Then,
Xn(t + 1)
= max [Xn(t) an(t)(Pn(t), hn(t))S(t), 0] + dn(t) (2)
subject to
Nn=1
an(t) 1, (3)
where an(t) {0, 1} is an allocation indicator variable. Whenthe channel is allocated to SU n at time slot t, an(t) = 1,otherwise, an(t) = 0. Constraint (3) states that at most oneSU can access the channel in any given time slot. The average
SU queue length is
X = limT
1
T
T1t=0
Nn=1
Xn(t). (4)
According to [12], the queue is stable if
limt
P(Xn(t) < x) = F (x), and limx
F (x) = 1. (5)
In this paper, we use Loynes theorem [13], which states that
when the arrival and service processes in a queueing system
are strictly stationary and ergodic, the queue is stable if and
only if the average arrival rate is strictly less than the average
service rate [14]. Let (t) be the collision indicator, where
(t) =
{1 if an SU transmits and S(t) = 0,
0 otherwise.(6)
Then the SU collision rate is
= limT
1
T
T1t=0
(t). (7)
Let the average SU transmit power be
P = limT
1
T
T1t=0
Nn=1
Pn(t) (8)
and the average bit rate for SU n be
n = limT
1
T
T1t=0
an(t) (Pn(t), hn(t))S(t), n N.
(9)
Our objective is to minimize the value of P
minP , (10)
subject to
n > n, n N (11)
P (collision|PU transmitting) . (12)
We assume the existence of a feasible solution space for the
problem in (10)-(12). In order to ensure that constraint (12)
is satised, we introduce a process Y (t), with Y (0) = 0,which evolves as follows: (1) If SUs do not transmit when
PU is transmitting, Y (t + 1) is lower than Y (t) by as longas Y (t) ; (2) if SUs transmit when PU is transmitting,Y (t + 1) is set higher than Y (t) by 1 when Y (t) and 1 when Y (t) < . Therefore, the dynamics of Y (t) canbe expressed as:
Y (t + 1) = max [Y (t) (1 S(t)), 0] + (t). (13)
When the system (13) is stable, i.e.,
q = limT
1
T
T1t=0
(1 S(t))
> limT
1
T
T1t=0
(t) = qP (collision|PU transmitting),
(14)
constraint (12) is satised. It can be seen that any feasible so-
lution to the constraints (11)-(12) should stabilize the systems
in (2) and (13).
Let Pdenote the optimal solution to the problem, where
= [1, 2, . . . , N ] is the data arrival rate vector. It isdifcult to solve the optimization problem in (10)-(12) due
to the fact that only partial channel state information (CSI)
is available to the SUs. Similar problems have been modeled
as partially observable Markov decision processes (POMDPs)
[7], [15]. However, nding an optimal solution for a general
POMDP problem can be computationally prohibitive. For
suboptimal algorithms with low complexity such as greedy
algorithm (i.e., an SU with the best channel quality in each
time slot has the highest priority to access the channel when
the channel is available) and longest queue algorithm (i.e.,
an SU with the longest queue length in each time slot has
the highest priority to access the channel when the channel is
available) cannot solve the problem effectively. For example,
an SU may not transmit any data when its channel quality is
always in Bad state.
In the following section, we study the impact of sensing
on the achievable rate region and the associated average
power consumption. The proposed access and power control
algorithm (APCA) is then discussed.
III. MAIN RESULTS
Note that sensing results at SUs have a signicant impact
on the optimal access rule. In this section, we rst study the
achievable rate region of (11)-(12) under a simple scenario
and transmission policy.
Scenario (S1): Sensing is conducted at SUs. The probability
of false alarm (PFA) is P (S(t) = 0|S(t) = 1) = and theprobability of miss detection (PMD) is P (S(t) = 1|S(t) =0) = , where S(t) denotes the sensing result at SUs at slott. S(t) = 1 when the channel is observed to be available, andS(t) = 0 otherwise.Transmission Policy (P1): When there is incoming data,
if S(t) = 1, a selected SU transmits data with probabilityP (SU transmitting|S(t) = 1) = > 0.
Lemma 1. Assume each SU transmits data at a xed power
P . The average data transmission rate of SU n is equal ton/(1 q) when the channel is available. Under Scenario S1and Transmission Policy P1, the maximum of for which con-straint (12) is satised is max = min(/, 1). The achievablerate region of constraints (11) and (12) is characterized by
Nn=1
nmax(1 )n
< 1. (15)
When is interior to , the system can be stabilized with
average power P =N
n=1[(1)(1q)+q]n
(1)nP .
Proof: For the constraint (12), it is required that the
inequality P (Transmitting|S(t) = 0) = P (Transmitting|S =1)P (S = 1|S(t) = 0) holds. If < , we get max = 1.If , we get / and derive max = /. Thus, weget max = min(/, 1).From transmission policy P1, we get
P (SU transmitting|S(t) = 0) = 0. Therefore,P (SU transmitting, S(t) = 1) = P (SU transmitting, S(t) =1, S(t) = 1) = P (SU transmitting|S(t) = 1)P (S(t) =1, S(t) = 1) = (1 q)(1 ). The average total number oftime slots needed for SU n in order to transmit n/(1 q)units of data is 1/((1q)(1)). The fraction of time that theserver is in busy state is equal to
Nn=1
n(1)n
and it should
be less than 1 for stability. Since P (SU transmitting, S(t) =0) = P (SU transmitting, S(t) = 1, S(t) = 0) =P (SU transmitting|S(t) = 1)P (S(t) = 1, S(t) = 0) = q,we get P (SU transmitting) = [(1 q)(1 ) + q]. Theactual average number of transmitting slots needed for SU nin order to transmit n/(1 q) units of data is
(1)(1q)+q(1)(1q) .
In this case, the fraction of time that the server is in busy
state is equal toN
n=1[(1)(1q)+q]n
(1)n. The associated
average power consumption isN
n=1[(1)(1q)+q]n
(1)nP .
Based on some assumptions (e.g., SUs transmit data with
xed power), Lemma 1 can not only provide the achievable
rate region for the problem in (10)-(12), but also give the av-
erage power consumption under scenario S1 and transmission
policy P1. If the assumption that SUs transmit data with xed
power is relaxed, its not easy to derive the corresponding
average power consumption. But when SUs transmit data
with dynamic power control, the average power consumption
should be less than or equal to that in the case of xed power
transmission.
In the following subsection, we assume that SUs transmit
under dynamic power control. We analyze the dynamics of
the systems in (2) and (13), and propose an APCA algorithm
based on Lyapunov optimization.
A. The Proposed Algorithm
Let V 0 be a xed control parameter. The proposedAPCA algorithm is based on solving the following optimiza-
tion problem
min f(Q(t),P(t), a(t))
= N
n=1
Xn(t) E [an(t)(Pn(t), hn(t))S(t)|Q(t)]
+ Y (t) E
[(1 S(t))
Nn=1
an(t)
]+ V E
[N
n=1
Pn(t)|Q(t)
],
(16)
with
Q(t) = [X1(t), X2(t), . . . , XN(t), Y (t)], (17)
where > 0 is a constant weighting factor, P(t) =[P1(t), . . . , PN (t)] denotes the power allocation vector anda(t) = [a1(t), . . . , aN (t)] is the channel allocation indicatorvector for time slot t. In each time slot, we need to determinethe values of an(t) and Pn(t) such that the objective functionis minimized. It can be seen that whether SUs access the
channel depends on not only the historical statistics of the
channel, but also on the current queue lengths.
SinceN
n=1 an(t) 1 and an(t) {0, 1} n N, we minimize the objective function in (16) by the
following steps: First, set an(t) = 1 and aj(t) =0, j N {n}. The corresponding objective func-tion is denoted by fn(Q(t),P(t), a(t)). In addition, weset an(t) = 0 n N. The corresponding objec-tive function is denoted by f0(Q(t),P(t), a(t)). Then wecompare fn(Q(t),P(t), a(t)), n = 0, 1, . . . , N . Assumefn(Q(t),P(t), a(t)) is the minimum. If n = 0, then we setan(t) = 1, and an(t) = 0, n N, otherwise. Specically, itcan be achieved by solving the following problem:
j = argminn
Un(t), (18)
where
Un(t) =
{n(t) if n(t) < 0
0 otherwise,
and
n(t) = minPn(t)
{Y (t)(1 S(t)) + V Pn(t)
Xn(t)S(t)(Pn(t), hn(t))}.
S(t) denotes the expectation of S(t) at time slot t. After theabove problem is solved, the access rule and power control
are as follows:
an(t) =
{1 if n = j and Un(t) < 0
0 otherwise,(19)
and
Pn(t) =
argminPn(t)
n(t) if an(t) = 1
0 otherwise.(20)
Based on the above discussion, we now use the following
theorem to illustrate how we derive the objective function in
(16) and characterize the performance of the APCA algorithm.
Specically, it provides an upper bound for the average queue
length and power consumption.
Theorem 1. Let be an achievable rate region of problem
(10)-(12). Assume is interior to . For any control pa-
rameter V > 0, the solution derived by the proposed APCAalgorithm can stabilize the system in (2) and (13) has the
following inequalities:
limT
sup1
T
T1t=0
E
[N
n=1
Xn(t) + Y (t)
]
B + V Pmaxmax
,
(21)
limT
sup1
T
T1t=0
E
[N
n=1
Pn(t)
]
B
V+ P, (22)
where B =(N + d2max + (
2 + 1)2)/2 and max is the
largest value such that (+ max) .
Proof: Suppose the state vector is dened as (17). In this
paper, we set = 5. The Lyapunov function can be denedas
L(Q(t)) =1
2Q(t)QT (t) =
1
2
[N
n=1
X2n(t) + 2Y 2(t)
],
(23)
where QT (t) denotes the transpose of Q(t). Assume H1is the best channel quality, then (Pn(t), hn(t)) =(Pmax, H1). Since (t) 1 and X
2n(t + 1) (Xn(t)
(Pn(t), hn(t)))2+d2n(t), the Lyapunov drift can be expressed
as
(Q(t)) =E [L(Q(t + 1)) L(Q(t))|Q(t)]
Nn=1
Xn(t) E[an(t)(Pn(t), hn(t))S(t)
dn(t)|Q(t)] + B
2Y (t) E [((1 S(t)) (t))|Q(t)] , (24)
where B =(N + d2max + (
2 + 1)2)/2. From (6), and
(19)-(20), we have
(t) =
Nn=1
an(t)1[n(Pn(t),h(t))>0](1 S(t))
=N
n=1
an(t)(1 S(t)),
where 1[] is an indicator function. 1[] = 1 when the condition[] is true, and 0 otherwise. Therefore, (24) can be changed to
(Q(t)) B +N
n=1
nXn(t) q2Y (t)
Nn=1
Xn(t) E [an(t)(Pn(t), hn(t))S(t)|Q(t)]
+ 2Y (t)
Nn=1
E [an(t)(1 S(t))|Q(t)] . (25)
Motivated by the Lyapunov optimization technique developed
in [4], we add V E[N
n=1 Pn(t)|Q(t)]with V > 0 to both
sides of (25), and obtain
(Q(t)) + V E
[N
n=1
Pn(t)|Q(t)
]
B +N
n=1
nXn(t) q2Y (t) + V E
[N
n=1
Pn(t)|Q(t)
]
N
n=1
Xn(t) E [an(t)(Pn(t), hn(t))S(t)|Q(t)]
+ 2Y (t)
Nn=1
E [an(t)(1 S(t))|Q(t)]
B(t)
Nn=1
Xn(t) E [an(t)(Pn(t), hn(t))S(t)|Q(t)]
+ 2Y (t)
Nn=1
E [an(t)(1 S(t))|Q(t)]
+ V E
[N
n=1
Pn(t)|Q(t)
], (26)
where B(t) = B +N
n=1 nXn(t) q2Y (t).
Since is interior to , there exist a strategy an(t), Pn (t)
such that
E
[N
n=1
Pn (t)
]= P +, (27)
E[an(t)(P
n (t), hn(t))S(t)
] n + , (28)
E[(t)
]+ q. (29)
Fig. 2: The Markov channel quality model.
Therefore, we get
(Q(t)) + V E
[N
n=1
Pn(t)|Q(t)
]
B +N
n=1
nXn(t) q2Y (t)
N
n=1
Xn(t) E[an(t)n(P
n (t), hn(t))|Q(t)
]
+ 2Y (t)N
n=1
E[an(t)(1 S(t))|Q(t)
]+ V P +
B
(N
n=1
Xn(t) + Y (t)
)+ V P +. (30)
Taking the expectation of (30) overQ(t) and summing it fromt = 0 to t = T 1 and divided by T , we get
L(Q(T )) L(Q(0))
T+ V
T1t=0
E
[N
n=1
Pn(t)
]
B + V P +
T
T1t=0
E
[N
n=1
Xn(t) + Y (t)
]. (31)
The above inequality always holds when + . Letmax be the largest value satisfying + max . Notethat
Nn=1 Pn(t) Pmax. By setting = 0 and = max,
inequalities (21) and (22) follow.
Theorem 1 indicates that the total average queue length
increases with V whereas the average power consumptionapproaches the minimum power as V increases. Therefore,adjusting the value of V allows a tradeoff between the queuelength and the power consumption.
IV. PERFORMANCE EVALUATION
In this section, we present computer simulation results to
illustrate the results in Section III. We consider a TDMA
cognitive radio system with N = 5 secondary link pairs. Thechannel state S(t) is characterized by a 0/1 Markov chain, asshown in Fig. 1, with = 0.7 and = 0.3. Data arrival ratefor every link is the same (i.e., 1 = 2 = = 5 = ). Thechannel quality for each secondary link pair is quantized into
three different states, namely {H1, H2, H3}, where H1 meansthe channel is in Good state, H2 is in Medium state andH3 is in Bad state. The channel quality model is the 3-state Markov chain shown in Fig. 2. The transition probability
0.05 0.1 0.15 0.2 0.250
5
10
15
20
25
30
Data Arrival Rate, (bits/slot)
Ave
rag
e S
U T
ran
sm
it P
ow
er,
P (
mW
)
V=0.1
V=1
V=5
V=20
Fig. 3: Average total SU power consumption versus arrival
data rate for different values of V .
matrix is
H =
0.4 0.4 0.20.3 0.5 0.2
0.2 0.5 0.3
.
Transitions between channel quality states for different link
pairs are assumed to be independent. We also set = 0.05,Pmax = 200 mW and N0 = 100 dBm. The simulation isconducted over T = 50000 time slots.The average total SU transmit power dened in (8) and the
average total SU queue length dened in (4) are plotted as
a function of the bit arrival rate, , in Figs. 3 and 4. FromFig. 3, it can be seen that for a given value of V , the averagetransmit power increases with the bit arrival rate due to the
increased transmission requirements. For a given bit arrival
rate, the average power decreases with the control parameter
V . It can be seen from Fig. 4 that the average queue lengthincreases with the bit arrival rate for a given value of V ; fora given value of , the average queue length increases withV . An appropriate value of V can be selected to achieve adesired tradeoff between the average transmit power and the
average queue length.
V. CONCLUSION
In this paper, the SU access and power allocation scheme
in a TDMA cognitive radio system was studied. An algo-
rithm was proposed which allows a tradeoff between power
consumption and queue length. The solution derived by the
algorithm can be a good solution when the control parameter
is chosen appropriately.
REFERENCES
[1] X. Lin and N. Shroff, The Impact of Imperfect Scheduling on Cross-Layer Congestion Control in Wireless Networks, IEEE/ACM Transac-tions on Networking, vol. 14, no. 2, pp. 302315, 2006.
[2] L. Georgiadis, M. Neely, and L. Tassiulas, Resource Allocation andCross-Layer Control in Wireless Networks, Foundations and Trends inNetworking, vol. 1, no. 1, 2006.
0.05 0.1 0.15 0.2 0.2510
1
102
103
104
Data Arrival Rate, (bits/slot)
Ave
rag
e S
U Q
ue
ue
Le
ng
th,
U (
bits)
V=0.1
V=1
V=5
V=20
Fig. 4: Average total SU queue length versus arrival data rate
for different values of V .
[3] L. Chen, S. Low, and J. Doyle, Joint Congestion Control and MediaAccess Control Design for Ad Hoc Wireless Networks, in 24th AnnualJoint Conference of the IEEE Computer and Communications Societies,vol. 3, pp. 22122222, 2005.
[4] C. Li and M. Neely, Energy-Optimal Scheduling with Dynamic Chan-nel Acquisition in Wireless Downlinks, IEEE Transactions on MobileComputing, pp. 527539, 2009.
[5] S. Sarkar and L. Tassiulas, End-to-End Bandwidth Guarantees throughFair Local Spectrum Share in Wireless Ad-Hoc Networks, in 42nd IEEEConference on Decision and Control, vol. 1, pp. 564569, 2003.
[6] Y. Yi and S. Shakkottai, Hop-by-Hop Congestion Control over aWireless Multi-Hop Network, in 32nd Annual Joint Conference of theIEEE Computer and Communications Societies, vol. 4, pp. 25482558,2004.
[7] Q. Zhao, L. Tong, A. Swami, and Y. Chen, Decentralized CognitiveMAC for Opportunistic Spectrum Access in Ad Hoc Networks: APOMDP Framework, IEEE Journal on Selected Areas in Communi-cations, vol. 25, no. 3, pp. 589600, 2007.
[8] W. Ren, Q. Zhao, and A. Swami, Power Control in Cognitive RadioNetworks: How to Cross a Multi-Lane Highway, IEEE Journal onSelected Areas in Communications, vol. 27, no. 7, pp. 12831296, 2009.
[9] R. Urgaonkar and M. Neely, Opportunistic Scheduling with ReliabilityGuarantees in Cognitive Radio Networks, IEEE Transactions on MobileComputing, pp. 766777, 2009.
[10] J. Huang and V. Krishnamurthy, Transmission Control in CognitiveRadio as a Markovian Dynamic Game: Structural Result on RandomizedThreshold Policies, IEEE Transactions on Communications, vol. 58,no. 1, pp. 301310, 2010.
[11] M. Neely, E. Modiano, and C. Rohrs, Dynamic Power Allocationand Routing for Time-Varying Wireless Networks, IEEE Journal onSelected Areas in Communications, vol. 23, no. 1, pp. 89103, 2005.
[12] W. Szpankowski, Stability Conditions for Some Distributed Systems:Buffered Random Access Systems, Advances in Applied Probability,vol. 26, no. 2, pp. 498515, 1994.
[13] R. Loynes, The Stability of a Queue with Non-Independent Inter-Arrival and Service Times, in Mathematical Proceedings of the Cam-bridge Philosophical Society, vol. 58, pp. 497520, Cambridge Univer-sity Press, 1962.
[14] W. Luo and A. Ephremides, Stability of N Interacting Queues inRandom-Access Systems, IEEE Transactions on Information Theory,vol. 45, no. 5, pp. 15791587, 1999.
[15] Y. Chen, Q. Zhao, and A. Swami, Joint Design and Separation Principlefor Opportunistic Spectrum Access in the Presence of Sensing Errors,IEEE Transactions on Information Theory, vol. 54, no. 5, pp. 20532071, 2008.