Vous êtes sur la page 1sur 5

Adaptive Scheduling of Spectrum Sensing Periods in Cognitive Radio Networks

Anh Tuan Hoang and Ying-Chang Liang


Institute for Infocomm Research - 21 Heng Mui Keng Terrace, Singapore 119613 Email: {athoang, ycliang}@i2r.a-star.edu.sg

Abstract We consider a cognitive radio (CR) network that makes opportunistic use of a set of channels licensed to some primary network. During its operation, the CR network is required to carry out spectrum sensing to detect active primary users, thereby avoiding interfering with them. However, spectrum sensing can cause negative effects to the performance of the CR network, as all CR communication has to be postponed during channel sensing. This paper focuses on adaptively scheduling the spectrum sensing periods so that negative impacts to the performance of the CR network are minimized. We rst consider the case when CR nodes always have data to transmit and experience timevarying channels. Based on knowledge of channel conditions, the sensing periods are adaptively scheduled to maximize the spectrum efciency of the CR operation. We then consider the case when CR nodes experience both stochastic data arrival and time-varying channels. By treating each sensing period as a virtual sensing packet, we convert the problem of joint data-transmission/spectrum-sensing scheduling into a standard queueing model. Based on that, efcient scheduling algorithms that take into account channel and queueing conditions of the CR network to achieve good quality of service are proposed.

I. I NTRODUCTION The traditional approach of xed spectrum allocation to licensed networks leads to spectrum under-utilization. In recent studies by the FCC, it is reported that there are vast temporal and spatial variations in the usage of allocated spectrum, which can be as low as 15% [1]. This motivates the concepts of opportunistic spectrum access that allows cognitive radio (CR) networks to opportunistically exploit under-utilized spectrum. On the one hand, opportunistic spectrum access can improve the overall spectrum utilization. On the other hand, transmission from CR networks can cause harmful interference to primary users of the spectrum. To mitigate such a problem, CR networks can frequently carry out spectrum sensing to detect primary users. Upon detecting a co-channel primary user, a CR network can either change its operating parameters, e.g., reduce its transmit powers, or move to another channel to avoid interfering with the primary user. In most cases, to achieve reliable spectrum sensing for a particular channel, a CR network has to postpone all of its transmissions on that channel. This leads to so called quiet sensing periods. It can be noted that quiet sensing periods result in negative impacts to various performance metrics of CR networks, such as throughput and latency. One approach to reduce these negative impacts is to design efcient spectrum sensing algorithms that require shorter sensing time. Given a spectrum sensing algorithm and a xed duration for each quiet sensing period, the impacts of spectrum sensing depend on the actual state of a CR network at the time. For example, scheduling of sensing periods reduces throughput of a CR network, however, this throughput depends on the quality of the channels when transmission or sensing activities are carried out. Similarly, scheduling a sensing period increases packet latency, however, the actual queuing delay depends on the buffer states of CR nodes. Despite the importance of spectrum-sensing scheduling, to the best of our knowledge, there has not been any work in the existing literature that addresses

this problem. This motivates us to study the problem of adaptive scheduling quiet sensing periods in order to optimize the performance of CR networks. We rst consider a CR network in which nodes always have data to transmit and experience time-varying channels. Data can be transmitted at different rates for different channel conditions. Assuming that instantaneous channel conditions are available, we adaptively schedule sensing periods so that the throughput of the CR network is maximized. This is based on the observation that transmitting data when the channels are good and carrying out spectrum sensing when the channels are in poorer conditions would increase the spectrum efciency of CR operation. We term this approach CSI-based sensing scheduling, where CSI stands for channel state information. For a nite state Markov channel (FSMC) model [2], we formulate the optimization problem as a nite-horizon dynamic programming and obtain optimal CSI-based sensing scheduling policies. Next, we consider a CR network in which users experience both stochastic data arrival and time-varying channels. Data packets of CR nodes have deadline constraints and the objective is to jointly scheduling data transmission and spectrum sensing activities of the CR network so that the packet loss due to deadline violation is minimized. Here, joint data-transmission/spectrum-sensing decisions are based on both channel and buffer conditions of the CR nodes. To tackle this problem, we introduce a novel idea of treating each quiet sensing period as a virtual sensing packet. Then, carrying out spectrum sensing on a particular channel is equivalent to transmitting a virtual sensing (VS) packet on that channel. This simple but useful abstraction enables us to translate the problem of joint data-transmission/spectrum-sensing scheduling into a parallel server allocation problem. Based on this, we propose efcient scheduling algorithms that take into account the channel and trafc conditions of the CR network to achieve good quality of service. We term this approach QCI-based scheduling, where QCI stands for queueing and channel information. The rest of this paper is organized as follows. In Section II, we discuss spectrum sensing operation in CR networks and describe our general system model. In Section III, the problem of adaptively scheduling sensing periods based on channel state information is considered. We then consider a scenario when CR nodes have stochastic data arrival in Section IV. Here, the sensing scheduling decisions are made based on both queueing and channel state information. Numerical results are presented in Section V. Finally, we conclude the paper in Section VI II. S PECTRUM S ENSING IN CR N ETWORKS There are various spectrum sensing algorithms that differ from each other in the required sensing duration, detection accuracy, and complexity. Let us briey go through important parameters that characterize spectrum sensing in CR networks.

3128 1930-529X/07/$25.00 2007 IEEE


This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2007 proceedings.

CH1

Comms.

Sensing

Comms.

Comms.

Sensing Comms.

CH2

Comms.

Sensing

Comms.

Comms.

Sensing

Comms.

CH3

Comms.

Sensing

Comms.

Comms.

Sensing

Comms.

Frame n

Frame n+1

Time

Fig. 1. An example of a CR network with multiple channels. Time is divided into frames. During each frame, one sensing period must be scheduled for each channel.

A. Parameters of Spectrum Sensing 1) Signal to Noise Ratio (SNR): When a primary user is active, the higher the SNR of the primary users signal at the receiver of a CR device, the easier it is to detect. We denote this SNR by . In practice, as is not exactly known at CR devices, sensing algorithms are usually designed for the worst-case SNR. 2) Probability of Detection Pd : This is the probability that a CR network accurately detects the presence of an active primary user. The higher the value of Pd , the better the protection for primary operation. 3) Probability of False Alarm Pf : This is the probability that a CR network falsely detects the presence of primary users when in fact none of them are active at the sensing time. From the CR network point of view, the lower the value of Pf , the higher the spectrum utilization. 4) Sensing Time Ts : This is the time that a CR network needs to postpone all communications to sense a channel. In general, the longer the value of Ts , the more accurate the sensing outcome. Note also that Ts may not need to be contiguous overtime. As an example, assume that: i) the primary signal is independent and identically distributed (i.i.d), complex PSK modulated, with zero mean; ii) the noise at a CR receiver is circular symmetric complex Gaussian with zero mean; and iii) the primary signal and noise are independent. Then, using energy detection algorithm, the minimum sensing time to achieve probabilities of detection and false alarm of Pd and Pf , respectively, is ( [3]): Ts = Q1 (Pf ) Q1 (Pd ) 2 2 + 1
2

point of view, the set of available channels will change if either of the following two events occur: E1: A primary user starts operation on a particular channel and is detected by the CR network. E2: No primary user starts operation, however, due to false alarm, the CR network mistakenly concludes that some channel is not available anymore. In this paper, we assume that events E1 and E2 occur so rarely that, when compared to the time-scale at which data transmission is carried out, we tend to have a quasi-static set of K available channels. We argue that this assumption holds in many practical cases. For example, the currently being developed IEEE 802.22 [4] standard let CR networks to operate on unused TV VHF/UHF channels, which do not change status on daily basis. Also, if the probability of false alarm Pf a is set to a low enough value, e.g., 1%, it takes hundreds of frames for an event E2 to occur. III. CSI- BASED S ENSING S CHEDULING In this section, we consider the case when the channel conditions of N CR nodes vary over time. The better the channel condition, the higher the rate a CR node can transmit at. We are focusing on adaptively scheduling quiet sensing periods, based on the channel conditions of CR nodes, so that the system throughput is maximized. We term this approach CSI-based sensing scheduling. A. Problem Formulation We assume that each frame of duration Tf is further divided into L slots of size Tc = Tf /L. Furthermore, during each frame, one slot must be scheduled for quiet sensing. The scheduling decisions are made in a slot-by-slot basis. 1) Channel Model: For each pair of user n and channel k, 1 n N, 1 k K, the channel model is represented by a nite state Markov chain (FSMC) [2]. In particular, the channel can take M possible states and when in state m, user n can transmit at rate rm on channel k. It is assumed that the channel state stays unchanged during each time slot and the probability of transitioning from state i into state j after each time slot is ij . For the sake of brevity and without loss of generality, we assume that the FSMC channel model has the same statistics for all pairs (n, k). 2) Control and Objective: Let mk (i) denote the state of channel n k for user n during slot ith , 1 i L, of a particular frame. Let M (i) = [mk (i)] be the N K matrix of channel states for all n pairs (n, k) during slot i. We dene the following CSI-based sensing problem. Denition 1: CSI-based Sensing Scheduling Problem - For each slot i within a frame, given that the channel state information M (i) is available, on each channel k, make a decision on whether to carry out sensing or to allow a user to carry out data transmission so that the achievable throughput is maximized. This is subject to the constraint that, for each channel, at least one slot within the frame must be scheduled for quiet sensing. For the above problem, it is clear that the following observations are true for an optimal control policy. First, exactly one sensing slot will be scheduled during each time frame for each channel. Second, if sensing is not carried out on a channel, then the channel should be assigned to the CR node that can transmit at the highest rate. Third, the problem can be decoupled into K control problems, one for each channel, and for channel k, the sensing scheduling decision for slot i is made based on the (N 1) vector of channel states m k (i) = [mk (i), mk (i), . . . mk (i)]. 1 2 N

(1)

where is the channel sampling interval and Q(.) is the complementary distribution function of a standard Gaussian variable, i.e., 1 Q(x) = 2
x

exp(

t2 )dt. 2

(2)

5) Detection Time Td : This is the time taken for a CR network to detect the presence of a primary user since it rst turns on. B. General System Model We consider a CR network consisting of N nodes. These CR nodes carry out opportunistic transmissions on a set of K channels that are licensed to some primary network. The operation of the CR network is depicted in Fig. 1. In particular, time is divided into frames of duration Tf seconds. During each frame, Ts seconds must be scheduled for spectrum sensing on each of the K channels. Spectrum sensing is carried out by at least one of the N CR nodes. We assume that Tf and Ts have been predetermined to meet the requirements for primary user detection, in terms of Pd , Pf a , and Td . The parts of each frame that are not used for channel sensing can be used for CR communications. At each time instance, each channel can only be used by at most one CR node. From the CR operation

3129 1930-529X/07/$25.00 2007 IEEE


This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2007 proceedings.

B. Optimal Scheduling Policies The CSI-based sensing scheduling problem for channel k can be formulated as a nite-horizon dynamic programing ( [5]) with the following components: 1) Control Horizon: The control horizon is L slots, i.e., one time frame. 2) System States: For each frame, let f k (i) be the sensing ag for slot i, where f k (i) = 1 indicates that sensing has been carried prior to slot i and f k (i) = 0 otherwise. The system state of our control problem comprises of the channel state vector m k (i) and the sensing ag f k (i), i.e., sk (i) = [f k (i), mk (i), mk (i), . . . mk (i)]. Note that 1 2 N sk (i) contains all information needed for making a CSI-based sensing scheduling decision for channel k during slot i. 3) Actions and Rewards: For each time slot, there are two control actions that can be taken, i.e., carrying out sensing, denoted by as , and carrying out transmission, denoted by at . Corresponding to each of these actions are the following rewards in throughput: R(sk (i), a) = maxn {k (i)}, n 0, if a = at if a = as , (3)
Spectrum Efficiency (b/s/Hz)

5.5 5 4.5 4 3.5 3 2.5 2 1.5 1 8

Optimal Greedy Fixed

Sensing time=20% of frame size

Sensing time=50% of frame size


10 12 14 16 18 20

Median SNR (dB)

Fig. 2. Comparing the spectrum efciency of different sensing scheduling policies. Slot 1 Slot 2 Slot 3 Slot 4 Slot 5 (last) SNR = 6 dB S S S S S SNR = 9 dB T S S S S SNR = 12 dB T T T S S SNR = 15 dB T T T T S SNR = 18 dB T T T T S TABLE I S TRUCTURE OF O PTIMAL CSI- BASED SENSING SCHEDULING POLICY.

where k (i) = rmk (i) is the transmission rate of user n on channel n n k during slot i. Note that, as one sensing slot must be carried out during each frame, if at slot L, the sensing ag f k (L) = 0 then only the sensing action as is allowed. 4) State Transition Probabilities: The transitioning probabilities of the system state sk (i) can be readily derived given the transitioning probabilities of channel states, i.e., ij . 5) Obtaining Optimal Solutions: Given the above nite-horizon dynamic programming (FHDP) setup, an optimal control policies that maximize the total reward (i.e., total throughput of each frame) can be obtained using standard techniques for solving FHDP. We refer the readers to [5] for an in-depth study of these techniques. C. Performance and Structure of CSI-based Sensing In order to assess the performance of the optimal CSI-based sensing scheduling approach, we compare that to the following control approaches. k Greedy Scheduling: in this scheme, if f (i) = 0 and slot i is not the last slot of the frame (i < L), then a quiet period is only scheduled when channel k is in the worst state. If slot i is the last slot in the frame (i = L) and f k (i) = 0, then a quiet sensing period must be scheduled. Fixed Scheduling: in this scheme, the rst slot of each frame is always scheduled for spectrum sensing. We obtain the performance of the three sensing scheduling algorithms, i.e., Optimal, Greedy, and Fixed, by setting the frame length to L = 2 and 5 slots, corresponding to the time taken for sensing equivalent to 50% and 20% of each frame. We consider a single-user, single-channel system with the channel being modelled by a 5-state FSMC. Let 1 , 2 , . . . 5 be the signal to noise ratio (SNR) of the ve channel states, we assume that m+1 = m + 3dB, m = 1, . . . 4. That means 3 is the median SNR. When the channel is in state m, the transmission rate is calculated using Shanon capacity, i.e., rm = log2 (1 + m ) (b/s/Hz). In Fig. 2, we plot the performance of three schemes, Optimal, Greedy, and Fixed, in terms of the average spectrum efciency when the median SNR of the channel is varied. As expected, the Optimal CSI-based sensing scheme always out-performs the other two schemes while Fixed scheduling of sensing slot location performs worst. The performance advantage of Optimal and Greedy, relative

to Fixed scheduling scheme is more prominent when sensing time takes more percentage per each frame duration. This is because the gain of CSI-based sensing scheduling is limited to the time when sensing is carried out, while what we plot here is average over the whole frame duration. It is also encouraging to observe that Greedy scheme performs very close to Optimal scheme. That means we can carry out CSI-based sensing using this simple Greedy scheme and still achieve good performance gain. To get a better understanding of the behavior of Optimal CSIbased sensing policies, in Table. I, we show the decisions made by such an optimal policy under different channel states and time instances within each frame. In the entries of this table, an S indicates the decision to carry out sensing (if sensing has not been done in the particular slot) and a T indicates the decision to carry out transmission and delay sensing until future slots. It is clear from Table. I that the optimal policy tends to delay spectrum sensing when the channel conditions are good for communications and when there are still multiple slots left in the frame for scheduling sensing later on. It can also be noted that the Greedy Scheduling approach possesses similar structural properties. That explains why the Greedy policies can achieve performance closed to that of the optimal CSI-based scheduling, as observed in Fig. 2. IV. QCI- BASED S ENSING S CHEDULING In this section, we assume that CR nodes experience time-varying channels, stochastic data arrival, and have quality of service (QoS) requirement. For that, we take a cross-layer control approach and adaptively schedule quiet sensing periods based on both the queueing and channel state information to achieve good QoS. We term this approach QCI-based sensing scheduling. A. Problem Formulation 1) Trafc Model: We assume that data packets randomly arrive to the buffers of N CR nodes. All data packets of node n have the same size of Pn bits and the same deadline of Dn seconds. If a packet is not transmitted by its deadline, it is dropped and considered lost. We also assume that the buffers of all CR nodes are long enough so that buffer overow is negligible.

3130 1930-529X/07/$25.00 2007 IEEE


This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2007 proceedings.

2) Channel Model: The channel model considered in this section is similar to that used in Section III, i.e., for each pair of user n and channel k, the channel condition can take one of the M possible states. However, we denote by k (t) the rate in packets per seconds at n which user n can transmit on channel k at time t. We further assume that this rate does not change during the transmission of each packet. 3) Control and Objective: Data transmission and spectrum sensing are carried out independently on different channels. For channel k, when sensing is not carried out, one of the N CR nodes will be scheduled to transmit one packet. After the transmission of this packet, channel k becomes available again and if sensing is not scheduled, some node will be allowed to transmit another packet on the channel. At time t, let wn (t) denote the total queueing time (sojourn time) of the packet at the front of the buffer of CR node n. We dene the following QCI-based sensing scheduling problem. Denition 2: QCI-based Sensing Scheduling Problem - For each channel k, 1 k K, supposing that the channel becomes available at time t and assuming that the knowledge of k (t) and wn (t), n n {1, 2, . . . N }, is available, select one of the following decisions: carrying out sensing on channel k, or allowing one of the N CR nodes to transmit one packet on channel k, so that the total packet loss rate due to deadline violation of N CR nodes is minimized. This is subject to the constraint that Ts seconds must be scheduled for sensing during each frame. We would like to highlight two important points in Denition 2. First, within each frame, control decisions are not made on a slot-byslot basis as in Section III, rather, this is done whenever a channel becomes available, i.e., after a packet transmission or a sensing period. Second, the constraint that Ts seconds must be scheduled for sensing during each frame does not require sensing to be carried out continuously. B. Spectrum Sensing as Virtual Transmission For each channel k, the effects of scheduling a sensing period for q seconds to the performance of N CR nodes are the same as the effects we would have if the channel was assigned to some virtual transmission for the same q seconds. This enables us to abstract the sensing activities by introducing the following specications. 1) Virtual Sensing Nodes: : K virtual sensing (VS) nodes are introduced, one for each of the K channels. When sensing is carried out on channel k, we say VS node k transmits a VS packet. 2) Virtual Sensing Packets: : The arriving patterns, lengths, and deadlines of VS packets are specied to match the sensing requirements. Suppose that, for each channel k, the sensing duration Ts is divided into S equal sub-periods of duration ts = Ts /S. Then, at the beginning of each frame, S VS packets arrive to the buffer of VS node k. All of these VS packets have deadline of Tf seconds. 3) Virtual Transmission Rates: : VS node k has VS packets for channel k, each packet take ts seconds to transmit. Therefore, the virtual transmission rate for VS user k on channel l is: l (t) = k 1/ts , 0, if k = l, otherwise. (4)

Data packets

Channels
Data packets

Data packets VS packets of Channel 1 VS packets of Channel 2

Fig. 3. Representing the joint transmission/sensing problem as a queueing system. Quiet sensing periods are represented by virtual sensing (VS) packets.

all VS packets meet their deadlines. However, for such a parallel queueing system, to the best of our knowledge, there has not been any scheduling scheme identied as optimal. We therefore select a suboptimal scheduling scheme that gives good performance. Let us start by reviewing some known efcient scheduling schemes. 1) G-c Rule: The Generalized-c rule (G-c) introduced in [6], [7] aims to minimize the queueing cost under heavy trafc conditions. The queueing cost is represented by convex functions that increase with the packet queueing time. For deadline-constrained trafc, the following cost function is proposed in [6]: Cn (w) = (w/Dn ) , (5)

where w is the sojourn time of the oldest packet in the buffer of user n, Dn is the deadline constraint, and > 1 is a selected value. The intuition behind this cost function is that, if the sojourn time (delay) w is kept smaller than the constraint Dn , then the queueing cost incured is mall, while this cost increases rapidly when w > Dn (as > 1). Applying to the queueing model in Fig. 3, when channel k is available at time t, the G-c rule will schedule a packet of node u that satises:
k u arg max {Cn (wn (t))n (t)}, nN V

(6)

where N and V are the set of data nodes and VS nodes, respectively, k Cn (.) is the rst order derivative of Cn (.), and n (t) = k (t) if n k k n N and n (t) = n (t) if n V . 2) M-LWDF Rule: In [8], Andrews et al. introduced the Modied Largest Weighted Delay First (M-LWDF) scheduling rule that, when applied to the queueing system in Fig. 3, assigns channel k to node u that satises:
k u arg max {n [wn (t)] n (t)}, nN V

(7)

C. QCI-based Scheduling Disciplines With the introduction of VS users and VS packets, our problem can be abstracted as a parallel queueing network as in Fig. 3. We aim to schedule the trafc in this parallel queueing system so that the packet loss due to deadline violation for N data nodes is minimized while

where n is a weighting factor that represents delay constraint of node n and is some selected constant. In [8], it is shown that M-LWDF rule is throughput optimal, i.e., M-LWDF stabilizes the queueing system if it is possible to do so with any other queueing discipline. It is further argued that by stabilizing the queues, the queueing latency experienced by packets is also kept bounded. 3) Proposed QCI-based Scheduling Rule: Note that both G-c and M-LWDF scheduling rules are only optimal under heavy trafc conditions. For our system, taking these two policies as guideline, we have tested several different dynamic scheduling rules that take wn (t), k (t), and Dn into account and found that selecting user u n to transmit on channel k according to the following rule gives the most favorable performance, in terms of reducing packet deadline violation: wn (t) k n (t)}. (8) u arg max { Dn nN V

3131 1930-529X/07/$25.00 2007 IEEE


This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2007 proceedings.

As can be seen, (8) is the same as (6) when Cn (w) = w2 /Dn and the same as (7) when n = 1/Dn and = 1. The intuitions behind (8) are that, in order to reduce packet loss due to deadline violation, higher priority should be given to packets with more stringent deadlines, longer sojourn time, and that have faster transmission rate. Note that when VS packets are scheduled together with data packets based on (8), there is no guarantee that all every VS packet will be transmitted by its deadline. Therefore, we introduce the following QCI-based sensing scheduling policy. First, we nd a user u according to (8). If u is a VS user, then transmit a VS packet of u on channel k, i.e., sensing is carried out on channel k. If u is one of the N CR nodes, we need to calculate the time that channel k will be occupied if a packet of u is transmitted. If this transmission time leads to deadline violation for any VS packets of channel k, then we do not allow u to transmit on the channel and schedule a sensing packet instead. If there is no deadline violation for any VS packets of channel k, then we allow data user u to transmit one packet. V. S IMULATION OF QCI- BASED S ENSING In this section, we obtain simulation results for the performance of the QCI-based sensing scheduling scheme. A. Simulation Model We consider a CR network of N = 10 nodes and K = 2 channels. Data packets arrive to the buffer of each CR node according to a Poisson distribution. All packets have the same length of 1000 bytes and deadline of 40 msecs. The channels are time varying and can take M = 5 possible states. Corresponding to the 5 channel states are transmission rates of 500 kbps, 1 Mbps, 2.75 Mbps, and 5.5 Mbps. On average, the channels are unchanged for = 5 msecs and 10 msecs. A channel changes from one state to all other states with equal probability. We set Tf = 200 ms and Ts = 20 ms. B. Results and Discussion In Fig. 4, we plot the packet loss rate due to deadline violation versus the trafc load for two approaches, QCI-based sensing and Fixed-location sensing. For QCI-based sensing scheduling, we also consider cases when the sensing time is divided into different number of sensing sub-periods (up to 8 sub-periods). As can be observed, there is signicant reduction in the amount of packet loss when QCIbased sensing is applied, relative to when Fixed-location sensing is applied. Furthermore, the gain becomes more prominent when the number of sensing slots, i.e., sensing sub-periods, increases. This is because by breaking long sensing slots into smaller sub-slots, we have more dynamic in scheduling them. At the same time, carrying out sensing in multiple short periods also lessen the negative impact on delay experienced by data packets. In Fig. 5, we plot the performance of QCI-based sensing and Fixedlocation Sensing when channels change less frequently. In particular, the time at which each channel stays unchanged is increased from 5 msecs (in Fig. 4) to 10 msecs. As can be observed, the performance trends are similar, however, the gain of QCI-based sensing, relative to Fixed-location Sensing, is less prominent. This can be explained by the fact that the advantage of QCI-based sensing is exploiting the dynamic of trafc and channel states. VI. C ONCLUSION In this paper, we consider two approaches, namely CSI-based sensing scheduling and QCI-based sensing scheduling, for scheduling quiet sensing periods in CR networks. The CSI-based approach adaptively schedule these sensing periods based on channel conditions experienced by CR nodes. By doing so, we improve the

Ratio of Packet Loss (Deadline Violation)

0.25 Fixedlocation Sensing QCIbased Sensing (1 sensing slot) QCIbased Sensing (2 sensing slots) QCIbased Sensing (4 sensing slots) QCIbased Sensing (8 sensing slots)

0.2

0.15

0.1

0.05

0 0.05

0.06

0.07

0.08

0.09

0.1

0.11

0.12

Arrival Rate (packets/ms/user)

Fig. 4. Performance of xed sensing location scheme and QCI-based sensing scheme. Channel changes every 5ms.
Ratio of Packet Loss (Deadline Violation)
0.3 Fixedlocation Sensing QCIbased Sensing (1 sensing slot) QCIbased Sensing (4 sensing slots) QCIbased Sensing (8 sensing slots)

0.25

0.2

0.15

0.1

0.05

0 0.05

0.06

0.07

0.08

0.09

0.1

0.11

0.12

Arrival Rate (packets/ms/user)

Fig. 5. Performance of xed sensing location scheme and QCI-based sensing scheme. Channel changes every 10ms.

spectrum efciency of CR operation, while still meeting the sensing requirement. In the QCI-based approach, the sensing scheduling decisions take into account both queueing and channel conditions of CR nodes. Our numerical results show that, for CR trafcs with deadline constraints, the QCI-based approach effectively reduce packet loss rate due to deadline violation. R EFERENCES
[1] FCC, Spectrum policy task force report, FCC 02-155. Nov. 2002. [2] H. S. Wang and N. Moayeri, Finite-state markov channel - a useful model for radio communication channels, IEEE Trans. on Vehicular Technology, vol. 44, no. 1, pp. 163171, Feb. 1995. [3] Y. C. Liang, Y. H. Zeng, E. Peh, and A. T. Hoang, Sensing-throughput tradeoff for cognitive radio networks, in Proc. of IEEE ICC07, Glasgow, Jun. 2007. [4] IEEE 802.22 Wireless RAN, Functional requirements for the 802.22 WRAN standard, IEEE 802.22- 05/0007r46, Oct. 2005. [5] D. P. Bertsekas, Dynamic Programming and Optimal Control: 2nd edition, vols. 1 and 2. Athena Scientic, 2001. [6] J. A. V. Mieghem, Due-date scheduling: Asymptotic optimality of generalized longest queue and generalized largest delay rules, Operations Research, vol. 51, no. 1, pp. 113122, Jan.-Feb. 2003. [7] A. Maldelbaum and A. L. Stolyar, Scheduling exible servers with convex delay costs: Heavy-trafc optimality of the generalized c-rule, Operation Research, vol. 52, no. 6, pp. 836855, Nov.-Dec. 2004. [8] M. Andrews, K. Kumaran, K. Ramanan, A. L. Stolyar, R. Vijayakumar, and P. Whitting, Scheduling in a queueing system with asynchronously varying service rates, Probability in the Engineering and Informational Sciences, vol. 18, pp. 191217, 2004.

3132 1930-529X/07/$25.00 2007 IEEE


This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2007 proceedings.

Vous aimerez peut-être aussi