Performance of Joint Spectrum Sensing and MAC Algorithms for Multichannel Opportunistic Spectrum...

Performance of Joint Spectrum Sensing andMAC Algorithms for Multichannel Opportunistic

Spectrum Access Ad Hoc NetworksJihoon Park, Student Member, IEEE, Przemyslaw Pawelczak, Member, IEEE, and Danijela �Cabri�c

Abstract—We present an analytical framework to assess the link layer throughput of multichannel Opportunistic Spectrum Access

(OSA) ad hoc networks. Specifically, we focus on analyzing various combinations of collaborative spectrum sensing and Medium

Access Control (MAC) protocol abstractions. We decompose collaborative spectrum sensing into layers, parametrize each layer,

classify existing solutions, and propose a new protocol called Truncated Time Division Multiple Access (TTDMA) that supports efficient

distribution of sensing results in “� out of N” fusion rule. In case of multichannel MAC protocols, we evaluate two main approaches of

control channel design with 1) dedicated and 2) hopping channel. We propose to augment these protocols with options of handling

secondary user (SU) connections preempted by primary user (PU) by 1) connection buffering until PU departure and 2) connection

switching to a vacant PU channel. By comparing and optimizing different design combinations, we show that 1) it is generally better to

buffer preempted SU connections than to switch them to PU vacant channels and 2) TTDMA is a promising design option for

collaborative spectrum sensing process when � does not change over time.

Index Terms—Opportunistic spectrum access, cognitive radio, ad hoc networks, medium access control, spectrum sensing.

Ç

1 INTRODUCTION

IT is believed that Opportunistic Spectrum Access (OSA)networks will be one of the primary forces in combating

spectrum scarcity in the upcoming years [1]. Therefore,OSA networks have become the topic of rigorous investiga-tion by the communications theory community. Specifically,the assessment of spectrum sensing overhead on OSAmedium access control (MAC) performance recently gaineda significant attention.

1.1 Research Objective

In the OSA network performance analysis, a description ofthe relation between the primary (spectrum) user (PU)network and the secondary (spectrum) user (SU) networkcan be split into two general models: macroscopic andmicroscopic. In the macroscopic OSA model [2], [3], it isassumed that the time limit to detect a PU and vacate itschannel is very long compared to the SU time slot, frame, orpacket length duration. Such a time limit is assumed to begiven by a radio spectrum regulatory organization. Also, inthe macroscopic model, it is assumed that the PU channelholding time, i.e., the time in which the PU is seen by the SUas actively transmitting, is much longer than the delayincurred by the detection process performed at the SU. As aresult, it can be assumed in the analysis that, given high PUdetection accuracy (which is a necessity), OSA networkperformance is determined by the traffic pattern of the SUs.That is, it depends on the total amount of data to be

transmitted by the SU network, the duration of individualSU data packets, and the number of SU nodes. In otherwords, the PU bandwidth resource utilization by the SU isindependent of PU detection efficiency.

In the microscopic OSA model, the detection time isshort in relation to the shortest transmission unit of the OSAsystem. Detection is also performed much more frequentlythan in the macroscopic model, i.e., for every SU packet [4]or in every time slot [5], [6], [7], [8], [9]. Also, themicroscopic model assumes much higher PU activity thanthe macroscopic model, which justifies frequent detectioncycles. Since the detection overhead is much larger than inthe macroscopic model, the analysis of utilization of PUresources by OSA network cannot be decoupled from theanalysis of the PU signal detection phase.

Therefore, while the distinction between macroscopicand microscopic models are somehow fluid, it is importantto partition the two cases and compare them in a systematicmanner. More importantly, the comparison should be basedon a detailed OSA multichannel and multiuser ad hocnetwork model [10, Section 7.4], which would not ignore theoverhead from both the physical layer (PHY) and MAClayers of different cooperative and distributed spectrumsensing strategies [10, Table 7.1] and, in case of microscopicmodel, account for different channel access procedures andconnection management strategies for the SUs upon PUdetection, like buffering or switching to a vacant channel.Finally, the comparison should be realized using tractableanalytical tools.

1.2 Related Work

The literature on this topic can categorized into threegroups: 1) performance analysis of general OSA networks,excluding a detailed model for spectrum sensing (mostlyfor the macroscopic model), 2) performance of spectrum

IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 10, NO. 7, JULY 2011 1011

. The authors are with the Department of Electrical Engineering, Universityof California at Los Angeles, 56-125B Engineering IV, Los Angeles, CA90095-1594. E-mail: {jpark, przemek, danijela}@ee.ucla.edu.

Manuscript received 24 Oct. 2009; revised 14 June 2010; accepted 12 Aug.2010; published online 20 Dec. 2010.For information on obtaining reprints of this article, please send e-mail to:[email protected], and reference IEEECS Log Number TMC-2009-10-0455.Digital Object Identifier no. 10.1109/TMC.2010.255.

1536-1233/11/$26.00 � 2011 IEEE Published by the IEEE CS, CASS, ComSoc, IES, & SPS

sensing isolated from MAC aspects of network collabora-tion, and 3) joint performance of spectrum sensing andnetworking for OSA.

One of the first works that gained insight into the generalperformance of OSA networks, considering impact of PUactivity on blocking and throughput of the SU network was[11] where the spectrum sharing gains for PU and SUnetworks were obtained for a distributed and multichannelad hoc OSA network. Unfortunately, a zero delay spectrumsensing process was assumed with genie-aided channelselection, i.e., in every time slot, the receiver knew of theexact channel the transmitter will use to send data [11,Section 3.3.1].

In later works, assumptions on the OSA network modelbecame more realistic. Specifically, Markovian analysis ofSU traffic buffering on the event of PU arrival waspresented for an SU exponential service time [12] and foran SU phase-type service time [13]. Unfortunately, theimpact of spectrum sensing detection time overhead on theOSA network performance was not investigated and theconnection arrangement process for new SU arrival wasassumed to be performed by a centralized entity.

A similar analysis, but with a different channelizationstructure, where the PU occupied more than one SU channel(contrary to [12], [13]) was performed in [14]. The authorsaddressed the cases of 1) connection blocking, and 2) channelreservation and switching of SU connections to emptychannels on PU arrival. This analysis was later extended tothe case of finite SU population and packet queuing [15], andbuffering and switching of SU connections preempted by PUarrivals [16]. Again, in all papers listed above, the spectrumsensing process was assumed to have no overhead andperfect reliability. Moreover, the connection arrangementprocess for SUs was not considered.

A system where the PU had to wait until an SU vacates achannel was analyzed in [17]. Both perfect and imperfectPU detection processes were considered; however, detec-tion overhead as well as a connection arrangement processfor the secondary system was not considered. The onlywork considering a microscopic model was [6], where arelation between sensing time, PU detectability, anddifferent connection arrangement processes was taken intoaccount. Detailed simulations and Markov analysis wereperformed; however, as noted in the paper, the proposedmodel did not yield accurate results over the ranges of allparameters considered, e.g., level of PU activity. Also, onlyone sensing strategy with SU connection buffering wasanalyzed for the case of different MAC protocols.

Considering a second group of papers (related to theperformance of spectrum sensing algorithms in isolationfrom higher protocol layers), in [5], the analysis of theaverage time consumed by two-stage spectrum sensing(proposed independently in [3] and [9]) was decoupledfrom SU traffic characteristics. Moreover, the delay causedby exchanging hard decision measurements in the coopera-tive sensing process based on the “and” rule was notincluded. The impact of sensing overhead on, e.g.,throughput, and energy consumption was explored in[18]. However, the analysis did not account for any OSAnetwork and SU traffic. Also, the relationship betweendetection time and detection quality was not investigated.

In [9], a microscopic model was analyzed with a sensingperiod every slot, synchronization between PU and SU, andPU stationary over the whole slot duration. Markovanalysis was performed only to evaluate the delay incurredwhile searching for unoccupied spectrum. The sensingprocess was not coupled to any of the known MACprotocols and SU connections. Also, only noncollaborativespectrum sensing was considered. Finally, in [2], most of theprocedures related to spectrum sensing were categorizedand divided into Open Systems Interconnection-like layers.Performance of the most common combinations of sensingalgorithms was assessed, but only for the macroscopicmodel. Our analysis is the microscopic treatment of [2].

Considering the final group of papers (related tocoupling spectrum sensing procedures with link layerprotocols), there is a fundamental trade-off between sensingtime, sensing quality, and OSA network throughput. Thishas been independently found for general OSA networkmodels with a single sensing band [4], multiple sensingbands [19] with and without cooperative detection andcentralized resource allocation, and in a context of MACprotocol abstraction [6] for a noncooperative sensing case.See also recent discussion in [10, Sections 2.3.1, 7.3, and10.2.4]. This trade-off will be especially clear, whileevaluating microscopic models, since the detection timecreates a significant overhead for the data exchange phase.

1.3 Our Contribution

In this paper, we present a unified analytical framework todesign the spectrum sensing and the OSA data MACjointly, for the macroscopic and microscopic cases. Thisdesign framework provides the 1) means of comparingdifferent spectrum sensing techniques plus MAC architec-tures for OSA networks and 2) spectrum sensing para-meters such as observation time and detection rate for givendesign options. As a metric for optimization and compar-ison, we consider the average link layer OSA networkthroughput. Our model accounts for the combined effects ofthe cooperative spectrum sensing and the underlying MACprotocol. For spectrum sensing, we consider severalarchitectures parametrized by sensing radio bandwidth,the parameters of the sensing PHY, and the parameters ofthe sensing MAC needed to exchange sensing data betweenindividual OSA nodes.

The rest of the paper is organized as follows: systemmodel and a formal problem description is presented inSection 2. Description of spectrum sensing techniques andtheir analysis is presented in Section 3. Analysis of MACstrategies is presented in Section 4. Numerical results forspectrum sensing process, MAC, and joint design frame-work are presented in Section 5. Finally, the conclusions arepresented in Section 6.

2 MODEL AND PROBLEM DESCRIPTION

2.1 Microscopic Model

For two multichannel MAC abstractions considered, i.e.,Dedicated Control Channel (DCC) and Hopping ControlChannel (HCC), both analyzed in [6] and [24], wedistinguish between the following cases: 1) when SU datatransfer interrupted by the PU is being buffered (or not) forfurther transmission and 2) when existing SU connection

1012 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 10, NO. 7, JULY 2011

can switch (or not) to a free channel on the event of PUarrival (both for buffering and nonbuffering SU connectioncases). Finally, we will distinguish two cases for DCC where1) there is a separate control channel not used by the PUand 2) when control channel is also used by the PU forcommunication. All these protocols will be explained indetail in Section 4.

We assume slotted transmission within the SU and PUnetworks, where PU and SU time slots are equal andsynchronized with each other. The assumptions on slottedand synchronous transmission between PU and SU arecommonly made in the literature, either while analyzingtheoretical aspects of OSA (see [4, Fig. 2], [6, Section 3], [8,Fig. 1]) or exploring practical OSA scenarios (see [7, Fig. 2] inthe context of secondary utilization of GSM spectrum or [20]in the context of secondary IEEE 802.16 resources usage).Our model can be generalized to the case where PU slots areoffset in time from SU slots; however, it would requireadditional analysis of optimal channel access policies, see,for example, [21], [22], which is beyond the scope of thispaper. We also note that the synchrony assumption allowsone to obtain upper bounds on the throughput whentransmitting on a slot-asynchronous interface [23].

The total slot duration is tt �s. It is divided in three parts:1) the detection part of length tq �s, denoted as quiet time,2) the data part of length tu �s, and if communication protocolrequires channel switching 3) switching part of length tp �s.The data part of the SU time slot is long enough to execute onerequest to send and clear to send exchange [6], [24]. For thePU, the entire slot of tt �s is used for data transfer, see Fig. 1a.

Our model assumes that there are M channels havingfixed capacity C Mbps that are randomly and indepen-dently occupied by the PU in each slot with probability qp.There are N nodes in the SU network, each one commu-nicating directly with another SU on one of the available PUchannels in one-hop fashion. Also, we assume no mergingof the channels, i.e., only one channel can be used by acommunicating pair of SUs at a time. SUs send packets withgeometrically distributed length with an average of 1=q ¼d=ðCtuÞ slots for DCC, and 1=q ¼ d=ðCftu þ tpgÞ slots forHCC [6, Sec. III-C4b], [24, Section 3.2.3], where d is theaverage packet size given in bits. Difference betweenaverage packet length for DCC and HCC is a result ofswitching time overhead for HCC, because during channelswitching, SUs do not transfer any data, even though theyoccupy the channel. We therefore virtually prolong datapacket by tp for HCC to keep the comparison fair.

Every time a node tries to communicate with anothernode, it accesses the control channel and transmits a controlpacket with probability p to a randomly selected andnonoccupied receiver. A connection is successful when onlyone node transmits a control packet in a particular timeslot. The reason for selecting a variant of S-ALOHA as a

contention resolution strategy was manyfold. First, inreality, each real-life OSA multichannel MAC protocolbelonging to each of the considered classes, i.e., HCC orDCC, will use its own contention resolution strategy.Implementing each and every approach in our analysis1) would complicate significantly the analysis, and mostimportantly 2) would jeopardize the fairness of the compar-ison. Therefore, a single protocol was needed for theanalytical model. Since S-ALOHA is a widespread and wellunderstood protocol in wireless networks and is a founda-tion of many other collision resolution strategies, includingCSMA/CA, it has been selected for the system model herein.

In each quiet phase, every SU node performs PU signaldetection based on signal energy observation. Since weassume that OSA nodes are fully connected in a one-hopnetwork, thus each node observes on average the samesignal realization in each time slot [9], [25]. PU channelsdetected by the SU are assumed as Additive White GaussianNoise with a channel experiencing Rayleigh fading. There-fore, to increase the PU detectability by the OSA network,we consider collaborative detection with hard decisioncombining in the detection process based on “� out of N”rule, as in [26]. Hence, we divide the quiet phase into 1) thesensing phase of length ts �s and 2) the reporting phase oflength tr �s. The sensing phase is of the same length for allnodes. For simplicity, we do not consider in this studysensing methods that adapt the sensing time to propagationconditions as in [27]. In the sensing phase, nodes performtheir local measurements. Then, during the reporting phase,nodes exchange their sensing results and make a decisionindividually by combining individual sensing results. Wewill analyze different PHY and MAC approaches tocollaborative spectrum sensing, especially 1) methods toassign sensing frequencies to users, 2) rules in combining thesensing results, and 3) multiple access schemes for measure-ment reporting. In this paper, we do not consider sensingstrategies applicable to single channel OSA networks [28],two-stage spectrum sensing [3], and sensing MAC protocolsbased on random access [29], due to their excessive delay.We will explain our spectrum sensing approaches in moredetail in Section 3. Further, we assume an error channel, forthe sensing layer as well as for data layer where probabilityof error during transmission is denoted as pe.

Finally, we consider two regulatory constraints underwhich the OSA network is allowed to utilize the PUspectrum provided the channel is idle: 1) maximumdetection delay td;max, i.e., a time limit within which an SUmust detect a PU, and 2) minimum detection probabilitypd;min, i.e., a probability with which an OSA system has todetect a PU signal with minimum signal to noise ratio �.Note that in the event of misdetection and subsequent SUtransmission in a channel occupied by PU, a packetfragment is considered successfully transmitted, since inour model, transmission power of SU is much higher thaninterference from PU, and regulatory requirements con-sidered here do not constrain SU transmission power1

PARK ET AL.: PERFORMANCE OF JOINT SPECTRUM SENSING AND MAC ALGORITHMS FOR MULTICHANNEL OPPORTUNISTIC SPECTRUM... 1013

Fig. 1. Difference between macroscopic and microscopic model in

1. The opposite case is to assume that a packet fragment is considered aslost and retransmitted. This approach however requires an acknowledgmentmechanism for a lost packet fragment, see, for example, [8, Section 2.2.3], [23,Section 2], that contradicts the model assumption on the geometricdistribution of SU packets.

(refer, for example, to IEEE 802.22 draft where Urgent

Coexistent Situation packets are transmitted on the same

channel as active PU [30]). Moreover, maximum transmis-

sion power is a metric specific to overlay OSA systems [10,

Sections 2.2.5 and 8.2.1] where typically no spectrum

sensing is considered. Also, we do not consider a metric

based on a maximum allowable level of collisions between

PU and SU.

2.2 Macroscopic Model

We assume the same system model as for the microscopic

case, except for the following differences. OSA performs

detection rarely, and the PU is stable for the duration of

OSA network operation, i.e., it is either transmitting

constantly on a channel or stays idle. In other words, quiet

period occurs for multiple time slots, see Fig. 1b. Also, since

the PU is considered stable on every channel, we do not

consider all types of OSA MAC protocols introduced for the

microscopic model. Instead, we use classical DCC and HCC

models proposed in [24] with the corrections of [6]

accounting for the incomplete transition probability calcu-

lations whenever OSA network occupied all PU channels

and new connection was established on the control channel.

2.3 Problem Description

To compute the maximum throughput for different

combinations of protocols and models, we define an

optimization problem. The objective is the OSA network

link layer throughput Rt. Therefore, considering the

regulatory constraints given above, we need to

maxRt ¼ � R subject to pd ¼ pd;min; td � td;max; ð1Þ

where td is the detection time, i.e., the time to process whole

detection operation as described in Section 3.4, R is the

steady state link layer throughput without sensing and

switching overhead, which will be computed in Section 4,

and 1� � is the sensing and switching overhead, see also

Fig. 1, where � is defined separately for two considered

cases as

�¼

tt � tq � tptt

; microscopic; DCC; channel switching;

td;max � tqtd;max

; macroscopic;

tt � tqtt

; otherwise:

8>>>>><>>>>>:

ð2Þ

Note that R in (1) is itself affected by pf , as it will be shown

in Section 4. Also note that tp is removed from second

condition of (2) since the switching time is negligible in

comparison to intersensing time.

3 SPECTRUM SENSING LAYER ANALYSIS

To design the spectrum sensing, we follow the approach of

Park et al. [2] in which the spectrum sensing process is

handled jointly by 1) the sensing radio, 2) the sensing PHY,

and 3) the sensing MAC. Using this layered model, we can

compare existing approaches to spectrum sensing and

choose the best sensing architecture in a systematic way.

3.1 Sensing Radio

The sensing radio scans the PU spectrum and passes thespectrum sensing result to the sensing PHY for analysis. Thesensing radio bandwidth is given as�Mb, where� is a ratio ofthe bandwidth of the sensing radio to the total PU bandwidthand b MHz is the bandwidth of each PU channel.2 With� > 1=M, node can sense multiple channels at once.However, the cost of such wideband sensing radio increases.

3.2 Sensing PHY

The sensing PHY analyzes the measurements from thesensing radio to determine if a PU is present in a channel.Independent of the sensing algorithm, such as energydetection, matched filter detection, or feature detection[31], [32], there exists a common set of parameters for thesensing PHY: 1) time to observe the channel by one nodete �s, 2) the PU signal to noise ratio detection threshold �,and 3) a transmit time of one bit of sensing informationta ¼ 1=C �s. We denote conditional probability of sensingresult pij; i; j 2 f0; 1g, where j ¼ 1 denotes PU presence andj ¼ 0 otherwise, and i ¼ 1 indicates the detection result ofPU being busy and i ¼ 0 otherwise. Observe that p10 ¼1� p00 and p11 ¼ 1� p01.

As noted in Section 2, we consider energy detection asthe PU detection algorithm since it does not require a prioriinformation of the PU signal. For this detection method inRayleigh plus Additive White Gaussian Noise channel p10 isgiven as [6, (1)] and p11 as [6, (3)], with � ¼ bte�Mbc as atime-bandwidth product defined therein. By definingG�ð�Þ ¼ p10 and � ¼ G�1

� ðp10Þ, we can derive p11 as afunction of p10 and te.

3.3 Sensing MAC

The sensing MAC is a process responsible for sensingmultiple channels, sharing sensing results with other users,and making a final decision on the PU presence. Because ofthe vast number of possibilities for sensing MAC algo-rithms, it is hard to find a general set of parameters. Instead,we derive cross-layer parameters for a specific option of thesensing MAC. This methodology can be applied to any newsensing MAC scheme. We now introduce classificationswhich will be used in the derivation of cross-layerparameters.

3.3.1 Sensing Strategy for Grouping Channels and

Users

Each SU has to determine which channel should be sensedamong the M channels. To reduce sensing and reportingoverhead, OSA system can divide users and channels intong subgroups [33]. Subgroup i 2 f1; . . . ; ngg is formed bynu;i users who should sense ms;i channels to make a finaldecision cooperatively. Assume that all users are equallydivided into groups, then ms;i 2 fbM=ngc; dM=ngeg andnu;i 2 fbN=ngc; dN=ngeg. Note that for M=ng 2 IN andN=ng 2 IN, all subgroups have the same nu;i ¼ N=ng andms;i ¼M=ng for all i. Given N and M, if ng is small, moreusers are in a group and the collaboration gain increases,


2. Note that C used later in calculating MAC throughput is an averagethroughput obtained using a certain modulation over a channel withbandwidth b.

but at the same time more channels must be sensed, whichresults in more time overhead for sensing. For large ng, this

relation is opposite.

3.3.2 Combining Scheme

By combining sensing results of other users, an OSA

network makes a more reliable decision on PU state. Asconsidered in [34], we will take � as a design parameter forthe sensing MAC and find an optimum value to maximize

the performance. Note that for the case of N usercooperation if � ¼ 1, the combining logic becomes the

“or” rule [10, Section 3.2], [25, Section 3.3] and if � ¼ N , itbecomes the “and” rule.

3.3.3 Multiple Access for Measurement Reporting

To transmit sensing results of multiple users through theshared media, a multiple access scheme is needed. Note

that this multiple access scheme is only for the reportingprocess, different from the multiple access for data transfer.

We consider the following approaches:Time division multiple access (TDMA). This is a static

and well-organized multiple access scheme for which a

designated one bit slot for sensing report transmission isassigned to each user [33].

Truncated time division multiple access (TTDMA). In

TDMA, when the SU receives all the reporting bits fromother users, the SU makes a final decision of presence of PU

on the channel. However, in OSA network using TTDMA,SUs may not need to wait until receiving the last reportingbit, because for the “� out of N” rule, a reporting operation

can stop as soon as � one bits denoting PU presence arereceived. This sensing MAC aims at reducing the reporting

overhead, but unfortunately, we have not seen any paperproposing and discussing TTDMA.

Single slot multiple access (SSMA). For this scheme,

known also as the boosting protocol [35], only one bit slot isassigned for reporting and all SUs use this slot as a common

reporting period. Any SU that detects a PU transmits one bitin the common designated slot. Otherwise, a user does not

transmit any bit in the designated slot. Then, reporting bitsfrom SUs who detect a PU are overlapped and as a result,

all power of the slot is summed up. By measuring thepower in the designated slot, an SU can determine whetherthe primary user exists or not. We assume perfect power

control and perfect synchronization. Even though this maynot be practical, because carrier frequency or the phase

offset cannot be avoided in real systems, this scheme servesas an upper bound for sensing MAC performance. For the

analysis of SSMA in isolation but in more realistic physicallayer conditions, the reader is referred to [36] and [37].

3.4 Cross-Layer Parameters

3.4.1 Detection Time td and Quiet Time tqDetection time td is defined as the time duration from thepoint that an SU starts to sense, to the point that an SUmakes a final decision on PU presence. Regardless of the

data transfer and spectrum sensing time overlap, the finaldetection decision is made only after combining the sensing

group’s reported information [38]. Thus, td is the time from

the start of the sensing phase to the end of the reportingphase, i.e., td ¼ ts þ tr.

Since the data transfer may not be possible duringsensing or reporting phases tq � td, depending on theapproach. When spectrum sensing and data transfer aredivided in time division manner tq ¼ ts þ tr. Note that threeother methods sharing the same problem are possible (theywill not be considered in the remainder of the paper):1) simultaneous reporting and data, which can be imple-mented by using the separate channel as in [39], for whichtq ¼ ts, 2) simultaneous sensing and data, implemented byusing the frequency hopping method as in [40], for whichtq ¼ tr, and 3) simultaneous sensing, reporting, and data forwhich tq ¼ 0. Conceptually, simultaneous sensing, report-ing, and data transfer are possible and seem most efficientbut we have not found any implementation of it in theliterature. Note that in order to implement simultaneoussensing and transmission, at least two radio front ends areneeded, which increases the total cost of the device.

Define �ms as the number of individual sensing events tocomplete sensing operation and �mr as the average numberof bits to report. Then, the sensing time and the reportingtime can be calculated as ts ¼ �mste and tr ¼ �mrta. Notethat �ms is affected by the bandwidth of the sensing radiobecause it can scan multiple channels at once if thebandwidth of the sensing radio is wide. For the case thatthe sensing radio is narrower than the bandwidth to sense,i.e., � < maxfms;1; . . . ;ms;ngg=M, we assume that an SUmonitors all channels by sequential sensing [19], becausethe reporting phase should be synchronized after allSUs finish the sensing phase. With this assumption�ms ¼ maxfms;1; . . . ;ms;ngg=�M

� �, because even though

the bandwidth to sense is less than that of the sensingradio, it still needs one sensing cycle to get information.For �mr, because there are ng groups in an OSA system,�mr ¼

Pngi¼1 �mr;i where �mr;i depends on the multiple access

schemes for reporting, which we compute below.TDMA. All nu;i users should transmit the sensing results

of ms;i channels. Thus, �mr;i ¼ nu;ims;i.TTDMA. For � < nu;i=2, if � of 1s is received, the

reporting process will end. We introduce a variable whichis the number of bits when the reporting process finishes.Thus, there should be �� 1 of 1s within � 1 bits and thenth bit should be 1. Because the range of is from � to nu;i,the average number of bits for this condition is derived as

m1;i ¼Xnu;i¼�

� 1

�� 1

� ��ð1� qpÞp��00 p�10 þ qpp��01 p�11

�: ð3Þ

Moreover, if the number of received 0s, denoting PU absence,equals to nu;i � �þ 1, the reporting process will stop becauseeven if the remaining bits are all 1, the number of 1s must beless than �. Then, the reporting process stops at th bit if � nu;i þ �� 1 bits of 1 are received within � 1 bits and 0 isreceived at th bit. The range of is from nu;i � �þ 1 to nu;i,and thus, the average number of bits for this condition is

m2;i ¼Xnu;i¼i

� 1

� i

� ��ð1� qpÞpi00p

�i10 þ qppi01p

�i11

�; ð4Þ

where i ¼ nu;i � �þ 1. Therefore, because there are ms;i

channels to sense in a group i, �mr;i ¼ ms;iðm1;i þm2;iÞ.


For the case � � nu;i=2, m1;i is calculated by counting 0s

and m2;i by counting 1s. Thus, we use �mr;i ¼ ms;iðm1;i þm2;iÞ again, by replacing � with nu;i � �þ 1, p00 with p10,

and p01 with p11.Because we assumed so far that � is known to each node in

the network, OSA nodes know when to stop reporting

measurements and start data communication without being

instructed by external parties. For comparison, we analyze

another type of TTDMA, denoted as �TTDMA, where a

cluster head node makes a decision to stop the reporting

phase in the OSA network. For example, this approach may

be necessary if the�value is updated in real time. In the worst

case scenario, this approach requires two bits to be reported

by the SU, i.e., one for sending sensing data and one for an

acknowledgment from the cluster head to report. Then, (3)

and (4) need to be modified by multiplying them by 2.SSMA. For this scheme, we need only one bit per

channel for reporting. Thus, �mr;i ¼ ms;i.

3.4.2 Total False Alarm Probability pf and Detection

Probability pdFinal probabilities pf and pd are obtained after cooperation,

and thus affected by the sensing MAC and sensing PHY.

Because each subgroup has a different number of users and

channels to sense, we have

pf ¼1

ng

Xngi¼1

pf;i; ð5Þ

where pf;i is the probability of false alarm of subgroup i.

Using (5), we can also derive pd by substituting pf;i with pd;i,

i.e., probability of detection of subgroup i. The definitions of

pf;i and pd;i for each protocol are as follows:TDMA. For this protocol, pf;i is derived as

pf;i ¼Xnu;i¼�

nu;i

� p10p

nu;i�00 ; ð6Þ

where px ¼ ð1� peÞpx þ peð1� pxÞ for px 2 fp10; p00g, while

pd;i is derived from (6) by substituting p10 with p11 and p00

with p01.TTDMA. In this case, SU does not need to receive nu;i

bits to make a final decision because the reporting phase is

ended when the number of 1s is �. To derive pf;i for this

case, we introduce a variable � denoting the number of 0s.

Then, total number of reporting bits is �þ � if the last bit

is 1 because otherwise, reporting phase will end at less

than �þ � bits. Therefore, there should be � of 0s in �þ� � 1 bits and �th bit should be 1. Because � can vary from

0 to nu;i � �

pf;i ¼Xnu;i��¼0

�þ � � 1

�

� �p�10p

�00: ð7Þ

Finally, pd;i is obtained from (7) by substituting p10 with p11

and p00 with p01.SSMA. Obviously, the process of the reporting informa-

tion for SSMA is the same as for TDMA. Therefore, pf;i and

pd;i are defined the same as for TDMA.

4 MULTICHANNEL OSA MAC PROTOCOL

ANALYSIS

4.1 Description of New Multichannel MAC Protocolsfor OSA

We consider two major groups of MAC protocols for OSA:1) those enabling buffering of the SU connections preemptedby the PU arrival, and 2) those enabling switching of the SUconnections to a vacant channel when preempted. In theformer group, when the PU arrives, the existing SUconnection will pause at the time of preemption and resumeon the same channel as soon as the PU goes idle. We assumethat the SU always waits for the PU to finish its transmission.The case where the buffered SU connection expires after apredefined time, not analyzed here, is presented in [12] forthe centralized network. We do not consider any channelreservation schemes for potential SU connections to bebuffered [14]. When buffering is not possible, the preemptedSU connection is considered as lost and a new connectionmust be established on the control channel. In the lattergroup, when the PU arrives, the existing SU connection willlook for a new empty channel, to continue transmission. Ifsuch a channel cannot be found, the connection is lost.Without channel switching, the exiting SU connection is lostas soon as the PU preempts the channel.

Obviously, we can have four combinations of thesegroups for OSA MAC, which have all been considered inthe analysis:

1. with no buffering and no channel switching [41]scheme denoted as B0S0, where SU connectionspreempted by PU are lost;

2. with no buffering and channel switching [14], [15],[42] denoted as B0S1, where SU connections pre-empted by PU switch to a free channel andconnections that cannot find a free channel areblocked;

3. with buffering and no channel switching [6], [12],[13] denoted as B1S0, where SU connections pre-empted by PU are being suspended from themoment of preemption until PU releases thechannel; and

4. with buffering and channel switching [43] denotedas B1S1, where SU connections preempted by the PUfirst look for free channels, and when no freechannels are found, the connections are beingbuffered until PU leaves the channel.

The detailed procedure of distributed channel selection onthe event of switching will be described in Section 4.3.2.Recall that the works referred above consider an OSAnetwork with centralized channel management, providingonly the upper bound on OSA network performance.

4.2 Multichannel MAC for OSA Analysis:Preliminaries

Usually, to compute the throughput of most non-OSAnetworks, it is assumed that a Markov chain characterizesthe network state defined as the current number of connec-tions used for data transfer [6, Section 3.3], [24, Section 3]. Thestate transition probability depends only on the networkusers’ traffic characteristics. However, in the OSA system, the


SU data transfer connections can be terminated or delayed if aPU is detected on their channel, and thus, the traffic generatedby PU also affects the state transition. Moreover, withconnection buffering enabled, when a PU is detected on thechannel, the SU does not terminate its connection but ratherwaits until the PU goes idle. Thus, the OSA networkthroughput is influenced by the number of channels thatare actually utilized by the SUs rather than solely by thenumber of SU data transfer connections.

We propose a three-dimensional Markov chain of whichthe state vector is given as ðXt; Yt; ZtÞ, where Xt; Zt 2 S ¼f0; 1; . . . ;minðbN=2c;MDÞg and Yt 2 f0; 1; . . . ;MDg, whereMD ¼M � 1 for DCC and MD ¼M for HCC. The elementsof the state vector are: 1) Xt, denoting the number ofchannels that are actually utilized by the SUs at time t, 2) Yt,denoting the number of channels on which the PU isdetected at time t, and 3) Zt, denoting the number ofconnections for the data transmission between the OSAusers at time t. This distinction allows to compute the exactchannel utilization, contrary to [12, Section 3] wherebuffered SU connections were also considered to beutilizing the PU channels.

Considering a real OSA system, there are conditions thatqualify valid states. With SU connection buffering-enabledMAC protocols for OSA, the number of connections cannotbe less than the number of channels utilized by SUs, i.e.,Xt � Zt. Additionally, SUs do not pause transmissions overunoccupied channels. Therefore, the number of SU connec-tions not utilizing a channel cannot exceed the number ofchannels occupied by PUs, i.e., Zt �Xt � Yt or Zt � Xt þ Yt.Finally, the sum of the channels utilized by PUs and the SUscannot be greater than MD, i.e., Xt þ Yt �MD. By combin-ing these conditions, we can compactly write them as

0 � Xt � Zt � Xt þ Yt �MD: ð8Þ

When connection buffering is disabled, the number of SUconnections must be the same as the number of channelsutilized by SUs, i.e., Xt ¼ Zt. Therefore, for nonbuffering SUconnection OSA MAC protocols ðXt; Yt; Zt ¼ XtÞ ) ðXt; YtÞ.

For the microscopic case, the average channel through-put, excluding switching and sensing overhead, is com-puted as

R ¼ CXsmx¼0

XMD

y¼0

Xsmz¼0

x�xyz; ð9Þ

where sm ¼ maxfSg and the steady state probability �xyz isgiven by

�xyz ¼ limt!1

PrðXt ¼ x; Yt ¼ y; Zt ¼ zÞ; ð10Þ

and the state transition probabilities to compute (10) will bederived in the subsequent section, uniquely for each OSAmultichannel MAC protocol.

Finally, for the macroscopic case, the average channelthroughput, excluding switching and sensing overhead, iscomputed as

R ¼ fqpð1� pdÞ þ ð1� qpÞð1� pfÞgRcC; ð11Þ

where Rc ¼Psm

i¼1 i�i and �i is a solution to a steady stateMarkov chain given by [6, (13)]. Since the macroscopic

model assumes no PU activity in each time slot, SUconnection buffering and switching is not needed. Notethat contrary to the incorrect assumptions of [6, (12)], [24, (7)and (9)], we compute R in (9) and (11) taking all thechannels into account, irrespective of the type of OSA MAC.This is because models in [6] and [24] considered only datachannels for the throughput investigation in DCC in thefinal calculation stage, assuming that no data traffic is beingtransmitted on control channel. However, the utilizationmust be computed over all channels, irrespective ofwhether one channel transmitted only control data or not.

4.3 Derivation of State Transition Probabilities forthe Microscopic Model

We denote the state transition probability as

pxyzjklm ¼ PrðXt ¼ x; Yt ¼ y; Zt ¼ zjXt�1 ¼ k; Yt�1 ¼ l; Zt�1 ¼ mÞ:

ð12Þ

Note that changes in Xt and Zt depend on the detection ofthe PU. In addition, changes in Zt depend on OSA trafficcharacteristics such as the packet generation probability p

and the average packet length 1=q. Also, note that thesteady state probability vector � containing all possiblesteady state probabilities �xyz is derived by solving � ¼ �P,where entries of right stochastic matrix P are defined as (12)knowing that

Px;y;z �xyz ¼ 1.

As a parameter to model PU state, pc denotes theprobability that an OSA network collectively detects a PUchannel as occupied,3 i.e.,

pc ¼ qppd þ ð1� qpÞpf : ð13Þ

We introduce two supporting functions. First, we denoteTðjÞk as the probability of termination of j SU connections at

time t given that k channels are utilized by the OSAnetwork at time t� 1, which is derived as [24, (2)]

TðjÞk ¼

kj

� qjð1� qÞk�j; k � j > 0;

0; otherwise:

(ð14Þ

Note that k in TðjÞk denotes the number of channels utilized

by OSA network rather than the number of SU connectionsbecause only active connections can be terminated at thenext time slot. And second, we denote SðjÞm as theprobability of j SU successful new connections at time t,given m connections were active at time t� 1. We need tomodify the definition of SðjÞm given in [24, (5) and (8)]considering PU detection on the control channel. If a PU isdetected on a control channel, an SU connection cannot begenerated because there is no chance to acquire a datachannel. We then have [6, (17)]

SðjÞm ¼

~Sð1Þm ; j ¼ 1 ðDCCÞ;~Sð1Þm

N � 2m� 1

N � 1

MD �mM

; j ¼ 1 ðHCCÞ;1� Sð1Þm ; j ¼ 0;0; otherwise;

8>>>><>>>>:

ð15Þ


3. Note that, contrary to [6, (8)], we do not consider packet capture effectsin the definition of pc as the packet capture model proposed in [6] was anapproximation.

where

~Sð1Þm ¼Sð1Þm ; PU free control channel; DCC;

ð1� pcÞSð1Þm ; otherwise;

ð16Þ

and Sð1Þm ¼ ðN � 2mÞpð1� pÞN�2m�1.Again, note that for the SU connection buffering

protocols, the subparameter m of SðjÞm is not the number of

channels utilized by SUs, but the number of SU connections.

This is because we assume that an SU that has a connection

but pauses data transmission due to the PU presence does

not try to make another connection. We can now derive the

transition probabilities individually for all four different

OSA MAC protocols.

4.3.1 Case B0S0

Recall that for nonbuffering OSA MAC protocols Zt ¼ Xt.Thus, pkljxy is defined as (12) without Zt. Because it isassumed that no more than one connection can begenerated in one time slot, it is impossible to transit fromk connections at time t� 1 to x > kþ 1 connections attime t. The state transition probability for this condition is 0.

For x ¼ kþ 1, only one SU connection is created, no

current connection is terminated and y PUs can appear on

the channels that are not utilized by SU. Thus, from MD � xchannels, a PU appears on y channels, so MD�x

y

� cases of

PU appearances are possible.Now, consider the case x < kþ 1. When an SU data

connection is terminated, there can be two possible reasons:1) an SU completes its transmission, or 2) a PU is detectedon a channel that is assigned to an SU for data transmissionbefore sensing. The former was analyzed [24, Section 3]. Tomodel the latter, we introduce variable i denoting thenumber of channels that were reserved for SU datatransmission before sensing but cannot be utilized due toPU detection. We have the following observation:

Observation 1. For multichannel OSA MAC without SU

connection buffering or channel switching, the number of

PU appearance combinations is xþii

� �MD�x�iy�i

� .

Proof. When the OSA network detects PU on i channels

from xþ i channels that are going to be utilized by SUs

before sensing, there can be xþii

� �possible combinations

for the PU appearance on the channels. For the

remaining MD � x� i channels, y� i channels should

be occupied by the PU because the total number of

channels in which a PU is detected at time t should be y.

Thus, there are MD�x�iy�i

� possible combinations for PU

appearance on unassigned channels. tuFor the SU traffic generation in the case of x < kþ 1, there aretwo possible cases: 1) no connection is created and k� xþ iconnections are terminated, and 2) one connection is createdand k� x� iþ 1 connections are terminated. Recall thatk connections at t� 1 are changed to xþ i connections at tbefore sensing. Also note that i 2 ½0; 1; . . . ;minðsm � x; yÞ�,where sm � x is the number of possible SU connections thatcan be terminated by PU appearance. By summing over allpossible i, we can compute the transition probability for thecase x < kþ 1.

In addition, we need to discuss the edge state4 whichconsiders two cases: 1) no more channels are available,either utilized by SUs or PUs, and 2) all possible SUconnections are established5 which we denote as “fullconnection state.” For the transition from full connectionstate to edge state, we have to consider the case that onenew connection is generated while any existing connectionis not terminated, which means a trial for the newconnection by the free SU is not established because therealready exist all possible connections.

Writing all conditions compactly, denote the indicatorfor the edge state

1x;y ¼1; xþ y ¼MD or x ¼ sm;0; otherwise;

ð18Þ

and define P ðiÞx;y ¼ xþii

� �MD�x�iy�i

� pycð1� pcÞ

MD�y, the complete

state transition probability is defined as

pxyjkl ¼0; x>kþ1

Tð0ÞkSð1ÞkPð0Þx;y ; x¼kþ1;Xim

i¼0

�Tðk�x�iÞk

Sð0ÞkþT ðk�x�iþ1Þ

kSð1Þk

�PðiÞx;y; x<kþ1; k<sm or 1x;y¼0;

Ximi¼0

�Tðk�x�iÞk


kSð1Þk

�� P ðiÞx;yþT

ð0ÞkSð1ÞkPð0Þ0;y; x<kþ1; k¼sm; 1x;y¼1;

8>>>>>>>>>>>><>>>>>>>>>>>>:

ð17Þ

where im ¼ minðsm � x; yÞ.

4.3.2 Case B0S1

Although in the SU connection nonswitching case, bothDCC and HCC can be considered, only DCC will be able toperform switching without any additional control dataexchange, which we prove formally.

Before going into detail of the derivation, note that forthe class of OSA MAC protocols with a dedicated controlchannel, every node can follow the connection arrangementof the entire network. Because the dedicated controlchannel is continuously monitored by all network nodesvia a separate front end [24, Section 2.2], each node canlearn the overall network configuration. Note that this alsoapplies to Split Phase Control Channel MAC (SPCC) [24,Section 2.4], [6, Fig. 2c] as well, since SPCC has a dedicatedcontrol channel phase. For HCC, as well as MultipleRendezvous Control Channel [24], it is impossible for asingle node to learn the whole network connectionarrangement since each sender receiver pair cannot listento others while following its own hopping sequence. Wenow present the following proof:

Theorem 1. Channel switching in DCC can be performedwithout any additional control message exchange.

Proof. We prove this by showing a possible distributedchannel switching process. Following earlier observation,


4. As shown in [6, Section 3.4], the edge state was not considered in [24,(6)], which resulted in an incorrect model.

5. If sm ¼MD, there can be many free SUs but no channel available. Onthe other hand, if sm ¼ bN=2c, there can be one free SU for even N or no freeSU for odd N .

in DCC, each node can trace the connection arrangementof others, i.e., which channel has been reserved by a senderreceiver pair. To distribute the switching events equallyamong SUs, each SU computes the priority level as

�i;t ¼ �i;t�1 þ 1p; ð19Þ

where

1p ¼1; preemption by PU;0; otherwise;

ð20Þ

and �i;t is the priority level of SU i at time t. For �i;0 62 IN,the priority is a MAC address of the SU, transformed intoa real number for each SU by a network-wide knownfunction. Now, having a set of priorities of all commu-nicating node pairs, the OSA network is able to select, ina distributed manner, a new set of communicationchannels upon PU arrival as

IfIa;t; Ib;t; . . . ; Ic;tg|fflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflffl}I

! UfUd;t; Ue;t; . . . ; Uf;tg|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}U

; ð21Þ

where jIj ¼ jUj ¼MD �Xt � Yt, ! is the mappingoperator denoting process of switching active SUconnection i to free channel j, Ii;t denotes index ofcommunicating SUs (transmitters) at time t, where�a;t > �b;t > � � � > �c;t, and Uj;t denotes free channelwith index j at t. tuNote that existing connections that have not been

mapped to a channel are considered blocked. Also notethat for the algorithm given in Theorem 1, connections arepreempted randomly with equal probability by PU. Sincenew SU connections are also assumed to use new channelsrandomly with equal probability, each SU connection isblocked with uniform probability.

To enable SU connection switching in HCC, one way isto augment it with a separate radio front end which wouldfollow the hopping sequences and control data exchange ofthe OSA network. Obviously, this increases the cost ofhardware and contradicts the idea of HCC, where allchannels should be used for data communication. There-fore, while evaluating OSA MAC protocols in Section 5.2,we will not consider SU connection switching for HCC.

We now define the state transition probability pxyjkl for the

considered OSA MAC protocol. Because x > kþ 1 is infea-

sible, the state transition probability for x > kþ 1 equals to

zero. For x ¼ kþ 1, y PUs can appear on any of MD channels

because even though a PU is detected, the SUs can still

transmit data by switching to the idle channels and the

possible number of PU appearances is MD

y

� . Note that the

possible number of PU appearances in the case B0S1 is alwaysMD

y

� , even for the edge state, because the data channel can

be changed by switching to a vacant channel after the PU

detection. Because it is impossible to create more than one

new connection at a time, the OSA connection creation

probabilities for x ¼ kþ 1 are the same as in (17), i.e., Tð0Þk S

ð1Þk .

For x < kþ 1, if SUs are not in a full connection state,there are two cases for the OSA traffic generation: 1) noconnection is created and k� x connections are terminated,and 2) one connection is created and k� xþ 1 connections

are terminated. On the other hand, for the state transition tothe edge state, we use variable i, just like in the case B0S0, toderive the probabilities because connections can be termi-nated by a PU in the full connection state. Furthermore, forthe transition from the full connection state to the fullconnection state, we should take into account T

ð0Þk S

ð1Þk again.

With all conditions, the state transition probabilities aredenoted compactly as

pxyjkl ¼0; x>kþ1;

Tð0ÞkSð1ÞkPð0Þ0;y; x¼kþ1;�

Tðx�kÞk

Sð0ÞkþT ðx�kþ1Þ

kSð1Þk

�Pð0Þ0;y; x<kþ1; 1x;y¼0;Xim

i¼0

�Tðk�x�iÞk


kSð1Þk

�Pð0Þ0;y; x<kþ1; k<sm; 1x;y¼1;

Ximi¼0

�Tðk�x�iÞk


kSð1Þk

�� P ð0Þ

0;yþT ð0Þ

kSð1ÞkPð0Þ0;y; x<kþ1; k¼sm; 1x;y¼1:

8>>>>>>>>>>>>>>><>>>>>>>>>>>>>>>:

ð22Þ

4.3.3 Case B1S0

Before we discuss this case, we present the followingobservation, which implicates the design of simulationmodels and derivation of pxyzjklm for SU connectionbuffering MAC protocols.

Observation 2. For all SU connection buffering OSA MACprotocols, the same average link level throughput results fromcreating a brand new connection or resuming a previouslypreempted and buffered connection on the arrival of PU on achannel.

Proof. Due to the memoryless property of the geometricdistribution

Prð1=qi > 1=qt1 þ 1=qt2 j1=qi > 1=qt1Þ ¼ Prð1=qi > 1=qt2Þ;

ð23Þ

where 1=qi is the duration of connection i, 1=qt1 is theconnection length until time t1 when it has beenpreempted by PU, and 1=qt2 is the remaining length ofthe connection after SU resumes connection at time t2.Since either a newly generated SU connection afterresumption, or the remaining part of a preemptedconnection needs a new connection arrangement on thecontrol channel, the number of slots occupied by eachconnection type is the same. tuHaving Observation 2, we can derive transition prob-

abilities. Because packet generation is affected by thenumber of connections, we use Zt to classify conditions toderive the state transition probabilities. Due to the assump-tion of a maximum number of one connection generation inone time slot, the state transition probability of the case ofz > mþ 1 is zero.

For z � mþ 1, a data connection is terminated only if the

transmitting node completes its transmission without PU

interruption. When a PU is detected in the channel, the SU

temporarily pauses communication without terminating

the connection. Among z SU connections, x connections

actually utilize channels for data transmission and z� x


connections are paused due to PU detection. Thus, zz�x� �

combinations of PU appearance are possible. At the same

time, because the total number of channels occupied by PUs

is y, the remaining y� zþ x PUs should appear on MD � zidle channels. Thus, there can be MD�z

y�zþx

� combinations for

PU appearance on idle channels.The SU connection generation probability for z ¼ mþ 1

is Tð0Þk Sð1Þm , just like in the case B0S0. For z < mþ 1: 1) no

connection is generated and m� z connections are termi-nated, and 2) one connection is generated and m� zþ 1

connections are terminated. For z ¼ mþ 1, one connection

is generated while no connections are terminated.For the transition between full connection states, i.e., the

state transition from m ¼ sm to z ¼ sm, we should take into

account the case that one connection is generated and noconnections are terminated because there is no available

resources for a new connection.

Finally, considering all these cases and defining

RðzÞx;y ¼ zz�x� �

MD�zy�zþx

� pycð1� pcÞ

MD�y, the state transition prob-

ability is given as

pxyzjklm ¼0; z>mþ1;

Tð0ÞkSð1Þm R

ðzÞx;y; z¼mþ1;�

Tðm�zÞk

Sð0Þm þT

ðm�zþ1Þk

Sð1Þm

�RðzÞx;y; z<mþ1; m<sm or z<sm;�

Tð0ÞkSð0Þm þT

ð1ÞkSð1Þm þT

ð0ÞkSð1Þm

�RðzÞx;y; z¼m¼sm;

8>>><>>>:

ð24Þ

Note that this OSA MAC has been previously analyzed in [6].As it has been pointed out, the model proposed did not

work well for the full range of parameters. This is due tothe following. A Markov model has been derived for

fXt; Ytg using unmodified transition probabilities of [24,(6)] used to calculate average throughput of networks based

on non-OSA multichannel MAC protocols. With thislimitation termination, the probability in [6, (14)], analog

to (14), included an aggregated stream of PU and SU traffic,

where PU traffic qp was later subtracted from steady statechannel utilization in [6, (10)], analog to (9). The approx-

imation of [6], although Markovian, worked reasonablywell only for a moderate values of PU activity qp.

4.3.4 Case B1S1

This OSA MAC from analysis perspective is the same as the

buffering OSA MAC with no channel switching, except for

the following two differences: first, if there is at least one

idle channel, an SU that has a connection but does not utilize

a channel cannot exist because this SU can switch to the idle

channel. Formally, the state transition to the state of z 6¼ xand xþ y < MD and the state transition from the statem 6¼ kand kþ l < MD are not possible. Second, in contrast to the

nonswitching OSA MAC, y PUs can appear in any of MD

channels because the SUs can switch to the idle channels in

this option, the same as for B0S1 case. Thus, the possible

number of cases of PU appearance is just MD

y

� . Therefore,

replacing RðzÞx;y with Rð0Þ0;y and adding conditions z 6¼

x; xþ y < MD;m 6¼ k and kþ l < MD to the condition z >

mþ 1 in (24) results in a complete definition of pxyzjklm. For

consistency, we present this state transition probability as

pxyzjklm ¼0; z>mþ1 or z 6¼x; xþy<MD

or m6¼k; kþl<MD;

Tð0ÞkSð1Þm R

ð0Þ0;y; z¼mþ1;�

Tðm�zÞk

Sð0Þm þT

ðm�zþ1Þk

Sð1Þm

�Rð0Þ0;y; z<mþ1; m<sm or z<sm;�

Tð0ÞkSð0Þm þT

ð1ÞkSð1Þm þT

ð0ÞkSð1Þm

�Rð0Þ0;y; z¼m¼sm:

8>>>>>><>>>>>>:

ð25Þ

4.3.5 Impact of Channel Error on the Throughput

Calculations

All previous analyses were done under the assumption of

the error-free channel. In this section, we will briefly discuss

the impact of channel error on the throughput calculations.Channel error impacts the throughput in two ways. First,

error affects throughput when SU involved in a connection

setup fails to receive a control message from the transmitter.

As a result, no connection is established. Second, error

affects throughput when SU not associated with the current

connection setup (which does not overhear the connection

setup from others) collides later with other users, believing

incorrectly it selected a free data channel for communica-

tion. The throughput of DCC is impacted by both effects.

On the other hand, HCC is influenced only by the first

effect. It is because HCC MAC protocol implementations do

not posses a separate control channel. Thus, no overhearing

of connection setup is possible and users with HCC MAC

protocol select data channel for communication automati-

cally according to a predefined hopping sequence. Because

of the prohibitive complexity of the analysis of the second

effect, we focus on the first error case and the HCC.For HCC, the control channel is selected as one of the

data channels by a hopping method. Thus, if we assume an

error on the control channel, it is reasonable to consider the

error on the data channel as well. For the control channel, if

an error occurs, a connection fails to be established. Thus, it

is modeled by multiplying Sm by 1� pe, where pe is a

probability of error in the current time slot. For the data

channel, different error handling strategies can be consid-

ered. We focus on the two following situations: 1) case E1

denoting packet punctured by unrecovered errors and

2) case E2 denoting transmission termination on error.Case E1. It can be assumed that when an error occurs on

a time slot, the SU simply discards that time slot and

resumes transmitting the remaining packet fragment from

the next correct time slot. This is modeled by replacing the

capacity C with Cð1� peÞ.Case E2. It can also be assumed that the connection

terminates when an error occurs. Thus, the probability that

the packet finishes transmitting, q, should be replaced by

q þ ð1� qÞpe. In addition, if the control channel hops to a

channel which is being utilized for data transmission but

error occurs, a new connection cannot be established. This is

modeled by multiplying Sm by ð1� peÞ2.


5 NUMERICAL RESULTS

We now present numerical results for our model. Due to avast combination of parameters to consider, we havedecided to follow the convention in [6] and [24] and focuson two general network setups (unless otherwise stated):1) small-scale network with M ¼ 3, N ¼ 12, d ¼ 5 kB and2) large-scale network with M ¼ 12, N ¼ 40, d ¼ 20 kB.

In this section, we will also compare the analyticalmodel of the sensing layer and OSA MAC protocols tosimulation results. The simulations were developed withMatlab and reflect exactly the sensing models and MACprotocols presented in this paper. Simulation results foreach system were obtained using the method of batchmeans for a 90 percent confidence interval. To evaluate thesensing protocols, each batch contained 100 events and thewhole simulation run was divided into 10 batches with nowarm up phase. When simulating the OSA MAC protocols,each batch contained 1,000 events while the wholesimulation was divided into 100 batches with the warmup period of 100 events.

5.1 Spectrum Sensing Architecture Performance

For all possible combinations of sensing architectures, wecompute the probability of false alarm for a wide range of tq.For two networks considered, we select a set of the followingcommon parameters: tt ¼ td;max ¼ 1 ms, C ¼ 1 Mbps, b ¼ 1MHz, qp ¼ 0:1 (which approximately corresponds to thelevel of actual measured PU occupancy on the channel from[44, Table 1]), � ¼ �5 dB, and � ¼ 1=M. In all sections,except for Section 5.1.2, we present results for an error-freechannel. Note that for all results presented in this section,simulation results for all protocols confirm the accuracy ofthe analytical model.

5.1.1 Measurement Reporting Protocol Performance

The results are presented in Fig. 2. For the same pdrequirement, SSMA results in the lowest pf for each valueof tq, while TDMA performs worst. The benefit ofintroducing TTDMA in comparison to TDMA is clearlyvisible for all network scenarios.

The advantage of TTDMA and SSMA can be shownmore clearly if we compare the results of different pd ¼pd;min requirements. We can observe that high detectionrequirement such as pd ¼ 0:99 makes the performanceworse, as generally known. However, if TTDMA or SSMAis applied, the performance for pd ¼ 0:99 can be higher than

that of TDMA for pd ¼ 0:9. For example, in the range oftq < 50 �s in Fig. 2a, SSMA for pd ¼ 0:99 outperformsTDMA for pd ¼ 0:9. Moreover, in Fig. 2b, for tq <� 550 �s,SSMA and TTDMA for pd ¼ 0:99 outperform TDMA forpd ¼ 0:9.

It is important to note that �TTDMA performs worse

than the rest of the protocols. It is due to excessive delay

caused by instant acknowledgment of reporting result to

the cluster head node. Note that �TTDMA is a lower bound

for the operation of TTDMA. Also note that when TDMA

needs to be equipped with acknowledgment function, as

�TTDMA, its performance would be degraded in the same

way as TTDMA. Since we analyze static network with

preset parameter values, e.g., � does not change over time,

in the following sections, we proceed with unmodified

TTDMA only.

5.1.2 Impact of Channel Errors during Reporting on PU

Detection Performance

The results are presented in Fig. 3. For small- and large-scale

network, and the same parameters as used in Section 5.1.1,

we have observed the probability of false alarm keeping

detection probability pd constant for varying quiet time tq.

First, it is obvious when comparing Figs. 2 (no channel

error) and 3 (channel error), the impact of error is clearly

visible, i.e., pf increases for every protocol. However, the

relation between individual protocols is the same since error

affects all protocols equally. Second, the effect of error on

the small-scale network is smaller than for the large-scale

network, compare Figs. 3a and 3b, since the probability that

SU will send a wrong report is larger for network with large

number of nodes. Lastly, for small values of �, probability of

false alarm stabilizes and never reaches zero. However,

large values of � reduce significantly the effect of channel

errors. It is because with high �, probability of making an

error decreases rapidly. With 20 percent of nodes partici-

pating in the cooperative agreement on PU state, � ¼ 2 for

small network and � ¼ 8 for large-scale network, effect of

error is reduced almost to zero.

5.1.3 Impact of Cooperation Level on PU Detection

Performance

The results are presented in Fig. 4. We have selectedTTDMA and set pd ¼ pd;min ¼ 0:99 as a protocol for further


Fig. 2. Performance of different measurement reporting protocols as afunction of tq and pd ¼ pd;min for (a) M ¼ 3, N ¼ 12, and (b) M ¼ 12,N ¼ 40. Common parameters: pe ¼ 0, � ¼ 1 (“or” rule), tt ¼ td;max ¼ 1ms, C ¼ 1 Mbps, b ¼ 1 MHz, qp ¼ 0:1, � ¼ �5 dB, � ¼ 1=M, and ng ¼ 1.

Fig. 3. The effect of channel errors on the performance of differentmeasurement reporting protocols as a function of tq and pd ¼ pd;min ¼0:99 for (a) M ¼ 3, N ¼ 12, � ¼ f1; 2g, and (b) M ¼ 12, N ¼ 40,� ¼ f1; 8g. All remaining parameters are the same as in Fig. 2 exceptfor pe ¼ 0:01.

investigation. We observe that for the small-scale network,see Fig. 4a, the performance for � ¼ 2 is the best, while forthe large-scale network, see Fig. 4b, the best performancecan be achieved when � ¼ 6 or 12 if pf < 0:1. Based on thisobservation, we conclude that for given detection require-ments, high detection rate of PU is obtained when � is wellbelow the total number of SUs in the network. While for theconsidered setup optimal � � 20%, this value might bedifferent for other network configurations.

5.1.4 Impact of Group Size on PU Detection

Performance

The results are presented in Fig. 5. To contrast the impact ofgroup size, we choose M ¼ 12, N ¼ 20 as the small-scalenetwork. We perform experiments only for the case when�ms equal for all groups, which means that the number ofgroups is the divisor of M, i.e., ng 2 f1; 2; 3; 4; 6; 12g.

An interesting observation is that the number of groups toachieve the best performance becomes larger as the numberof usersN increases. For the small-scale network, see Fig. 5a,the best performance is observed for ng ¼ 2 or ng ¼ 3, whilefor large-scale network, Fig. 5b, ng ¼ 6 is the best. This isbecause for the large-scale network, the reporting overheadcaused by large number of users offsets the performanceimprovement achieved by large cooperation scale.

5.1.5 Impact of � on PU Detection Performance

The results are presented in Fig. 6. For two network sizes,large and small, three sensing MAC protocols and fixedvalue of pd, we vary � to see its impact on the sensing MACprotocol performance. We observe probability of false alarmpf as well as change in quite time tq. First, we notice thatvarying � does not change probability of false alarm for anyprotocol, in both network configurations. Moreover, thelowest probability of false alarm is obtained when small

number of users agree on the PU state. The larger the

channel number, the larger the range of � when network

obtains the lowest probability of false alarm, compare

Figs. 6a and 6b. The trends of pf for both network

configurations are the same, since all protocols keep false

alarm rate on the same level irrespective of the parameter

change. In case of quiet time, TDMA and SSMA have qpconstant and independent from �, which differs them from

TTDMA whose operation strictly depends on the value of �

considered. And again, when comparing Figs. 6c and 6d,

the optimal value of tq for TTDMA is in the same range as

pf which proves the optimality of the design.

5.2 OSA MAC Protocol Performance

To evaluate the effectiveness of all proposed and analyzed

MAC protocols, we have fixed C ¼ 1 Mbps, p ¼ e�1=N ,

tq ¼ tp ¼ 100 �s, tt ¼ td;min ¼ 1 ms, pd;min ¼ pd ¼ 0:99, and

pf ¼ 0:1. Note that we do not relate pd and pf with the

actual spectrum sensing process at this moment (this will

be done in Section 5.3), assuming that spectrum sensing

layer is able to obtain such quality of detection. Again, as in

Section 5.1, results are presented separately for error-free

and error channel.

5.2.1 Impact of PU Activity Level on OSA MAC

Protocols

The results are presented in Fig. 7. We observe that PU

activity degrades DCC and HCC for B0S0, irrespective of

other network parameters. Their performances are compar-

able in this case. DCC and HCC perform best with B1S0.

The results show that the nonbuffering OSA MAC protocols

are very sensitive to qp where the greatest throughput

decrease is visible at low ranges of PU activity. On the other

hand, with connection buffering, we observe a linear

relation between qp and Rt.


Fig. 4. Performance of TTDMA as a function of tq and � for (a) M ¼ 3,N ¼ 12, and (b) M ¼ 12, N ¼ 40. Rest of the parameters are the sameas in Fig. 2, except for pd;min ¼ 0:99.

Fig. 5. Performance of TTDMA as a function of tq and ng for (a) M ¼ 12,N ¼ 20, and (b) M ¼ 12, N ¼ 40. Rest of the parameters are the sameas in Fig. 2, except for pd;min ¼ 0:99.

Fig. 6. The effect of varying � on (a), (b) false alarm probability, and (c),(d) quiet time of different measurement reporting protocols as a functionfor pd ¼ pd;min and (a), (c) small-scale network, M ¼ 3, N ¼ 12, and (b),(d) large-scale network, M ¼ 12, N ¼ 40. Rest of the parameters are thesame as in Fig. 2 except for te ¼ 20 �s.

5.2.2 Impact of SU Packet Size on OSA MAC Protocols

The results are presented in Fig. 8. Obviously, for larger SUpacket size, the OSA network is able to grab more capacity.However, when packets become excessively large, thethroughput saturates. It remains that with no bufferingand no channel switching, protocols obtain the lowestthroughput, no matter what network setup is chosen.Interestingly, although intuitively B1S1 should obtain thehighest channel utilization, it does not perform better thanB1S0 due to large switching time. With tp approaching zero,DCC B1S1 would perform best, irrespective of the networksetup as we discuss below.

5.2.3 Impact of Switching Time on OSA MAC Protocols

The results are presented in Fig. 9. In this experiment, weverify that for small tp, DCC B1S1 outperforms DCC B1S0.However, there is no huge difference between theirperformances even at tp ¼ 10 �s. This is because connectionswitching does not seriously impact the data throughputfor the network setups in which the number of channels isless than the number of possible connections, i.e., M < 2N .For this network setup, all channels are utilized in most oftime, and therefore, there may not exist many idle channelsto switch. The performance of DCC B0S1 is also improvedfor small tp, and we observe that DCC B0S1 outperformsHCC B1S0 for large-scale network, see Fig. 9b.

5.2.4 Relation between Number of SUs and PU

Channels

Finally, we want to explore the relationship between thenumber of OSA network users and the number of availablePU channels. The results are presented in Fig. 10.

With increasing ratio N=M, we observe an increasingthroughput, where at some point, all protocols almost

saturate. Again, because of the high switching penalty, DCCwith B1S1 is inferior to B1S0. For small-scale network, asshown in Fig. 10a, a separate comment is needed for HCCB1S0. For small N=M, DCC with B1S1 and B1S0 obtainshigher throughput than HCC B1S0. However, for highN=M, HCC B1S0 achieves the highest Rt of all protocols.For large-scale network as shown in Fig. 10b, comparingchannel switching and buffering options, we conclude thatmuch more channel utilization is obtained by connectionbuffering than by channel switching alone when N=M > 1.

Note that for all cases described in this section, simulationresults agree with our analytical model. Comparing ourmodel and analytical results in [6] for DCC B1S0, seeFig. 10b, we observe that prior analysis overestimated theperformance resulting in more than 2 Mbps difference atN=M ¼ 1. Interestingly, if we consider the same set ofparameters as in Section 5.2.1, then the model of Pawełczaket al. [6] almost agrees with the model of our paper. Since theset of parameters that has been chosen in 5.2.1 are similar to[6], we remark that the observations on the performance ofthis OSA MAC in [6] were reflecting the reality.

5.2.5 Impact of Channel Errors on the OSA Multichannel

MAC Performance

To observe the impact of channel errors on the MACprotocol throughput, we have set up the following experi-ment. For HCC and both network sizes, small and large, wehave observed the average throughput for different SUpacket lengths and channel error probabilities. The resultsare presented in Fig. 11. For comparison in Fig. 11, wepresent the system with no errors, denoted as E0. We keptvalues of pe realistic, not exceeding 1 percent. Obviously,system with punctured errors E1 obtains much higherthroughput than system E2, since more data can be


Fig. 7. Performance of OSA MAC protocols versus PU activity level for(a) M ¼ 3, N ¼ 12, d ¼ 5 kB and (b) M ¼ 12, N ¼ 40, d ¼ 20 kB.Common parameters: pe ¼ 0, p ¼ e�1=N , pd ¼ pd;min ¼ 0:99, pf ¼ 0:1,tq ¼ tp ¼ 100 �s, tt ¼ td;max ¼ 1 ms, and C ¼ 1 Mbps.

Fig. 8. Performance of OSA MAC protocols versus packet size d for(a) M ¼ 3, N ¼ 12, and (b) M ¼ 12, N ¼ 40. Rest of the parameters arethe same as in Fig. 7, except for qp ¼ 0:1.

Fig. 9. Performance of OSA MAC protocols versus channel switchingtime tp for (a) M ¼ 3, N ¼ 12, d ¼ 5 kB and (b) M ¼ 12, N ¼ 40,d ¼ 20 kB. Rest of the parameters are the same as in Fig. 7, except forqp ¼ 0:1.

Fig. 10. Performance of OSA MAC protocols for ratio of number of SUsto number of PU channels for (a) M ¼ 3, d ¼ 5 kB and (b) M ¼ 12,d ¼ 20 kB. Rest of the parameters are the same as in Fig. 7, except forqp ¼ 0:1.

potentially sent after one control packet exchange. Again,buffering allows to obtain higher throughput in comparisonto nonbuffered case, even with the data channel errorspresent. Note that system E2 is more prone to errors thanE1, observe Figs. 11a and 11b for B1S0 E1 and B1S0 E2.

5.2.6 Impact of PU Channel Occupancy Distributions on

the OSA Multichannel MAC Performance

All previous analyses were done under the assumption thattraffic generated by SU and channel occupancy of PU can bedescribed by the geometric process. This assumption holdsgenerally either for SU traffic or PU channel occupancystatistics. For example, it has been shown recently in [45]that geometric process constitutes more than 60 percent ofthe measured PU traffic in GSM 900 uplink, GSM 1,800downlink, DECT, and 2.4 GHz UNII channels. It isimportant however to see the behavior of the considereddata MAC protocols with other traffic distributions. Sincethe impact of different SU packet length distributions hasbeen investigated in [24, Section 5.2], concluding thatcomparable throughput of multichannel MAC protocols isobtained, we focus on the impact of different PU trafficdistributions on the OSA network performance. Due to vastnumber of combinations of protocol and traffic distribu-tions, we have narrowed our presentation to DCC and thefollowing distributions: 1) discrete uniform (denotedsymbolically as U), 2) log-normal (denoted symbolically asL), and for comparison 3) geometric (denoted symbolicallyas E) used in the analysis. We have tested the protocolperformance for different combinations of “on” and “off”times of PU activity. These were EE, LE, EL, LL (all possiblecombinations of “on” and “off” times obtained in [45,Tables 3 and 4]) and additionally, EU, UU, where first andsecond letter denote selected distribution for “on” and “off”times, respectively. Due to the complexity of the analysis,we show only the simulation results using the samesimulation method of batch means, with the same para-meters as described at the beginning of Section 5.

The parameter of each distribution was selected such that

the mean value of each distribution was equal to 1=pc for “on”

time and 1� 1=pc for “off” time. The uniform distribution has

a noncontinuous set of mean values, ðab þ anÞ=2, where

ab; an 2 IN denoting lower and upper limit of the distribution,

respectively, which precludes the existence of every mean

“on” or “off” value for pc 2 ð0; 1Þ. To solve that problem, a

continuous uniform distribution with required mean was

used and rounded to the highest integer. This resulted in a

slightly lower last peak in the probability mass function at anfor 1=pc 62 IN or 1� 1=pc 62 IN. In case of log-normal distribu-

tion, because it is continuous, it was rounded it to the nearest

integer as well, with scale parameter ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffilogðvl=c2

l þ1Þq

and

location parameter �¼ logðc2l

ffiffiffiffiffiffiffiffiffiffiffiffivlþc2

l

q �1

Þ, where cl ¼ 1=pc,

vl ¼ ð1� pcÞ=p2c is the mean and variance of the resulting

discretized log-normal distribution. Note that the variance of

the used discretized log-normal distribution is equal to the

variance of geometric distribution for the same mean value.

The variance of resulting discretized uniform continuous

distribution could not be equal to the variance of the

geometric distribution due the reasons described earlier.


Fig. 11. Throughput of HCC OSA MAC as a function of packet size d for(a) M ¼ 3, N ¼ 12, pe ¼ 0:1 and (b) M ¼ 12, N ¼ 40, pe ¼ 0:01, and twodistinct error handling strategies. Rest of the parameters are the sameas in Fig. 7, except for qp ¼ 0:1. E1 and E2 denote error modelsdescribed in Section 4.3.5. E0 denotes the system with pe ¼ 0.

Fig. 12. Impact of different PU “on” and “off” times distributions on OSADCC multichannel MAC performance; (a)-(d) small-scale network,(e)-(h) large-scale network, as described in as in Fig. 7. E, U, and Ldenote geometric, uniform, and log-normal distribution, respectively,where the first and second parameter in the legend denote “on” and “off”time, respectively. (a) DCC B0S0 SN. (b) DCC B0S1 SN. (c) DCC B1S0 SN.(d) DCC B1S1 SN. (e) DCC B0S0 LN. (f) DCC B0S1 LN. (g) DCC B1S0 LN.(h) DCC B1S1 LN. SN denotes small network and LN denotes largenetwork.

The results are presented in Fig. 12. We focus on twonetwork types, as indicated earlier: 1) large-scale and2) small-scale, with the assumed parameters as in Fig. 7.We select four values of qp for the clarity of the presentation.The most important observation is that irrespective of theconsidered distribution, DCC obtains relatively the samethroughput and the same relation between different proto-col options exists as it was shown analytically in Fig. 7. Ifone wants to select the distribution combination with thehighest throughput, it would be LE and LL, while thethroughput obtained being almost equal to the one obtainedvia analysis for the geometric distribution. The distributionwith the lowest throughput is UU and EU, due to thedifference of the second moment between the other twodistributions for the “on” time. The difference in through-put between UU, EU, and the remaining distributions ismore visible for the large network. The most surprisingresult of this investigation is that any DCC MAC protocoloption with buffering removes the impact of distributiontype on the obtained performance, compare Figs. 12a and12c, or 12e with 12h for any value of qp.

5.3 Performance of Joint Spectrum Sensing andOSA MAC Protocols

Having results for spectrum sensing protocol and OSAMAC, we join these two layers to form a complete OSAnetwork stack. By means of exhaustive search, we solve theoptimization problem of (1). We will also investigate the setof parameters that maximize Rt for small- and large-scalenetwork.

We divide our analysis in macroscopic and microscopiccase observing Rt for small-scale network with M ¼ 3,N ¼ 12, d ¼ 5 kB, and large-scale network with M ¼ 12,N ¼ 40, d ¼ 20 kB. For each case, we select a set of spectrumsensing and OSA MAC protocols that are possible and, aswe believe, most important to the research community. Fora fixed set of parameters C ¼ 1 Mbps, b ¼ 1 MHz,p ¼ e�1=N , td;max ¼ 1 ms (microscopic case), td;max ¼ 2 s(macroscopic case), � ¼ 1=M, tt ¼ 1 ms, pd;min ¼ 0:99, � ¼�5 dB, qp ¼ 0:1, and tp ¼ 100 �s, we leave �, te, ng, and pf asoptimization variables.

5.3.1 Microscopic Model

Here, we focus only on DCC protocol, since collaborativespectrum sensing is only possible via a PU free control

channel, which is inefficient to accomplish with HCC. Also,for sensing measurement dissemination, we do not considerSSMA, which would be most difficult to implement inpractice. The results are presented in Fig. 13.

DCC B1S0 with TTDMA is the best option, both forsmall-scale and large-scale networks, see Figs. 13a and 13b,respectively. Because of relatively high switching time, B1S1

performs slightly worse than B1S0, for small- and large-scale network. DCC B0S0 with TDMA is the worst protocolcombination, which confirms earlier results from Sections5.1 and 5.2. Irrespective of network size, it is always betterto buffer SU connections preempted by PU than to look forvacant channels, compare again B1S0 and B0S1 in Figs. 13aand 13b. The difference between B0S0 and B0S1 is mostlyvisible for a large network scenario, see Fig. 13, since with alarge number of channels, there are more possibilities tolook for empty channels.

For all protocol combinations and both network sizes,� ¼ 2 maximizes throughput performance, see Fig. 13a.Interestingly, network size dictates the size of a sensinggroup. For small-scale network, ng ¼ 1 is the optimal value,see Fig. 13a, but for a large network, Rt is maximized whenng ¼ 3 (for B0S0) and ng ¼ 4 (for the rest). We can concludethat with a small network, it is better to involve all nodes insensing, while for larger networks, it is better to dividethem into groups, which agrees with the observation fromSection 5.1.4. Moreover, we observe that the performancedifference between TTDMA and TDMA is not as big as inFig. 2 when parameters are optimized.

The most interesting result is observed for pf . With theincrease of protocol complexity, false alarm increases aswell. Also, with an increase of pf , quiet time is decreasing.Because buffering and switching improves the performance,there can be more margin to design the spectrum sensing.

5.3.2 Macroscopic Model

For the macroscopic model, we explore both non-OSA DCCand HCC with TDMA and TTDMA as sensing protocols.The results are presented in Fig. 14.

DCC obtains higher throughput than HCC for a small-scale network, and vice versa, compare Figs. 14a and 14b,respectively. This confirms the observations of [6, Fig. 3],[24, Fig. 3]. Just like in Fig. 13a, for small-scale network,� ¼ 2 and ng ¼ 2 are the ones that maximize Rt. For thelarge-scale network, however, � ¼ 3 and ng ¼ 3 is optimalfor TDMA, and � ¼ 4 and ng ¼ 4 for TTDMA. This meansthat for large networks, it is beneficial to split the network


Fig. 13. Optimization result of the selected protocol combination withDCC for the microscopic model for (a) M ¼ 3, N ¼ 12, and (b) M ¼ 12,N ¼ 40. Common parameters: pe ¼ 0, d ¼ 5 kB, C ¼ 1 Mbps, b ¼ 1MHz, p ¼ e�1=N, td;max ¼ 1 ms, � ¼ 1=M, tt ¼ 1 ms, pd;min ¼ 0:99, � ¼�5 dB, qp ¼ 0:1, and tp ¼ 100 ms.

Fig. 14. Optimization result of the selected protocol combination for themacroscopic model for (a) M ¼ 3, N ¼ 12, and (b) M ¼ 12, N ¼ 40.Common parameters are the same as in Fig. 13, except for td;max ¼ 2 s.

into smaller groups. Again, this confirms our findings fromSection 5.3.1. For both network scenarios, pf and te arerelatively the same for all protocols considered.

Note that for the large-scale network in the macroscopicmodel, an SU takes more time to detect a PU than in themicroscopic model because large td;max reduces the timeoverhead. The release of time restriction impacts the large-scale network by requiring greater value of � to achieve themaximum throughput.

6 CONCLUSION

We have presented a comprehensive framework enablingassessment of the performance of joint spectrum sensingand MAC protocol operation for OSA networks. In themodel we have proposed, we focused on the link layerthroughput as the fundamental metric to assess perfor-mance. We have parametrized spectrum sensing architec-tures for energy detection-based systems with collaborativemeasurements combining. We have proposed a novelspectrum sensing MAC denoted Truncated Time DivisionMultiple Access. We have also categorized multichannelMAC protocols for OSA networks based on their ability tobuffer and switch existing SU connections on the arrival of aPU. Our analysis is supported by simulations which provethe accuracy of the obtained expressions.

Our future task will be to investigate the delayexperience by using any of OSA MAC protocols proposed.We plan to develop a comprehensive simulation softwarewhich will implement features not covered by our model,like queue per each SU.

REFERENCES

[1] R.V. Prasad, P. Pawełczak, J. Hoffmeyer, and S. Berger, “CognitiveFunctionality in Next Generation Wireless Networks: Standardi-zation Efforts,” IEEE Comm. Magazine, vol. 46, no. 4, pp. 72-78,Apr. 2008.

[2] J. Park, R. Jain, and D. �Cabri�c, “Spectrum Sensing DesignFramework Based on Cross-Layer Optimization of DetectionEfficiency,” Proc. IEEE Int’l Conf. Comm. (ICC ’09), June 2009.

[3] M. Timmers, S. Pollin, A. Dejonge, L. Van der Perre, and F.Catthoor, “A Distributed Multichannel MAC Protocol for Multi-hop Cognitive Radio Networks,” IEEE Trans. Vehicular Technology,vol. 59, no. 1, pp. 446-459, Jan. 2010.

[4] Y.-C. Liang, Y. Zeng, E.C. Peh, and A.T. Hoang, “SensingThroughput Tradeoff in Cognitive Radio Networks,” IEEE Trans.Wireless Comm., vol. 7, no. 4, pp. 1326-1337, Apr. 2008.

[5] W.S. Jeon, D.G. Jeong, J.A. Han, G. Ko, and M.S. Song, “AnEfficient Quiet Period Management Scheme for Cognitive RadioSystems,” IEEE Trans. Wireless Comm., vol. 7, no. 2, pp. 505-509,Feb. 2008.

[6] P. Pawełczak, S. Pollin, H.-S.W. So, A. Bahai, R.V. Prasad, and R.Hekmat, “Performance Analysis of Multichannel Medium AccessControl Algorithms for Opportunistic Spectrum Access,” IEEETrans. Vehicular Technology, vol. 58, no. 6, pp. 3014-3031, July 2009.

[7] P. Papadimitratos, S. Sankaranarayanan, and A. Mishra, “ABandwidth Sharing Approach to Improve Licensed SpectrumUtilization,” IEEE Comm. Magazine, vol. 43, no. 12, pp. S10-S14,Dec. 2005.

[8] A.T. Hoang, Y.-C. Liang, D.T.C. Wong, Y. Zeng, and R. Zhang,“Opportunistic Spectrum Access for Energy-Constrained Cogni-tive Radios,” IEEE Trans. Wireless Comm., vol. 8, no. 3, pp. 1206-1211, Mar. 2009.

[9] X.Y. Wang, A. Wong, and P.-H. Ho, “Extended Knowledge-Based Reasoning Approach to Spectrum Sensing for CognitiveRadio,” IEEE Trans. Mobile Computing, vol. 9, no. 4, pp. 465-478,Apr. 2010.

[10] E. Hossain, D. Niyato, and Z. Han, Dynamic Spectrum Access andManagement in Cognitive Radio Networks. Cambridge Univ., 2009.

[11] S. Srinivasa and S.A. Jafar, “How Much Spectrum Sharing IsOptimal in Cognitive Radio Networks?” IEEE Trans. WirelessComm., vol. 7, no. 10, pp. 4010-4018, Oct. 2008.

[12] S. Tang and B.L. Mark, “Modelling and Analysis of OpportunisticSpectrum Sharing with Unreliable Spectrum Sensing,” IEEE Trans.Wireless Comm., vol. 8, no. 4, pp. 1934-1943, Apr. 2009.

[13] S. Tang and B.L. Mark, “Analysis of Opportunistic SpectrumSharing with Markovian Arrivals and Phase-Type Service,” IEEETrans. Wireless Comm., vol. 8, no. 6, pp. 3142-3150, June 2009.

[14] X. Zhu, L. Shen, and T.-S.P. Yum, “Analysis of Cognitive RadioSpectrum Access with Optimal Channel Reservation,” IEEEComm. Letters, vol. 11, no. 4, pp. 304-306, Apr. 2007.

[15] E.W.M. Wong and C.H. Foh, “Analysis of Cognitive RadioSpectrum Access with Finite User Population,” IEEE Comm.Letters, vol. 13, no. 5, pp. 294-296, May 2009.

[16] Y. Zhang, “Dynamic Spectrum Access in Cognitive RadioWireless Networks,” Proc. IEEE Int’l Conf. Comm. (ICC ’08), May2008.

[17] D.T.C. Wong, A.T. Hoang, Y.-C. Liang, and F.P.S. Chin, “DynamicSpectrum Access with Imperfect Sensing in Open SpectrumWireless Networks,” Proc. IEEE Wireless Comm. and NetworkingConf. (WCNC ’08), Mar./Apr. 2008.

[18] B. Hamdaoui, “Adaptive Spectrum Assessment for OpportunisticAccess in Cognitive Radio Networks,” IEEE Trans. Wireless Comm.,vol. 8, no. 2, pp. 922-930, Feb. 2009.

[19] W.-Y. Lee and I.F. Akyildiz, “Optimal Spectrum Sensing Frame-work for Cognitive Radio Networks,” IEEE Trans. Wireless Comm.,vol. 7, no. 10, pp. 3845-3857, Oct. 2008.

[20] P. Grønsund, H.N. Pham, and P.E. Engelstad, “TowardsDynamic Spectrum Access in Primary OFDMA Systems,” Proc.IEEE Personal, Indoor and Mobile Radio Comm. Symp. (PIMRC ’09),Sept. 2009.

[21] S. Geirhofer, L. Tong, and B.M. Sandler, “Dynamic SpectrumAccess in the Time Domain: Modeling and Exploiting WhiteSpace,” IEEE Comm. Magazine, vol. 45, no. 5, pp. 66-72, May 2007.

[22] S. Huang, X. Liu, and Z. Ding, “Optimal Sensing-TransmissionStructure for Dynamic Spectrum Access,” Proc. IEEE INFOCOM,Apr. 2009.

[23] J. Gambini, O. Simeone, Y. Bar-Ness, U. Spagnolini, and T. Yu,“Packet-Wise Vertical Handover for Unlicensed Multi-StandardSpectrum Access with Cognitive Radios,” IEEE Trans. WirelessComm., vol. 7, no. 12, pp. 5172-5176, Dec. 2008.

[24] J. Mo, H.-S.W. So, and J. Walrand, “Comparison of MultichannelMAC Protocols,” IEEE Trans. Mobile Computing, vol. 7, no. 1,pp. 50-65, Jan. 2008.

[25] F.E. Visser, G.J.M. Janssen, and P. Pawełczak, “MultinodeSpectrum Sensing Based on Energy Detection for DynamicSpectrum Access,” Proc. IEEE Vehicular Technology Conf. (VTC-Spring), May 2008.

[26] K.B. Letaief and W. Zhang, “Cooperative Communications forCognitive Radio Networks,” Proc. IEEE, vol. 97, no. 5, pp. 878-893,May 2009.

[27] W. Zhang, A.K. Sadek, C. Shen, and S.J. Shellhammer, “AdaptiveSpectrum Sensing,” Proc. Information Theory and ApplicationsWorkshop, Jan./Feb. 2010.

[28] H. Jiang, L. Lai, R. Fan, and H.V. Poor, “Optimal Selection ofChannel Sensing Order in Cognitive Radio,” IEEE Trans. WirelessComm., vol. 8, no. 1, pp. 297-307, Jan. 2009.

[29] H. Li, H. Dai, and C. Li, “Collaborative Quickest SpectrumSensing via Random Broadcast in Cognitive Radio Systems,” Proc.IEEE Global Comm. Conf. (GLOBECOM), Nov./Dec. 2009.

[30] C.R. Stevenson, G. Chouinard, Z. Lei, W. Hu, S.J. Shellhammer,and W. Caldwell, “IEEE 802.22: The First Cognitive RadioWireless Regional Area Network Standard,” IEEE Comm. Maga-zine, vol. 47, no. 1, pp. 130-138, Jan. 2009.

[31] R. Tandra, S.M. Mishra, and A. Sahai, “What Is a Spectrum Holeand What Does It Take to Recognize One?” Proc. IEEE, vol. 97,no. 5, pp. 824-848, May 2009.

[32] T. Yucek and H. Arslan, “A Survey of Spectrum SensingAlgorithms for Cognitive Radio Applications,” IEEE Comm.Surveys and Tutorials, vol. 11, no. 1, pp. 116-13, Jan. 2009.

[33] A.R. Biswas, T.C. Aysal, S. Kandeepan, D. Kilazovich, and R.Piesiewicz, “Cooperative Shared Spectrum Sensing for DynamicCognitive Radio Network,” Proc. IEEE Int’l Conf. Comm. (ICC ’09),June 2009.


[34] C. Sun, W. Zhang, and K.B. Letaief, “Cluster-Based CooperativeSpectrum Sensing in Cognitive Radio Systems,” Proc. IEEE Int’lConf. Comm. (ICC ’07), June 2007.

[35] T. Weiss, J. Hillenbrand, A. Krohn, and F.K. Jondral, “EfficientSignaling of Spectral Resources in Spectrum Pooling Systems,”Proc. IEEE Symp. Comm. and Vehicular Technology (SCVT ’03), Aug.2003.

[36] F. Li and J.S. Evans, “Optimal Strategies for Distributed Detectionover Multiaccess Channels,” Proc. IEEE Int’l Conf. Acoustics, Speechand Signal Processing (ICASSP ’08), Mar./Apr. 2008.

[37] S. Zhang, T. Wu, and V.K.N. Lau, “A Low-Overhead EnergyDetection Based Cooperative Sensing Protocol for Cognitive RadioSystems,” IEEE Trans. Wireless Comm., vol. 8, no. 11, pp. 5575-5581,Nov. 2009.

[38] Y. Yuan, P. Bahl, R. Chandra, P.A. Chou, J.I. Ferrell, T.Moscibroda, S. Narlanka, and Y. Wu, “KNOWS: KognitivNetworking over White Spaces,” Proc. IEEE Int’l Symp. NewFrontiers in Dynamic Spectrum Access Networks (DySPAN ’07), Apr.2007.

[39] L. Ma, X. Han, and C.-C. Shen, “Dynamic Open Spectrum SharingMAC Protocol for Wireless Ad Hoc Networks,” Proc. IEEE Int’lSymp. New Frontiers in Dynamic Spectrum Access Networks(DySPAN ’05), Nov. 2005.

[40] W. Hu, D. Willkomm, G. Vlantis, M. Gerla, and A. Wolisz,“Dynamic Frequency Hopping Communities for Efficient IEEE802.22 Operation,” IEEE Comm. Magazine, vol. 45, no. 5, pp. 80-87,May 2007.

[41] H. Lee and D.-H. Cho, “VoIP Capacity Analysis in CognitiveRadio System,” IEEE Comm. Letters, vol. 13, no. 6, pp. 393-395, June2009.

[42] M.A. Kalil, H. Al-Mahdi, and A. Mitschele-Thiel, “Analysis ofOpportunistic Spectrum Access in Cognitive Ad Hoc Networks,”Proc. Int’l Conf. Analytical and Stochastic Modeling Techniques andApplications (ASMTA ’09), June 2009.

[43] W. Zhang, K. Malik, and K.B. Letaief, “Cooperative SpectrumSensing Optimization in Cognitive Radio Networks,” Proc. IEEEInt’l Conf. Comm. (ICC ’08), May 2008.

[44] P. Pawełczak, S. Pollin, H.-S.W. So, A. Bahai, R.V. Prasad, and R.Hekmat, “Quality of Service of Opportunistic Spectrum Access: AMedium Access Control Approach,” IEEE Wireless Comm. Maga-zine, vol. 15, no. 5, pp. 20-29, Oct. 2008.

[45] M. Wellens, J. Riihijarvi, and P. Mahonen, “Empirical Time andFrequency Domain Models of Spectrum Use,” Elsevier PhysicalComm. J., vol. 2, nos. 1/2, pp. 10-32, Mar.-June 2009.

Jihoon Park received the MSc degree inelectrical engineering from Sogang University,Seoul, Korea, in 1998. From 1998 to 2005, hewas a member of mobile terminal R&D Depart-ment of Samsung Electronics, Seoul, Korea,where he developed software for more than20 commercial Samsung cell phones, includingthe world’s first IS-2000 cell phone and the firstKorean WCDMA cell phone. Since 2005, he hasbeen working toward the PhD degree in

electrical engineering at the University of California, Los Angeles. Hiscurrent research focuses on the cross-layer modeling methods ofpractical opportunistic spectrum access networks. He is a studentmember of the IEEE.

Przemyslaw Pawelczak received the MScdegree from the Wroclaw University of Technol-ogy, Poland, in 2004, and the PhD degree fromDelft University of Technology, The Netherlands.From 2004 to 2005, he was a staff member ofSiemens COM Software Development Center,Wroclaw, Poland. During Fall 2007, he was avisiting scholar at the Connectivity Lab, Uni-versity of California, Berkeley. Since 2009, hehas been a postdoctoral researcher at the

Cognitive Reconfigurable Embedded Systems Lab, University ofCalifornia, Los Angeles. His research interests include cross-layeranalysis of opportunistic spectrum access networks. He is a vice chair ofthe IEEE DySPAN Standardization Committee. He was a coordinatorand an organizing committee member of cognitive radio workshopscollocated with IEEE ICC in 2007, 2008, and 2009. Since 2010, he hasbeen a cochair of the demonstration track of IEEE DySPAN. He was therecipient of the annual Telecom Prize for Best PhD Student inTelecommunications in The Netherlands in 2008 awarded by the DutchRoyal Institute of Engineers. He is a member of the IEEE.

Danijela �Cabri�c received the Dipl. Ing. degreefrom the University of Belgrade, Serbia, in 1998,and the MSc degree in electrical engineeringfrom the University of California, Los Angeles, in2001. She received the PhD degree in electricalengineering from the University of California,Berkeley, in 2007, where she was a member ofBerkeley Wireless Research Center. In 2008,she joined the faculty of Electrical Engineering ofthe University of California, Los Angeles, as an

assistant professor. Her key contributions involve novel radio architec-ture, signal processing, and networking techniques to implementspectrum sensing functionality in cognitive radios. She has written threebook chapters and more than 25 major journal and conference papers inthe fields of wireless communications and circuits and embeddedsystems. She was awarded Samueli Fellowship in 2008 and OkawaFoundation research grant in 2009.

. For more information on this or any other computing topic,please visit our Digital Library at www.computer.org/publications/dlib.


Performance of Joint Spectrum Sensing and MAC Algorithms for Multichannel Opportunistic Spectrum...

Documents

Transcript of Performance of Joint Spectrum Sensing and MAC Algorithms for Multichannel Opportunistic Spectrum...