WIRELESS spectrum is currently regulated by governmental

Transcription

1 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 4, APRIL Spectrum-Aware Moblty Management n Cogntve Rado Cellular Networks Won-Yeol Lee, Student Member, IEEE, and Ian F. Akyldz, Fellow, IEEE Abstract Cogntve rado (CR) networks have been proposed as a soluton to both spectrum neffcency and spectrum scarcty problems. However, they face several challenges based on the fluctuatng nature of the avalable spectrum, makng t more dffcult to support seamless communcatons, especally n CR cellular networks. In ths paper, a spectrum-aware moblty management scheme s proposed for CR cellular networks. Frst, a novel network archtecture s ntroduced to mtgate heterogeneous spectrum avalablty. Based on ths archtecture, a unfed moblty management framework s developed to support dverse moblty events n CR networks, whch conssts of spectrum moblty management, user moblty management, and ntercell resource allocaton. The spectrum moblty management scheme determnes a target cell and spectrum band for CR users adaptvely dependent on tme-varyng spectrum opportuntes, leadng to ncrease n cell capacty. In the user moblty management scheme, a moble user selects a proper handoff mechansm so as to mnmze a swtchng latency at the cell boundary by consderng spatally heterogeneous spectrum avalablty. Intercell resource allocaton helps to mprove the performance of both moblty management schemes by effcently sharng spectrum resources wth multple cells. Smulaton results show that the proposed method can acheve better performance than conventonal handoff schemes n terms of both cell capacty as well as moblty support n communcatons. Index Terms Cogntve rado, spectrum pool, handoff, ntercell resource allocaton, spectrum moblty management, user moblty management. Ç 1 INTRODUCTION WIRELESS spectrum s currently regulated by governmental agences and s assgned to lcense holders or servces on a long-term bass over vast geographcal regons. Recent research has shown that a large porton of the assgned spectrum s used sporadcally leadng to under utlzaton and wastage of valuable frequency resources [1]. To address ths crtcal problem, the Federal Communcatons Commsson (FCC) has recently approved the use of unlcensed devces n lcensed bands [14]. Ths new area of research foresees the development of cogntve rado (CR) networks to further mprove spectrum effcency. The basc dea of CR networks s that the unlcensed devces (also called CR users) share the lcensed spectrum wthout nterferng wth the transmsson of other lcensed users (also known as prmary users) [2]. If ths band s found to be occuped by a lcensed user, the CR user moves to another spectrum hole to avod nterference, whch s referred to as spectrum moblty. Ths concept has been wdely nvestgated to solve the exponental data traffc growth n the current cellular network [4], [5]. The man dfference between classcal and CR cellular networks les n spectrum moblty, whch gves rse to a new type of handoff n CR cellular networks, the so-called spectrum handoff. In [6], a proactve spectrum handoff approach s proposed where CR users predct. The authors are wth the Broadband and Wreless Networkng Laboratory, School of Electrcal and Computer Engneerng, Georga Insttute of Technology, 777 Atlantc Dr., NW, Atlanta, GA E-mal: {wylee, an}@ece.gatech.edu. Manuscrpt receved 26 May 2009; revsed 25 Oct. 2010; accepted 10 Jan. 2011; publshed onlne 17 Mar For nformaton on obtanng reprnts of ths artcle, please send e-mal to: tmc@computer.org, and reference IEEECS Log Number TMC Dgtal Object Identfer no /TMC future spectrum avalablty based on the past channel hstores, and ntellgently swtch the channel pror to the appearance of prmary users, whch mnmzes dsruptons to prmary users and mantans relable communcaton at CR users. A spectrum handoff scheme based on a dscretetme Markov chan s developed to reduce a forced termnaton probablty of CR transmsson [7]. IEEE , the frst CR standard for CR networks, mantans a backup channel lst so as to provde the hghest probablty of fndng an avalable spectrum band wthn the shortest tme [3]. In [8], an algorthm for updatng the backup channel lst s developed to fnd the dle spectrum bands fast and relably by cooperatng wth neghbor CR users. However, recent research mentoned above has manly focused on spectrum moblty, but does not consder the effect of moble users across multple cells. In the cellular network, moblty management, especally a handoff scheme s one of the most mportant functons. Thus, much research on cellular networks has explored the handoff ssues, manly focusng on cell selecton and resource management n the last couple of decades [9]. Although dverse cell selecton methods have been proposed to support seamless handoff schemes whle maxmzng the network capacty [10], [11], [12], [13], all of them are based on the classcal multcell-based networks and do not consder the fluctuatng nature of spectrum resource n CR networks. Especally, no specal attenton s gven to ether tme and locaton-varyng spectrum avalablty or swtchng delay n traversng the spectrum dstrbuted over a wde frequency range, whch makes conventonal handoff schemes nfeasble n CR cellular networks. As a result, spectrum moblty and user moblty must be jontly consdered n desgnng a moblty management scheme for CR cellular networks, whch consttutes an /12/$31.00 ß 2012 IEEE Publshed by the IEEE CS, CASS, ComSoc, IES, & SPS

2 530 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 4, APRIL 2012 mportant but unexplored topc n CR networks to date. There are the followng challenges:. Heterogeneous moblty events. CR networks are requred to provde two dfferent handoff types: classcal ntercell handoff due to physcal user moblty and spectrum handoff due to spectrum moblty. Thus, t necesstates a unfed moblty management framework to explot two handoff types adaptvely to moblty events.. Dynamc spectrum avalablty. Spectrum avalablty n CR networks vares over tme and space, makng t more dffcult to provde seamless and relable communcatons to moble users traversng across multple cells. For effcent moblty management, CR networks need to mtgate ths dynamc spectrum avalablty by performng moblty management adaptvely dependent on the heterogeneous network condtons.. Broad range of avalable spectrum. In CR networks, avalable spectrum bands are not contguous and found over a wde frequency range. Thus, when CR users swtch ther spectrum bands, they need to reconfgure the operatng frequency of the rado frequency (RF) front-end so as to tune to a new spectrum band, leadng to a swtchng delay much longer than that n classcal wreless networks. To address the challenges mentoned above, we propose a spectrum-aware moblty management scheme for CR cellular networks. Frst, we propose a novel CR cellular network archtecture based on the spectrum poolng concept, whch mtgates the heterogeneous spectrum avalablty. Based on ths archtecture, a unfed moblty management framework s defned so as to support dverse moblty events n CR networks, consstng of ntercell resource allocaton, spectrum moblty management, and user moblty management functons. Through ntercell resource allocaton, each cell determnes ts spectrum confguraton to mprove moblty as well as total capacty. To support spectrum moblty whle mantanng maxmum cell capacty, the spectrum moblty management functon determnes a proper handoff type and target cell for CR users experencng PU actvtes by consderng both spectrum utlzaton and stochastc connectvty models. On the other hand, user moblty management manly focuses on spectrum heterogenety n space, and offers a swtchng cost-based handoff decson mechansm to mnmze servce qualty degradaton n moble users. The rest of the paper s organzed as follows: Sectons 2 and 3 present the proposed network archtecture and moblty management framework for CR networks, respectvely. Handoff models n the proposed framework are presented n Secton 4. In Sectons 5 and 6, novel spectrum and user moblty management methods are proposed, respectvely. Performance evaluaton and smulaton results are presented n Secton 7. Fnally, conclusons are presented n Secton 8. 2 PROPOSED NETWORK ARCHITECTURE The cellular network s the most successful wreless technology, but currently suffers from ncreasng data traffc as data hungry smartphones prolferate. The CR technology s consdered as a promsng soluton to ths data exploson problem n the current cellular network. The CR cellular network s supposed to be deployed n several ways. Frst, t can be appled to the unused TV spectrum bands, the so-called TV whte spaces, as the FCC recently allowed unlcensed devces to use them [5], [14]. Second, whle the ultrabroadband cellular technology such as 3rd Generaton Partnershp Project (3GPP) Long-Term Evoluton-Advanced (LTE-Advanced) requres up to 100 MHz per channel [15], the amount of wdeband spectrum s lmted. The CR technology enables bandwdth aggregaton by sharng spectrum owned by other cellular operators, or opportunstcally utlzng unused spectrum bands lcensed to other servces such as dgtal TV, and publc safety [16]. Fnally, n the current cellular networks, the base staton (BS) has only an RF unt, and a dgtal unt for all communcaton functonaltes s mplemented n a separate central server [17]. As a result, the cost of BSs wll be cheap enough for anybody to nstall anywhere. Ths allows a new type of a moble vrtual operator based on CR, whch operates ts own BSs n a local area wthout spectrum lcenses. In ths paper, we manly focus on the moblty ssues n CR cellular networks. The system model used n ths paper s descrbed n the followng sectons. 2.1 Basc System Model In ths paper, we consder nfrastructure-based CR networks consstng of multple cells. Each cell has a sngle BS and ts CR users. In ths archtecture, CR users observe ther rado envronments and report the results to the BS. Accordngly, the BS determnes proper actons n support of a upper-level control node, such as the moblty management entty (MME) n 3GPP LTE [15]. CR users have a sngle wdeband RF transcever that can sense multple contguous spectrum bands at the same tme wthout RF reconfguraton. Each CR user m needs K m channels to satsfy ts QoS requrement. All spectrum bands are assumed to be lcensed to dfferent prmary networks. Furthermore, each spectrum can have multple prmary networks that are operated ndependently n the dfferent regon, called PU actvty regon. For example, the cell coverage of each BS s consdered as the PU actvty regon. Snce most of prmary networks such as cellular networks or TV broadcastng networks have a fxed servce coverage, the PU actvty regon s generally assumed to be fxed [18]. Generally, each PU actvty regon n the lcensed band has ON and OFF states where ON (Busy) state represents the perod used by prmary users and an OFF (Idle) state represents the unused perod. In ths paper, we assume that the length of ON and OFF perods at PU actvty regon k n spectrum band j s exponentally dstrbuted wth means 1=ðj; kþ and 1=ðj; kþ, respectvely [19], [20], [21]. Then, the dle probablty n the lcensed band P off ðj; kþ can be expressed as ðj; kþ=ððj; kþþðj; kþþ. Snce the state transtons are detected by perodc sensng operatons n every nterval of length t, the PU actvty n the lcensed band s modeled as a two-state Markov chan.

3 LEE AND AKYILDIZ: SPECTRUM-AWARE MOBILITY MANAGEMENT IN COGNITIVE RADIO CELLULAR NETWORKS 531 Fg. 1. Spectrum pool-based CR network archtecture. 2.2 Spectrum Pool Structure In ths secton, we present the proposed spectrum pool and cell archtectures and a correspondng capacty model Spectrum Pool Archtecture for Spectrum Moblty To prevent ntercell nterference, classcal cellular networks generally adopt an nterference coordnaton scheme where each cell uses dfferent spectrum band wth ts neghbor cells [15]. Ths concept can be also appled to CR cellular networks. If the spectrum bands are contguous and located n a relatvely narrow frequency range, moble users can swtch the spectrum wthout changng ther RF front-ends. However, snce avalable spectrum bands n CR networks are spread over a wde frequency range as explaned n Secton 1, CR users need to reconfgure the operatng frequency of ther RF front-end whenever they detect PU actvtes n the current band, resultng n a sgnfcant swtchng latency. To solve the problem n spectrum moblty, we modfy a conventonal spectrum poolng concept, known as the most sutable structure to adapt to the dynamc rado envronment n CR networks [22], [23], for handlng both spectrum and user mobltes n a multcell envronment. The man components of the spectrum pool are defned as follows (Fg. 1):. Spectrum pool: A set of contguous lcensed spectrum bands, each of whch conssts of multple channels.. Spectrum band: A basc bandwdth unt for operatng a certan wreless access technology such as 5 MHz WCDMA band operatng at 2.1 GHz.. Channel: A mnmum logcal unt of wreless resource that moble users can access through multple access schemes. Each channel has the dentcal capacty Cell Archtecture for User Moblty In the proposed archtecture, spectrum pools are assgned to each cell exclusvely wth ts neghbor cells wth a predetermned frequency reuse factor, f, as shown n Fg. 1. Although ths archtecture supports a seamless transton between spectrum bands wthn the pool, we stll has dffculty n provdng seamless communcaton to CR users movng across dfferent cells. To address ths problem, we Fg. 2. Dfferent handoff types n CR networks. ntroduce two dfferent types of cell coverage as depcted n Fg. 1:. Basc area (BA): A basc coverage not overlapped wth that of neghbor cells.. Extended area (EA): A larger coverage overlapped wth the basc areas of neghbor cells. Based on ths archtecture, each cell has multple basc spectrum bands accessble only wthn BA and at least one extended spectrum band supportng both BA and EA, as shown n Fg. 2. The EA of the current cell s adjacent to those of other cells havng the same spectrum pool, referred to as extended neghbors. The use of EA helps to mprove the moblty performance sgnfcantly by mantanng the operatng frequency of moble users Spectrum and Cell Capactes The basc spectrum offers N max ðjþ channels n BA. To support the same number of channels n the larger coverage, the extended spectrum should use a hgher transmsson power than the basc spectrum. Assume that the extended spectrum j at spectrum pool supports N max ðjþ channels for the users n EA. Then t can support more channels, N max ðjþ, to the users n BA due to the shorter dstant from the BS where s greater than unty and s determned dependent on the transmsson power and the mnmum sgnal strength for decodng. The cell capacty s defned as the number of currently all avalable channels n the cell,.e., the sum of channels n all spectrum bands currently not occuped by the prmary networks n the cell. The use of the extended spectrum helps to mprove the cell capacty. However, snce the extended neghbors are located wthn the nterference range of the extended spectrum n the current cell, the extended spectrum of the current cell cannot be used n ts extended neghbors so as to avod ntercell nterference. 2.3 Handoff Types Moblty management n classcal cellular networks s closely related to user moblty. However, CR networks have another unque moblty event, the so-called spectrum moblty. By takng nto account both moblty events based on the proposed network archtecture, we defne four dfferent types of handoff schemes as shown n Fg. 2:

4 532 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 4, APRIL 2012 performs an ntercell resource allocaton through coordnaton wth ts neghbor cells. Through ths operaton, the cell can obtan the addtonal spectrum pool. Ths concept has been wdely studed n [24], [25], and [26]. To manly focus on the performance of moblty management, we assume that each cell has a sngle spectrum pool. Besdes the above spectrum sharng capablty, ntercell resource allocaton necesstates a unque functonalty to mantan the herarchcal cell structure, whch s explaned n the followng secton. Fg. 3. The proposed moblty management framework.. Intracell/ntrapool handoff. The CR user moves to a spectrum band n the same spectrum pool wthout swtchng a servng BS.. Intercell/ntrapool handoff. The CR user swtches ts servng BS to a neghbor BS wthout changng the spectrum pool.. Intercell/nterpool handoff. The CR user swtches ts servng BS to a neghbor BS, whch has a dfferent spectrum pool.. Intracell/nterpool handoff. The CR user changes ts spectrum bands from one spectrum pool to another wthn the current cell. Each handoff type s related to dfferent moblty event, and ts performance s manly dependent on both network and user condtons, such as resource avalablty, network capacty, user locaton, etc. Thus, CR networks requre a unfed moblty management scheme to explot dfferent handoff types adaptvely to the dynamc nature of underlyng spectrum bands, whch wll be explaned n Secton 3. 3 MOBILITY MANAGEMENT FRAMEWORK 3.1 Overvew Compared to the classcal cellular network, the CR network requres more complcated moblty management functonaltes due to the dynamc spectrum envronment and heterogeneous handoff types, as shown n Fg. 3. These functonaltes are ntated by three dfferent events: user moblty, spectrum moblty, and qualty degradaton. Here, user moblty s defned as the event that a moble CR user transfers an ongong connecton from one BS to another as t approaches the cell boundary. On the contrary, spectrum moblty s referred to as the event that CR users swtch ther spectrum due to the PU actvty. Each BS detects one of these events by montorng current spectrum avalablty and the qualty varaton of current transmssons and performs a proper moblty management functon accordngly. In the case of user and spectrum moblty events, CR networks decde on a proper handoff type for ther moble users by performng user and spectrum moblty management functons, respectvely. Accordng to the decson, CR users need ether to select a target cell (cell selecton) orto determne the best avalable spectrum (spectrum allocaton). If a current cell does not have enough spectrum resource due to ether PU actvty or ncrease n users, the BS 3.2 Intercell Resource Allocaton Snce each cell has tme-varyng wreless resource because of the dynamc nature of underlyng spectrum n CR networks, t cannot have a permanent extended spectrum. Furthermore, as explaned n Secton 2, the extended spectrum of the current cell cannot be used n ts extended neghbors, leadng to decrease n ther capacty. As a result, CR networks necesstate an ntercell resource allocaton scheme to select and mantan the extended spectrum. Although global optmzaton n every spectrum change acheves optmal allocaton, t requres a huge computatonal complexty and also causes a hgh communcaton overhead due to frequent spectrum swtchng. Instead, we consder the stochastc characterstcs of spatal and temporal spectrum avalabltes, and develop a dstrbuted ntercell resource allocaton method, whch mproves total network capacty as well as moblty support,.e., the avalablty of the extended spectrum. The followng are the procedures of the proposed method: 1. Intally, all avalable spectrum bands n current cell are consdered as basc spectrum bands. 2. CR users can access the spectrum band only when all PU actvty regons n the BA of the current cell are dle. Thus, the expected capacty of spectrum band j at cell s defned as follows: C ðjþ ¼N max ðjþ Y k2a B ðjþ P off ðj; kþ; where A B ðjþ s a set of the PU actvty regons of spectrum j n the BA of cell. 3. Once the extended spectrum s lost due to the PU actvty, ntercell spectrum sharng s performed to fnd a new spectrum, whch takes tme due to nformaton exchange wth ts neghbor cells. Thus, the relablty of the extended spectrum can be expressed as the rato of an average dle tme n the extended spectrum band to total tme ncludng an ntercell spectrum sharng delay as follows: R ðjþ ¼ P k2a E P 1 1 k2a E ðj;kþ ðjþ ðjþ ðj;kþ þ T nter ; ð2þ ð1þ where A E ðjþ s a set of the PU actvty regons of spectrum j n the EA of cell, 1= P k2a E ðjþ ðj; kþ represents the average dle perod of the extended spectrum j at cell, and T nter s the ntercell resource allocaton delay.

5 LEE AND AKYILDIZ: SPECTRUM-AWARE MOBILITY MANAGEMENT IN COGNITIVE RADIO CELLULAR NETWORKS 533 TABLE 1 Symbols Used for the Analytcal Modelng 4. Each cell prefers an extended spectrum wth hgher relablty. However, once the current cell determnes the extended spectrum, ts extended neghbors cannot use that spectrum, and hence lose ther capacty. To descrbe these features, we develop a novel metrc for the expected gan, whch can be expressed as the product of the spectrum relablty of the extended spectrum n the current cell and a rato of the capacty gan n the current cell to the sum of capacty loss n extended neghbors as follows: G ðjþ ¼R ðjþ Nmax ðjþ Q k2a E ðjþ P off ðj; kþ C ðjþ P 0 2N E ½C 0 ðjþn maxðjþš ; ð3þ 0 where N E s a set of the extended neghbors of cell. 5. The current cell consders the expected capacty gan over all avalable spectrum bands and chooses the extended spectrum j to satsfy the followng condton: j ¼ arg max G ðjþ; ð4þ j2s where S s a set of currently dle spectrum at cell. In the followng sectons, we ntroduce handoff models accordng to the swtchng latency and then propose spectrum and user moblty management schemes. For ease of presentaton, the mportant symbols used n the subsequent dscusson are summarzed n Table 1. 4 SPECTRUM HANDOFF MODELING Accordng to the moblty events, each handoff scheme necesstates dfferent strateges as follows:. Proactve handoff. When CR users detect handoff events, they perform handoff procedures whle mantanng communcatons. After CR users make all decsons on the handoff, they cut off communcaton channels and swtch to a new spectrum band or a new BS. User moblty and cell overload are the examples of proactve handoff events. Most of classcal handoff schemes are based on the proactve approach.. Reactve handoff. CR users should stop the transmsson mmedately n the reactve handoff event. Then, they make decsons and perform the handoff. As a result, ths handoff has an addtonal handoff delay, unlke the proactve approach. In the case that the prmary user appears n the spectrum, the CR network should ntate the reactve handoff by mmedately vacatng the spectrum to avod nterference, and then performng decson on a new avalable band. Based on these strateges, the handoff schemes defned n Secton 2 can be modeled as follows: 4.1 Intracell/Intrapool Handoff Intracell/ntrapool handoff occurs when prmary users are detected n the spectrum. Thus, t s mplemented n a reactve approach. Frst, ths handoff approach requres a preparaton tme to determne the handoff type ðd prep Þ. After that, for sensng operatons, CR users need to wat for the next sensng cycle, called a sensng synchronzaton Þ. Then, they sense the spectrum bands n the pool ðd sen Þ and determne the proper spectrum ðd dec Þ. Fnally, CR users move to a new spectrum band and resume transmsson after the synchronzaton to the transmsson schedule tme ðd sen syn on that spectrum ðd tx synþ. Snce spectrum bands n the pool are contguous, CR users can swtch the spectrum wthout reconfgurng ther RF front-ends, and hence the physcal spectrum swtchng delay s neglgble. In summary, the latency for ntracell/ntrapool handoff (Type 1) can be expressed as follows: D 1 ¼ d prep þ d sen syn þ d sen þ d dec þ d tx syn : 4.2 Intracell/Interpool Handoff If CR BSs can explot multple spectrum pools, ntracell/ nterpool handoff may happen n PU actvty. If the current spectrum pool does not have enough spectrum resource due to PU actvty, CR users detectng PU actvtes swtch to another spectrum pool n the current cell. Ths s also a reactve handoff. Thus, ts handoff latency s smlar to that of the ntracell/ntrapool handoff as follows (Type 2): D 2 ¼ d prep þ d recfg þ d sen syn þ d sen þ d dec þ d tx syn : ð6þ However, unlke the ntracell/ntrapool handoff, ths scheme requres the reconfguraton of RF front end snce each spectrum pool s placed n the dfferent frequency range. Usually, reconfguraton takes longer tme than other delay components. 4.3 Intercell/Interpool Handoff Ths handoff scheme s smlar to that n classcal cellular networks, whch s requred for CR users movng across multple cells. To determne a target cell, a moble CR user needs to observe the sgnals from neghbor cells durng ts transmsson. However, snce neghbor cells use dfferent spectrum pools, the moble CR user should stop ts transmsson and reconfgure ts RF front-end n every observaton of neghbor cells, whch s a tremendous overhead n handoff. Thus, nstead of ths moble statoncontrolled method, a network-controlled approach s more feasble for ntercell/nterpool handoff, where the BS determnes the target cell based on the stochastc user nformaton, whch s explaned n Secton 5. Consequently, moble CR users need a sngle reconfguraton tme. In ths case, the BS prepares the handoff n advance accordng to user moblty. Thus, ths s a proactve handoff and does not ð5þ

6 534 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 4, APRIL 2012 requre the handoff preparaton tme d prep used n prevous reactve handoff types as follows (Type 3): D 3 ¼ d recfg þ d sen syn þ d sen þ d dec þ d tx sych : ð7þ Furthermore, PU actvtes can ntate ths handoff scheme n specal reactve events. Frst, when all spectrum pools n the current cell are overloaded due to PU actvty, the BS forces CR users to move to neghbor cells. Ths s exactly the same procedure as the ntracell/nterpool handoff, and requres D 2 handoff latency. Second, f a PU actvty s detected n the extended spectrum, CR users n the extended spectrum do not have other avalable band n the current cell, and hence should swtch to the neghbor cells. In ths case, they lose a control channel as well. To solve ths problem, the BS determnes handoff nformaton and sends t to a selected target cell. Then, the target cell broadcasts the advertsement message for the CR user through ts control channel. In ths scenaro, CR users need one or more reconfguratons of the RF front end untl t hears the advertsement message. Also n every reconfguraton, CR users montor the control channel for a certan tme (d ls ). The latency n ths case (Type 4) can be expressed as follows: D 4 ¼ d prep þ ðd recfg þ d ls Þþd sen syn þ d sen þ d dec þ d tx syn : ð8þ Due to multple reconfguratons, ntercell/nterpool handoff n ths case shows the worst performance n terms of swtchng latency. s dependent on the searchng order of neghbor cells. In ths paper, the order s randomly chosen, and hence s consdered as ðf þ 1Þ=2 on average where f s a frequency reuse factor. 4.4 Intercell/Intrapool Handoff Ths handoff happens when moble CR users n EA successfully swtch to the extended neghbors. Ths s also a proactve handoff. Furthermore, a new target cell s an extended neghbor that uses the same spectrum pool as the current cell, and hence reconfguraton s not requred. Therefore, the latency for ntercell/ntrapool handoff scheme (Type 5) can be expressed as follows: D 5 ¼ d sen syn þ d sen þ d dec þ d tx syn : ð9þ In ths handoff, the latency s sgnfcantly reduced compared to that n other cases. Thus, ths type of handoff s more advantageous to moble CR users, and hence mproves moblty n CR networks. 5 SPECTRUM MOBILITY MANAGEMENT IN COGNITIVE RADIO NETWORKS 5.1 Overvew Spectrum moblty s the unque characterstc n CR networks. When prmary users appear n the spectrum, CR users generally change ts spectrum band wthout swtchng the BS. However, snce CR networks have tmevaryng spectrum avalablty, each cell may not have enough spectrum band to support current users. To solve ths problem, an admsson control scheme s proposed n [27]. However, n the proposed archtecture, CR users can have another opton, cell swtchng, because of the herarchcal spectrum structure descrbed n Secton 2.2. Here, we propose a spectrum moblty framework by consderng both spectrum and cell swtchng methods. When the PU actvty s detected n the cell, the BS needs to check f t has enough spectrum resource for ntracell/ ntrapool handoff. If the cell has enough spectrum resource, the BS performs the ntracell/ntrapool handoff for all users requrng new spectrum bands (Type 1). Otherwse, some of current users are forced to move to the neghbor cells. If the PU actvty s detected n the extended spectrum, all users n EA need to perform ntercell/ nterpool handoff (Type 2) regardless of current spectrum resource snce they cannot fnd other avalable spectrum bands for swtchng n that area. After the user selecton, the selected users need to fnd the proper target cell. Unlke the classcal handoff, CR users cannot observe the sgnal strength from other neghbor cells whle mantanng the connecton to current cells. Instead, CR users select the new BS based on the stochastc connectvty model, whch s explaned n Secton 5.3. The ntracell/ntrapool handoff s exactly same as the spectrum decson proposed n [27], and hence out of scope n ths paper. In the followng sectons, we wll descrbe the user and cell selectons for the ntercell/nterpool handoff scheme n more detal. 5.2 User Selecton Let S be a set of the currently avalable spectrum n the cell. Then, the number of unused channels n the avalable spectrum bands at the cell, N av, can be expressed as Pj2S N max ðjþ ðn b þ N eþ. Here, Nb and N e are the numbers of channels used n the BA and EA of cell, respectvely, and can be obtaned as follows: N b ¼ X m2u b K m ; N e ¼ X K m ; m2u e ð10þ where U b and U e are the sets of users n BA and EA, respectvely. If the number of new channels requred for spectrum swtchng N req s less than N av,.e., the cell s not overloaded, CR users just perform the ntracell ntrapool handoff. As explaned n Secton 5.1, the users n EA should move to the neghbor cells when they detect the PU actvty, and hence ther channels are not counted n N req. When detectng cell overload, the current cell forces some of ts users to be out to neghbor cells. In ths case, to mnmze the number of users to be swtched, the overloaded cell chooses users occupyng more spectrum resources untl t becomes free from cell overload as follows:. If N av <N req N av þ Nj e, CR users usng dnreq N av e channels n EA need to be selected and moved to the neghbor cells. As the users stay n EA for a longer tme, cell capacty becomes lower. Also, these users have a hgher probablty to be nterrupted by the PU actvty. Furthermore, the users havng more channels reduce the number of users that the current cell can admt. Thus, the BS selects the users n EA wth the longest expected stayng tme as well as the hghest channel occupancy. As a result, a user

7 LEE AND AKYILDIZ: SPECTRUM-AWARE MOBILITY MANAGEMENT IN COGNITIVE RADIO CELLULAR NETWORKS 535 selecton metrc can be obtaned as K m d m =v m where d m s the expected movng dstance of user m to the cell boundary, whch s dependent on the user moblty model. v m s the velocty of user m. The BS chooses users n EA wth the largest decson metrc, repeatedly untl t can avod the cell overload state.. If N req >N av þ N e, t s not enough to select all users n EA. To avod droppng or blockng connectons due to the cell overload, the BS hands over some of users n BA to ts neghbor cells. Unlke the prevous case, the BS selects CR users usng N req ðn av þ N e Þ channels wth the shortest expected stayng tme n BA snce they are hghly lkely to move to EA, whch wll requre more spectrum resource. Smlar to the prevous case, t s more advantageous to hand over the users wth more channels. Thus, the BS chooses CR users n BA wth the smallest decson metrc, d m =ðv m K m Þ. 5.3 Cell Selecton One of man challenges n CR moblty management s how to determne a proper target cell. Snce each spectrum pool s dstrbuted over a wde frequency range, CR users need to reconfgure ther RF front-ends for montorng the sgnals from neghbor cells, leadng to relatvely long temporary dsconnecton of the transmsson. In ths paper, nstead of the receved sgnal strength, we ntroduce a stochastc connectvty estmaton for selectng a proper target cell. The user connectvty to the BS s manly related to the dstance from the transmtter. Furthermore, stochastc factors such as shadowng and multpath fadng nfluence the connectvty. If the receved sgnal needs to be greater than p 0;dB for decodng data relably, the connecton probablty can be obtaned as follows [28]: P w ¼ 1 Y!! ð1 P c ðjþþ 1 Nb þ N e j2s Pj2S N max : ð12þ ðjþ Note that the target cell should be selected among canddate cells that have the cell connectvty P c supportng the mnmum QoS for a moble user. 6 USER MOBILITY MANAGEMENT IN COGNITIVE RADIO NETWORKS 6.1 Overvew The user moblty s another man reason to ntate handoff n CR networks, whch happens at the boundary of ether BA or EA.. When CR users approach the boundary of EA, they check the feasblty of ntercell/ntrapool handoff (Type 5) frst. Unlke the ntercell/nterpool handoff, CR users can measure the sgnal strength from other BS drectly, whch s exactly same as classcal handoff schemes. If CR users cannot fnd a proper target cell for ntercell/ntrapool handoff, they need to perform the ntercell/nterpool handoff to fnd a cell havng a dfferent spectrum pool. Ths procedure s same as the cell selecton scheme but does not requre a preparaton tme (Type 3), whch s explaned n Secton When CR users approach the basc cell boundary, they need to determne whether they wll stay n the EA of the current cell. For moble users, a larger cell coverage s generally known to be much more advantageous snce t reduces the number of handoffs [10]. However, n CR networks, the large cell coverage s not always desrable for moble users. As the cell coverage becomes larger, the PU actvty becomes hgher snce t s more hghly probable to nclude multple PU actvty regons. The PU actvty n EA results n a sgnfcantly long swtchng latency, as descrbed n Secton 4. In addton, snce CR users n BA are allowed to have a hgher prorty n channel access, as presented n Secton 5, cell overload also nfluences the use of extended spectrum band. As a result, CR networks need a sophstcated algorthm to select the best handoff type for moble users at the boundary of BA. Thus, n ths secton, we focus on the moblty management n the boundary of BA. When CR users become closer to the boundary, the BS ntates the handoff procedures and gather the neghbor cell nformaton from a central network entty. Based on the nformaton, the BS estmates the connectvty of the canddate cells and determnes the handoff tmng t and target cell as follows: P c ðjþ ¼Pr½p t;db L 0;dB 10 log 10 E½ 2 Š 10 log 10 D X s p 0;dB Š ¼ 1 2 ð1 erf½ð10 log ð11þ 10 D þ p 0;dB p t;db p L 0;dB 10 log 10 E½ 2 ŠÞ= ffffff 2 s ŠÞ; where p t;db s the transmsson power, L 0;dB s the average path loss at the reference dstance, D s the dstance from the BS, 10 log 10 E½ 2 Š s the average multpath fadng n db, s the path loss exponent, X s s shadowng, 2 s multpath fadng, and erf½zš s the error functon defned by R z p2ffff 0 e x2 dx. Snce the spectrum pool conssts of multple spectrum bands, the connectvty of spectrum pool, P c, can be defned as the probablty that at least one spectrum band provdes the vald connecton, whch can be expressed as 1 Q ½t ; Š¼ arg P c j2s ð1 P cðjþþ c c d 0 c þ v r c t max P c d 0 þ v r t ; ðjþ s the connecton probablty 2C;t>0 2C ð13þ of spectrum j n pool. Besdes the connectvty, where P c s a connectvty of cell, whch s a functon of the spectrum utlzaton s also an mportant factor n determnng the target cell. Thus, CR users select the target cell and the relevant velocty vr dstance d 0 to ts BS. C s a set of canddate cells, c represents the current cell, and t s the wth the hghest weghted connectvty, P w, whch can be obtaned by consderng both connectvty and spectrum utlzaton as follows: movng tme. Once a target cell s determned, the BS determnes the handoff type by consderng the expected swtchng costs of

8 536 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 4, APRIL 2012 EA. Based on these observatons, we derve the probablty that no prmary user can be detected durng r sensng perods as follows: P av ð1þ ¼ Y 0 2N E 1 P f 0 Y k2a E ðjþ e ðj;kþt ; ð14þ P av ðrþ ¼ Y 1 P f r Y e ðj;kþrt P 0 under ðrþ; 0 2N E k2a E ðjþ r ¼ 2; 3;...;R; ð15þ Fg. 4. The nfluence of prmary user actvtes n the extended area. both ntracell/ntrapool (Type 1) and ntercell/nterpool handoff schemes (Type 3) at the boundary of BA. The expected swtchng costs can be determned by estmatng the probablty of moblty events after the decson. After the decson, CR users may experence the unexpected ntercell/nterpool handoff due to the followng reasons: 1) PU actvty n EA, 2) capacty overload n EA, and 3) capacty overload n BA. In the followng sectons, frst, we analyze these future events after the decson and accordngly propose an ntellgent handoff decson scheme. 6.2 Prmary User Actvty n the Extended Area If CR users determne to perform the ntracell/ntrapool handoff (Type 1) at the boundary of BA, they can stay n the current cell, whch does not requre a long swtchng latency for ntercell/nterpool handoff. However, n EA, CR users may experence the moblty events that cause ntercell/nterpool handoff (Type 4). One of those events s the PU actvty. Snce CR users n EA cannot fnd other avalable bands when they detect the PU actvty, the ntercell/nterpool handoff s nevtable. As shown n Fg. 4, more PU actvty regons can be nvolved n determnng spectrum avalablty n EA, whch leads to a hgher PU actvty. Furthermore, the nterference range of the extended spectrum s larger than ts coverage, and hence s overlapped wth the BAs of the extended neghbors. Thus, for the accurate detecton, all extended neghbors need to be nvolved n detectng the PU actvty wth ts own detecton and false alarm probabltes. Assume that cooperatve detecton s performed accordng to the data fuson by OR rule [29], [30]. Then, a cooperatve detecton probablty converges to 1 as the number of cells ncreases. Thus, the detecton probablty can be gnored when estmatng the spectrum avalablty. On the contrary, the false alarm probablty ncreases as the number of cells ncreases, whch nfluences the spectrum avalablty sgnfcantly n EA. Even though a spectrum band s dle, t s determned to be unavalable f the false alarm s detected. Thus, n order to avod the ntercell/nterpool handoff, any PU actvtes or false alarms should not be detected n where R ¼dT m =te where T m s the expected tme for user m to stay n EA. The frst term represents the probablty that all extended neghbors do not generate any false alarms durng r sensng perods. Ths s based on the OR rule n the decson fuson, and wll change f other cooperatve decson crtera are used. Here, the sensng operaton s assumed to be performed n every t sensng perod. The second term denotes the probablty that no PU actvty appears n EA durng r sensng slots. P under ðrþ s the probablty that the cell does not experence the capacty overload durng r sensng slots, whch s derved n the followng secton. Then, the probablty of ntercell/nterpool handoff due to the PU actvty n EA can be obtaned as follows: P h pu ¼ XR r¼2 P av ðr 1Þð1 P av ð1þþ: ð16þ The frst term n the summaton represents the probablty that the extended spectrum s avalable wthout cell overload durng r 1 sensng slots, whch s multpled by the probablty that the PU actvty s detected at slot r. 6.3 Capacty Overload n the Extended Area As explaned n Secton 5, when a current cell s overloaded, CR users n EA may need to perform an ntercell/nterpool handoff (Type 2). In ths secton, we derve a probablty of cell overload. The PU actvty n the extended spectrum leads to the ntercell/nterpool handoff regardless of the cell overload, whch s already consdered n Secton 6.2. Thus, we assume that the cell overload results from PU actvtes only n basc spectrum bands, and the extended spectrum s consdered to be dle n ths case. Frst, snce each PU actvty regon n the spectrum has two states, busy and dle, we can model a transton matrx Xðj; kþ wth followng transton probabltes: x 1;1 ðj; kþ ¼e ðj;kþt ; x 1;2 ðj; kþ ¼1 e ðj;kþt ; x 2;1 ðj; kþ ¼1 e ðj;kþt ; x 2;2 ðj; kþ ¼e ðj;kþt ; ð17þ where x 1;1 ðj; kþ and x 1;2 ðj; kþ are the transton probabltes from dle to dle and from dle to busy, respectvely. x 2;1 ðj; kþ and x 2;2 ðj; kþ represent the transton probabltes from busy to dle and from busy to busy, respectvely. From ths, the transton matrx after rt can be obtaned as ½Xðj; kþš r. Let x 0 ðj; kþ 2fð1; 0Þ; ð0; 1Þg be an ntal vector

9 LEE AND AKYILDIZ: SPECTRUM-AWARE MOBILITY MANAGEMENT IN COGNITIVE RADIO CELLULAR NETWORKS 537 to descrbe a current spectrum status where ð1; 0Þ and ð0; 1Þ denote that PU actvty regon k at spectrum j s currently dle and busy, respectvely. Then, the dle probablty of regon k after rt, P dle ðj; k; rþ s the frst element of the vector, x 0 ½Xðj; kþš r, whch can be obtaned by (18) [31]. ðj; k; rþ ¼ 8 P dle x 2;1 ðj; kþ x 1;2 ðj; kþþx 2;1 ðj; kþ þð1 x 1;2ðj; kþ x 2;1 ðj; kþþ r x 1;2 ðj; kþ >< x 1;2 ðj; kþþx 2;1 ðj; kþ ; x 2;1 ðj; kþ x 1;2 ðj; kþþx 2;1 ðj; kþ ð1 x 1;2ðj; kþ x 2;1 ðj; kþþ r x 2;1 ðj; kþ >: x 1;2 ðj; kþþx 2;1 ðj; kþ ; x 0 ¼ð1; 0Þ x 0 ¼ð0; 1Þ ð18þ Based on dle probabltes at each PU actvty regon, we derve the dle and busy probabltes of spectrum j. Assume that a current cell has multple PU actvty regons n spectrum j. Then, t can use that spectrum only when all of these regons are dle, and hence the dle and busy probabltes are expressed as follows: P dle ðj; rþ ¼ Y P dle ðj; k; rþ; k2a B ðjþ ð19þ P busy ðj; rþ ¼1 P dle ðj; rþ; where A B ðjþ s the set of PU actvty regons of BA n spectrum j at cell. Based on both probabltes of each spectrum n the pool, we derve the overload probablty of cell as follows: assume that S ~ s a set of all spectrum bands assgned to the current cell, whch are used by ether prmary or CR users. Snce the extended spectrum s not consdered as explaned earler, the cell has 2 j S ~ j 1 states accordng to spectrum avalablty. Among them, some states cannot satsfy capacty requrements to support current users n the cell, resultng n cell overload. These states can be obtaned as follows: ( " # ) X L E ¼ arg ðjþ <N b þ N e ; for 8n ; ð20þ n N max j2i n where I n s a set of dle spectrum bands at state n (n ¼ 1;...; 2 j S ~ j 1 ). To resolve cell overload at each state n L E, the cell needs to obtan addtonal channels by swtchng some CR users to neghbor cells. The followng s the probablty that a moble user c n EA s swtched to the neghbor cell due to cell overload at state n 2L E : u h E ðr; nþ ¼mn N b þ N e Bmax ðnþ;n e N e K ð21þ cðd c =v c r tþ Pm2U K ; md e m =v m where B max ðnþ s the number of avalable channels n cell at state n, whch can be obtaned as P j2i n N max ðjþ. N b þ ðnþ represents the number of channels requrng N e Bmax the ntercell/nterpool handoff to prevent cell overload. K c and d c =v c are the QoS requrement of the current moble user, and ts expected stayng tme n EA, respectvely. The frst term n (27) represents the rato of the overloaded capacty n EA. Ths rato s multpled by a weghted coeffcent to consder the spectrum moblty functon where the cell selects users havng a hgher selecton metrc K m d m =v m, as presented n Secton 5.2. As shown n the second term of (27), the weghted coeffcent s defned as the rato of user c s selecton metrc to the sum of those of all users n EA, and decreases as the user approaches the boundary of EA. Let I n be a set of dle spectrum bands at state n 2L E. Then, the probablty of cell overload at rth sensng slot can be obtaned as follows: P over ðrþ¼ X " Y P dle ðj; rþ Y # P busy ðj; rþu h Eðr; nþ : ð22þ n2l E j2i n j62i n Accordngly, the probablty that the cell s not overloaded P under ðrþ s derved as 1 P over ðrþ. To avod ntercell/nterpool handoff n EA, the cell should not experence any cell overload as well as any PU actvty durng ther stayng tme. Thus, ts probablty durng rt can be expressed as follows: P h no ðrþ ¼ P av ðrþ P h over ðrþ ;r¼1; 2;...;R: ð23þ Here, P hover ðrþ s the probablty of ntercell/nterpool handoff n EA caused by cell overload durng rt, whch s expressed as follows: P hover P hover ðrþ ¼P av ð1þ ¼P av ð1þp over ð1þ; ð24þ ð1þp hno over ðr 1ÞP ðrþ ð25þ þ P hover ðr 1Þ; r¼ 2; 3;...;R: The frst term n (25) represents the probablty that ntercell/nterpool handoff due to cell overload occurs n sensng slot r, whch s obtaned by multplyng the probablty that no PU actvty s detected n slot r, the probablty that no ntercell/nterpool handoff occurs durng r 1 slots, and the cell overload probablty n slot r. 6.4 Capacty Overload n the Basc Area If the BS determnes to perform the ntercell/nterpool handoff (Type 3) at the boundary of BA, moble CR users may experence the capacty overload n the BA of target cell, whch causes ntercell/nterpool handoff. Ths cell overload probablty can be determned wth the smlar procedures used n dervng PE over n Secton 6.3. Frst, the spectrum states detectng cell overload n BA can be derved as follows: ( " # ) L B ¼ arg n X N max j2i n ðjþ <N b ; for 8n : ð26þ Based on overload states n 2L B, we derve a probablty of ntercell/nterpool handoff n BA to resolve cell overload as follows:

10 538 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 4, APRIL 2012 B ðr; nþ ¼mn N b N max ðnþ;n b N b! K c 1 d c=v c r t P m2u K : md e m =v m u ho ð27þ TABLE 2 Handoff Delay Components Used n Smulatons The frst term s the rato of the number of channels n BA requred to resolve cell overload to total channels of current users n BA. In the case of cell overload n BA, users havng a shorter expected stayng tme are selected for ntercell/ nterpool handoff, as explaned n Secton 5.2. Thus, the weghed coeffcent n the second term of (27) s expressed as one mnus the rato of user c s selecton metrc to the sum of those of all users n BA, whch s dfferent from (27). The probablty of cell overload n BA, PB; over ðrþ, can be obtaned by replacng L E and u h E ðnþ n (22) wth L B and u h BðnÞ, respectvely. Accordngly, the probablty that the CR users n BA perform ntercell/nterpool handoff durng r slots, P h over B ðrþ, s estmated as follows: P hover B; P h over B; ð1þ ¼PB; over ð1þ; ð28þ ðrþ ¼ 1 P hover B; ðr 1Þ PB; over ðrþ ð29þ þ P h over B; ðr 1Þ; r¼ 2; 3;...;R: Unlke (25), we consder all spectrum bands ncludng the extended spectrum n ths case. Thus, we do not need to consder the probablty of the spectrum avalablty n EA, P av ðrþ. 6.5 Swtchng Cost Accordng to the probablty on future moblty events, we estmate the swtchng cost of two possble optons n the boundary of BA. Frst, when CR users stay n the current cell by performng ntracell/ntrapool handoff to EA, the expected swtchng cost T EA can be obtaned as follows: T EA ¼ D 1 þ P h over þ 1 P hover ðrþd 2 þ P hpu D 4 ðrþ P h pu D5 : ð30þ The total delay ncludes the ntracell/ntrapool handoff when the CR user swtches to the extended spectrum, ntercell/nterpool handoffs due to the cell overload and and PU actvty, and ntercell/ntrapool handoff when t s successfully handed over to the extended neghbors. Second, when CR users move to the neghbor cell by performng ntercell/nterpool handoff, the expected swtchng cost can be expressed as the sum of the nstant swtchng delay and the expected swtchng delay due to the overload n that neghbor cell as follows: T m T BA ¼ D 3 þ D 1 þ P h over B; ðrþd 2 : ð31þ T off; þ D 1 The latency n ths case ncludes the ntercell/nterpool handoff to a new target cell, ntracell/ntrapool handoff n the target cell, and ntercell/nterpool handoff due to cell overload. Here, the average number of ntracell/ntrapool handoff s obtaned as T m =ðt off; þ D 1 Þ. T off; s the average dle perod of the spectrum n cell, whch s expressed as the average of 1= P k2a B ðjþ ðj; kþ over all spectrum j. Based on the analyss above, the BS determnes the handoff type wth the lower expected spectrum cost. 7 PERFORMANCE EVALUATION 7.1 Smulaton Setup In order to evaluate the performance of the proposed moblty management framework, we mplement a network smulator to support the network topology consstng of multple cells n 10 km 10 km area. Here, we assume 59 cells that have dfferent channel utlzaton. The transmsson range of each cell s set to 750 m. The nterference range s set to twce larger than the transmsson range. The transmsson range of the extended spectrum s also twce larger than that of basc spectrum. Furthermore, we consder four spectrum pools, each of whch conssts of 10 spectrum bands. The basc and extended spectrum bands can support 10 and 40 channels for users n BA, respectvely (.e., s set to 4). Furthermore, each band has three to fve PU actvty regons, whch have dfferent PU actvtes, ðj; kþ and ðj; kþ unformly dstrbuted n [0.01, 0.05]. The sensng nterval t s 0.1 sec. The BSs are assumed to generate a false alarm every 2 hours on average when they sense the avalablty of each spectrum. Furthermore, based on the delay components n Table 2, the handoff delays defned n Secton 4, D 1, D 2, D 3, D 4, and D 5 are set to 0.2, 0.5, 0.4, 1.2, and 0.1 sec, respectvely. An operatonal delay for ntercell resource allocaton, T nter,s assumed to be 5 sec. In a power attenuaton model, a channel gan s set to 31:54 db, a reference dstance s 1 meter, and a path loss coeffcent s 3.5. Shadow fadng s and multpath fadng of spectrum bands n each cell are randomly dstrbuted n [3, 7] and [2, 4] db, respectvely. The BS uses 56:21 dbm=hz transmsson power on average for the basc spectrum and 47:18 dbm=hz for the extended spectrum. Nose power n the recever s 174 dbm=hz. The mnmum decodable SNR s set to 0 db. To descrbe user moblty, we consder a Gauss-Markov moblty model proposed n [32], where the memory level parameter s set to 0.5, and asymptotc standard devatons are 1.2, 6, 12, and 20 m/sec for average veloctes 6, 30, 60, and 100 km/h, respectvely. Note that any moblty model can be appled to the proposed method, but accordng to ts accuracy, the performance of the proposed method changes sgnfcantly. 7.2 Performance of Intercell Resource Allocaton for Extended Spectrum Bands In ths smulaton, we evaluate the performance of the proposed ntercell resource allocaton by comparng wth the followng methods:

11 LEE AND AKYILDIZ: SPECTRUM-AWARE MOBILITY MANAGEMENT IN COGNITIVE RADIO CELLULAR NETWORKS 539 Fg. 5. Average channel avalablty: (a) best case and (b) worst case.. Classcal handoff scheme. Ths scheme supports only the basc spectrum bands, and does not have an extended spectrum. Thus, each cell s able to access all avalable spectrum bands n the pool wthout nfluence on ts extended neghbors.. Hghest capacty preferred scheme. The BS selects the extended spectrum so as to maxmze the total number of avalable channels n the network. A decson prncple of ths scheme s smlar to (3), but does not consder a relablty metrc R ðjþ.. Hghest avalablty preferred scheme. The spectrum wth the hghest dle probablty s selected for the extended spectrum,.e., G ðjþ s set to Q off k2a E ðjþ P ðj;kþ n (4).. Fxed allocaton. Each cell s assgned to the predetermned extended spectrum bands based on the proposed method n (3), but does not change them regardless of tme-varyng spectrum avalablty. In Fg. 5, we nvestgate total spectrum avalablty,.e., the total network capacty of each scheme. If the cell has the extended spectrum, total network capacty s dependent on the locaton of users. Fgs. 5a and 5b show the best case (.e., all users usng the extended spectrum are located on BA), and the worst case (.e., all users n the extended spectrum are located on EA), respectvely. In the best case, both proposed and fxed methods show slghtly hgher capacty than the classcal approach snce the extended spectrum supports more channels for users n BA although t restrcts the use of that spectrum n ts extended neghbors. On the contrary, n the worst case, the classcal method has much more avalable channels because the use of the extended spectrum n both proposed and fxed methods reduces the channel utlzaton n extended neghbor cells whle users n EA requre more channel resource for the same qualty of servce as users n BA. In ths case, snce the proposed method has a hgher utlzaton of the extended spectrum, t shows the lowest number of avalable channels. In Fg. 6, we compare the proposed method wth other decson prncples n terms of total capacty (best case) and the avalablty of the extended spectrum. The avalablty of the extended spectrum s defned as the rato of the tme that the extended spectrum s vald for the cell to total smulaton tme. The hghest capacty preferred method shows the hghest total channel avalablty by reducng the effect on the rest of networks, but has trouble wth fndng more relable extended spectrum. The hghest avalablty preferred scheme shows lower capacty snce t causes an adverse nfluence on neghbor cells. In addton, snce t only focuses on the overall dle probablty of the spectrum wthout the consderaton of ntercell operatonal delay, t may choose the spectrum requrng more frequent swtchng, leadng to lower relablty n the extended spectrum than the proposed method. On the contrary, the proposed method shows the hghest avalablty of the extended spectrum whle mantanng hgher capacty compared to the hghest avalablty preferred and fxed methods by jontly consderng a capacty gan and relablty n the extended spectrum. In summary, the use of the extended spectrum leads to lower network capacty compared wth the classcal methods but hgher avalablty n the extended spectrum. However, t mproves moblty performance n CR cellular network, and hence allows the proposed method to acheve hgher actual total capacty by reducng the adverse effects of dynamc network envronments, whch s shown n the subsequent smulatons. 7.3 Performance of Spectrum and User Moblty Management Schemes In ths smulaton, we nvestgate transmsson statstcs n moble users under dfferent network envronments to evaluate the performance of both spectrum and moblty management schemes. To ths end, we perform 20 1-hour smulatons for each case and obtan average values. Here, we analyze the performance of moblty management n terms of three factors: user QoS requrement (.e., the number of channels requred for a current communcaton), current network load (.e., the number of channels currently occuped by other users), and the velocty of Fg. 6. Intercell resource allocaton: (a) total avalable channels and (b) avalablty n extended spectrum bands.

12 540 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 4, APRIL 2012 Fg. 7. Handoff types: (a) user capacty, (b) lower cell occupancy, (c) hgher cell occupancy, and (d) user velocty. moble users. In the smulaton, we compare the proposed method wth the classcal handoff scheme and two other methods as follows:. BA-preferred. Ths scheme has the extended spectrum, the use of whch ams at ncrease n the number of channels n BA. At the cell boundary, a moble user swtches to ts neghbor cell f a vald target cell exst. Otherwse, t moves to EA.. EA-preferred. Ths scheme manly focuses on enhancng moblty. Thus, a moble user at the cell boundary tres to stay n the EA of the current cell f the extended spectrum s avalable. Fg. 7 shows the number of dfferent moblty events n the proposed method. As the user QoS requrement ncreases, the number of each handoff type decreases snce t reduces the probablty to fnd enough resource n BA and EA (Fg. 7a). Fgs. 7b and 7c show the changes n handoff types accordng to network load. If the network s underloaded, the cell capacty does not nfluence the number of each handoff type sgnfcantly. On the contrary, n a hghly loaded network, as the cell capacty ncreases, the number of all types of handoff decreases and conversely a drop rate ncreases because ncrease n the cell overload probablty reduces transmsson opportunty, whch s explaned n Fg. 7c. Furthermore, the underloaded case can mantan a hgher avalablty of the extended spectrum due to a lower cell overload probablty, and hence shows more Type 5 handoffs than the hghly loaded case. If user velocty ncreases, Type 5 handoff to EA ncreases to reduce the abrupt qualty degradaton due to the frequent ntercell/nterpool handoff. In all cases, the proposed method sustans the lowest number of the worst handoff (Type 4) by ntellgently choosng proper handoff types based on the expected swtchng delay. One of the most mportant statstcs n moblty management s a call drop probablty. The call drop occurs when a moble user cannot fnd any avalable spectrum n both servng and target cells. Here, we do not consder a call blockng probablty. Fg. 8 shows smulaton results on a drop rate under dfferent stuatons. To show the relablty of smulaton results, we ndcate 95 percent confdence nterval on the graph of the proposed method. In Fg. 8, the proposed method shows better performance n the drop rate than classcal and other handoff methods. As shown n Fg. 8a, although the user QoS requrement ncreases, the proposed method mantans a certan level of drop rate through spectrum moblty management. If the network load ncreases, a drop rate becomes hgher due to the lack of avalable spectrum resource, but s stll lower than other methods by selectng the handoff type adaptvely to cell condtons. Furthermore, the proposed method allows moble users to adaptvely swtch to ether BA or EA whle reducng the number of ntercell/nterpool handoff. As a result, the proposed method sustans a lower drop rate although a moble user traverses across more cell boundares wth a hgher velocty, as shown n Fg. 8c. Fg. 9 shows the lnk effcency, whch s defned as a real transmsson tme over an entre smulaton tme. In ths smulaton, the classcal method shows a lower lnk effcency over all cases due to qualty degradaton caused by frequent ntercell/nterpool handoffs. Furthermore, Fg. 8. Drop rate: (a) user capacty, (b) cell occupancy, and (c) user velocty.

13 LEE AND AKYILDIZ: SPECTRUM-AWARE MOBILITY MANAGEMENT IN COGNITIVE RADIO CELLULAR NETWORKS 541 Fg. 9. Lnk effcency: (a) user capacty, (b) cell occupancy, and (c) user velocty. when the current cell s overloaded, some moble users cannot use spectrum resources untl spectrum avalablty changes or they move nto a new target cell, whch also reduces the lnk effcency. Both EA and BA preferred methods also show lower lnk effcences snce they do not consder the expected swtchng latency caused by future moblty events. On the contrary, the proposed method shows a hgher lnk effcency by ntellgently determnng the handoff type to reduce the latency as well as the drop rate. From these smulatons, we can see that the proposed method acheves more actual transmsson opportunty as well as less qualty degradaton durng the transmsson regardless of user and network condtons, although t shows lower network capacty theoretcally due to the use of extended spectrum bands. 8 CONCLUSIONS In ths paper, we present a spectrum-aware moblty management scheme for CR cellular networks. Avalable spectrum bands n CR cellular networks vary over tme and space and are dstrbuted dscontnuously over a wde frequency range. Frst, we propose the spectrum poolbased network archtecture, whch mtgates the heterogeneous spectrum avalablty. Based on ths archtecture, a unfed moblty management framework s defned so as to support dverse moblty events n CR networks, consstng of ntercell resource allocaton, and spectrum and user moblty management functons. Through ntercell resource allocaton, each cell determnes ts spectrum confguraton to mprove moblty as well as total capacty. For the PU actvty, spectrum moblty management s developed where the network determnes a proper spectrum band and target cell accordng to both current spectrum utlzaton and stochastc connectvty model. In user moblty management, the swtchng cost-based handoff decson mechansm s proposed so as to mnmze qualty degradaton caused by user moblty. Smulaton results show that the proposed method provdes maxmum cell capacty whle provdng mnmum qualty degradaton n moble users. ACKNOWLEDGMENTS Ths materal s based upon work supported by the US Natonal Scence Foundaton under Grant No. CNS REFERENCES [1] Federal Communcatons Comsson, Spectrum Polcy Task Force Report, ET Docket No , Nov [2] I.F. Akyldz, W.Y. Lee, M.C. Vuran, and S. Mohanty, Next Generaton/Dynamc Spectrum Access/Cogntve Rado Wreless Networks: A Survey, Computer Networks, vol. 50, pp , Sept [3] IEEE P802.22/D , IEEE WG, Draft Standard for Wreless Regonal Area Networks Part 22: Cogntve Wreless RAN Medum Access Control (MAC) and Physcal Layer (PHY) Specfcatons, IEEE, Sept [4] M.M. Buddhkot, Cogntve Rado, DSA and Self-X: Towards Next Transformaton n Cellular Networks, Proc. IEEE Symp. New Fronters n Dynamc Spectrum (DySPAN 10), Apr [5] J. Sachs, I. Marc, and A. Goldsmth, Cogntve Cellular Systems wthn the TV Spectrum, Proc. IEEE Symp. New Fronters n Dynamc Spectrum (DySPAN 10), Apr [6] L. Yang, L. Cao, and H. Zheng, Proactve Channel Access n Dynamc Spectrum Network, Proc. Int l ICST Conf. Cogntve Rado Orented Wreless Networks (CROWNCOM 07), July [7] X. Zhu, L. Shen, and T.P. Yum, Analyss of Cogntve Rado Spectrum Access wth Optmal Channel Reservaton, IEEE Comm. Letters, vol. 11, no. 4, pp , Apr [8] H. Km and K.G. Shn, Fast Dscovery of Spectrum Opportuntes n Cogntve Rado Networks, Proc. IEEE Symp. New Fronters n Dynamc Spectrum Access Networks (DySPAN 08), Oct [9] I.F. Akyldz, J. McNar, J.S.M. Ho, H. Uzunaloglu, and W. Wang, Moblty Management n Next-Generaton Wreless Systems, Proc. IEEE, vol. 87, no. 8, pp , Aug [10] P.A. Ramsdale and W.B. Harrold, Technques for Cellular Networks Incorporatng Mcrocells, Proc. IEEE Int l Symp. Personal, Indoor and Moble Rado Comm. (PIMRC 92), pp , Oct [11] S.V. Hanly, An Algorthm for Combned Cell-Ste Selecton and Power Control to Maxmze Cellular Spread Spectrum Capacty, IEEE J. Selected Areas n Comm., vol. 13, no. 7, pp , Sept [12] A. Sang, X. Wang, M. Madhan, and R.D. Gtln, Coordnated Load Balancng, Handoff/Cell-Ste Selecton, and Schedulng n Mult-Cell Packet Data Systems, Proc. ACM MobCom, pp , Sept [13] D. Amzallag, R. Bar-Yehuda, D. Raz, and G. Scalosub, Cell Selecton n 4G Cellular Networks, Proc. IEEE INFOCOM, pp , Apr [14] Federal Communcatons Comsson, In the Matter of Unlcensed Operaton n the TV Broadcast Bands: Second Report and Order and Memorandum Opnon and Order, FCC , Nov [15] 3GPP TSG-RAN, Requrements for Further Advancements for Evolved Unversal Terrestral Rado Access (E-UTRA), 3GPP TR V9.0.0, Dec [16] M.M. Buddhkot, I. Kennedy, F. Mullany, and H. Vswanathan, Ultrabroadband Femtocells va Opportunstc Reuse of Mult- Operator and Mult-Servce Spectrum, Bell Labs Techncal J., Specal Issue on 4G Networks, vol. 13, pp , Feb [17] 3GPP TSG RAN WG4, RF Requrements for Multcarrer and Mult-RAT BS (Release 9), 3GPP TR V1.0.0, Sept

14 542 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 4, APRIL 2012 [18] R. Tandra, S.M. Mshra, and A. Saha, What Is a Spectrum Hole and What Does It Take to Recognze One? Proc. IEEE, vol. 97, no. 5, pp , May [19] K. Srram and W. Whtt, Characterzng Superposton Arrval Processes n Packet Multplexers for Voce and Data, IEEE J. Selected Areas n Comm., vol. 4, no. 6, pp , Sept [20] C. Chou, S. Shankar, H. Km, and K.G. Shn, What and How Much to Gan by Spectrum Aglty? IEEE J. Selected Areas n Comm., vol. 25, no. 3, pp , Apr [21] Q. Zhao, L. Tong, A. Swam, and Y. Chen, Decentrallzed Cogntve Mac Opportunstc Spectrum Access n Ad Hoc Networks: A Pomdp Framework, IEEE J. Selected Areas n Comm., vol. 25, no. 3, pp , Apr [22] T.A. Wess and F.K. Jondral, Spectrum Poolng: An Innovatve Strategy for the Enhancement of Spectrum Effcency, IEEE Comm. Magazne, vol. 42, no. 3, pp. 8-14, Mar [23] D. Cabrc, S. Mshra, D. Wllkomm, R. Brodersen, and A. Wolsz, A Cogntve Rado Approach for Usage of Vrtual Unlcensed Spectrum, Proc. 14th IST Moble and Wreless Comm. Summt, June [24] L. Cao and H. Zheng, Dstrbuted Spectrum Allocaton va Local Barganng, Proc. IEEE Sensor and Ad Hoc Comm. and Networks (SECON 05), pp , Sept [25] W.Y. Lee and I.F. Akyldz, Jont Spectrum and Power Allocaton for Inter-Cell Spectrum Sharng n Cogntve Rado Networks, Proc. IEEE Symp. New Fronters n Dynamc Spectrum Access Networks (DySPAN 08), Oct [26] C. Peng, H. Zheng, and B.Y. Zhao, Utlzaton and Farness n Spectrum Assgnment for Opportunstc Spectrum Access, ACM Moble Networks and Applcatons, vol. 11, pp , [27] W.Y. Lee and I.F. Akyldz, A Spectrum Decson Framework for Cogntve Rado Networks, IEEE Trans. Moble Computng, vol. 10, no. 2, pp , Feb [28] W. Wang and M. Zhao, Jont Effects of Rado Channels and Node Moblty on Lnk Dynamcs n Wreless Networks, Proc. IEEE INFOCOM, pp , Apr [29] Y.C. Lang, Y. Zeng, E. Peh, and A.T. Hoang, Sensng- Throughput Tradeoff for Cogntve Rado Networks, IEEE Trans. Wreless Comm., vol. 7, no. 4, pp , Apr [30] S.M. Mshra, A. Saha, and R.W. Brodersen, Cooperatve Sensng among Cogntve Rados, Proc. IEEE Int l Conf. Comm. (ICC 06), vol. 4, pp , June [31] D. Kannan, An Introducton to Stochastc Processes. North Holland, [32] D. Lang and Z.J. Haas, Predctve Dstance-Based Moblty Management for PCS Networks, Proc. IEEE INFOCOM, vol. 3, pp , Mar Won-Yeol Lee receved the BS and MS degrees from the Department of Electronc Engneerng, Yonse Unversty, Seoul, Korea, n 1997 and 1999, respectvely. He receved the PhD degree n electrcal and computer engneerng from the Georga Insttute of Technology, Atlanta, n 2009 under the gudance of Prof. Ian F. Akyldz. From 1999 to 2004, he was the senor research engneer of the Network R&D Center and Wreless Multmeda Servce Development Dvson at LG Telecom, Seoul, Korea. Currently, he s the deputy drector of the Next Moble Technology Group, Network R&D Laboratory, Korea Telecom (KT), Seoul, Korea. Hs current research nterests nclude cogntve rado networks, next generaton wreless systems, and wreless sensor networks. He receved the 2008 Researcher of the Year Award from the Broadband Wreless Networkng Laboratory, School of Electrcal and Computer Engneerng, Georga Insttute of Technology. He s a student member of the IEEE. Ian F. Akyldz s the Ken Byers dstngushed char professor wth the School of Electrcal and Computer Engneerng, Georga Insttute of Technology. Snce June 2008, he has been an honorary professor wth the School of Electrcal Engneerng at the Unverstat Poltecnco de Catalunya, Barcelona, Span. Also snce March 2009, he has been an honorary professor wth the Department of Electrcal, Electronc and Computer Engneerng at the Unversty of Pretora, South Afrca. He s the edtor-n-chef of the Computer Networks (COMNET) journal as well as the foundng edtor-n-chef of the Ad Hoc Networks journal and the Physcal Communcaton journal, all wth Elsever. Hs current research nterests are n cogntve rado networks, wreless sensor networks, and nanocommuncaton networks. He has receved numerous awards, ncludng the 1997 IEEE Leonard G. Abraham Prze Award (IEEE Communcatons Socety) for hs paper enttled Multmeda Group Synchronzaton Protocols for Integrated Servces Archtectures publshed n the IEEE JSAC n January 1996; the 2002 IEEE Harry M. Goode Memoral Award (IEEE Computer Socety) wth the ctaton for sgnfcant and poneerng contrbutons to advanced archtectures and protocols for wreless and satellte networkng ; the 2003 IEEE Best Tutoral Award (IEEE Communcaton Socety) for hs paper enttled A Survey on Sensor Networks, publshed n IEEE Communcatons magazne n August 2002; the 2003 ACM Sgmoble Outstandng Contrbuton Award wth the ctaton for poneerng contrbutons n the area of moblty and resource management for wreless communcaton networks ; the 2004 Georga Tech Faculty Research Author Award for hs outstandng record of publcatons of papers between 1999 and 2003; the 2005 Dstngushed Faculty Achevement Award from School of Electrcal and Computer Engneerng, Georga Tech; the 2009 Georga Tech Outstandng Doctoral Thess Advsor Award for hs 20+ years of servce and dedcaton to Georga Tech and producng outstandng PhD students ; and the 2009 Electrcal and Computer Engneerng Dstngushed Mentor Award from the School of Electrcal and Computer Engneerng, Georga Tech. He has been a fellow of the ACM snce He s a fellow of the IEEE.. For more nformaton on ths or any other computng topc, please vst our Dgtal Lbrary at