The Impact of Spectrum Sensing Frequency and Packet- Loading Scheme on Multimedia Transmission over Cognitive Radio Networks

Transcription

1 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. The Impact of Spectrum Sensng Frequency and Pacet- Loadng Scheme on Multmeda Transmsson over Cogntve ado Networs Xn-Ln Huang, Student Member, IEEE, Gang Wang, Fe Hu, Member, IEEE, and Sunl Kumar, Senor Member, IEEE Abstract ecently, multmeda transmsson over cogntve rado networs (CNs) becomes an mportant topc due to the C s capablty of usng unoccuped spectrum for data transmsson. Conventonal wor has focused on typcal qualty-of-servce (QoS) factors such as rado ln relablty, mum tolerable communcaton delay, and spectral effcency. However, there s no wor consderng the mpact of C spectrum sensng frequency and pacet-loadng scheme on multmeda QoS. Here the spectrum sensng frequency means how frequently a C user detects the free spectrum. Contnuous, frequent spectrum sensng could ncrease the MAC layer processng overhead and delay, and cause some multmeda pacets to mss the recevng deadlne, and thus decrease the multmeda qualty at the recever sde. In ths research, we wll derve the math model between the spectrum sensng frequency and the number of remanng pacets that need to be sent, as well as the relatonshp between spectrum sensng frequency and the new channel avalablty tme durng whch the CN user s allowed to use a new channel (after the current channel s re-occuped by prmary users) to contnue pacet transmsson. A smaller number of remanng pacets and a larger value of new channel avalablty tme wll help to transmt multmeda pacets wthn a delay deadlne. Based on the above relatonshp model, we select approprate spectrum sensng frequency under sngle channel case, and study the trade-offs among the number of selected channels, optmal spectrum sensng frequency and pacet-loadng scheme under mult-channel case. The optmal spectrum sensng frequency and pacet-loadng solutons for mult-channel case are obtaned by usng the combnaton of Hughes-Hartogs and Dscrete Partcle Swarm Optmzaton (DPSO) algorthms. Our experments of JPEG2 pacet-stream and H.264 vdeo pacet-stream transmsson over CN demonstrate the valdty of our spectrum sensng frequency selecton and pacet-loadng scheme 1. Index Terms Cogntve ado Networs (CN), Multmeda Transmsson, Spectrum Sensng Frequency, Pacet-Loadng, Hughes-Hartogs, Dscrete Partcle Swarm Optmzaton (DPSO). I. INTODUCTION eal-tme multmeda applcatons requre strngent qualty of servce (QoS) performance such as enough bandwdth and strct delay constrants. Lmted avalable bandwdth s consdered to be one of the major bottlenecs for hgh-qualty multmeda transmsson over wreless lns [1-3]. ecently, Cogntve ado (C) technology has emerged to ntellgently dentfy and use free spectrum as long as those spectrum bands are not occuped by prmary users. By opportunstcally usng unoccuped spectrum, C users can use more rado bandwdth whle not volatng FCC regulatons [4-6]. In a CN, a secondary user s devce has 1 Manuscrpt receved n May 21. Copyrght (c) 21 IEEE. Personal use of ths materal s permtted. However, permsson to use ths materal for any other purposes must be obtaned from the IEEE by sendng a request to pubs-permssons@eee.org. Xn-Ln Huang and Gang Wang are wth the Communcaton esearch Center, Harbn Insttute of Technology, Harbn 11, P.. Chna. (e-mal: xlhtcrc@163.com, gwang51@ht.edu.cn). Fe Hu s wth the department of Electrcal and Computer Engneerng, The Unversty of Alabama, Tuscaloosa, AL USA (e-mal: fe@eng.ua.edu). Sunl Kumar s wth the department of Electrcal and Computer Engneerng, San Dego State Unversty, San Dego, CA USA (e-mal: sumar@mal.sdsu.edu). 1 Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.

2 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. spectrum-agle rado transcever that can sense the avalable spectrum band (such an operaton s called spectrum sensng), automatcally confgure the rado frequency, and swtch to the selected frequency band [7-8]. elated wor: In the followng dscusson we wll summarze the exstng wor from two aspects that are closely related to our research: (1) Multmeda transmsson over CN: esearch on multmeda over CN has not caused much attenton so far. Optmal channels allocaton and effcent pacets schedulng are two mportant ssues at the MAC layer to ensure that multmeda data to be sent wthn ther delay bound. ecently, most of research efforts n CNs focus on spectrum sensng and management ssues. Only recently some wor has been performed to nvestgate approprate technques to enable multmeda delvery n CNs. A relable multmeda transmsson over CNs usng fountan codes s proposed n [1]. Such codes are used to delver multmeda data to free spectrum bands, and those codes are also used to compensate for the data loss caused by the prmary user s rado nterference. Some schemes have been proposed n [1] to balance the performance metrcs such as ln relablty, spectral effcency and codng overhead. A dynamc channel-selecton scheme for delay-senstve multmeda transmsson over CNs s proposed n [9]. It consders the vsble multmeda replay effects, varous data rate requrements, and delay performance of heterogeneous multmeda streams. (2) Spectrum sensng: In CNs, C users are responsble for detectng the actvty of prmary users (PUs) and avodng nterference to PUs. However, current C's F front-ends can not perform spectrum sensng and data transmsson smultaneously [1]. Moreover, multmeda transmsson s extremely senstve to wreless condton fluctuatng [11]. Thus, sensng accuracy has been consdered as the most mportant factor to determne the performance of multmeda transmsson over CNs. In survey [12], several typcal spectrum sensng methods are dscussed: energy detector-based sensng, waveform-based sensng, cyclostatonarty based sensng, rado dentfcaton based sensng, matched flter based sensng. In [1, 13], they have nvestgated the relatonshp among sensng perod, probablty of detecton and probablty of false alarm based on sensng effcency. However, there s no wor consderng how C spectrum sensng frequency (.e. how frequently a C user detects the avalable spectrum) mpacts on multmeda transmsson. Ths s one of the ey technques to enable multmeda servces n C lns [3]. Snce the C transcever can only do one tas at a tme, contnuous, frequent spectrum sensng can ncrease the MAC (medum access control) layer processng overhead and channel access delay. Moreover, t can cause some multmeda pacets to mss the transmsson deadlne and thus decrease the multmeda qualty at a recever sde. Transmttng multmeda content wth the support of strct QoS requrements (such as delay, throughput, jtter, etc.) s a challengng topc, and the followng aspects should be consdered for multmeda transmsson over CNs [3,14]: (1) There are serous resource constrants n wreless networs such as spectrum bandwdth, transmsson power and data rate. There also exst F (rado frequency) nterference, shadowng and mult-path fadng. (2) Multmeda data may be encoded wth dfferent profles and prortes. Loss of certan mportant pacets may degrade the mage/vdeo qualty at the recever sgnfcantly. Delay caused by spectrum sensng and channel swtchng may cause mage/vdeo data to mss the tme deadlne. (3) Upon arrval of prmary users, secondary users wth multmeda traffc could be affected more serously than the users wth non-multmeda traffc due to the strct QoS requrements n multmeda applcatons. Therefore, an ntegrated desgn to consder approprate spectrum sensng, resource allocaton and pacet schedulng, s requred to fnd optmal soluton for multmeda transmsson. Snce the pacets from secondary user may get lost due to the prmary user s arrval, we ntroduce two parameters, namely, new channel avalablty tme and remanng pacets, to descrbe the mpacts of 2 Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.

3 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. spectrum re-occupancy by a prmary user. The new channel avalablty tme refers to the allowable tme n the new channel for the contnuaton of pacet transmsson. A secondary user swtches to a new channel when the channel s re-occuped agan by a prmary user. Therefore, the avalable tme n the new channel could be very lmted. The remanng pacets mean the rest of pacets (n a multmeda stream) watng to be transmtted n the new channel. A longer new channel avalablty tme and a smaller value of remanng pacets can help to acheve the successful pacet delvery wthn a mum tolerable delay. Moreover, a smaller rato of remanng pacets to new channel avalablty tme means a lower requrement for the qualty of a new reestablshed ln. Those two parameters are determned by spectrum sensng frequency, the number of selected communcaton channels, and prmary user arrval model. We wll nvestgate the math relatonshp among them, and see the optmal spectrum sensng frequency and resource allocaton strateges. The rest of ths paper s organzed as follows. In Secton II, system models are descrbed. In Secton III, we dscuss the mpact of spectrum sensng frequency on multmeda transmsson va strct math models. In Secton IV, we provde the optmal soluton to the proposed math models. In Secton V, experment results are demonstrated to show the effcency of our algorthms n terms of supportng multmeda QoS n CNs. In Secton VI, we dscuss the applcaton of our proposed scheme to dfferent traffc types, and conclude the paper n Secton VII. II. SYSTEM MODEL In ths secton, we frst ntroduce the concept of spectrum pool. It s used for searchng unoccuped channels for multmeda transmssons. Later on, we wll dscuss the prmary user arrval model and multmeda transmsson structure. A. Spectral Avalablty For C users (also called secondary users n ths paper), the channel avalablty depends on the prmary users channel usage patterns. In order to transmt multmeda data n a CN, secondary users should frst fnd out the spectrum holes unoccuped by prmary users. The spectrum pool concept [1, 15] could be used to descrbe the model of secondary users total avalable spectrum holes [16]. After the avalable spectrum holes are detected, secondary users dvde them nto small channels, each of whch has a bandwdth of W (Hz). Then, secondary users select a set of avalable channels (denoted as S) to form a F ln n a specal way to ensure hgh performance and low nterrupted rato. For example, a good channel selecton strategy can mae sure that the arrval of a prmary user wll not cause the complete falure of the secondary user s rado ln. In other words, the rado ln can be establshed by mult-channel, whch may not be reoccuped by prmary users smultaneously. B. Prmary User Actvty Model Snce the secondary users wor on unlcensed channels, t should vacate the channel mmedately when prmary user appears. We consder each prmary channel as an ON-OFF model [17]. An ON/OFF state represents the case whether or not a prmary user s occupyng a channel, and the channel state alternates between state ON (actve) and state OFF (nactve). The secondary user can only utlze the OFF tme slot to transmt ther own pacets. We assume that each prmary channel changes ts state ndependently. For smplcty, we assume the prmary user s pacet arrval rate follows Posson process n ths paper. Then, prmary user nterarrval tme follows the exponental dstrbuton [1, 17]. 3 Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.

4 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. C. Multmeda Transmsson Structure Here we assume that a multmeda data stream to be transmtted conssts of a group of frames (each frame could be a pcture). A frame may ft nto one or multple networ pacets. At the start of every pcture, a secondary user s F ln s set up by selectng a set of S avalable channels from dfferent lcensed bands n the spectrum pool. Channel selecton crteron s based on F ln qualty, whch s mpacted by the prmary user s arrval tme. For nstance, f a prmary user qucly comes bac and re-occupes that channel, the F ln wll drop the current communcaton pacets. When the multmeda data s transmtted over CNs, the secondary user performs spectrum sensng at the start of each pcture, and senses spectrum agan after transmttng f pacets over channel, as shown n Fg.1. After the secondary user s ln s establshed, data s transmtted wthn certan delay bound. Once a channel fals due to prmary user s arrval, t s consdered not avalable durng the pcture transmsson. In the rest of the paper, we use f to ndcate spectrum sensng frequency level snce the tme perod of sendng f pacets can ndrectly determne the duraton of spectrum sensng. For nstance, a hgher value of f means that we have less tme spent n spectrum sensng. Pcture Duraton Spectrum Sensng pacets f Spectrum Sensng pacets f Spectrum Sensng t T sensng t t Kt t D Fg. 1: Multmeda transmsson structure over channel. D. Secondary User Actvty Model The secondary user s actvty s nfluenced by the prmary user s actvty snce the secondary user should vacate the frequency mmedately when a prmary user re-occupes the channel. We defne the secondary user s actvty model as shown n Fg.2. The secondary user s schedule conssts of spectrum sensng tme and multmeda data transmsson tme. Assume that a secondary user detects that channel s avalable and transmts pacets over t. When a prmary user re-occupes channel at tme, the secondary user swtches to an alternate channel j to contnue pacet transmsson. In Fg.2, we have mared the concepts of remanng pacets and new channel avalablty tme. They reflect the demands on channel j to satsfy QoS requrement. Note that we may not be able to fnsh the transmsson of all remanng pacets wthn the new channel avalablty tme. Ths depends on how soon a prmary user wll tae the channel bac. In next secton we wll deduce the models for those two concepts. Channel Spectrum Sensng tt sensng t Channel j pacets f Spectrum Sensng t Pcture Duraton pacets f Spectrum Sensng Xt ( X 1) t Spectrum Sensng emanng Pacets transmtted over channel j Kt New Channel Avalablty Tme td Occuped by Prmary Users Fg. 2: Secondary user actvty model. 4 Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.

5 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. Wthout loss of generalty, the prmary user arrval process n channel, ( 1, 2, 3,, S, S s the number of avalable channels n spectrum pool), s modeled as a Posson process wth the parameter (.e., the arrval rate). Such a model has also been assumed n many other references such as [1, 13]. Obvously, the prmary user nter-arrval tme wll be an exponental dstrbute wth the mean arrval tme 1/, as shown n Fg.2. Each channel has a capacty of. Here, the channel capacty s defned as total number of pacets that can be transmtted durng D. (Ths defnton s the same as the one n reference [1]). It s ponted out n [3] that the ey metrcs n spectrum sensng are the probablty of correct spectrum detecton, probablty of false alarm P, and probablty of mss detecton P. To nvestgate the mpacts of P n a clearer way, we assume PCD 1 and PMD. MD III. IMPACT OF SPECTUM SENSING FEQUENCY AND PACKET-LOADING SCHEME ON QOS In ths secton, we wll dscuss the mpact of spectrum sensng frequency and pacet-loadng scheme on multmeda transmsson performance, consderng sngle-channel and mult-channel ln, respectvely. The optmal spectrum sensng frequency s solved n theory under sngle-channel ln stuaton. For mult-channel case, optmal spectrum sensng frequency and pacet-loadng are solved by usng Hughes-Hartogs and Dscrete Partcle Swarm Optmzaton (DPSO) algorthms. A. Sngle Channel case We frst nvestgate the sngle-channel case, that s, each secondary user can only use sngle channel n each ln. In secton B, we wll extend our dscussons to mult-channel case. We frst ntroduce two parameters, namely, remanng pacets and new channel avalablty tme, to llustrate the mpact of spectrum sensng frequency on multmeda transmssons. From Fg.2, we can see that the sum of tme consumpton for f pacets transmsson and one perod of spectrum sensng can be represented as follows: D t f Tsensng (3.1) where D s the mum tolerable delay and Tsensng s spectrum sensng tme. Frst, we consder that only one channel s selected to transmt multmeda stream. The total number of pacets that are successfully transmtted can be expressed as: N X f (3.2) where the nteger (a random varable) X,1,2,, K, and the upper bound,.e., the mum value of X, s: K N/ f (here N s the total number of pacets requred to transmt on secondary user s ln). Obvously, from Fg.2, we can see that K cannot be larger than D / t, consderng equaton (3.1), we have: D K D /( f Tsensng ) (3.3) From Fg.2, we get K / N f. Consderng (3.3), we have: N f NTsensng /[ D (1 )] (3.4) Snce the prmary user nter-arrval tme s exponentally dstrbuted, we derve the dstrbuton functon of P CD X as: 5 Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.

6 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. Then, we get PX { } 1 t PX { 1} (1 P) e PX { 2} (1 P) e K PX { K} (1 P) e 2 2t Kt PX { } 1 (1 P ) e PX { 1} (1 P ) e (1 P ) e PX { 2} (1 P ) e (1 P ) e K Kt PX { K} (1 P) e t t 2 2t 2 2t 3 3t The remanng pacets, means the number of pacets that are not transmtted yet due to a prmary user s arrval. Snce the secondary user must vacate a channel when a prmary user s detected, the rest of pacets need to be transmtted on another channel after the F ln s reestablshed. Let of N represent the number of pacets to be sent on another channel. A larger value means that the C transcever needs to transmt more pacets over a re-establshed ln, and thus more delay s caused. Obvously, we get N K EN ( ) ( Nnf) PX { n} n (3.5) (3.6) N ( N X f) (3.7) K K aa K N(1 a ) f f( K 1) a 1a (3.8) K a(1 a ) N f ( f K s a contnuous varable) 1a t where a (1 P) e, and EN ( ) represents the expectaton value of remanng pacets N. d( E( N )) Let, we have df where BE E( BE CD) f CE B P T P T sensng sensng C (1 P ) D / D N / D NP e D N / E 1e Please refer to Appendx 1 for detaled nformaton. In Fg.3, we set up the total number of pacets that are requred to transmt on the secondary user s ln, as N 8. And we assume that the channel capacty s (3.9) (3.1) 1 pacets, spectrum sensng tme s Tsensng.1 seconds, and the probablty of false alarm s P.1. Fg.3 shows the relatonshp between remanng pacets and spectrum sensng frequency for three channels wth dfferent Posson parameters. Fg.3 also mars (wth ) the optmal values of spectrum sensng frequency f that brng the lowest remanng pacets N. efer bac to Fg.1 we see that a smaller f (.e., less pacets to send, or, less tme spent n data transmsson) mples more tme consumpton n spectrum sensng. On the other hand, however, a larger f means that more tme s spent n data transmsson, whch means the secondary user does not have enough tme to perform spectrum sensng (.e., detect the prmary user s arrval events). For both cases the user has a larger number of pacets ( N ) watng for retransmsson n the 6 Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.

7 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. new channel (see Fg.2). From Fg.3, we can conclude that a better channel (.e., a smaller, whch means a larger prmary user nter-arrval tme) wll produce a smaller N and thus larger optmal spectrum sensng frequency f. The Expectaton Number of emanng Pacets Transmtted Over a New Ln E( N ) =.1 =.2 =.3 Optmal Value: theory Sensng Frequency f Fg. 3: The relatonshp between EN ( ) and f. When a new ln s re-establshed (.e., the secondary user swtches to a new channel), we not only expect the number of remanng pacets N to be as small as possble, but also wsh the new channel avalablty tme (.e., the data transmsson tme avalable n new channel) to be as long as possble. Here, we defne new channel avalablty tme T as: T D ( X 1) t (3.11) K D ET ( ) [ D( n1) ( f Tsensng)] PX { n} n (3.12) K 1a D t 1a Snce the equaton (3.12) has smlar expresson as equaton (3.8) (please refer to Appendx 2), we use equaton (3.9) to calculate the Pea (optmal) Value. Fg.4 shows the relatonshp between ET ( ) and f. The optmal value s also shown accordng to equaton (3.9) and (3.12). The parameter settng s the same as before. From Fg.4, we can see that a better channel (t has smaller ) has a smaller new channel avalablty tme, snce the prmary user s nter-arrval tme s larger, whch gves the secondary user more tme to transmt multmeda data. For each channel, there also exsts an optmal spectrum sensng frequency f that mzes the new channel avalablty tme. It s natural to use the rato of EN ( ) to ET ( ) to reflect the QoS performance n a new re-establshed ln. A smaller EN ( )/ ET ( ) can better support QoS requrement. In Fg.5, the relatonshp between EN ( )/ ET ( ) and f s gven (the parameter settng s the same as Fg.3). From Fg.5, we can see that a better channel (wth a smaller ) wll produce smaller EN ( )/ ET ( ) (thus better QoS). In summary, we can see that the pea values n theory match very well wth the smulaton results (see Fgs.3, 4, 5). 7 Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.

8 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. The Expectaton of emanng Tme for Transmttng Over a New Ln E( T ) (seconds) =.1 =.2 =.3 Optmal Value: theory Sensng Frequency f Fg. 4: The relatonshp between ET ( ) and f. The ato E( N )/E( T ) (pacets per second) =.1 =.2 =.3 Optmal Value: theory Sensng Frequency f Fg. 5: The relatonshp between EN ( )/ ET ( ) and f. B. Mult-Channel Case Now we dscuss mult-channel case,.e., the secondary user can transmt data n multple channels smultaneously. In C networs, the C transcever can wor on non-contguous channels. In [18], a mult-band (MB) OFDM system s mplemented n C. The MB-OFDM based transcever can detect prmary user actvty and transmt data over dfferent channels smultaneously. For a mult-channel case, a rado ln could transmt data n dfferent channels. Ths maes a routng path have certan dversty, whch can help to acheve dstrbuted multmeda streamng va mult-paths. Therefore, mult-channel communcaton can mprove the networ throughput. As a mult-channel communcaton example, n [19] multple streamng servers are used to provde robust delvery even f one of the communcaton channels fals due to networ congeston. The mult-band dversty can also mprove secondary users communcaton relablty. For nstance, f a prmary user arrves n a partcular channel, the secondary user could mmedately vacate ths band and stll be able to use other channels to mantan stable data communcatons [1]. We assume S channels (denoted as 1, 2, 3, S ) are selected to transmt N multmeda pacets. Channel s lcensed by prmary user, wth arrval rate ( 1, 2, 3, S ) n that channel. We also assume N pacets are loaded to S channels, and the number of pacets loaded to each channel s denoted as N, N, N, N 1 2 S. For smplcty, we assume each channel has the same 8 Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.

9 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. capacty and spectrum sensng frequency f. Snce equaton (3.8) and (3.12) obtan the pea value at approxmately the same spectrum sensng frequency, we wll see for optmal spectrum sensng frequency and pacet-loadng scheme n theory based on equaton (3.8). The optmzaton functon s defned as: S K K aa K E( N ) N(1 a ) f f( K 1) a (3.13) 1 1a NTsensng In equaton (3.13), there are S 1varables, N1, N2,, NS and f, (and / f ). It s dffcult to drectly SD ND solve the optmal values of these S 1 varables when mnmzng equaton (3.13). We thus propose to use two optmzaton algorthms, Hughes-Hartogs and Dscrete Partcle Swarm Optmzaton (DPSO), to effcently fnd those values. For a fxed spectrum sensng frequency f, the Hughes-Hartogs algorthm can fnd optmal pacet-loadng results N1, N2,, NS, (detals later). Here pacet-loadng s smlar to the concept of bt-loadng n OFDM systems, whch tres to fnd out how many bts should be allocated to each channel dependng on the channel qualty. After the optmal pacet-loadng result s obtaned for a fxed spectrum sensng frequency f, we can calculate the optmzaton functon values n equaton (3.13), and use DPSO algorthm to fnd the optmal partcle n each teraton. After several teratons, the optmal pacet-loadng results sensng frequency f can be obtaned. N, N,, NS and spectrum In next secton, we wll gve the optmal soluton to the selecton of spectrum sensng frequency and pacet-loadng scheme usng Hughes-Hartogs and Dscrete Partcle Swarm Optmzaton (DPSO) algorthms. 1 2 IV. OPTIMAL SOLUTION TO MULTI-CHANNEL CASE A. Hughes-Hartogs algorthm Hughes-Hartogs algorthm s used wdely accordng to water-fllng prncple n OFDM system for optmal bt-loadng [2]. Snce the equaton (3.8) has smlar meanng as water-fllng prncple n OFDM [21], we can use Hughes-Hartogs algorthm to obtan the optmal pacet-loadng results. The water-fllng prncple used n our problem s shown n Fg.6. From Fg.6, we can K see that a larger f a(1 a ) /(1 a) wll produce a smaller remanng pacets EN ( ) and vce versa (under a fxed N ). Hughes-Hartogs algorthm s a greedy algorthm to see optmal results. Here we use t to fnd optmal pacet-loadng strategy (such as how many pacets should be loaded to each channel) for the mult-channel case, under a fxed spectrum sensng frequency f. The detaled steps are gven below: 1) Intally, assume all channels load zero pacets, that s N, 1, 2,, S 2) Calculate the ncrement of EN ( ), denoted as EN ( ), when one bloc (contanng f pacets) s loaded to the channel wth channel ID. Here E( N ) s defned as: K1 E( N ) ( N f nf ) P{ X n} ( N nf ) P{ X n} n n D If D ( f Tsensng )( K 1), load one multmeda bloc (contanng f pacets), set K K 1 for ths channel and go to step 2, otherwse go to step 4. 9 K 3) Fnd out the channel ID of mnmal E( N ), and calculate the new channel avalablty tme based on the followng relatonshp: D D f T K 1) ( sensng )( Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.

10 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. 4) Set EN ( ) for the channel wth mnmal pacets. a(1 a K ) f 1 a N The expectaton of remanng pacet EN ( ) B. Dscrete Partcle Swarm Optmzaton (DPSO) algorthm Avalable channel Spectrum band Fg. 6: The water-fllng prncple used n our problem. Partcle swarm optmzaton (PSO) s a populaton-based optmzaton method [22]. We have used t to model arcraft networ routng control [23]. PSO s nspred by brds flocng behavor n two-dmensonal space. It ntroduces the populaton of partcles for system optmzaton. Each partcle has a random ntalzaton status, and the followng teratons are used to update ts poston and velocty: v v cr( p x ) c r ( p x ) (4.1) 1 d d 11 d d 2 2 bd d x x v (4.2) 1 1 d d d where {1, 2,, N s }, d{1, 2,, D}, D s the search space s dmenson sze, and s the generaton of evoluton process. Note that we can use the optmzaton functon of (3.13) to determne each partcle s ftness value. Thus we are able to retreve the partcle s drecton, poston, and dstance for partcle update. The above two equatons can ensure that each partcle converges to an optmal status va an evolutonary algorthm. In (4.1), we have used two varables, p and p, to represent the personal best poston (PBP) and the global best poston (GBP), respectvely. In the PBP, an ndvdual partcle has the smallest cost functon (.e., error). Whle n the GBP, the lowest error s acheved among all the p s outputs. As we can see from (4.1), by tracng p and p, PSO can successfully update each partcle. d d bd In (4.1), we have also used two pseudorandom sequences, and r, both of whch follow unform dstrbuton n the range of [, 1], to acheve the stochastc characterstcs of PSO algorthm. In (4.1), the two constants, and c, called acceleraton coeffcents, are used to control how far a partcle moves n a sngle update round. In many cases we can smply set them to a value of 2. In (4.2), r1 2 d bd c1 2 x d s the partcle s current poston. It provdes the optmzaton soluton f PSO operates n contnuous space. However, the above model s for contnuous PSO update. If PSO operates n dscrete space, we cannot use (4.1) and (4.2) to drectly solve the poston from velocty value. In dscrete space, only or 1 values are allowed for the followng varables: x,, and. Therefore, n dscrete PSO we should use a bnary verson (dscrete) of PSO (DPSO) [24-25], to update each d p d p bd partcle, That s, v v cr( p x ) c r ( p x ) 1 d d 11 d d 2 2 bd d (4.3) 1 f ( Sg( v ) r) then x 1 d d where Sg( v ) [1/1 exp( v )] d dstrbuton n the range of [,1]. d 1 else x d (4.4) s partcle evoluton probablty, and r s a pseudorandom sequence that follows an unform 1 Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.

11 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. If we represent the spectrum sensng frequency f wth log2 bts, then we can see that a partcle contans log2 bnary numbers,.e. D log 2. Our proposed algorthm to see for the optmal soluton of spectrum sensng frequency selecton based on DPSO s descrbed below. 1) Set, and produce the poston x and velocty for each partcle, where x {,1}, v [ V, V ], 1 d D. d v d 2) For each partcle, fnd out the optmal pacet-loadng results accordng to Secton IV Part A, and determne the ftness value through the optmzaton functon (3.13). Then we get p [ x, x,, x ], and p [ x, x,, x ], where b s the ndex of the ftness partcle. vd 1 2 D d d b b1 b2 bd 3) Let 1, and change the status of v accordng to equaton (4.3). If v V, then v V ; If v V, then V. d 4) Generate a stochastc number r that follows unform dstrbuton n [,1], and use (4.4) to update the status of 5) For each partcle, we use the optmzaton functon (3.13) to determne the ftness value. If the new ftness value s less than, set p [ x, x,, x ], otherwse, set 1 p 1 2 D,, 1 p [ x 1, x 2 x ], else p p. b D b b p 1 p d. Smlarly, f the new ftness value s less than 6) If the number of teratons reaches a threshold, we should stop, otherwse, go bac to step (3). Intally, produce a random poston and velocty for each partcle d d x d. 1 p b, we set For a partcle (correspondng to sensng frequency f) Fnd out optmal pacet-loadng scheme usng Hughes-Hartogs algorthm Calculate optmzaton functon n equaton (3.13) Update the poston and velocty of partcle accordng to DPSO algorthm Iteraton tmes=? No Yes Fnd out the optmal partcle (correspondng to optmal sensng frequency f) and pacetloadng scheme among all partcles Fg. 7: Flow chart of fndng out optmal soluton. Fg.7 shows the chart of jont Hughes-Hartogs and DPSO algorthms to obtan the optmal soluton n mult-channel case. For each partcle (correspondng to a spectrum sensng frequency f), we use Hughes-Hartogs algorthm to fnd the optmal pacetloadng results over each avalable channel. Then, the optmzaton functon value n equaton (3.13) can be calculated. After each 11 Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.

12 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. partcle performs the above two procedures, they update the states (ncludng the poston and velocty) accordng to DPSO algorthm. When the teraton tmes reaches a threshold (.e., the algorthm converges), the optmal/suboptmal spectrum sensng frequency and pacet-loadng results are then obtaned. C. Computaton Complexty Analyss Snce our algorthm s a combnaton of Hughes-Hartogs and DPSO algorthms, the computaton complexty depends on both algorthms. From Secton IV part A, we can see that the computaton complexty of Hughes-Hartogs algorthm s a lnear functon of the number of pacets N, whch determne the teraton tmes [2]. Whle the computaton complexty of DPSO algorthm s a lnear functon of the number of partcles proposed algorthm s a lnear functon of NN. N s and the teraton tmes [24]. Hence, the total computaton complexty of our s V. EXPEIMENTAL ESULTS Frstly, we conduct our experments wth JPEG2 pacet stream. The mage name s lena.bmp, wth mage sze 512 by 512 and 8-lever pxel depth (.e., the value of each pxel ranges from to 255). Next, to llustrate the mpact of the number of selected channels, optmal spectrum sensng frequency and pacet-loadng scheme on multmeda transmsson, we use vdeo stream Bus n our experments. We mplement our proposed optmal spectrum sensng frequency selecton and pacet-loadng schemes n Matlab. A. Theoretcal Performance Study For the general case where all channels have dfferent prmary user arrval rates, we assume 1 channels are avalable n the spectrum pool wth ndependent prmary user arrval events whose rates are as follows: [.1,.2,.3,.4,.5,.6,.7,.8,.9,1] and other parameters are set up as follows: 1 pacets, D 1 second, N 3 pacets, P.1, Tsensng.1 seconds. Obvously, the qualty of channel s determned by prmary user arrval rate. A smaller arrval rate means a better channel qualty snce ths means that a prmary user does not arrve (and re-occupy the channel) so qucly. Thus a secondary user would have more tme to fnsh multmeda transmsson. Fg.8 (a)~(g) show the smulaton results of optmal spectrum sensng frequency and pacet-loadng results for dfferent number of selected channels (from S = 4 to S = 1). Note that here we used Hughes-Hartogs and DPSO algorthm ( Ns 3 partcles and 1 rounds) to obtan the optmal spectrum sensng frequency and pacet-loadng results. From Fg.8, we can see that a better channel wll be allocated more JPEG2 pacets to transmt on; and the spectrum sensng frequency f wll reduce as more channels partcpate n pacet transmsson, whch s caused by three reasons: (1) when the number of selected channels S s larger, some worse channels wll be selected to partcpate n pacet transmsson; (2) Meanwhle, the spectrum sensng s smultaneous for all selected channels; (3) And a worse channel need to be sensed frequently to detect prmary user state, thus a smaller f (please refer to Secton III part A). Hence, a larger number (S) of selected channels ntroduce more MAC overhead and transmsson delay. However, total pacets are loaded to a larger number of selected channels, and thus we have a smaller pacet amount n each channel, whch may brng n a hgher probablty of successful pacets delvery and hgher QoS performance. 12 Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.

13 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton The Number of Selected Channels: S=4 Optmal Sensng Frequency: f =24 Pacet-Loadng esults Avalable Channel (a) The Number of Selected Channels: S=5 Optmal Sensng Frequency: f =189 8 Pacet-Loadng esults Avalable Channel (b) 1 9 The Number of Selected Channels: S=6 Optmal Sensng Frequency: f =156 8 Pacet-Loadng esults Avalable Channel (c) 13 Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.

14 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. 1 9 The Number of Selected Channels: S=7 Optmal Sensng Frequency: f =156 8 Pacet-Loadng esults Avalable Channel (d) 1 9 The Number of Selected Channels: S=8 Optmal Sensng Frequency: f =156 8 Pacet-Loadng esults Avalable Channel (e) 1 9 The Number of Selected Channels: S=9 Optmal Sensng Frequency: f =132 8 Pacet-Loadng esults Avalable Channel (f) 14 Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.

15 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. 1 9 The Number of Selected Channels: S=1 Optmal Sensng Frequency: f =132 8 Pacet-Loadng esults Avalable Channel (g) Fg. 8: Trade-offs among the number of selected channels, optmal spectrum sensng frequency and pacet-loadng results. E( N ) pacets Our Proposed Pacet-Loadng Equal Pacet-Loadng Optmal Value: theory E( N ) pacets Our Proposed Pacet-Loadng Equal Pacet-Loadng Optmal Value: theory Sensng Frequency f (S=4) Sensng Frequency f (S=5) E( N ) pacets Our Proposed Pacet-Loadng Equal Pacet-Loadng Optmal Value: theory E( N ) pacets Our Proposed Pacet-Loadng Equal Pacet-Loadng Optmal Value: theory Sensng Frequency f (S=8) Sensng Frequency f (S=1) Fg. 9: EN ( ) comparson between our proposed pacet-loadng and equal pacet-loadng. Fgs.9~11 shows the performance comparson between our proposed pacet-loadng scheme and conventonal equal pacetloadng, whch smply loads the same amount of JPEG2 pacets to each channel [1]. We can see that our proposed algorthm s better than equal pacet-loadng scheme, snce our algorthm has fully utlzed the heterogeneous feature of avalable channels to dstrbute dfferent amount of pacets n each channel. We can also see that a larger number of selected channels produce a smaller requred optmal spectrum sensng frequency, whch s also reflected n Fg Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.

16 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. E( T ) seconds Our Proposed Pacet-Loadng.5 Equal Pacet-Loadng Optmal Value: theory Sensng Frequency f (S=4) E( T ) seconds Our Proposed Pacet-Loadng 1.2 Equal Pacet-Loadng Optmal Value: theory Sensng Frequency f (S=5) E( T ) seconds Our Proposed Pacet-Loadng 3 Equal Pacet-Loadng Optmal Value: theory Sensng Frequency f (S=8) E( T ) seconds Our Proposed Pacet-Loadng 4 Equal Pacet-Loadng Optmal Value: theory Sensng Frequency f (S=1) Fg. 1: ET ( ) comparson between our proposed pacet-loadng and equal pacet-loadng. From Fg.11, we can nfer that our proposed algorthm has a hgher successful pacet delvery rato and thus a better QoS performance. The optmal values n theory are also shown n each fgure, whch match very well wth the smulaton results. E( N )/E( T ) (pacets per second) Our Proposed Pacet-Loadng Equal Pacet-Loadng Optmal Value: theory Sensng Frequency f (S=4) E( N )/E( T ) (pacets per second) Our Proposed Pacet-Loadng Equal Pacet-Loadng Optmal Value: theory Sensng Frequency f (S=5) E( N )/E( T ) (pacets per second) Our Proposed Pacet-Loadng Equal Pacet-Loadng Optmal Value: theory Sensng Frequency f (S=8) E( N )/E( T ) (pacets per second) Our Proposed Pacet-Loadng Equal Pacet-Loadng Optmal Value: theory Sensng Frequency f (S=1) Fg. 11: EN ( )/ ET ( ) comparson between our proposed pacet-loadng and equal pacet-loadng. B. The Impact of Spectrum Sensng Frequency and Pacet-Loadng Scheme on Image Transmsson To better see the performance of multmeda content transmttng n CNs, we consder a standard pcture lena.bmp whch s to be sent. We use the compresson rate of 1/16. We use JPEG2 standard compresson software Jasper to produce codng stream. The codng stream s dvded nto 816 pacets. Ten channels are avalable n the channel pool. The prmary users arrval rate changes as follows: [5,6,7,8,9,1,11,12,13,14]. Frstly, we consder sngle-channel case, and 1 pacets, D 1 second, N 816 pacets, P.1, Tsensng.1 seconds. The smulaton results are shown n Fgs.12~ Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.

17 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. Fg. 12: Orgnal compressed mage (compresson rate=1/16) wth PSN=34.5dB. Fg. 13: Only one channel s selected to transmt pacets, and no spectrum sensng operates durng transmsson. 195 pacets are successfully transmtted and reconstructed (PSN=28.79dB). Fg. 14: One channel s selected to transmt pacets before prmary user arrval, and spectrum sensng (accordng to equaton (3.9)) operates durng transmsson. 526 pacets are successfully transmtted and reconstructed (PSN=32.69dB). We can see that our proposed algorthm performs better than the stuaton n whch no spectrum sensng operates durng transmsson. Here we mplement no-spectrum-sensng scenaro by smply stoppng spectrum sensng durng pacet transmsson. 17 Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.

18 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. Fg.12 s the orgnal compressed mage wth pea sgnal-to-nose rato (PSN) 34.5 db. Snce our proposed algorthm uses optmal spectrum sensng frequency, t can well detect prmary user s actvty and select a new channel for ln reestablshment when prmary user arrves. Ths can be seen from Fg.14. Our scheme performs better than no spectrum sensng case (see Fg.13) for more than 3.9 db. The smulaton results llustrate that our proposed algorthm s more sutable to cogntve rado envronment, and supports a hgher QoS applcaton for secondary users, whch also proves the correctness of our theory analyss n Secton III Part A. Next we consder the stuaton that mult-channel s used. The prmary user s arrval models are set to dfferent values as follows: [5, 6, 7,8, 9,1,11,12,13,14], and 2 pacets, D.2 seconds, N 816 pacets, P.1, Tsensng.1 seconds. The conventonal pacet-loadng method (such as [1]) just selects the frst fve channels and loads pacets equally. For comparson, we select the frst fve channels (S=5) and load pacets accordng to our algorthm. If one of them s occuped by prmary user, we wll select another channel to re-establsh a new channel. We stll use JPEG2 standard compresson software Jasper to produce codng stream. The DPSO algorthm uses Ns 3 partcles and 1 rounds. Fg. 15: Only fve channels are selected to transmt pacets and no spectrum sensng operates durng transmsson. 142 pacets are successfully transmtted and reconstructed (PSN=27.36dB). Fg. 16: Fve channels are selected to transmt pacets before prmary user arrval. Optmal spectrum sensng frequency and pacet-loadng schemes are done wth our proposed algorthm (Hughes-Hartogs and DPSO algorthms). 383 pacets are successfully transmtted and reconstructed (PSN=31.43dB). 18 Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.

19 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. The smulaton results are shown n Fg.15 and Fg.16. Our algorthm (see Fg.16) performs better than Fg.15 case for over 4.7 db, and provdes better QoS performance for secondary users multmeda transmsson, whch also proves that our theory analyss s correct (n Secton III Part B). Note that Fg.13 and Fg.15 are for the equal loadng scheme. Because the mum tolerable delay s reduced from 1 second to.2 seconds, Fg. 15 has less number of pacets sent and thus has poorer performance than Fg. 13. Smlarly, Fg. 16 has less number of pacets sent and poorer performance than Fg. 14. Our above smulaton results llustrate the mportance of selectng optmal spectrum sensng frequency and pacet-loadng scheme to mage transmsson n CNs. The mage transmsson s one of the major cases n multmeda applcatons; however, t s dfferent from vdeo transmsson. In mage transmsson (especally usng JPEG2 compresson) system, one pacets loss may mae the followng pacets unrecoverable due to error propagaton n decoder; however, n vdeo transmsson (especally H.264 compresson) system, some pacets loss wll not mpact the decodng of the followng vdeo stream. In the followng smulaton, we wll use vdeo transmsson wth our proposed algorthm and equal-pacet loadng algorthm, respectvely. C. Trade-offs Among the Number of Selected Channels, Optmal Spectrum Sensng Frequency and Pacet-loadng Scheme, and Ther Impacts on Multmeda Transmsson To further show the trade-offs among the number of selected channels S, optmal spectrum sensng frequency f and pacetloadng results [ N1, N2,, N S ] and ther mpacts on multmeda transmsson, we use vdeo pacets stream n our experments. We also consder the stuaton that mult-channel s used n vdeo transmsson. The prmary user s arrval models are set as follows: [1,2,3,4,5,6,7,8,9,1], and 1 pacets, P.1, Tsensng.1 seconds. The DPSO algorthm uses N 3 partcles and 1 rounds. The vdeo streamng applcaton s Bus (CIF format, 15 frames and transmsson delay deadlne D seconds), whch s coded by H.264/AVC JM 12.2 software [26] and produces N 3417 pacets wth three prortes.5 (they contan 15, 1221 and 1146 pacets, respectvely). The conventonal pacet-loadng method (such as [1]) just allocates the total pacets equally to each selected channel. For comparson, we select dfferent number of channels S for pacets transmsson and load pacets accordng to our proposed algorthm and conventonal equal pacet-loadng scheme, respectvely. The smulaton results are shown n Table 1. From Table 1, we can see that: (1) A better channel could be allocated more pacets; (2) When the number of selected channels S ncreases, the spectrum sensng frequency f and the number of vdeo pacets allocated to each channel are reduced. Furthermore, the conventonal equal pacet loadng algorthm has a lower pacet loss rate (PL) when the selected channels S ncreases. Snce no spectrum sensng occurs durng pacet transmsson, larger number of selected channels wll have less pacets allocated on each channel and thus we have a hgh probablty of successfully pacet delvery [1] (.e., low pacet loss rate); However, the reconstructed vdeo qualty (we use average Y-PSN value n our smulaton) s not mproved sharply when more channels are used. Ths stuaton s consdered as low spectral effcency n [1]. However, our proposed spectrum sensng and pacet-loadng scheme have a lower vdeo qualty when more channels are used. Ths s because our pacet-loadng scheme allocates dfferent number of pacets to dfferent channels based on channels condtons, and spectrum sensng operaton can detect the prmary user actvtes and swtch channels when necessary. Hence, our method s greedy and can ultmately utlze the avalable channel resource step by step, whch ndcates a hgher spectral effcency. When the number of selected channels S s ncreased to S (S =1 n our smulaton), our proposed algorthm performs worse, snce spectrum sensng s operated frequently 19 D s Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.

20 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. and no alternatve channels can be used when any channel s re-occuped by prmary user (the secondary user actvty model s based on channel swtchng when the prmary user s detected, please refer to Fg.2). Number of selected channels (S) Table 1: The mpact of the number of selected channels, optmal spectrum sensng frequency and pacet-loadng results on multmeda transmsson. Our proposed algorthm wth optmal spectrum sensng frequency and pacet-loadng schemes Optmal Sensng frequency (f) Pacet-loadng scheme [,,, ] N1 N2 N S PL (%) Average Y-PSN (db) The equal pacet-loadng algorthm, wthout spectrum sensng durng pacet transmsson Pacet-loadng scheme [,,, ] N1 N2 N S S =4 18 [9,9,9,717] [855,854,854,854] S =5 146 [876,876,73,497,438] [684,684,683,683,683] S =6 121 [847,847,65,484,363,271] [57,57,57,569,569,569] S =7 15 [84,84,525,42,315,267,21] [489,488,488,488,488,488,488] S =8 15 [84,735,525,372,315,21,21,21] [428,427,427,427,427,427,427,427] S =9 15 [84,735,525,315,315,21,21,162,15] [38,38,38,38,38,38,379,379,379] S =1 91 [819,728,455,364,273,182,182,182,141,91] [342,342,342,342,342,342,342,341,341,341] PL Average Y-PSN (db) Table 2: The mpact of transmsson delay deadlne D on multmeda transmsson over CNs wth our proposed algorthm and the equal pacet-loadng algorthm respectvely. Number of selected channels (S) D.35 seconds, 7 pacets D.4 seconds, 8 pacets D.45 seconds, 9 pacets Our proposed algorthm PL (%) Average Y-PSN (db) The equal pacetloadng algorthm Average PL Y-PSN (%) (db) Our proposed algorthm PL (%) Average Y-PSN (db) The equal pacetloadng algorthm Average PL Y-PSN (%) (db) Our proposed algorthm PL (%) Average Y-PSN (db) The equal pacetloadng algorthm Average PL Y-PSN (%) (db) S = S = S = S = S = (a) (b) (c) Fg.17 The qualty of reconstructed vdeo frame: (a) orgnal vdeo frame No.18; (b) reconstructed frame usng our proposed algorthm; (c) reconstructed frame usng the equal-pacet loadng algorthm wthout spectrum sensng durng pacet transmsson. (a) (b) (c) Fg.18 The qualty of reconstructed vdeo frame: (a) orgnal vdeo frame No.14; (b) reconstructed frame usng our proposed algorthm; (c) reconstructed frame usng the equal-pacet loadng algorthm wthout spectrum sensng durng pacet transmsson. 2 Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.

21 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. For comparson, Fg.17 and Fg.18 show the reconstructed vdeo frames of two algorthms, where S = 4 channels are selected to transmt vdeo pacets (The frst case n Table 1). We can see that our proposed method performs better than the algorthm that uses equal pacet loadng, and provde a hgher QoS to secondary users. Snce multmeda transmsson s not only senstve to avalable spectrum band, but also to transmsson delay deadlne, we study the mpact of transmsson delay deadlne on multmeda transmsson over CNs. For comparson, we fx the optmal spectrum sensng frequency and pacet-loadng scheme n Table 1, and set the transmsson delay deadlne D as {.35,.4,.45} seconds, respectvely. The smulaton results are shown n Table 2. From Table 2, we can see that: (1) Under the same number of selected channels, a larger D wll produce a hgher average Y-PSN, and our proposed algorthm provdes sharply mproved performance snce more spectrum resource can be used (when the allocated spectrum bands s small,.e., S=4 and 5); (2) We can use more spectrum bands (more channels) to overcome the strngent transmsson delay deadlne D, and the dsadvantage of smaller D s reduced when lots of channels are selected; (3) Our proposed algorthm s better than the algorthm that uses equal-pacet loadng (t does not have spectrum sensng operaton durng pacet transmsson) under dfferent D, when the number of selected channels S s small. VI. APPLICATION TO DIFFEENT TAFFIC TYPES By selectng the optmal spectrum sensng frequency and pacet-loadng scheme, our proposed algorthm explots the channel characterstcs to delver more pacets wthn a mum tolerable delay. We have focused on mage and vdeo transmssons n our theoretcal analyss and smulaton results. However, our proposed spectrum sensng frequency and pacet-loadng scheme can also be useful to other types of traffc. We dscuss three popular traffc types and ther applcatons whch can be supported n CN by our proposed algorthm. (1) Vdeo traffc: Vdeo applcatons nclude remote vdeo montorng, vdeo-on-demand, and Internet Protocol Televson (IPTV). Snce these applcatons have strct QoS requrements (.e., delay and throughput), we need to carefully select the spectrum sensng frequency and pacet-loadng scheme to support hgh qualty vdeo servce. In ths paper, we have explctly provded the system model, algorthm and expermental results for vdeo transmsson. Specfcally, we derved the optmal spectrum sensng frequency for sngle-channel case (n Secton III Part A), and optmal spectrum sensng frequency and pacetloadng scheme for mult-channel case (n Secton III Part B and Secton IV). Smulaton results show that the reconstructed vdeo qualty s better than the conventonal equal pacet loadng scheme. (2) Audo traffc: Audo applcatons nclude voce over IP (VoIP), musc-on-demand (.e., Tunes), and other sound/speechorented servces. In general, audo traffc requres lower bt rate than vdeo traffc. However, audo applcatons also have strct delay requrements le vdeo. Our proposed scheme can be drectly appled to audo traffc. (3) Data/Text traffc: Typcal data/text applcatons nclude many non-real tme servces, such as document transmsson (.e., Fle Transfer Protocol (FTP)), emal, and database update. Dfferent from vdeo and audo servces, these applcatons usually do not have strngent delay requrements. We can therefore set the mum tolerable delay to a larger value (.e., D ) n the optmzaton functon (n equaton (3.13)). Then, our proposed algorthm can derve the optmal spectrum sensng frequency and pacet-loadng scheme to mze the throughput performance for data/text traffc. 21 Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.

22 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. VII. CONCLUSIONS AND FUTUE WOK Snce the C transcever can only do one tas at a tme, contnuous sensng may ncrease the MAC overhead and delay, and cause some multmeda pacets to mss the transmsson deadlne, and thus the multmeda qualty decreases at a recever sde. Meanwhle, because the channels features are dfferent from each other, the pacet-loadng must adapt to the heterogeneous channels and mprove the qualty of servce (QoS). In ths paper, we have ntroduced two QoS parameters to measure the performance of multmeda transmsson n cogntve rados,.e., new channel avalablty tme and remanng pacets, and have derved the relatonshp between spectrum sensng frequency, pacet-loadng and the above two parameters. Then, we used ths relatonshp to select optmal spectrum sensng frequency and pacet-loadng scheme through two algorthms (Hughes-Hartogs and Dscrete Partcle Swarm Optmzaton (DPSO) algorthms), consderng a Posson-based model to descrbe prmary user traffc. Smulaton results showed that our algorthm could well utlze the heterogeneous channels and perform a better QoS under the stuatons wth sngle channel and mult-channel lns. Future nvestgatons n ths area nclude vdeo transmsson from multple secondary users smultaneously, and mult-hop multmeda transmsson n CNs. A multmeda-orented routng protocol should be desgned consderng spectrum sensng frequency and pacet-loadng scheme along the path from source to destnaton to support the QoS for secondary users. ACKNOWLEDGEMENT The authors sncerely than all anonymous revewers for ther valuable comments on ths wor. The authors also than Seethal Palur for her help n vdeo smulaton. Part of ths research wor was done when the frst author was n Dr. Fe Hu s research lab n the Unversty of Alabama. d( E( N ( ) )) BE E BE CD 1. Proof: f. df CE We consder APPENDIX K as a contnuous varable, and K N / f. Then, K K a a EN ( ) N(1 a ) f f( K 1) a 1 a K a(1 a ) N f 1 a a (1 P ) e 1 a 1 (1 P ) e D ( f Tsensng ) D ( f Tsensng ) (1 P ) [1 ( f Tsensng )] 1 (1 P ) [1 ( f Tsensng )] D D D 1 ( P T P T ) (1 P ) D P Tsensng P Tsensng (1 P ) f sensng sensng f 1BC f BC f K 22 Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.

23 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. d( E( N )) Let, then df K f (1 a ) f f {(1 P ) e } 2 2 D fk K f f (1 P ) e e N f f (1 P ) e f D N D D N N (1 e ) f NP e E f D D ( f Tsensng ) K CE f 2 BCE fcdb(1be ) BE E( BE CD) f CE 2. Proof: equaton (3.8) and equaton (3.12) have the smlar characterstcs. D Snce t f Tsensng, equaton (3.8) can be further rewrtten as: and equaton (3.12) can be rewrtten as: K a(1 a ) EN ( ) N f 1a 2 K N ( t Tsensng ) ( a a a D 2 K N t ( a a a ) D K 1 a ET ( ) D t 1 a D t aa a 2 K 1 (1 ) D t a t aa a K 2 K (1 ) ( ) D t aa a 2 K ( ) Tsensng K We assume t >> T (pacets transmsson tme t larger than spectrum sensng tme T n each perod) and D >> t sensng (total multmeda transmsson tme D larger than each tme slot t ). Hence, we can conclude that equaton (3.8) and equaton (3.12) have smlar characterstcs. ) sensng EFEENCES [1] Hareshwar Kushwaha, Ypng Xng, ajarathnam Chandramoul, and Harry Heffes. elable Multmeda Transmsson Over Cogntve ado Networs Usng Fountan Codes, Proceedng of The IEEE, vol. 96, no. 1, pp , 28. [2].-T. Sheu and J.-L.C. Wu. Performance Analyss of ate Control Wth Scalng QoS Parameters for Multmeda Transmssons, Proc. of IEE conference n Communcatons, vol. 15, no. 5, pp , 23. [3] Norazzah M. Arpn, ozeha A. ashd, Norulhusna Ahmad, Ada E.M Hamzah, Norshela Fsal, and Sharfah K. Syed Yusof. Cross Layer Desgn of Multmeda Transmsson Over Cogntve ado UWB Multband OFDM System, Int l Graduate Conference n Engneerng & Scence (IGCES), Johor Bharu, pp. 1-1, 28. [4] Andrew Mchael James. A Ln-qualty-aware Graph Model for Cogntve ado Networ outng Topology Management, Thess, ochester Insttute of Technology, pp. 1-4, Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.

24 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. [5] Hcham Khalfe, Satyajeet Ahuja, Naceur Malouch, Marwan Krunz. Probablstc Path Selecton n Opportunstc Cogntve ado Networs, Proc. of IEEE Globecom conference, pp. 1-5, 28. [6] Mao Pan, ongsheng Huang, and Yuguang Fang. Cost Desgn for Opportunstc Mult-Hop outng n Cogntve ado Networs, Proc. of IEEE MILCOM conference, pp. 1-7, 28. [7] Ian F. Ayldz, Won-Yeol Lee, Mehmet C. Vuran, and Shantdev Mohanty. Next Generaton/Dynamc Spectrum Access/Cogntve ado Wreless Networs: a Survey, Computer Networs, vol. 5, no. 13, pp , 26. [8] Guome Zhu, Ian F. Ayldz, and Gengsheng Kuo. STOD: A Spectrum-Tree Based On-Demand outng Protocol for Mult-hop Cogntve ado Networs, Proc. of IEEE Globecom conference, pp. 1-5, 28. [9] Hsen-Po Shang and Mhaela van der Schaar. Queung-Based Dynamc Channel Selecton for Heterogeneous Multmeda Applcatons Over Cogntve ado Networs, IEEE Transactons on Multmeda, vol. 1, no. 5, pp , 28. [1] Won-Yeol Lee and Ian. F. Ayldz. Optmal Spectrum Sensng Framewor for Cogntve ado Networs, IEEE Transactons on Wreless Communcatons, vol. 7, no. 1, pp , 28. [11] We Wang, Mchael Hempel, Dongmng Peng, Honggang Wang, Hamd Sharf, and Hsao-Hwa Chen. On Energy Effcent Encrypton for Vdeo Streamng n Wreless Sensor Networs, IEEE Transactons on Multmeda, vol. 12, no. 5, pp , 21. [12] Tevf Yüce and Hüseyn Arslan. A Survey of Spectrum Sensng Algorthms for Cogntve ado Applcatons, IEEE Communcatons Survey & Tutorals, vol. 11, no. 1, pp , 29. [13] Yng-Chang Lang, Yong-Hong Zeng, Edward C. Y. Peh, and Anh Tuan Hoang. Sensng-Throughput Tradeoff for Cogntve ado Networs, IEEE Transactons on Wreless Communcatons, vol. 7, no. 4, pp , 28. [14] Norazzah Mohd Arpn, Norshela Fsal, and ozeha A. ashd. Performance Evaluaton of Vdeo Transmsson Over Ultwdeband WPAN, Proc. of 29 Thrd Asa Internatonal Conference on Modelng & Smulaton, pp [15] Hareshwar Kushwaha and. Chandramoul. Secondary Spectrum Access Wth LT Codes for Delay-Constraned Applcatons, Proc. of IEEE Consum. Commun. Netwo, 27. [16] Danjela Čabrć, Shrdhar Mubaraq Mshra, Danel Wllomm, obert Brodersen, and Adam Wolsz. A Cogntve ado Approach for Usage of Vrtual Unlcensed Spectrum, Proc. of 14 th IST Moble Wreless Commun., 25. [17] Hang Su and X Zhang. Opportunstc MAC Protocols for Cogntve ado Based Wreless Networs, Proc. of CISS 41 st Annual Conference on Informaton Scences and Systems, pp , 27. [18] Tmo A. Wess and Fredrch K. Jondral. Spectrum Poolng: An Innovatve Strategy for The Enhancement of Spectrum Effcency, IEEE Communcatons Magazne, vol. 42, pp. 8-14, 24. [19] John Apostolopoulos, Tna Wong, Wa-Tan Tan, and Suse Wee. On Multple Descrpton Streamng Wth Content Delvery Networs, Proc. of IEEE INFOCOM conference, vol. 3, pp , 22. [2] Xn-Ln Huang, Gang Wang, Yong-Ku Ma, Cheng-Wen Zhang and Hao Jang, An Effcent Bt Loadng Algorthm for OFDM System, Journal of Harbn Insttute of Technology, vol. 42, no. 9, pp , 21. [21] Dr Hughes Hartogs and Morgan Hll, Ensemble Modem Structure for Imperfect Transmsson Meda, U. S. Patent Number: (May 1989). [22] James Kennedy and ussell Eberhart. Partcle Swarm Optmzaton, Proc. of IEEE Neural Networs Conference, vol. 1995, no. 4, pp , [23] Xn-Ln Huang, Gang Wang, Fe Hu, and Sunl Kumar, Stablty-Capacty-Orented outng for Hgh Moblty, Mult-Hop Cogntve ado Networs, IEEE Transactons on Vehcular Technology, submtted n August Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.

25 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. [24] Zhjn Zhao, Shyu Xu, Shlan Zheng, and Yangxao Nu. Cogntve ado Decson Engne Based on Bnary Partcle Swarm Optmzaton, Acta Physca Snca, vol. 58, no. 7, pp , 29. [25] Yan Lu and Xue-Png Gu. Seleton-Networ econfguraton Based on Topologcal Characterstcs of Scale-Free Networs and Dscrete Partcle Swarm Optmzaton, IEEE Transactons on Power Systems, vol. 22, no. 3, pp , 27. [26] Please refer to: Authors Bographes: Xn-Ln Huang s worng for Ph.D. degree n communcaton engneerng at Communcaton esearch Center, Harbn Insttute of Technology, Harbn, P.. Chna. Hs research focuses on jont source-channel codng, OFDM technology and Cogntve ado Networs. He was a recpent of Chnese Government Award for Outstandng Ph.D. Students n 21. He s a student member of IEEE and paper revewer for IEEE Communcatons Letters, Wreless Personal Communcatons, and Internatonal Journal of Communcaton Systems. Gang Wang, as professor, he s now wth the Communcaton esearch Center, Harbn Insttute of Technology, Harbn, P.. Chna. Hs general nterests nclude jont source-channel codng, ISDN, and wreless communcatons. He has publshed over 5 research papers and four boos n these felds. He was a recpent of Natonal Grade II Prze of Scence and Technology Progress and Natonal Grade III Prze of Scence and Technology Progress. Fe Hu s an assocate professor n the Department of Electrcal and Computer Engneerng at the Unversty of Alabama, Tuscaloosa, AL, U.S.A. Hs research nterests are wreless networs, wreless securty and ther applcatons n Bo-Medcne. Hs research has been supported by NSF, Csco, Sprnt, and other sources. He obtaned hs frst Ph.D. degree at Shangha Tongj Unversty, Chna n 1999, and second Ph.D. degree at Clarson Unversty, U.S.A. n 22, all n the feld of Electrcal and Computer Engneerng. He has publshed over 1 journal/conference papers and boo (chapters). He s also the edtor for over fve nternatonal journals. Sunl Kumar s an Assocate Professor and Thomas G. Pne Faculty Fellow n the Electrcal and Computer Engneerng department at San Dego State Unversty, San Dego, Calforna, U.S.A. From August 22 to July 26, he was an Assstant Professor n Electrcal and Computer Engneerng at Clarson Unversty, Potsdam, NY. He receved M.E. and Ph.D. n Electrcal and Electroncs Engneerng from Brla Insttute of Technology and Scence, Plan (Inda) n 1992 and 1997, respectvely. Hs research nterests nclude QoS-aware Cross-layer Protocols for Multmeda Traffc n Wreless Networs, and obust Multmeda Compresson technques. He has publshed more than 95 research artcles n nternatonal journals and conferences, ncludng two boo/boo chapters. Hs research has been supported by U.S. Natonal Scence Foundaton (NSF), U.S. Ar Force, DOE, Csco and Sprnt. 25 Copyrght (c) 211 IEEE. Personal use s permtted. For any other purposes, Permsson must be obtaned from the IEEE by emalng pubs-permssons@eee.org.