760 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 30, NO. 4, MAY 2012

Transcription

1 760 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 30, NO. 4, MAY 2012 Cooperatve Communcaton Aware Lnk Schedulng for Cogntve Vehcular Networks Mao Pan, Student Member, IEEE, PanL,Member, IEEE and Yuguang Fang, Fellow, IEEE Abstract Throughput maxmzaton s a key challenge for wreless applcatons n cogntve Vehcular Ad-hoc Networks (C- VANETs). As a potental soluton, cooperatve communcatons, whch may ncrease lnk capacty by explotng spatal dversty, has attracted a lot of attenton n recent years. However, f lnk schedulng s consdered, ths transmsson mode may perform worse than drect transmsson n terms of end-to-end throughput. In ths paper, we propose a cooperatve communcaton aware lnk schedulng scheme and nvestgate the throughput maxmzaton problem n C-VANETs. Regardng the features of cooperatve communcatons and the avalablty of lcensed spectrum, we extend the lnks nto cooperatve lnks/general lnks, defne extended lnk-band pars, and form a 3-dmensonal (3- D) cooperatve conflct graph to characterze the conflct relatonshp among those pars. Gven all cooperatve ndependent sets n ths graph, we mathematcally formulate an end-to-end throughput maxmzaton problem and near-optmally solve t by lnear programmng. Due to the NP-completeness of fndng all ndependent sets, we also develop a heurstc prunng algorthm for cooperatve communcaton aware lnk schedulng. Our smulaton results show that the proposed scheme s effectve n ncreasng end-to-end throughput for the sesson n C-VANETs. Index Terms Throughput Maxmzaton, Cogntve Vehcular Ad-hoc Networks, Cooperatve Communcatons, Lnk Schedulng. I. INTRODUCTION WITH the maturty of road nfrastructure and the ncreasng number of motorsts, hghway travelng has become a part of lfe for people n US and many other countres. Varous broadband vehcular communcaton applcatons n Vehcular Ad-hoc Networks (VANETs), whch can entertan passengers and make long journeys enjoyable, are envsoned to be prevalent n the near future. However, prolferaton of vehcular applcatons beyond safety requres addtonal rado resources to support, whch makes the already crowed lcensed spectrum even worse. Meanwhle, for all these passengerorented applcatons [1] [3], no matter vehcle-to-vehcle (V2V) communcaton based applcatons (e.g., network gamng among passengers n dfferent cars, fle transfers, vrtual Manuscrpt receved 15 May 2011; revsed 27 October Ths work was partally supported by the U.S. Natonal Scence Foundaton under grants CNS / and ECCS / The work of Y. Fang was also partally supported by the Natonal Natural Scence Foundaton of Chna under grant M. Pan and Y. Fang are wth the Department of Electrcal and Computer Engneerng, Unversty of Florda, Ganesvlle, FL 32611, USA. (e-mal: maopan@ufl.edu; fang@ece.ufl.edu). Y. Fang was also a Changjang Scholar Char Professor wth the Natonal Key Laboratory of Integrated Servces Networks, Xdan Unversty, X an, Chna. P. L s wth the Department of Electrcal and Computer Engneerng, Msssspp State Unversty, Msssspp State, MS USA (e-mal: l@ece.msstate.edu). Dgtal Object Identfer /JSAC /12/$25.00 c 2012 IEEE meetngs among coworkers, etc.) or vehcle-to-roadsde (V2R) communcaton based ones (e.g., web browsng, cooperatve downloadng, onlne vdeo, etc.), the most crtcal and essental requrement s the data transmsson wth hgh end-to-end throughput, whch s also a challengng task n VANETs. In vew of the rado spectrum demands from VANETs, Federal Communcatons Commsson (FCC) opens the underutlzed lcensed TV spectrum (.e., the UHF televson frequency spannng over MHz) and allows the opportunstc accessng of unlcensed users. By explotng cogntve rado (CR) technology, the vehcles/nodes (the words vehcles/nodes wll be used n ths paper nterchangeably) as well as the roadsde unt (RSU) n VANETs can sense the vacant spectrum and opportunstcally use these lcensed bands temporally/geographcally, when/where prmary servces are not actve. We call such a VANET wth CR capablty [3], [4] as a cogntve VANET (C-VANET). On the other hand, by employng multple antennas, e.g., multple-nput and multple-output (MIMO), spatal dversty has been shown to be effectve n lowerng bt error rate, enhancng power effcency and mprovng throughput n VANETs. However, equppng a wreless node wth multple antennas may not always be practcal. To acheve spatal dversty wthout requrng multple transcever antennas on the same node, the so-called cooperatve communcatons has been ntroduced n [5], [6]. The dea of cooperatve communcatons can be best llustrated by a three-node example [5], [6] shown n Fg. 1(a). In ths sub-fgure, node transmts to node j va one-hop, and node r acts as a cooperatve relay node. Cooperatve transmsson from to j s done on a frame-byframe bass. Wthn each frame, there are two tme slots [1], [5], [7] [9]. In the frst tme slot (sold lnes), makes a transmsson to destnaton j. Due to the broadcast nature of wreless transmssons, transmsson by s also overheard by relay node r. In the second tme slot (dash lnes), r forwards the data t overheard n the frst tme slot to j. Thus, under cooperatve communcatons, each node s equpped wth only a sngle antenna and reles on the antennas of neghborng cooperatve nodes to acheve spatal dversty. If the cooperatve relay node s approprately selected, cooperatve communcatons can effectvely ncrease the lnk capacty [7], [10]. However, f we take tme-frame based lnk schedulng nto consderaton, cooperatve communcatons s not necessarly helpful to mprovng the end-to-end throughput. Take the toy topology shown n Fg. 1(b) as an example. If node drectly transmts packets to node j, lnk (, j) wll have no nterference wth lnk (u, v), so that they can be scheduled to transmt smultaneously. By contrast, f (, j) employs r for

2 PAN et al.: COOPERATIVE COMMUNICATION AWARE LINK SCHEDULING FOR COGNITIVE VEHICULAR NETWORKS 761 j (a) A 3-node schematc for cooperatve communcatons. Fg. 1. r... s r j r (b) A schematc for the nterference ncurred by cooperatve communcatons. Illustratve toy topologes for cooperatve communcatons. cooperatve communcatons, (, j) wll conflct wth (u, v) snce the transmssons of cooperatve relay r cast nterference on the recevng node v of (u, v). As a result, (, j) and (u, v) cannot be scheduled to transmt smultaneously, whch may decrease the end-to-end throughput from s r to d t. In terms of throughput, the beneft brought by cooperatve communcatons may be offset, or even overwhelmed by the loss of opportuntes for schedulng more lnks to transmt at the same tme. Based on that observaton, there appear several nterestng questons for the throughput maxmzaton problem n C-VANETs: When lnk schedulng s consdered, does there exst an optmal approach to maxmze the beneft brought by cooperatve communcatons n terms of the endto-end throughput? Does the avalablty of lcensed bands have any mpact on transmsson mode selecton (.e., drect transmssons or cooperatve communcatons) as well as the throughput? Can we fnd a smple and feasble way to solve ths problem n practce? To address these ssues, n ths paper, we propose a cooperatve communcaton aware lnk schedulng scheme, wth the objectve of maxmzng the throughput for a sesson n C-VANETs. We let the RSU schedule the mult-hop data transmssons among vehcles on hghways by sendng smallsze control messages. Jontly consderng avalablty of lcensed spectrum, transmsson modes and lnk schedulng, we mathematcally formulate the throughput maxmzaton problem, near-optmally solve t by lnear programmng, and provde a smple heurstc algorthm to gve feasble results. Our salent contrbutons are summarzed as follows. Regardng the features of cooperatve communcatons, we novelly extend a lnk usng cooperatve communcatons nto a cooperatve lnk. To keep notaton consstent, we leverage a dummy cooperatve relay and extend a lnk usng drect transmssons nto a general lnk. Inspred by the lnk conflct graph n pror work [11] [15], we propose a 3-dmensonal (3-D) cooperatve conflct graph to descrbe the nterference relatonshp among the extended lnks n C-VANETs. Smlar to the methodology used n [13] [15], we nterpret each vertex n the graph as a basc resource pont for schedulng and represent each resource pont wth an extended lnkband par. Based on these extended lnk-band pars, we establsh the 3-D cooperatve conflct graph and re-defne the cooperatve ndependent sets and conflct clques. Wth the help of 3-D cooperatve conflct graph, the RSU can mathematcally formulate the throughput maxmzaton problem under multple constrants (.e., avalablty of bands, selecton of transmsson modes and v u d t lnk schedulng). Gven all cooperatve ndependent sets n C-VANETs, the RSU can relax the nteger varables n the formulaton, solve the optmzaton problem by lnear programmng and obtan the optmal end-to-end throughput between the source and destnaton nodes. Snce t s NP-complete to fnd all the cooperatve ndependent sets n C-VANETs [12] [16], we employ a number of maxmum cooperatve conflct clques and develop a heurstc prunng algorthm to approxmate the optmal end-to-end throughput. We let the RSU select the band and transmsson mode for the extended lnk-band pars n those clques, prune the pars not selected and update clque transmsson tme untl the largest clque transmsson tme among all clques cannot be further decreased. The throughput s estmated as the recprocal of the largest clque transmsson tme. By carryng out numercal smulatons, we demonstrate the mpact of the number of avalable bands and the dstance between source and destnaton nodes on the performance of throughput n C-VANETs. We also show that ) the CR capablty creates more opportuntes for usng cooperatve communcatons; ) the performance of cooperatve communcaton aware lnk schedulng s better than that purely relyng on one transmsson mode; ) the proposed prunng algorthm s close to the optmal one n terms of end-to-end throughput n C-VANETs. The rest of the paper s organzed as follows. In Secton II, we ntroduce the settngs and related models n C-VANETs. In Secton III, we descrbe the 3-D cooperatve conflct graph and present the concept of cooperatve ndependent sets and conflct clques. In Secton IV, we mathematcally formulate the throughput maxmzaton problem n C-VANETs and nearoptmally solve t by lnear programmng. In Secton V, we develop a heurstc prunng algorthm for cooperatve communcaton aware lnk schedulng. Fnally, we conduct smulatons and analyze the performance results n Secton VI, and draw concludng remarks n Secton VII. II. NETWORK MODEL A. Network Settng of C-VANETs We consder a mult-hop C-VANETs [3], [4] consstng of multple vehcles operatng on dfferent vacant lcensed frequency bands and a RSU (e.g., a base staton (BS), a gateway, an access pont (AP), etc.) who serves ths group of nodes N = {1, 2,,n,,N} on (one way) hghways. Let s r /d t denote the source/destnaton node for a sesson n C-VANETs. Our objectve s to maxmze end-to-end throughput of ths sesson. By exchangng small-sze control messages wth the vehcles, the RSU 1 can schedule the transmssons of largesze data packets for mult-hop V2V communcatons [4]. The schedulng perod s set to τ consderng the vehcles mergng nto/extng from the hghway as well as the avalablty of lcensed bands. Suppose that the set of lcensed spectrum bands B = {1, 2,,b,,B} have the dentcal bandwdth, where the sze of the bandwdth s equal to W. Both drect transmssons and cooperatve communcatons can be used for 1 The RSU can also be nterpreted as a group of assocated RSUs connected by the backbone network, f the length of the path s long.

3 762 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 30, NO. 4, MAY 2012 packets delvery. To dstngush two types of relay nodes [10] n C-VANETs, we call a relay node used for cooperatve communcatons purpose as a cooperatve relay and a relay node used for mult-hop relayng n the tradtonal sense as mult-hop relay 2. Consderng the concept of cooperatve communcatons as well as the nherent hardware lmtaton of CR devces, we also assume that each node has only one rado, but the rado can be tuned nto any avalable frequency band for packet delvery. Each node N employs certan spectrum sensng technques (e.g., [17], [18]) to dentfy a set of avalable lcensed bands, whch are not occuped by prmary servces. Dependng on the geographcal locatons of nodes, the avalable bands at one node may be dfferent from another one n C-VANETs. To put t n a mathematcal way, let B Brepresent the set of avalable lcensed bands at CR node N. B may be dfferent from B j,wherej s not equal to, andj N,.e., possbly B B j. For a lnk (, j) usng cooperatve relay r, we assume the transmsson from to j and the transmsson from r to j use the same band. Thus, we have B (,r,j) = B (,j) = B Bj. Besdes, the tme share 3 assgned by the RSU wll be measured n tme frames, and each tme frame wll be equally dvded nto two tme slots for the transmsson from to j and that from r to j, f cooperatve communcatons s employed. B. Transmsson Modes In ths subsecton, we gve expressons for achevable data rate under dfferent transmsson modes. For cooperatve communcatons, we consder both AF and DF modes [5], [7]. 1) Amplfy-and-Forward (AF): Under ths transmsson mode, cooperatve relay r receves, amplfes and forwards the sgnal from node to node j [5], [7], [10]. Let h j, h r, h rj capture the effects of path-loss, shadowng and fadng between nodes and j, and r, andr and j, respectvely. Denote z j and z r the zero-mean background nose at nodes j and r, wth varance σj 2 and σ2 r, respectvely. Besdes, denote P and P r the transmsson powers at nodes and r, respectvely. Snce the results are vald for all the bands, we omt the band notatons n ths subsecton. Followng the same notatons n [5] [7], [10], the achevable data rate under AF can be expressed as C AF (, r, j) =W I AF (, r, j), (1) ( ) where I AF (, r, j) = 1 2 log 2 1+SNR j + SNRr SNRrj SNR r+snr rj+1, SNR j h j 2,SNR r = P σr 2 h rj 2, = P σ 2 j h r 2,SNR rj = Pr σ 2 j and W s the avalable bandwdth at nodes and r. 2) Decode-and-Forward (DF): Under ths transmsson mode, relay node r decodes and estmates the receved sgnal from node n the frst tme slot, and then transmts the estmated data to node j n the second tme slot [5] [7]. As 2 Note that a cooperatve relay operates at the physcal layer whle a multhop relay operates at the network layer. 3 In ths paper, tme perod refers to the schedulng perod,.e., τ; tme share refers to the actve tme scheduled for an ndependent set,.e., λ mτ, as llustrated n Sec. IV-B; tme frame refers to the basc unt of tme for lnk schedulng; tme slot refers to the two tme slots defned n cooperatve communcatons [5], [7]. n [5] [7], [10], the achevable data rate under DF transmsson mode s gven as C DF (, r, j) =W I DF (, r, j), (2) where I DF (, r, j) = 1 2 mn{log 2 (1+SNR r), log 2 (1+SNR j + SNR rj )}. Note that I AF ( ) and I DF ( ) are ncreasng functons of P and P r, respectvely. Ths suggests that, n order to acheve the maxmum data rate under ether mode, both node and node r should transmt at the maxmum power. In ths paper, we let P = P r = P. 3) Drect Transmsson: When cooperatve communcatons s not used, the achevable data rate from node to node j s C DTx (, j) =W log 2 (1 + SNR j ). (3) Based on the above results, we have two observatons. Frst, comparng C AF (or C DF )toc DTx, t s hard to say that cooperatve communcaton s always better than the drect transmsson. In fact, a poor choce of relay node could make the achevable data rate under cooperatve communcatons be lower than that under drect transmssons [7]. Second, although AF and DF are dfferent mechansms, the capactes for both of them have the same form,.e., a functon of SNR j, SNR r,andsnr rj. Therefore, a cooperatve communcaton aware lnk schedulng algorthm desgned for AF can be readly extended for DF. Therefore, t s suffcent to focus on one of them, where we choose AF n ths paper. C. Transmsson/Interference Range The nterference n wreless networks can be defned accordng to the protocol model or the physcal model [21]. In protocol model [12], [21], there wll be a fxed transmsson range and a fxed nterference range, where the nterference range s typcally 1.5 to 3 tmes of the transmsson range. These two ranges may vary wth the frequency bands. Let T b denote the set of neghborng nodes wthn node s transmsson range over lcensed band b B. For a lnk (, j) usng r for cooperatve communcatons over band b, wehave r j and r T(,j) b = T b T b j. III. COOPERATIVE CONFLICT GRAPH, CONFLICT CLIQUES AND INDEPENDENT SETS IN C-VANETS In ths secton, we frst extend the lnks n C-VANETs nto cooperatve lnks/general lnks wth respect to (w.r.t.) the specal features of cooperatve communcatons. Then, we establsh a 3-D cooperatve conflct graph to descrbe the nterference relatonshp among these extended lnks. Besdes, we also re-defne ndependent sets and conflct clques [11], [12] to show whch lnks can be actvated at the same tme and whch lnks cannot, when cooperatve communcatons s nvolved n C-VANETs. A. Extendng Lnks nto Cooperatve/General Lnks For a lnk (, j), f node r s the best cooperatve relay for t, we calculate the achevable data rate for cooperatve communcatons (.e., C AF (, r, j)) as llustrated n (1). If C AF (, r, j) >C DTx (, j), we can extend lnk (, j) nto (, r, j)

4 PAN et al.: COOPERATIVE COMMUNICATION AWARE LINK SCHEDULING FOR COGNITIVE VEHICULAR NETWORKS 763 and defne (, r, j) as a cooperatve lnk. To keep the lnk notaton consstent, we explot (, φ, j) to denote a lnk usng drect transmssons, where φ s a dummy cooperatve relay, and defne (, φ, j) as a general lnk. The same procedure can be done for each lnk n the C-VANET. Defne R b (,j) = {φ} T(,j) b. Then, we can extend each lnk (, j) nto the form of (, r, j) over band b, wherer R b (,j). Note that for a lnk qualfed to be a cooperatve lnk, the RSU can choose to use t as a cooperatve lnk or a general lnk, when the RSU consders the nterference relatonshp among dfferent lnks and schedules the transmssons over these lnks. B. Establshng the 3-D Cooperatve Conflct Graph Regardng the avalablty of lcensed bands and the features of cooperatve communcatons, we ntroduce a 3-D cooperatve conflct graph to characterze the nterference relatonshp among multple lnks n C-VANETs. Specfcally, n a 3-D cooperatve conflct graph G(V, E), each vertex corresponds to an extended lnk-band par, where a extended lnk-band par s defned as ((, r, j),b). The lnkband par ndcates that the extended lnk (, r, j) operates on avalable lcensed band b. Note that t ncludes the general lnk when the cooperatve relay r = φ, and ncludes the cooperatve lnk when the cooperatve relay r φ. Italso ncludes cooperatve communcatons n sngle-rado snglechannel networks as a specal case when the number of avalable lcensed bands B =1. Two extended lnk-band pars are defned to nterfere wth each other, f any of the followng condtons s true: Condton 1: Two dfferent extended lnk-band pars have nodes n common. Condton 2: If the two extended lnk-band pars are usng the same band, ther transmssons nterfere wth each other when ether the recevng node or the cooperatve relay node of one par s n the nterference range of ether the transmttng node or the cooperatve relay node n the other par. Based on these condtons, we connect two vertces n V wth an undrected edge n G(V, E), f ther correspondng lnk-band pars nterfere wth each other. Note that cooperatve communcaton may ncrease the achevable data rate of a lnk, but t also ncurs addtonal nterferences. The reason s that we must consder both the nodes wthn the nterference range of the transmttng node and the nodes wthn the nterference range of the cooperatve relayng node, when ths cooperatve lnk-band par s scheduled for transmssons 4. C. Cooperatve Independent Sets and Conflct Clques Gven a 3-D cooperatve conflct graph G =(V, E) representng C-VANETs, we descrbe the mpact of vertex u V on vertex v V as follows, { 1, (f there s an edge between vertex u and v) w uv = 0, (f there s no edge between vertex u and v), (4) where the two vertces correspond to two lnk-band pars. Provded that there s a vertex/extended lnk-band set I V and an extended lnk-band u Isatsfyng v I,u v w uv < 4 For specfc examples, please refer to the techncal report posted at Fg. 2. u j v u v j j u v Possble cases for relay selecton collsons w.r.t. lnk schedulng. 1, the transmsson at lnk-band par u wll be successful even f all the other lnk-band pars belongng to the set I are transmttng at the same tme. If any u Isatsfes the condton above, we can schedule the transmssons over all these extended lnk-band pars n I to be actve smultaneously. Such a vertex/extended lnk-band par set I s called a cooperatve ndependent set. If addng any one more extended lnk-band par nto a cooperatve ndependent set I results n a non-ndependent one, I s defned as a maxmum cooperatve ndependent set. Besdes, f there exsts a vertex/extended lnkband par set Z Vn G and any two extended lnk-band pars u and v n Z satsfyng w uv 0(.e., vertex u and v cannot be scheduled to transmt successfully at the same tme.), Z s called a cooperatve conflct clque. If Z s no longer a conflct clque after addng any one more extended lnk-band par, Z s defned as a maxmum cooperatve conflct clque. IV. OPTIMAL COOPERATIVE COMMUNICATION AWARE LINK SCHEDULING FOR HIGH END-TO-END THROUGHPUT After we construct the 3-D cooperatve conflct graph, n ths secton, we frst dscuss the possble collsons of relay selecton w.r.t. lnk schedulng n C-VANETs. Then, we address how to calculate the path capacty and descrbe flow routng constrants for the sngle-rado based nodes. Accordng to the cross-layer constrants, we mathematcally formulate the throughput maxmzaton problem n C-VANETs and nearoptmally solve t by lnear programmng. A. Collsons of Relay Selecton w.r.t. Lnk Schedulng Before we dscuss cooperatve communcaton aware lnk schedulng, we need to clarfy two ssues related to the collsons of relay selecton w.r.t. lnk schedulng. As ntroduced n [10], two knds of relay selecton collsons may happen when cooperatve communcatons s ncorporated nto multhop wreless networks. The frst one s the collson between cooperatve relay selecton and mult-hop relay selecton (.e., a node s chosen both as a cooperatve relay and a multhop relay), as shown n Case 1 and 2 n Fg. 2; the second one s the collson among dfferent lnks for cooperatve relay selecton (.e., dfferent lnks choose the same node as cooperatve relay), as shown n Case 3 n Fg. 2. If there s only one band avalable n the network, t can easly be proved that the relay selecton collsons can never happen w.r.t. lnk schedulng 5. However, f there are multple bands avalable n the network (e.g. n C-VANETs), 5 The hnt s that for any two lnks havng relay selecton collson, these two lnks nherently nterfere wth each other f there s only one band avalable. They cannot be scheduled to transmt smultaneously.

5 764 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 30, NO. 4, MAY 2012 both collsons exst as shown n Fg. 2. Fortunately, the 3- D cooperatve conflct graph can perfectly descrbe all the relay selecton collsons n C-VANETs (e.g., all three cases n Fg. 2 satsfyng nterference Condton 1), so that the RSU can explot t to conduct the cooperatve communcaton aware lnk schedulng. Note that a node n C-VANETs can alternate ts role between cooperatve relay and mult-hop relay at dfferent tme shares, whch s dfferent from the node s fxed role n [10]. B. Path Capacty wth Lnk Schedulng Consderaton For a gven path P, we can establsh the 3-D cooperatve conflct graph G P =(V P, E P ) followng the same approach llustrated n Sec. III-B. Then, we can lst all ndependent sets as I P = {I 1, I 2,, I m,, I M },wherem s I P,and I m V P for 1 m M. Although t s a NP-complete problem to fnd all ndependent sets [13], [16], [22], some brute-force algorthm can fnsh t n polynomal tme f the number of extended lnk-band pars n V P s not large [12]. At any tme, at most one ndependent set can be actvated to transmt packets for all lnk-band pars n that set. Let λ m 0 denote the tme share scheduled to ndependent set I m,and λ m 1, λ m 0(1 m M). (5) 1 m M Let r b (,r,j) (I m) be the data rate for the extended lnk (, r, j) over band b, wherer b (,r,j) (I m)=0f lnk-band par ((, r, j),b) I m.otherwse,f(, r, j) s a cooperatve lnk and ((, r, j),b) I m, r b (,r,j) (I m) s the achevable data rate for (, r, j) over band b when cooperatve communcatons s leveraged. Under AF transmsson mode, r b (,r,j) (I m) can be calculated from (1); f (, r, j) s a general lnk and ((, r, j),b) I m, r b (,r,j) (I m) s the achevable data rate for (, r, j) over band b usng drect transmssons, whch can be calculated as llustrated n (3). By explotng the ndependent set I m,theflow rate that an extended lnk (, r, j) can support over band b n the tme share λ m s λ m r b (,r,j) (I m). Lets represent the aggregated traffc demands. Consderng the avalablty of lcensed bands n C- VANETs, the traffc s feasble at the extended lnk (, r, j) f there exsts a schedule of the ndependent sets satsfyng s s (,r,j) = I P m=1 λ m B (,r,j) b=1 r b (,r,j) (I m). (6) To maxmze the path capacty of P, wehave C P =max mn s (,r,j). (7) (,r,j) P C. Flow Routng Constrants n C-VANETs As for routng, the RSU wll help the source node to fnd the avalable paths to the destnaton node for data delvery. Smlar to the modelng n [19], we mathematcally present those routng constrants as follows. Let f(,r,j) b represent the flow rate of the extended lnk (, r, j) over band b, where N, j T b, r Rb (,j) and r j. If node s the source node,.e., = s r,then b B (j,r,) r,r R b (j,) f b (j,r,) =0. (8) Regardng the sngle-rado requrement of cooperatve communcatons and the nherent sngle-rado constrant of CR devces, we focus on the uncast and sngle-path routng problem. Thus, we have b B (,r,j) r j,r R b (,j) f b (,r,j) δb (,r,j) = s, (9) where δ(,r,j) b ndcates that the extended lnk (, r, j) can only have a nonzero flow at a tme due to the sngle-rado constrant,.e., b B (,r,j) r j,r R b (,j) δ(,r,j) b 1, δb (,r,j) {0, 1}. (10) If node s an ntermedate mult-hop relay node (not a cooperatve relay node),.e., s r and d t,then b B (,r,j) = b B (j,q,) r j,r R b (,j) q, q R b (j,) f b (,r,j) δb (,r,j) f b (j,q,) δb (j,q,). (11) If node s the destnaton node,.e., = d t,then b B (j,r,) r,r R b (j,) f b (j,r,) δb (j,r,) = s. (12) D. Maxmzng the Throughput under Multple Constrants To maxmze the end-to-end throughput between the source node and the destnaton node, the RSU needs to fnd a feasble soluton to jontly assgnng the avalable frequency bands, conductng cooperatve communcaton aware lnk schedulng bands, and routng the traffc for transmsson and recepton n mult-hop C-VANETs. Thus, the end-to-end throughput maxmzaton problem under multple constrants n C-VANETs can be formulated as follows. Maxmze s s.t. (8), (9), (10), (11), (12), (5), and 0 B (,r,j) b=1 I f(,r,j) b m=1 λ m B (,r,j) b=1 r b (,r,j) (I m) ( N,j T b,r R b (,j),b B (,r,j) and I m I ), (13) where (8), (9), (10), (11), and (12) specfy that there s at most one outgong lnk from each node wth a nonzero flow, and that there s a path selected by the RSU between the source and the destnaton; (5) and (13) ndcate that the flow rate of traffc over (, r, j) cannot exceed the capacty of ths extended lnk, whch s obtaned from the cooperatve communcaton aware lnk schedulng as llustrated n Sec. IV-B. Note that I ncludes all ndependent sets n C-VANETs. Gven all ndependent sets 6 n the network, we fnd that the 6 That s a general assumpton used n exstng lterature [11] [15] for obtanng throughput bounds or performance comparson, where both lnk schedulng and flow routng are consdered.

6 PAN et al.: COOPERATIVE COMMUNICATION AWARE LINK SCHEDULING FOR COGNITIVE VEHICULAR NETWORKS 765 formulated optmzaton s a mxed-nteger lnear programmng problem snce δ j only has bnary values. It can nearoptmally be solved n polynomal tme by some typcal algorthms (e.g., sequental fxng algorthm [19], [20], branch and bound [23], etc.) or softwares (e.g., CPLEX [24]), provded that all the cooperatve ndependent sets can be found n G(V, E). V. A HEURISTIC PRUNING ALGORITHM FOR COOPERATIVE COMMUNICATION AWARE SCHEDULING As we know, to fnd all cooperatve ndependent sets n G(V, E) s NP-complete [11] [14], [25]. Compared wth complex path selecton n other wreless networks, t s much more smple n C-VANETs because there are only a few paths between the source and destnaton nodes due to the lmted spatal redundancy and fxed drecton of hghways 7.However, even for a gven path, t s too complex for the RSU to fnd all cooperatve ndependent sets along the path, f the number of extended lnks or the number of avalable lcensed bands along the path s large. Instead of usng cooperatve ndependent sets, n ths secton, we employ a number of maxmum cooperatve clques and propose a 7-step prunng algorthm to approxmate the maxmum throughput for a sesson n C-VANETs. Step 1: Establshng the 3-D cooperatve conflct graph Gven a canddate path P, wefrst set up a 3-D cooperatve conflct graph G P (V P, E P ) as llustrated n Sec. III-B. Step 2: Searchng for the maxmum conflct clques Wth the establshed 3-D cooperatve conflct graph of the gven path P, wetrytofnd all the maxmum cooperatve conflct clques n G P (V P, E P ) and form the set Z consstng of the maxmum cooperatve conflct clques. If P nvolves wth too many extended lnks or avalable bands, and t s mpossble to fnd all the maxmum clques, we can employ K maxmum clques for approxmaton when K s large enough. Step 3: Calculatng the conflct clque transmsson tme Then, we let the RSU employ the maxmum cooperatve conflct clques to estmate the benchmark path capacty for the path P. Smlar to the llustraton n [11], [12], we defne the cooperatve conflct clque transmsson tme T Z for a cooperatve conflct clque Z as T Z = T ((,r,j),b) (14) ((,r,j),b) Z where T ((,r,j),b) s the transmsson tme for one unt of traffc over the extended lnk (, r, j) usng the avalable lcensed band b. Specfcally, T ((,r,j),b) can be wrtten as T ((,r,j),b) = r b (,r,j) 1 (15) (Z), where r(,r,j) b (Z) s equal to the achevable data rate of lnk (, r, j) over band b, f((, r, j),b) Z). Otherwse, r(,r,j) b (Z) =. Step 4: Sortng the maxmum cooperatve conflct clques For Z Z, we sort the maxmum cooperatve conflct clques n terms of the cooperatve conflct clque transmsson 7 In [8], Dng and Leung even employ strng topology to nvestgate the cross-layer routng problem n VANETs. tme T Z.LetT P be the maxmum value of the transmsson tme for all cooperatve conflct clques. T P can be wrtten as T P =max T Z. (16) Z Z Consderng an extended lnk-band par ((, r, j),b) n Ẑ = argmax (T Z ) and one unt of traffc successfully delvered Z Z from the source to the destnaton, t takes tme T P to travel through all the extended lnk-band pars n Ẑ, and((, r, j),b) cannot be scheduled to do any other transmsson durng T P. That ndcates that the throughput at the extended lnk-band par ((, r, j),b) s less than or equal to 1 T P. Snce the end-toend throughput cannot be larger than the throughput of any lnk along the path, the benchmark path capacty C P can be estmated as 8 C P = 1 T P. Step 5: Selectng the optmal band for the hgh throughput If there are multple avalable lcensed bands for an extended lnk to access, one of them must be chosen due to the sngle-rado constrant. From (14) and (16), we fnd that f the sze of Ẑ shrnks, the throughput of the path may ncrease. It s obvous that f some of the co-band nterference between the extended lnks can be mtgated, the sze of Ẑ can be effectvely reduced. As we know, the CR devces can be tuned nto dfferent frequences and allow the extended lnks to operate on dfferent bands. Ths specal CR feature wll help to reduce co-band nterference between the extended lnks so that the end-to-end throughput may be mproved. Followng ths thread, we conduct the optmal band selecton as follows. Frst, for an extended lnk (, r, j) wth multple accessng bands, we randomly select an extended lnk-band par ((, r, j),b) n Ẑ and temporarly delete other ((, r, j), ) parsaswellastheconflct edges assocated wth ((, r, j), ). Then, we fnd the maxmum cooperatve conflct clque n the leftover graph cut from Ẑ and calculate the clque transmsson tme T ((,r,j),b) as n (14). For b B Ẑ (,r,j), the same process s conducted and the values of clque transmsson tme are stored. After that, we update TẐ as TẐ =mn{t ((,r,j),1) Ẑ,T ((,r,j),2) Ẑ,,T ((,r,j), B (,r,j) ) }. (17) We dentfy the band reachng the value of TẐ, put that band nto (, r, j) s usage and prune all the other ((, r, j), ) parsaswellastheconflct edges assocated wth ((, r, j), ). The same procedure above s repeated by all the extended lnk-band pars n Ẑ one after another, and the T Ẑ s contnuously updated. If all the avalable lcensed bands are dentcal to an extended lnk n terms of band condton (.e., bandwdth, the propagaton gan, etc.), t wll be much more smple to select the optmal band for ths extended lnk. As for such an extended lnk, we just need to keep the extended lnk-band par wth the least conflct edges and elmnate the other lnkband pars assocated wth ths extended lnk. Meanwhle, we also prune the correspondng conflct edges and update TẐ based on the leftover graph cut from Ẑ. 8 Actually, the benchmark path capacty C P should be upper-bounded by 1,.e., C T P 1. The equal sgn holds f there are no odd cycles [22] P T P n G P as llustrated n [12]. In ths paper, we just consder the general paths wthout odd cycles. Ẑ

7 766 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 30, NO. 4, MAY 2012 Step 6: Prunng the cooperatve/general lnk-band pars After the band selecton for an extended lnk, t s necessary to determne whch type of transmsson (.e., cooperatve communcatons or drect transmsson) should be used by ths extended lnk. In Ẑ, there may be two coupled lnk-band pars extended from the same lnk (, j): a general lnk-band par ((, φ, j),u) and a cooperatve lnk-band par ((, r, j),v), (r R (,j) and r φ), where u and v are the avalable bands selected for (, φ, j) and (, r, j) n Step 5, respectvely. From (1), (2), (3), (15) and (14), we can easly calculate the transmsson tme for the clque Ẑ\{((, φ, j),u)} and Ẑ\{((, r, j),v)},.e., T Ẑ\{((,φ,j),u)} and T Ẑ\{((,r,j),v)}, respectvely. We compare TẐ\{((,φ,j),u)} and TẐ\{((,r,j),v)} and make the decson of prunng the cooperatve/general lnkband pars as follows. If TẐ\{((,φ,j),u)} >TẐ\{((,r,j),v)}, the RSU wll keep the general lnk-band par ((, φ, j),u) and prune the cooperatve lnk-band par ((, r, j),v) as well as the conflct edges assocated wth ((, r, j),v). Thats,theRSU chooses the drect transmsson nstead of cooperatve communcatons for the lnk (, j). In addton, the RSU wll update TẐ by settng TẐ = TẐ\{((,r,j),u)}. If TẐ\{((,φ,j),u)} TẐ\{((,r,j),v)}, the RSU wll keep the cooperatve lnk-band par ((, r, j),v) and prune the general lnk-band par ((, φ, j),u) as well as the conflct edges assocated wth ((, φ, j),u). That s, the RSU chooses cooperatve communcatons nstead of the drect transmsson for the lnk (, j). In addton, the RSU wll update TẐ by settng TẐ = TẐ\{((,φ,j),u)}. The same procedure s repeated by any two coupled extended lnk-band pars n Ẑ assocated wth the same lnk, and the TẐ s contnuously updated. Step 7: Iteratng the procedure and estmatng the throughput Jump back to Step 4, resort the maxmum cooperatve conflct clques n terms of the cooperatve conflct clque transmsson tme (wth the updated TẐ), fnd new Ẑ and terate the followng steps wth ths clque. Iteratons contnue untl ˆTP cannot be decreased further. Then, the RSU can set TP = T P = TẐ and estmate the throughput of P as CP = 1 T. Smlarly, the RSU can maxmze the throughput P of the other paths va cooperatve communcaton aware lnk schedulng, and select the one wth the hghest throughput 9. VI. PERFORMANCE EVALUATION We consder a C-VANET consstng of N = 30 vehcles randomly dstrbuted along a 3 km two lane straght hghway. All the vehcles are movng n the same drecton. The bandwdth for each band W s set to be 8 MHz, schedule perod τ s set to be 10 s, the maxmum transmsson power at each node s set to be 5 W, the transmsson range s set to be 250 m, and the nterference range 10 s set to be 400 m. For smplcty, we assume that h j only ncludes the propagaton gan between node and j and s gven by 9 For specfc examples, please refer to the techncal report posted at 10 As llustrated n [20], the transmsson range and nterference range can be determned by the recever senstvty and the threshold of nterference tolerance, respectvely. Achevable Rate (Mbps) CC Dtx Dstance from Source Node to Cooperatve Relay (m) Fg. 3. Comparson between cooperatve communcatons and drect transmssons for a three-node schematc. h j 2 = d(, j) 4,whered(, j) s the dstance (n meters) between nodes and j and path loss ndex s 4. For the AWGN channel, we assume the varance of nose s W at all nodes. Besdes, we set K = 200,.e., f the total number of maxmum cooperatve clques n G P s less than or equal to 200, we employ all the maxmum cooperatve clques; otherwse, we employ 200 maxmum cooperatve clques for approxmaton. For llustratve purposes, we nvestgate the throughput maxmzaton problem n C-VANETs wth the followng two scenaros: ) all the vehcles move at the speed of 75 mph (.e., km/h, the typcal speed lmt); ) vehcle speed follows a Gaussan dstrbuton wth a mean of 75 mph and a standard devaton of 10 mph (.e., km/h). By fxng the leftmost node as the source and the rghtmost node as the destnaton, we compare the results of dfferent throughput maxmzaton algorthms. These results nclude the optmal throughput consderng both transmsson mode selecton and band selecton (.e., Optmal CC/Dtx w/ CR ), the throughput obtaned from the proposed prunng algorthm (.e., Prunng CC/Dtx w/ CR ), the optmal throughput consderng band selecton under dfferent transmsson modes (.e., Optmal CC w/ CR and Optmal Dtx w/ CR ) and the sngle-band based optmal throughput under dfferent transmsson modes (.e., Optmal CC w/o CR and Optmal Dtx w/o CR ) [12]. Note that gven the ndependent sets, we can employ CPLEX [24] to solve the optmzaton problems and obtan near-optmal results. Besdes, we demonstrate the mpact of the number of avalable lcensed bands on the throughput n C-VANETs and present the results n Fg. 4. For the sessons from the source node to all the other nodes along the hghway, we also conduct smulatons to evaluate the mpact of dstance wth dfferent throughput maxmzaton algorthms and show the correspondng results n Fg. 5. In Fg. 3, we compare two transmsson modes n terms of lnk capacty. Here, we assume the transmtter, the cooperatve relay and the recever are on the same lane, and the dstance between the transmtter and the recever s 250 m. We fnd that cooperatve communcatons s not necessarly better than drect transmssons n terms of lnk capacty, and the beneft brought by cooperatve communcatons hghly depends on the locaton of the cooperatve relay. Fgure 4 demonstrates the mpact of the number of avalable lcensed bands on the end-to-end throughput n C-VANETs. From the results shown n Fg. 4(a) and Fg. 4(b), four observatons can be made n order. Frst, Optmal CC/Dtx w/ CR and the heurstc prunng algorthm outperform the other algorthms n terms of end-to-end throughput. It s not

8 PAN et al.: COOPERATIVE COMMUNICATION AWARE LINK SCHEDULING FOR COGNITIVE VEHICULAR NETWORKS 767 End to end Throughput (Mbps) Optmal CC/Dtx w/ CR Prunng CC/Dtx w/ CR Optmal Dtx w/ CR Optmal CC w/ CR Optmal Dtx w/o CR Optmal CC w/o CR End to end Throughput (Mbps) Optmal CC/Dtx w/ CR Prunng CC/Dtx w/ CR Optmal Dtx w/ CR Optmal CC w/ CR Optmal Dtx w/o CR Optmal CC w/o CR 2.5 End to end Throughput (Mbps) Number of Avalable Lcensed Bands (a) Scenaro 1: vehcle speed s 75 mph. Optmal CC/Dtx w/ CR Prunng CC/Dtx w/ CR Optmal Dtx w/ CR Optmal CC w/ CR Optmal Dtx w/o CR Optmal CC w/o CR End to end Throughput (Mbps) Node ID (n order of dstance from the source node) (a) Scenaro 1: vehcle speed s 75 mph. Optmal CC/Dtx w/ CR Prunng CC/Dtx w/ CR Optmal Dtx w/ CR Optmal CC w/ CR Optmal Dtx w/o CR Optmal CC w/o CR Number of Avalable Lcensed Bands (b) Scenaro 2: vehcle speed follows Gaussan dstrbuton wth a mean of 75 mph and a standard devaton of 10 mph. Fg. 4. Impact of the number of avalable lcensed bands on the end-to-end throughput n C-VANETs Node ID (n order of dstance from the source node) (b) Scenaro 2: vehcle speed follows Gaussan dstrbuton wth a mean of 75 mph and a standard devaton of 10 mph. Fg. 5. Impact of dstance between the source and destnaton nodes on the end-to-end throughput n C-VANETs. surprsng because both of them have a jont consderaton of transmsson mode selecton and the band selecton, when the transmssons are scheduled. In addton, the throughput obtaned from the proposed prunng algorthm s close to that from the optmal one. Second, consderng lnk schedulng, cooperatve communcatons may ncur extra nterference and hnder the end-to-end throughput, especally when the number of avalable bands s lmted. Thrd, the CR capablty of the nodes creates more opportuntes to use cooperatve communcatons and therefore mprove the throughput. As for those algorthms consderng the CR capablty of nodes, the endto-end throughput ncreases as the number of avalable bands ncreases. The reason s that more lcensed bands avalable gve more opportuntes for nodes accessng, so that more cooperatve lnks can be utlzed wthout ncurrng addtonal nterference and more lnks can be actvated for transmsson smultaneously. The ncrement of throughput stops when the number of avalable bands s large enough,.e., the throughput cannot be further ncreased snce both cooperatve communcatons and lnk schedulng are fully exploted. Fourth, the devaton of vehcle speed leads to performance degradaton of lnk schedulng. That s because speedng up/slowng down may result n certan changes of network topology (e.g., overtakng) n C-VANETs. Fgure 5 shows the mpact of dstance between the source and destnaton nodes on the throughput n C-VANETs. For the smplcty of computng ndependent sets [13], we assume there are 2 lcensed bands avalable n the network. Except for the observatons we already have made n Fg. 4, we fnd that the longer dstance the path spans, the more lkely the throughput s affected by the band selecton, transmsson mode selecton and lnk schedulng. For a short-dstance path whch ncludes only a few lnks, cooperatve communcatons s always preferred snce there s no lnk schedulng nvolved. By contrast, a long-dstance path ncludes more lnks, whch mples that more lnks could be scheduled to transmt at the

9 768 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 30, NO. 4, MAY 2012 same tme. Thus, the end-to-end throughput maxmzaton of such a path depends more on band selecton, transmsson mode selecton and lnk schedulng. VII. CONCLUSION In ths paper, we have studed the throughput maxmzaton problem n C-VANETs under multple constrants (.e., CR devces nherent sngle-rado constrant, the avalablty of lcensed spectrum, transmsson mode selecton and lnk schedulng). Consderng the specal features of cooperatve communcatons, we frst extend the lnks and classfy them nto cooperatve lnks/general lnks. Then, dependng on the avalable bands at dfferent extended lnks, we defne extended lnk-band pars and form a 3-D cooperatve conflct graph to descrbe the conflct relatonshp among those pars. After that, we mathematcally formulate the end-to-end throughput maxmzaton problem. Gven all cooperatve ndependent sets n C-VANETs, we can relax the formulated optmzaton problem and near-optmally solve t by lnear programmng. Due to the NP-completeness of fndng all ndependent sets, we provde a heurstc prunng algorthm for the cooperatve communcaton aware lnk schedulng as well. By numercal smulatons, we demonstrate that: ) the CR capablty creates more opportuntes for usng cooperatve communcatons; ) the performance of lnk schedulng wth approprately selected transmsson mode s better than that purely relyng on one transmsson mode. REFERENCES [1] H. Ilhan, M. Uysal, and I. Altunbas, Cooperatve dversty for ntervehcular communcaton: Performance analyss and optmzaton, IEEE Trans. Veh. Technol., vol. 58, no. 7, pp , Februrary [2] B. Jarupan and E. Ekc, A survey of cross-layer desgn for VANETs, Ad Hoc Networks, vol. 9, no. 5, pp , July [3] K. Tsukamoto, S. Matsuoka, M. Tsuru, Y. Oe, and O. Altntas, Dstrbuted channel coordnaton n cogntve wreless vehcle-to vehcle communcatons, n Proc. of Internatonal Conference on Wreless Access n Vehcular Envronments, Dearborn, MI, December [4] D. Nyato, E. Hossan, and P. Wang, Optmal channel access management wth QoS support for cogntve vehcular networks, to appear n IEEE Trans. Moble Computng. [5] J. N. Laneman, D. N. C. Tse, and G. W. Wornell, Cooperatve dversty n wreless networks: effcent protocols and outage behavor, IEEE Trans. Inf. Theory, vol. 50, no. 12, pp , December [6] J. N. Laneman, Cooperatve dversty n wreless networks: Algorthms and archtectures, PhD s Thess, Massachusetts Insttute of Technology, USA, September [7] S. Sharma, Y. Sh, Y. T. Hou, and S. Kompella, An optmal algorthm for relay node assgnment n cooperatve communcatons, IEEE/ACM Trans. Netw., vol. 19, no. 3, pp , June [8] Z. Dng and K. K. Leung, Cross-layer routng usng cooperatve transmsson n vehcular ad-hoc networks, IEEE J. Sel. Areas Commun., vol. 29, no. 3, pp , March [9] L. Dng, T. Meloda, S. Batalama, and J. Matyjas, Dstrbuted routng, relay selecton, and spectrum allocaton n cogntve and cooperatve ad hoc networks, n SECON, 2010, Boston, MA, June [10] S. Sharma, Y. Sh, Y. T. Hou, H. D. Sheral, and S. Kompella, Cooperatve communcatons n mult-hop wreless networks: Jont flow routng and relay node assgnment, n INFOCOM 2010, SanDego, CA, March [11] F. Chen, H. Zha, and Y. Fang, Avalable bandwdth n multrate and multhop wreless ad hoc networks, IEEE J. Sel. Areas Commun., vol. 28, no. 3, pp , Aprl [12] H. Zha and Y. Fang, Impact of routng metrcs on path capacty n multrate and multhop wreless ad hoc networks, n ICNP 2006, Santa Barbara, CA, November [13] H. L, Y. Cheng, C. Zhou, and P. Wan, Mult-dmensonal conflct graph based computng for optmal capacty n MR-MC wreless networks, n ICDCS 2010, Genoa, Italy, June [14] J. Tang, G. Xue, and W. Zhang, Interference-aware topology control and qos routng n mult-channel wreless mesh networks, n MobHoc, 2005, Urbana-Champagn, IL, May [15] J. Tang, S. Msra, and G. Xue, Jont spectrum allocaton and schedulng for far spectrum sharng n cogntve rado wreless networks, Computer Networks (Elsever) Journal, vol. 52, no. 11, pp , August [16] M. R. Garey and D. S. Johnson, Computers and Intractablty; A Gude to the Theory of NP-Completeness. New York, NY: W. H. Freeman & Co., [17] H. Km and K. G. Shn, Effcent dscovery of spectrum opportuntes wth mac-layer sensng n cogntve rado networks, IEEE Trans. Moble Computng, vol. 7, no. 5, pp , May [18] W.-Y. Lee and I. F. Akyldz, Optmal spectrum sensng framework for cogntve rado networks, IEEE Trans. Wreless Commun., vol. 7, no. 10, pp , October [19] M. Pan, C. Zhang, P. L, and Y. Fang, Jont routng and lnk schedulng for cogntve rado networks under uncertan spectrum supply, n INFOCOM 2011, Shangha, Chna, Aprl [20] Y. T. Hou, Y. Sh, and H. D. Sheral, Spectrum sharng for mult-hop networkng wth cogntve rados, IEEE J. Sel. Areas Commun., vol. 26, no. 1, pp , January [21] P. Gupta and P. R. Kumar, The capacty of wreless networks, IEEE Trans. Inf. Theory, vol. 46, no. 2, pp , March [22] R. Destel, Graph Theory. Sprnger, [23] Y. Pochet and L. A. Wolsey, Producton Plannng by Mxed Integer Programmng. Secaucus, NJ, USA: Sprnger-Verlag New York, Inc., [24] IBM ILOG CPLEX Optmzer. [Onlne]. Avalable: bm.com/software/ntegraton/optmzaton/cplex-optmzer/ [25] K. Jan, J. Padhye, V. N. Padmanabhan, and L. Qu, Impact of nterference on mult-hop wreless network performance, n Mobcom 2003, San Dego, CA, September 2003.