MITSUBISHI ELECTRIC RESEARCH LABORATORIES http://www.mer.com Dynamc Resource Contro for Hgh-Speed Downn Pacet Access Wreess Channe Hua-Rong Shao, Cha Shen, Daqng Gu, Jnyun Zhang, Php Or TR2003-60 May 2003 Abstract It s a chaengng tas to provde Quaty of Servce (QoS) contro for a shared hgh-speed downn pacet access (HSDPA) wreess channe. In ths paper, we frst propose a new dynamc resource contro framewor ntegrated wth adaptve moduaton and codng (AMC) and hybrd automatc repeat request (H-ARQ) to support cass-based mutmeda appcatons over HSDPA wreess channes. Then we present a new schedung agorthm, Deay-senstve Dynamc Far Queueng (DSDFQ), to meet deay requrements of mutmeda appcatons as we as mantan hgh networ effcency. The proposed approach can easy adapt to oad fuctuatons of dfferent traffc casses and varyng wreess channe condtons caused by user mobty, fadng and shadowng. Performance evauaton shows the advantage of our proposed approach. 23rd Internatona Conference on Dstrbuted Computng Systems, Internatona Worshop on Mobe and Wreess Networs Ths wor may not be coped or reproduced n whoe or n part for any commerca purpose. Permsson to copy n whoe or n part wthout payment of fee s granted for nonproft educatona and research purposes provded that a such whoe or parta copes ncude the foowng: a notce that such copyng s by permsson of Mtsubsh Eectrc Research Laboratores, Inc.; an acnowedgment of the authors and ndvdua contrbutons to the wor; and a appcabe portons of the copyrght notce. Copyng, reproducton, or repubshng for any other purpose sha requre a cense wth payment of fee to Mtsubsh Eectrc Research Laboratores, Inc. A rghts reserved. Copyrght c Mtsubsh Eectrc Research Laboratores, Inc., 2003 201 Broadway, Cambrdge, Massachusetts 02139
MERLCoverPageSde2
Dynamc Resource Contro for Hgh-Speed Downn Pacet Access Wreess Channe Hua-Rong Shao, Cha Shen, Daqng Gu, Jnyun Zhang, Php Or Mtsubsh Eectrc Research Laboratores, 201 Broadway, Cambrdge, MA, 02139 {shao, shen, dgu, zhang, por}@mer.com ABSTRACT It s a chaengng tas to provde Quaty of Servce (QoS) contro for a shared hgh-speed downn pacet access (HSDPA) wreess channe. In ths paper, we frst propose a new dynamc resource contro framewor ntegrated wth adaptve moduaton and codng (AMC) and hybrd automatc repeat request (H-ARQ) to support cass-based mutmeda appcatons over HSDPA wreess channes. Then we present a new schedung agorthm, Deay-senstve Dynamc Far Queueng (DSDFQ), to meet deay requrements of mutmeda appcatons as we as mantan hgh networ effcency. The proposed approach can easy adapt to oad fuctuatons of dfferent traffc casses and varyng wreess channe condtons caused by user mobty, fadng and shadowng. Performance evauaton shows the advantage of our proposed approach. 1. INTRODUCTION Future wreess ceuar networs w support ntegrated mutmeda appcatons wth varous quaty of servce (QoS) requrements. Provdng QoS dfferentaton n hghspeed wreess data networs s consdered as a promsng souton for the ncreasng mutmeda demands from wreess end users. Due to the dfference between traffc characterstcs of pacet data and tradtona crcutswtched voce, dedcated channes are aocated for data servces n many wreess communcaton systems and standards such as Hgh Data Rate (HDR) systems (a..a. 1xEV-DO) [1], 1xTREME (a..a. 1xEV-DV) of 3rd Generaton Partnershp Proect 2 (3GPP2) [2], and Hgh Speed Downn Pacet Access (HSDPA) of 3rd Generaton Partnershp Proect (3GPP) [3]. In HSDPA, a hgh-speed downn data channe s shared by mutpe users wthn the same ce. Many technooges are currenty under consderaton for HSDPA system to enhance the system performance. They ncude adaptve moduaton and codng (AMC), hybrd automatc repeat request (H-ARQ), fast ce seecton (FCS) and mutpe-nput-mutpe-output (MIMO) technooges. AMC offers a n adaptaton method that can dynamcay adapt moduaton-codng scheme to current channe condtons for each user (nown as UE (user equpment)). In a system wth AMC, users cose to the base staton (BTS) (nown as node B n HSDPA) usuay have good rado n and are typcay assgned hgher order moduatons and hgher code rates (e.g. 64 QAM wth R=3/4 turbo codes). The moduaton-order and/or code rate w decrease as the dstance of a user from BTS ncreases. H-ARQ provdes a retransmsson mechansm for the ost or erroneous nformaton. There are many schemes for mpementng H-ARQ - Chase combnng, Rate compatbe Punctured Turbo codes and Incrementa Redundancy. An mportant research ssue s how to perform resource contro and management and ntegrate rado resource contro schemes wth these new technooges. As a smpe nstance, data transmsson capacty at a base staton w vary accordngy wth AMC schemes used by users. Gven the same amount of spreadng code (code space) and tme sot (tme space) resources, an mobe user wth hgher order moduaton scheme and hgher codng rate usuay can obtan hgher data rate than those wth ower order moduaton schemes and ower codng rate. Ths paper focuses on the resource aocaton wth QoS contro n HSDPA system wth AMC and H-ARQ mechansms. Pacet schedung s one of the most mportant QoS contro approaches for wreess mutmeda communcatons. Many proposas on pacet schedung n wreess envronments have been proposed, for exampe, CSDPS (Channe state dependent pacet schedung)[5], IWFQ (Ideazed Wreess Far Queueng Agorthm) [6], CIF-Q (Channe-Condton Independent Far Queueng) [7], SBFA (Server Based Farness Agorthm) [8], I-CSDPS (Improved channe state dependent pacet schedung) [9], CAFQ (Channe adaptve far queueng) [10], M-LWDF (Modfed Largest Weghted Deay Frst) [11], and CDGPS (Code-dvson generazed processor sharng) [12]. Most approaches [5-9] assume a smpe two-state Marov mode: the scheduer smuates an error-free system runnng a wre-ne pacet schedung agorthm when the sessons have good (or perfect) channe states (effectve throughput s mum). When the sesson that s schedued to transmt data encounters a bad channe state, t w gve the transmsson opportunty to other error-free sessons (.e., wth a good channe state). Then these errorfree sessons w gve ther transmsson rghts bac to the error sesson once the error sesson escapes from a bad
channe state. These agorthms many am to provde farness and aso soft QoS guarantees. In [10], a new defnton of farness and a schedung agorthm adaptng to severa channe condtons s proposed. However, t doesn t provde expct QoS guarantee. In [11], a user schedung scheme based on the tradeoff between deay and throughput s presented. Ths approach assumes that each user can ony support one QoS traffc cass at a tme. GPS (Generazed processor sharng) scheme s dynamcay apped to spreadng codes rather than tme sots for dfferent mobe users n [12]. In ths paper, we propose a resource schedung mechansm that s cosey ntegrated wth some HSDPA technooges such as AMC and H-ARQ. Because AMC scheme for a user s changed dynamcay accordng to ts channe condton, the scheduer can obtan the channe condton nformaton of each user through ts AMC scheme. Therefore, the proposed schedung mechansm s not based on the smpe On-Off two state Marov mode. H-ARQ ntroduces extra traffc oad nto wreess networs. Prevous wor ddn t consder ths serousy. In the proposed scheme, we dfferentate the orgna pacets and the re-transmtted pacets by pacng them at dfferent queues. In the future, one mobe devce shoud be abe to support mutpe streams wth dfferent QoS requrements smutaneousy. For nstance, one user can get streamng vdeo from a vdeo server and get text fe from a FTP server at the same tme. Thus the scheduer at base staton needs to hande both QoS traffc casses and dfferent mobe users sharng the same HSDPA downn channe. Prevous wor usuay consders how to schedue resources to dfferent mobe users and assgn each user a separate queue. In the proposed scheme, queues are assgned to dfferent QoS casses rather than dfferent users. Queue parameters such as queue ength are specfed accordng to deay and pacet oss requrements. We propose DSDFQ (Deay-senstve Dynamc Far Queueng) schedung scheme that dynamcay adust the queue weghts wth traffc dynamcs and queue status. Therefore, expct QoS can be provded wth the proposed scheme. Each cass s then sub-dvded nto sub-casses accordng to each user s AMC scheme. It s that, n each cass, mobe users wth the same AMC are grouped nto the same sub-cass. Users wth better channe condton (ndcated by AMC schemes) are gven hgher schedung prorty than those wth worse channe condton. It s acheved by settng dfferent RED (random eary Detecton) for dfferent AMC schemes. Because of ths nd of sub-cass mechansm, a good system throughput can be mantaned by constranng the mobe users wth bad channe condton. The rest of the paper s organzed as foows. Secton 2 presents our new dynamc resource contro framewor. Secton 3 descrbes the proposed pacet schedung agorthm and then gves a theoretca anayss. Secton 4 shows the smuaton resuts. Fnay, secton 5 concudes the paper. Fg.1: Resource contro framewor for HSDPA 2. RESOURCE CONTROL FRAMEWORK Fg. 1 ustrates the proposed dynamc resource contro framewor for HSDPA. We many consder how to perform resource contro at node B n HSDPA systems. It s assumed that mutpe mobe or fxed UEs exst n one ce, and each UE can support mutpe streams or traffc casses. Four traffc casses are specfed n 3GPP or HSDPA: Conventona cass voce traffc; Streamng cass audo and vdeo traffc; Interactve cass web browsng, database read types of traffc; and Bacground cass best effort traffc. Compared wth other s wor, our approach not ony consders how to schedue pacets to dfferent UEs, but aso schedue dfferent traffc casses n a mutpe UE envronment. Wthn the framewor, transmsson requests are sent to the Admsson Contro modue. The Admsson Contro modue maes the decson whether or not to admt new transmsson streams by computng the avaabe resource from physca-ayer resource measurement and exstng traffcs. Once admtted, traffc streams enter the Traffc Shapng and Cassfcaton modue, and then are passed to dfferent queues accordng to traffc casses to whch traffc streams beong. Traffc streams are cassfed accordng to deay and pacet oss requrements. These QoS parameters decde the queue ength, the weght of a queue and RED (Random Eary Detecton) confguraton. Because H-ARQ retransmsson may ntroduce arge amount of extra traffc oad when wreess channes are n bad condton, we assgn two queues to each traffc cass: orgna transmsson queue and re-transmsson queue. For each cass, sub-casses (dfferent coors) are specfed accordng to MCS (Moduaton and codng scheme). Both spreadng codes and tme frames are schedued to UEs.
3GPP HSDPA specfes that tme frame ength (or TTI: Transmsson Tme Interva) s fxed and equa to 2ms, and spreadng codes are orthogona codes. SF (Spreadng Factor) s fxed and equa to 16, so that the tota spreadng code number s 16. We assume that one UE can be assgned wth mutpe spreadng codes n the same TTI. WFQ weghts are dynamcay adusted accordng to the queue status of dfferent casses. Detaed schedung w be presented n Secton 3. As ustrated n Fg. 1, each UE montors the channe status and feedbac carrer-to-nterference (C/I) nformaton to node B perodcay. Then node B estmates the avaabe resource and decdes moduaton and codng scheme for data transmsson to the UE. We assume there are N MCS eves avaabe, and et R n ( n N ) data rate wth MCS eve n, FER (,n) 1 be γ denote frame error rate (FER) for a gven sgna-to-nose rato of channe γand MCS eve n. Then, the effectve data transmsson rate s R n 1 FER γ, n. A UE seects MCS eve equa to ( ( ) so as to mze the effectve data rate for the measured γ. = R 1 FER γ, n where s the That s, { ( ( ))} n n seected MCS eve. The protoco data unt (PDU) at a rado nterface s the data traffc carred durng a tme frame (or TTI). Snce data rate vares wth MCS eve, PDU sze aso vares. An IP pacet arrved at node B s segmented accordng to the aowabe PDU szes. A UE acnowedges the recepton of a PDU. If node B doesn t receve an acnowedgement wthn a tme perod, t retransmts the PDU. If the mum number of retransmsson tras for a PDU s M, a PDU can be transmtted up to M+1 tmes ncudng the orgna transmsson. Durng a retransmsson sequence, MCS eve s ept the same as that for the orgna transmsson. To mprove n utzaton and adapt to our schedung mechansm, we propose dfferent mum number of retransmsson (M) for PDUs wth dfferent MCS eves. The hgher MCS eve s, the arger M vaue s gven. 3. DSDFQ Schedung We propose a new schedung scheme caed Deay- Senstve Dynamc Far Queueng (DSDFQ) for HSDPA systems. The scheduer s ocated at MAC ayer to schedue PDUs carred at dfferent TTIs and wth dfferent spreadng codes. For convenence, we st ca these nds of PDUs as pacets. From Secton 2, t can be seen the proposed scheduer s cosey ntegrated wth AMC and H- ARQ. The ey pont of the schedung scheme s dynamc adustment of the weght of each queue accordng to the current deay. We use coored toens to dstngush pacets wth dfferent MCS eves wthn one queue. Moreover, we use Weghted RED as the pacet drop scheme. 3.1. Schedung Agorthm The purpose of Deay-senstve Dynamc Far Queueng s to mantan a dynamc farness accordng to the deay status of each queue. Une most far queueng schemes, ths scheme s not an approxmate approach to GPS (Generazed Processor Sharng), though some deas are from the common Far Queueng schemes. In [13], a sorted prorty queue mechansm that s commony used by Vrtua Coc, WFQ and WF2Q was dscussed. The framewor of Deay-senstve Dynamc Far Queueng s aso based on ths mechansm. In ths mechansm, there s a state varabe assocated wth each connecton to montor and enforce ts traffc. Here, we use the same state varabe as WFQ that s caed the Vrtua Fnsh Tme. The vrtua fnsh tme of a pacet s defned as: 1 L F = { F, V( t)} +, where F s the vrtua fnsh tme of the th pacet of cass, V (t) s the vrtua tme when the th pacet arrves, s the weght of cass and L s the pacet sze of the th pacet measured n byte. In WFQ, weght s fxed and does not refect the current stuaton of the queue. In the proposed Deay-senstve Dynamc Far Queueng, the weght of each queue,, s a varabe. When an event occurs, the deay of ths pacet ( deay ) s changed and s recacuated as foows: 0 = f ( deay ) = mn( + deay where 0 s the basc weght of cass,, ) s the mum weght of cass, deay s queueng deay of the th pacet n cass, and s the adustment parameter. Let S denote the vrtua tme when pacet n cass starts to be served, and F denote the vrtua tme when pacet n cass fnshes servng. When a pacet s enterng a queue, the processng steps are: 1) Get the ma deay of the queue, 2) Update vrtua tme,.e., get S, then F, and determne the dequeue order. Mathematcay, Steps 1) and 2) can be respectvey represented as deay and = deay 1 + L 1 1 /( Bw)
F where 1 = ( F, V ( tast ) + ( t tast ) / ) + L /( Bw), 1 L s the ength of the th pacet n cass, Bw s the bandwdth of the output n, and t ast s the rea tme when the vrtua tme ast updates (the weght s atest change). Accordng to the order of fnsh tme, nsert ths pacet to the outgong queue. 3) Next, change the weght. Accordng to the new deay vaue, cacuate the current weght as foows: 0 = f ( deay ) = + g ( deay ), 0 where, s the basc weght of cass, and g ( deay ) can be treated as a curve caed Weght Curve. 4) Determne the rea tme at whch the next pacet w depart the queue. Then schedue the next departure of the queue. (Ths can aso be processed when dequeueng a pacet f the processng speed s fast enough). Denote the rea tme as Next(t). Then at tme Next(t), the obectve s to change the weght and deay of each cass as foows. deay = deay L /( Bw), At the output n, the system smpy forwards the pacets. Thus, when a pacet P enters the queue, ts deay s frst cacuated, and then the weght s ncreased f the deay s arger than before. So the guaranteed rate of cass s temporary ncreased. Snce the pacets of each cass arrve ndependenty and the weght changes each tme when a pacet arrves, the dynamc baance s mantaned and the bandwdth aocaton coud be much more far than other statc schedung schemes. We ca such farness Dynamc Farness. 3.2. Theoretca Anayss We anayze the deay bound of Deay-senstve Dynamc Far Queueng when g (deay ) s not zero. In the foowng anayss, we assume that the traffc has the constrant smar to those by eay bucet shapng. Thus, the traffc enterng the networ s shaped as foows: A ( τ, t) σ + ρ ( t τ ), 0 τ t, where A ( τ, t) s the amount of traffc for cass that enters the networ durng tme nterva [ τ,t]. Toens are generated at a fxed rate ρ, and pacets can be reeased nto the networ ony after acqurng the requred number of toens from the eay bucet. The eay bucet contans at most σ bts toens. In our scheme, the cacuaton of deay of pacet n cass s a ey probem. As mentoned earer, we cacuate the deay when a pacet enqueues and update t when a pacet dequeues the system. Thus the deay that we get woud be smaer than the rea deay by a vaue wthn (1,L/Bw) where L s the pacet ength and Bw s the bandwdth of the output n. We use another method to get the deay of a pacet. Suppose that the system starts ts busy perod from tme zero. Let D ( t) = L / Bw (, ) A denote the tota deay of pacet of cass from tme zero. The rea deay of pacet of cass can be represented as: deay = D ( t) t = L / Bw t, (, ) A where t s the rea tme when pacet of cass arrves. Further, we assume that the traffc s greedy and the servce start tme of a pacet s ust the tme that the prevous pacet fnshes ts servce. Thus, we have 1 S = F, and F 1 = S + L /( Bw) = F + L /( Then we can get: D D 1 = + Bw). L (, ) B(, ) / Bw and 1 B(, ) = {(, ) : F F < L /( Bw)}. The number of eements n set B (, ) can be nterpreted as N N, B, B < where, whch satsfes L ( / L ) ρ mn Bw L 1 < ( ρ ) / Bw, Lmn ρ s the average rate of cass, L mn s the mnmum sze of a pacet, and s the mum weght of cass. Now, we can see 1 L L D < D + Bw L mn ( ρ 1 ) / Bw 2 1 1 L 1 1 D < D + ( ). 2 ρ = D + Cs Bw Lmn Consequenty, we can derve the deay as foows: D < Cs, 0 2 1 1 D < D + Cs, 1 f ( D ) 1 1 and 1 1 D < D + Cs. 1 f ( D )
The deay of pacet n cass s: deay Suppose the buffer of cass s w be dropped when the queue sze q m, where Qen Qen = = 1 = 1 ( t = 1 ( t t ( t = D t. q m, then the pacet Qen s equa to Qen can be restrcted by the foowng: t 1 t 1 ) 1 ) Bw B w, f ( deay ) ) Bw. 4. Smuaton Evauaton We smuated one ce case and ddn t consder nterference between ces n the smuaton. At ntazaton, K UEs are unformy dstrbuted n the ce under observaton, and ther movng drectons are randomy chosen. The speed of a UE s decded as unformy dstrbuted wth the mean of 3 m/h and standard devaton of 1 m/h. Every 5 seconds, a UE changes ts speed accordng to the same manner as at the ntazaton. At that tme, the UE ether updates ts movng drecton wth probabty of 0.3 or does not change ts movng drecton wth probabty of 0.7. If t decdes to change ts movng drecton, the drecton s chosen as unformy dstrbuted wthn [-π/2, π/2]. We use the foowng sgna path oss mode that s defned n 3GPP as the channe mode. L = 128.1 + 37.6 og 10 D, where D s the dstance (n m) between node B and a UE. We consder the foowng correaton mode for shadow C( d ) = e d n2 d cor S = C( d ) S * + 1 2 [ C( d )] N (0, σ ). where C(d) s a normazed autocorreaton functon of shadowng, dcor s the de-correaton ength, and d s the dstance that a UE has moved snce the ast cacuaton of shadowng. dcor was set to 5 m n the smuaton. The ognorma shadowng n ogarthmc scae s characterzed by a Gaussan dstrbuton wth zero mean and a standard devaton σ. The ognorma shadowng wth zero mean and a standard devaton of 8 db was consdered. S s the shadowng vaue represented n db and s updated usng the ast cacuated shadowng vaue S*. In smuaton, we update shadowng vaue every 5 sec. A four traffc casses were nputted nto the smuaton envronment: E-ma for best effort cass, web browsng for nteractve cass, vdeo for streamng cass and PCM voce for conventona cass. As for vdeo, we smuated both CBR vdeo and sef-smar VBR vdeo wth Pareto mode, and three dfferent bt rates were consdered: 56bps, 384bps and 1.5Mbps. We compared our DSDFQ scheme wth WFQ and FIFO (no QoS) schemes. Fg. 2 and Fg. 3 demonstrate the deay performance of those three schemes.. Frsty, our scheme can acheve much smaer deay than WFQ and FIFO for both VBR and CBR vdeos. In the case of CBR, our scheme can contro deay n 20 ms, but vdeo deay wth WFQ s amost 20 tmes arger than our scheme. In the case of VBR, our scheme reduces the deay n haf when compared wth WFQ. Secondy, our scheme can aso acheve good deay performance for other traffc casses when compared wth WFQ and FIFO. fadng. Fg.2: Deay Comparson for CBR vdeo and other traffc casses
Fg.3: Deay Comparson for VBR vdeo and other traffc casses 5.Concuson and Future Wor Ths paper proposes a new dynamc resource contro framewor to support mutpe traffc casses over the shared hgh-speed wreess pacet downn channe of next generaton wreess networ. In ths framewor, combned wth AMC and H-ARQ, the proposed DSDFQ scheduer aocates both spreadng codes and tme sots to dfferent UEs and traffc casses to meet the deay requrements from dfferent mutmeda appcatons. The proposed approach can aso be easy adapted to oad fuctuatons from dfferent traffc casses and dynamc wreess envronment. The future wor ncudes how to reduce sgnang oad between UEs and node B by decreasng spreadng code swtch tmes for each UE at the scheduer. 6. References [1] P. Bender, et a. CDMA/HDR: A bandwdth effcent hgh speed data servce for nomadc users, IEEE Commun. Mag., Vo. 38, No. 7, pp. 70 77, Juy 2000. [2] Motoroa and Noa, 3GPP2 1xTREME Presantaton C00-20000327-003, Mar. 2000. [3] Motoroa, Feasbty study of advanced technque for Hgh Speed Downn Pacet Access, TSG-R WG1 document, R1-556, Apr, 2000, Seou, Korea. [4]Y. Cao, V. L, Schedung Agorthms n Broad-Band Wreess Networs, IEEE PROCEEDINGS OF THE IEEE, p. 76~87, VOL. 89, NO. 1, JANUARY 2001. [5] C. Fragou, V. Svaraman, and M. Srvastava, Controed mu-tmeda wreess n sharng va enhanced cass-based queueng wth channe-state dependent pacet schedung, n Proc. IN-FOCOM 98, vo. 2, Mar. 1998, pp. 572 580. [6] S. Lu and V. Bharghavan, Far schedung n wreess pacet net-wors, IEEE/ACM Trans. Networng, vo. 7, no. 4, pp. 473 489, 1999. [7] T. S. Eugene Ng, I. Stoca, and H. Zhang, Pacet far queueng agorthms for wreess networs wth ocatondependent errors, n Proc. INFOCOM98, Mar. 1998, pp. 1103 1111. [8] P. Ramanathan and P. Agrawa, Adaptng pacet far queueng agorthms to wreess networs, n Proc. ACM MOBICOM 98, Oct. 1998. [9] J. Gomez, A. T. Campbe, and H. Morawa, The Havana frame-wor for supportng appcaton and channe dependent QoS n wreess networs, n Proc. ICNP 99, Nov. 1999, pp. 235 244. [10] L Wang, Yu-Kwong Kwo, Wng-Cheong Lau, and Vncent K. N. Lau, Channe Capacty Far Queueng n Wreess Networs: Issues and A New Agorthm, ICC 2002, Apr, 2002, New Yor, U.S.A. [11] M. Andrews, K. Kumaran, K. Ramanan, A. Stoyar, P. Whtng, and R. Vayaumar, Provdng quaty of servce over a shared wreess n, IEEE Communcatons Magazne, vo. 39, no. 2, pp. 150 154, Feb. 2001. [12] L. Xu, X. Shen, J. Mar, Dynamc bandwdth aocaton wth far schedung for WCDMA systems, IEEE Wreess Communcatons, Apr, 2002. [13] Hu Zhang, Servce Dscpne For Guaranteed Performance Servce n Pacet-Swtchng Networs, Proceedngs of IEEE, 83(10), Oct. 1995. [14] W. Jeon, D. Jeong, B. Km, Desgn of Pacet Transmsson Scheduer for Hgh Speed Downn Pacet Access Systems, Proc. Of the IEEE VTC 2002, Sprng.