Token Enabled Multple Access (TEMA) for Packet Transmsson n Hgh t Rate Wreless Local Area etworks Saed Akhavan Taher Elecrcal & Computer Eng. Dept., Unversty of ew Mexco Emal: saed@eece.unm.edu Anna Scaglone School of Elecrcal & Computer Eng., Cornell Unversty Emal: anna@ece.cornell.edu Abstract In ths paper we llustrate a novel Wreless Local Area etwork Multple Access Control (WLA MAC) protocol whose allocaton polcy adopts a strategy smlar to the token rng: the moble statons access the channel upon recevng the token and release the channel by passng the token to the next user. The traffc s dvded n two classes: Guaranteed andwdth (G) and est Effort (E). The base staton ndrectly controls the traffc access by allottng the tokens and the bandwdth s shared by the same users wth a decentralzed contenton-less protocol n case of the G users and on a contenton bass for the E users. The MAC protocol proposed the physcal layer adopts a Multcarrer Code Dvson Multple Access multplexng technque, whch mtgates the lnear dstortons of broadband transmsson and allows to dvde the frequency and tme resources n elementary unts that can be assocated to the token. The specfcaton on the average and peak delay and rate are guaranteed by lmtng the number of users passng the token and fxng a deadlne to release the token. At the same tme, smlar to a reservaton based protocol, the token strategy avods collsons and s able to guarantee effcently the ntegrty of the transmtted data. Index Terms Wreless etworks, Multple Access I. ITRODUCTIO I case of bursty data sources, multple access schemes that perform a statc channel assgnment, such as TDMA, FDMA, CDMA, SDMA, are known to be neffcent compared to contenton based multple access protocols. ALOHA-type random access protocols have been proposed to solve the problem of contenton based systems []. However, n hgh traffc condtons ALOHA-type (slotted and un-slotted) random access protocols yeld relatvely low throughput, due to the hgh collson probablty. The carrer Sense Multple Access Protocols (CSMA wth collson detecton, non-persstent, - persstent and p-persstent) offer a more effcent alternatve to ALOHA protocols. In CSMA the staton lstens to the carrer and when the channel s dle, transmts a packet []. In partcular, CSMA wth collson detecton (CD), whch s used as a standard Ethernet protocol, mnmzes the channel downtme. CSMA/CD yelds a hgh throughput when the propagaton delays (τ d ) are relatvely small n comparson to the packet duraton (τ d < T p ) []. An alternatve scheme s the so called Demand Assgned Multple Access (DAMA) where a separate request channel s used by the users to request a specfc rate. The request s processed by a central master staton or a common algorthm runnng n each termnal []. DAMA s a dynamc reservaton-based algorthm and t s effectve when the set of actve transmtters changes rapdly. Schedulng at packet level accordng to the demanded QoS have been consdered n several works: a comprehensve survey on packet schedulng n wreless networks s provded n [3]. In ths paper we propose a dstrbuted multple access control scheme, named Token Assgnment Multple Access (TEMA), desgned to handle effcently heterogenous classes of traffc n heavy traffc load. Our scheme manly s based on the pollng and the token rng concept. Pollng multple access protocols have been consdered for local communcaton networks for wred meda [3], [8], [9]. In these protocols the termnals are nterrogated n a cyclc order, by ether a central controller or, n a dstrbuted fashon, by usng specalzed pollng query/response (token) packets [4]. The statons are organzed nto a rng (they are, n fact, named token-rng networks and physcally they can be organzed n a tree shaped cable) and each staton knows the addresses of the statons at ts left and rght. When the logcal rng s ntalzed, the hghest numbered staton sends the frst frame. Subsequently, the staton passes the permsson to ts mmedate neghbor by sendng to the neghbor a specal control frame called a token. Token rng are one of the most relable LA n the ndustry [7]. However, wred token rng networks have the dsadvantage that a broken secton n the rng cable causes an outage of the whole network. Also, n comparson to the other MAC protocols, the token-passng MAC has control overhead [0]. To analyze the TEMA network performance we extend the prorty-based token rng to the wreless settng and borrow results from [7], [8], whch derve the performance of prorty pollng schemes desgned to provde sochronous and asynchronous type servces for wred meda. For the QoS specfcatons we provde a coarser separaton by dentfyng only two classes of servce, Guaranteed andwdth (G) and est Effort (E) classes. TEMA s ntalzed s wth a random access protocol: Accordng to the demanded QoS the base staton ntalzes the rngs, by allottng a subchannel (token) to a set of users (rng). The demanded contenton free part for hgh QoS users s constructed n a dstrbuted manner, by passng the token from user to user. Usng the rest of the resources through contenton based scheme we do not waste bandwdth. The rest of ths paper s organzed as follows. Secton I s dedcated to the descrpton of the basc TEMA dea and s followed by the descrpton of the specfc MC-CDMA physcal layer that we choose to support the TEMA archtecture. Secton IV shows the performance analyss and comparson between exstng multple access modes. II. TEMA: TOKE EALED MULTIPLE ACCESS Consder users U ordered by, =,..., wshng to communcate wth a base staton through a multpleaccess communcaton channel. Our dea s to logcally organze groups of users n subnetworks (nterconnected wreless 0-7803-7400-/0/$7.00 00 IEEE 93
token rngs), so that wth the pollng dscplne whch s not centralzed each user polls the consecutve user n the carousel (see fgure ). Each of the users n the rng are quered by an algorthm n a dstrbuted fashon. If the answer to ths query s postve, the channel s completely assgned to the user for data transmsson. Upon completon the transmsson or gvng negatve answer to the query, the control polls the next user n order. A staton s served untl ether ) ts buffer s empted, or ) a specfed tme slot to the user s over, whchever occurs frst. Each token corresponds to a physcal MC-CDMA subchannel whch s guaranteed to have a certan average rate and satsfy a probablty of error bound, as descrbed n detal n Secton III. "! $#&%' one rotaton of the hgh-prorty token n the tme left the low-prorty token wll rotate. C. on-homogenous traffc, hgh qualty of servce and bursty traffc. In ths scheme the token passng overhead s avoded for the bursty data by alternatng the users that pass the token wth a contenton based perod where all bursty data sources contend the bandwdth wth a CSMA strategy. Fgure shows the actvty of the channel for these three denote the servce tme for th user n kth recepton of token, be the length of the kth token cycle for user and be the so called walk tme, whch ncludes the token passng overhead. Smlar to [3], we defne Q (K) to be transmsson quota,.e., for each token cycle the th user schemes. Let message duraton s lmted, Q (K). From fg., denotng by Ŵ = we can derve the followng relatons between and for each of the proposed schemes: For scheme shown n Fg. A), = Ŵ + M j= j. () Fg.. Pollng wth cyclc vst orderng for users and one server. For scheme shown n Fg. ), To smplfy the scheme s analyss and havng analogy wth the exstng schemes for fber dstrbuted data nterface (FDDI) [5], only two dfferent classes of QoS are consdered: the guaranteed bandwdth (G) class, for whch the requrement s ether that P r(delay > D) < ɛ or P r(end to end packet loss) < ɛ for a gven ɛ, and the best effort (E) class, for whch there s no requrement. Several prorty-based token schemes for wred networks have been studes [7]. In partcular, one of our schemes (the second one presented below) shares smlartes wth the delayconstraned scheme n [7], where messages are dvded n hgh prorty and low prorty messages. y defnng two types of tokens, hgh level and low level, the system allows all users n the rng to send ther hgh level messages at frst and then low level messages. Ths scheme s not well talored for wreless communcatons, snce the type of load and bandwdth usage and probablty of error are dfferent. We propose three dfferent token-schemes applcable to the wreless meda for: A. Homogenous traffc; n ths scheme all users demandng a sngle type of servce wth hgh qualty (e.g., broadband multmeda). Hence, only one type of token (prorty) s consdered. To satsfy the tght delay requrements, a lmted dwell tme polcy s appled to ths scheme.. on-homogenous traffc; n ths scheme there are non-bursty to moderately-bursty users whch requre hgher and lower QoS respectvely. Ths scheme s appled when the users can have two dfferent type of demanded QoS. In the rng frst all hgh-prorty users are served whle the remanng resources are dedcated to the low prorty packets. In ths scheme, hghprorty and low-prorty tokens are employed. After = Ŵ + j= M Fnally, for scheme 3 shown n Fg. C), = Ŵ + j= F j= T (CSMA) j. () j. (3) Each token s not only assocated to a physcal set of subcar- -M M -M M Q Q Q -M-s -M-s+ Q Q -M-s+ -M M Q Q Q Q Q -M-s -M-s+ -M-s+ -M -M M Q Q -M-s Q A) G traffc. ) G traffc wth two dfferenent prortes -M T C) G and E traffc ----Random Access--- Fg.. Tmng dagram for dfferent token passng approaches; Guaranteed andwdth (G) and est Effort (E) traffc. 94
rers but also to a fxed frame duraton, the token cycle. The maxmum number of users supported n a rng and the token cycle depend on the QoS that the rng has to support and on the bandwdth lmtatons of the physcal layer. Wth reference to Fg. A, to guarantee P r(delay < D) < ɛ, we assgn users to be the prmary members of the rng, but ndeed only M of them are usng resources. M s a random varable ndcatng the number of users are slent n the rng. The prmary rng members are guaranteed to receve the token after at most the tme equal to +Q (K) +. In scheme 3, the M vacances goes for the M other members, that have no QoS guarantees, share the remanng tme usng random access. They can experence any delay or packet loss, but ths s complant wth the type of servce that they requre. In practce, all the tme whch s left over by the prmary users, s flled wth the E users data perodcally on a contenton bass. In fact, even f the number of rng prmary users reaches the maxmum number for whch t can be guaranteed that P r(delay < D) < ɛ) the G users wll not frequently use ther entre quota Q (K), unless they are Constant t Rate (CR) sources. As a result, the fracton of total tme of the token cycle n whch the G users are dle,.e.: (K) := (4) can be drectly utlzed by the E users. Due to the dstrbuted nature of the control mechansm, and because of the QoS and buffer lmtatons, t s mpossble to handle all user s demand wthn a sngle rng. The logcal opton s usng a mult-rng system, where several non-conflctng rngs organze the access to the resources. III. THE MC-CDMA PHYSICAL LAYER Although TEMA can be adapted to other physcal layer platforms, we wll see next that MC-CDMA offers greater flexblty n the assgnment of the transmsson resources. Multcarrer communcatons seem to offer a cost effectve soluton for the new generaton of broadband wreless communcatons [4]. MC-CDMA combnes the good propertes of the two emergng multplexng and modulaton technques for wdeband transmsson, namely: CDMA, chosen for 3G cellular communcatons (UMTS and IMT-000) and OFDM, whch adopted n the wreless LA standards IEEE 80. []and HIPERLA [8]. A good comparson between the dfferent multplexng alternatves and the advantages of the MC-CDMA approach can be found []. In III-A we derve the basc MC-CDMA desgn parameters that need to be used to determne the physcal resources assocated to one token. Our approach allows to construct tokens that delver the desred amount of bts/sec. satsfyng the error rate bound, despte the frequency selectve tmevaryng nature of the wreless medum, enhancng the robustness of the proposed strategy. A. The Token-Codes It s convenent to utlze a block transmsson technque and assocate to each token a set of sgnature codes that spread one block of Q data at a tme. To explot fully the flexblty of the Fg. 3. /M Ts PTs Partton of the resources n the MC-CDMA system. MC-CDMA strategy we dstrbute the spread data over the entre grd n Fg. 3. If P s the number of slots n one frame there are P M subcarrers per frame. To elmnate possble nterference between dfferent rngs and thus the need of power control, we partton the entre set of subcarrers T n Fg. 3 mappng each token onto a dstnct subset S,.e. such that: S T and Ths strategy s llustrated by Fg. correspond dfferent codes: Fg. 4. /M T s PTs Dfferent MC-CDMA Token-Codes. S { }. (5) 4, where dfferent colors Let us denote by F the K matrx of MC-CDMA codes and by s [k] the block of nformaton symbols transmtter at tme k over the th token. The superposton of the spread data s: x [k] = F s [k], (6) and the entres x [k] are transmtted on the set S contanng K subcarrers. From our dscusson t follows that, K P M, (7) and the tme necessary to transmt the symbols s at most the frame duraton T f : P (M + L) T f = P T s = (8) where, as mentoned before, s the total avalable bandwdth M s the number of subcarrers and L s the duraton of the cyclc prefx. L has to be greater then the channel delay spread and maxmum asynchronsm between users. Fxng M to be greater than the delay spread of the channel, L = M would sacrfce effcency but render block synchronzaton unnecessary. Assumng the recever carrer s synchronzed, the outputs of the FFT flters correspondng to the token-subcarrers are the entres of the followng K vector: y [k] = H F s [k] + v [k], (9) 95
where v s addtve whte gaussan nose (AWG) wth varance 0 and H s a dagonal wth dagonal entres {H } k,k = H (f k ) wth f k S. Assumng that the channel s known only at the recever sde and usng symbols that belong to the same constellaton of sze D, the token bt rate s (servce rate) µ = log (D )/T f bts/sec, (0) and f the target rate s µ µ then: D µ T f /. () The Qos specfcaton n terms of ER can be satsfed assumng that the ISI s approxmately Gaussan and usng the followng bounds on symbol the error probablty, where β s the sgnal to nose rato at the detecton level: P s () QAM < e 3 β (D ). () Thus, usng (), we can enforce the ER constrant by: P s () < P s β 3 ( µ T f / ) log Ps (3) To derve β we can assume that: ) the {F } k,q are..d. ± random varables; ) the channel coeffcent are known at the recever sde and the data are processed through a matched flter; 3) the spreadng factor K ; 4) the symbols are..d.. Under these assumptons, the sgnal to nose rato s: ( K ) P k= H (f k ) β = K 0 k= H (f k ) + P( ) K k= H (f k ), 4 (4) whch depends on the specfc channel parameters. However, modellng the channel as random and assumng that the frequences f k are suffcently spaced so that H (f k ) are approxmately uncorrelated, for K we can use the law of large numbers and have: β PK 0 + P( ) 0 κ H (5) where, wth E[ ] denotng the expectaton, κ H E[ H(f k ) 4 ] E[ H (f k ) ], depends on the fadng statstcal characterstcs only (for example, n Raylegh fadng κ H = ). Fnally, usng (5) n (3) we obtan a constrant that relates wth K that can be used to determne the necessary spreadng and block sze: K + P 0 κ H ( ) ( µ T f / ) log Ps. (6) P 3 0 IV. DELAY AALYSIS The token passng process can be nterpreted as a form of dstrbuted pollng where the token passng scheme s classfed as lmted servce system [6]. The traffc s generated by dentcal users whch are modelled by queues q through q. Each queue has a fnte buffer of sze, and ndependent Posson message arrvals at a rate λ message/second. The servce (transmsson) average rate of the token code µ s normalzed to one. The servce admnstrator serves staton only for a lmted tme Q (t) f there s any demand from staton. k messages out of those found at the poolng nstant are served contnuously at staton. A staton s served untl ether ) the buffer s empted, or ) a specfed number of messages or packets are served, whchever occurs frst. Then, the token n fnte reply nterval w goes to nspect staton +. We denote by x and by x respectvely the mean and the second order moment of the servce tme for the packets n the the queue. The utlzaton for user ρ and the system utlzaton ρ are: ρ λ x and ρ ρ ρ. (7) To obtan the average packet delay, we employ basc queung analyss methodologes, smlar to those used n [9], [] and extend them to the case where there are two classes of servce. Frst we calculate the average number of packets n each user s queue and then apply Lttle s law to fnd the average delay. Our proposed token-based MAC scheme can be modelled as a M/G// queue system wth vacatons (for detals see e.g.[5]) and the lmted number of watng places for each user s queue. To smplfy the analyss we set =. From the user s sde, the servce to the user contnues tll the buffer becomes empty or the quota tme reaches, whchever comes earler, then the server starts a vacaton. The remanng perod or server vacaton tme n the kth cycle s a random varable denoted by V (k) second moment v = C (k) T (k) wth the mean tme v and the. Let us denote by l(k) (t) the number of messages buffered at staton at tme t and defne the system s state at tme t to be L (k) (t) [ (t), l(k) (t),..., l(k) (t)]. For scheme A (see Fg. A)), n a rng where M users are actve wth constant walk tme w, the vacaton tme for staton to staton + can be expressed as: (t) = ( M)w + x = j=+ j (t) + x l (k+) j (t). (8) For scheme (see Fg. )), snce there are two classes of servces, there are two types of vacaton tmes. For the th staton they are respectvely:, (t) = ( M)w + Mw + x M + x j (t) + x l (k+) j=+ j (t) j (t), (9), (t) = ( M)w + x j (t) + Mw + x M j=+ j (t) + x l (k+) j (t). (0) 96
Smlarly, for the scheme C the vacaton tme s:, (t) = ( M)w + x j=+ j (t) + x l (k+) j (t) + RAP () where RAP stands for Random Access Part. The necessary condton for the rng stablty s that ρ <,.e. the average load arrvng n a cycle has to be less than the maxmum server attendance tme, thus: ρ(q max + v) < Q max, or ρ ] < Q max, () where Q max s the maxmum quota and C (k) s the cycle tme. Consderng the stablty crteron for the average cycle tme for scheme A we can wrte, ] = Mw + Solvng (3) we obtan: M = λ ] x. (3) ] = Mw ρ. (4) Smlarly for the schemes and C we have respectvely: ]=, ] + (M )w E[C(k), ] = + w,(5) ρ ρ Mw + RAP ]=. (6) ρ The Pollaczek-Knchne (P-K) formula [0] provdes the followng expresson for l o the average number of packets n the queue and at server: lo = λ x + (λ x ) ( ρ). (7) v v + The average remanng vacaton tme for prorty s v. Ths s applcable to all three schemes except for the traffc wth second prorty of the scheme C where there s not queung polcy. y Lttle s law the average number of arrvng packets for a user durng ths vacaton tme s, v lv = λ( + v ). (8) v The average total work ū done by the server for the user s ū = ū l + ū o + ū v, where ū l s the amount of work left behnd when vacaton starts, ū o s the amount of work at arbtrary tme for the M/G/ queue wth no vacaton, and u v s the amount of work durng the vacaton. We can also wrte ū = l x = ( l o + l v ) x. Followng [], ū for the users wth prorty s λ ū = x ( ρ ) + ū p ρ v + ρ ( + v ), (9) v v 97 where u p s the average amount of work at the pollng nstant. The approxmate value of u p can be derved solvng the followng non-lnear equaton []: ū p = ρ [ v + ū p ( e Qmax up )] + ūp e Qmax ūp (30) The average amount of work seen by Posson arrvals s the sum of l q x, ( l q average number of packets over tme n the queue) plus the resdual servce tme for the packet: ū = l q x + ρ x x (3) Snce ū = l x and snce from Lttle s law we have l = λ D the average delay s D = l λ = ū xλ. Hence, from (9), the average packet delay n the th queue wth prorty s s: D s = x ( ρ ) +ūp, v +( v s + v t ), (s, t) = (, ), (, ). ρ v s (3) REFERECES []. Abramson, Multple Access n Wreless Dgtal etworks, Proc. IEEE, vol. 8, pp. 360-369, sep. 994. [] A. aocch, F. Cuomo and S. olognes, IP QoS Delvery n roadband Wrelss Local Loop: MAC Protocol Defnton and Performance Evaluaton, IEEE JSAC, vol. 8, no. 9, Sept. 000. [3] Y. Cao and V. K. L, Schedulng Algorthms n road-and Wreless etworks, Proc. of the IEEE, vol. 89, no., 00. [4] J. Chuang, L. J. Cmn, G. Y. L,. Mcar,. Sollemberg, H. Zhao, L. Ln, M. Suzuk, Hgh Speed Wreless Data Access ased on Combng EDGE wth Wdeband OFDM, IEEE Comm. Mag., pp. 9-98, ov. 999. [5] M. Andrews, K. Kumaran, K. Ramanan, A. Stolyar, P. Whtng, and R. Vjayakumar, Provdng qualty of servce over a shared wreless lnk, IEEE IEEE Comm. Mag., Feb. 00. [6] C.H. Foh and M. Zukerman, CSMA wth Reservatons by Interruptons (CSMA/RI): A ovel Approach to REduce Collsons n CSMA/CD, IEEE JSAC, vol. 8, no. 9, 000. [7] J. Follows, Token Rng Solutons, IM Internatonal Techncal Support Organzaton, avalable at www.redbook.bm.com. [8] ETSI, Rado Equpment and Systems, HIgh Performance Rado Local Area etwork (HIPERLA) Type, ETS 300-65, Oct. 996. [9] L. Klenrock and M. Scholl, Packet Swtchng n Rado Channels: ew Conflct-Free Multple Access Scheme, IEEE Trans. on Comm., vol. Com 8, no.7, Jul. 980. [0] L. Klenrock, Performance Evaluaton of Dstrbuted Communcaton Systems, n Queung Theory and ts Applcatons, Eds, O.J. axma and R. Sysk, orth-holland, 988. [] IEEE Protocol 80.. Draft Standard for wreless LA: Medum Access Control (MAC) and Physcal layer (PHY) specfcatons, IEEE July 996. [] K.K. Leung and M. Esenberg, A Sngle Server Queue wth Vacatons and Gated Tme-Lmted Servce, IEEE Trans. on Comm., vol. 38, no. 9, 990. [3] J. W. M. Pang, F. A. Tobag and S. oyd, Generalzed Access Control Strateges for Integrated Servces Token Passng Systems, IEEE Trans. on Comm., vol. 4, no.8, 994. [4] I. Rubn and L.F.M. Demoraes, Message Delay Analyss for Pollng and Token Multple-Access Schemes for Local Communcaton etworks, IEEE JSAC, vol., no. 5, ov. 983. [5] I. Rubn and J.C. Wu, Analyss of an M/G// Queue wth Vacatons and Its Iteratve Applcaton on FDDI Tmed-Token Rngs, IEEE/ACM Trans. on etworkng, vol. 3, no. 6, Dec. 995. [6] H. Takag, Analyss of Pollng System, The MIT Press, 986. [7] Z. Tsa and I. Rubn, Performance of Token Schemes Supportng Delay- Constraned Prorty Traffc Streams, IEEE Trans. on Comm.vol. 38, no., 990. [8] S. Zhang and A. urns, An optmal synchronous bandwdth allocaton scheme for guaranteeng synchronous message deadlnes wth the tmedtoken MAC protocol, IEEE/ACM Trans. on etworkng, vol. 3, no. 6, 995.