Architectures for Resource Reservation Modules for Optical Burst Switching Core Nodes *

Transcription

1 4. ITG-Fachagung Phoonic Neworks, May , 2003, Leipzig, Germany Archiecures for Resource Reservaion Modules for Opical Burs Swiching Core Nodes * Sascha Junghans, Chrisoph M. Gauger Universiy of Sugar, Insiue of Communicaion Neworks and Compuer Engineering (IKR) Pfaffenwaldring 47, Sugar, Germany {junghans,gauger}@ikr.uni-sugar.de Absrac Opical Burs Swiching (OBS) is a promising candidae for a more dynamic opical nework layer. One cenral ask is he reservaion of resources in he core nodes. Reservaion modules in he oupu pors of OBS nodes will perform his reservaion ask. As hese modules should no limi he performance of he node, reservaion mus be performed in very shor ime inervals. This paper surveys reservaion schemes in OBS and discusses relevan scenarios wih respec o offse values. Then, wo archiecures for reservaion modules supporing he jus-enough-ime (JET) reservaion scheme are presened. Finally, heir implemenaion complexiy and heir scalabiliy are analyzed. 1 Inroducion Opical burs swiching seems o be a promising candidae o bring more dynamic ino he opical layer of ranspor neworks [2]. I can be regarded as an inermediae soluion on he evoluion pah owards all opical packe swiching neworks. As echnological limiaions do no allow swiching of individual IP packes, several packes wih he same roue hrough he nework are assembled ino burss and swiched hrough he nework all opically. For he conrol of daa burs swiching an ou-of-band signaling scheme is used. A conrol packe is sen hrough he nework and iniiaes he reservaion of opical resources for he corresponding burs. The conrol sysem of each oupu por of a node processes all conrol packes for he conneced link and reserves wavelenghs (WL) for he requesed burs ransmission inervals. For successful reservaions new conrol packes mus be generaed and sen o he nex node. The reservaion module has o find idle wavelenghs for he burss and, afer he assignmen of a burs o a wavelengh, he module has o ake care, ha no oher burs will be assigned o he same wavelengh a he same ime. In his paper we inroduce firs approaches for designing reservaion modules for OBS nodes and analyze heir implemenaion complexiy and scalabiliy. The remainder of his paper is srucured as follows: In Secion 2 OBS and differen reservaion schemes are inroduced. Secion 3 presens wo archiecures for reservaion modules which suppor he JET reservaion scheme and in Secion 4 resuls for he scalabiliy of he reservaion modules are shown. 2 Reservaion schemes in opical burss swiching 2.1 Inroducion of OBS As ransmission speed on opical links grows much faser han he elecrical processing speed, opical burs swiching represens an ideal soluion for closing he gap beween opical circui swiching and opical packe swiching. I allows a finer granulariy han circui swiching and leads o more flexibiliy in he opical domain bu does no require he high complexiy of he conrol par for opical packe swiching sysems. Two ypes of nodes can be found in an OBS nework: In he edge nodes, he incoming raffic from access or mero neworks is colleced, sored in elecrical buffers and assigned o larger daa unis, called burss. The core nodes serve as ransi nodes. They swich he incoming burss direcly in he opical domain o he appropriae oupu wihou elecrical buffering or complex daa processing. For he swiching of he daa burss, a conrol packe is generaed and sen separaely from is corresponding daa burs. This conrol packe mus be convered back ino he elecrical domain in each core node in order o exrac he informaion necessary for swiching he burs. A reservaion module in he corresponding oupu por of he node mus search a wavelengh which will be free during he ransmission ime of he burs. As he conrol packe is processed in each node, i propagaes slower hrough he nework han he burs iself. Therefore suiable precauions mus be aken. A common way in OBS scenarios is he inroducion of an offse ime. * This work was parly funded wihin he TransiNe projec under conrac No. 01AK020C and wihin he MuliTeraNe programme by he German Bundesminiserium für Bildung und Forschung.

2 The conrol packe is sen ou a cerain ime (he offse ime) before he daa burs iself. A core node uses he offse inerval for processing he conrol packe, seing up he swiching marix and generaing a new conrol packe o inform he nex nodes on he pah. A key concep of OBS is one-pass reservaion, i.e. a burs is no delayed unil an acknowledgemen of successful reservaion of an end-o-end pah is received bu is sen ou as soon as he burs is assembled. Several algorihms for he reservaion of wavelenghs were inroduced in lieraure [2, 6]. The JET (jus enough ime) algorihm [1] promises a good uilizaion of he links and enables he usage of qualiy of service (QoS) mechanisms by he inroducion of addiional QoS offses beween burs and conrol packe [3, 7]. In order for he conrol packe no o ge caugh by he burs, processing of burs reservaion and seing up he swiching marix has o be very fas. Alhough here is no unique definiion of OBS, he separaion of conrol informaion and daa in ime and space seems o be widely agreed. Also, he lengh of he burss is no limied, hus we assume variable lengh burss. Several approaches for offse sizes are inroduced in lieraure. A he edge nodes, a specific offse is deermined for he burss. On heir way hrough he nework, his offse can be reduced by core nodes. In he following secions we caegorize differen approaches in edge and core node and idenify scenarios for more deailed sudies abou he realizaion of he according reservaion modules in core nodes. 2.2 Deerminaion of offses a he edge When a fully assembled burs in an edge node is ready for ransmission, he respecive conrol packe mus be generaed and ransmied. The offse inerval beween conrol packe and daa burs can be deermined in hree ways: E1 No offse is used. Conrol packe and daa burs are ransmied in parallel on wo differen channels. In his case no addiional complexiy for delaying he daa burs a he edge or predicing burs properies for sending he conrol packe in advance would be necessary. Of course, he processing ime of he conrol packe in he core nodes mus be compensaed differenly by he core nodes, i. e. by adding a fibre delay line (FDL) for delaying he daa burss. E2 The offse could be proporional o he hop coun. For each core node, he burs will pass, he edge node adds an offse for compensaing he processing ime of he conrol packes in he core node. This offse for compensaion processing delays is called basic offse. E3 The suppor of Qualiy of Service (QoS) mechanisms can be provided by using differen offses for differen service classes. The earlier a reservaion is performed, he higher is he probabiliy of geing he desired resources. Thus, high prioriy burss are sen wih greaer offses han low prioriy burss. The offse for differeniaing service classes is called QoS-offse. The combinaion of he wo offse based mechanisms E2 and E3 is possible bu can easily lead o unexpeced ineracion and unfairness among burss in he same service class [7]. 2.3 Offse adapaion by core nodes The conrol packes arriving a a core node mus be convered ino he elecrical domain, forwarded o he corresponding processing modules and finally processed. The las processing sage mus generae a new conrol packe and send i o he nex node on he pah. All hese seps lead o a saisical delay of he conrol packe flow which is no experienced by he daa burss. The core nodes can influence he offse lengh in several ways: C1 The offse lengh can be kep consan. As he elecrical processing will always ake ime, he daa burs mus be delayed in his scenario. Fibre delay lines (FDL) can be applied o delay he opical signals for a consan period. Conrol packes mus be processed in a deerminisic ime in order o keep he offse consan. If processing akes variable ime wih an upper limi which corresponds o he FDL lengh, he new conrol packes mus be sen ou precisely so ha he original offse is reached. This scenario also copes wih an offse ime of zero, which means ha daa burs and conrol packe are sen a he same ime. C2 Offse reducion by a consan value. Like before, he processing ime of conrol packes mus be limied by an upper bound and he conrol packe is sen ou afer a fixed period of ime afer is arrival ime. C3 Offse reducion by a variable value. The core node sends he conrol packe o he nex node immediaely afer processing i. As processing ime can vary from packe o packe, he offses are decreased jus as much as necessary. 2.4 Offses in nework scenarios The scenarios for deerminaion of offse imes in edge nodes and he influence on he offse adapaion in core nodes described in he previous wo secions can be combined in several ways. Of course, no all combinaions are useful or even funcioning. Bu a large number of hem is

3 considered in lieraure and some more could be imagined. From he poin of view of reservaion modules for core nodes, he combinaions can be classified in hree groups, deermined by he propery of he offse ime disribuion. Unique, nework wide offse value If all offse imes should be kep consan in he nework, he core nodes mus no change he offse imes and need a possibiliy o delay he daa burss for he processing ime of he conrol packes (C1). All edge nodes send conrol packes wih a nework wide consan offse (E1), which can also be zero, owards he core nodes. Daa burss pass he node in he same order as he corresponding conrol packes. Each conrol packe arrival ime is deerminisically correlaed o he sar ime of burs ransmission. The reservaion module mus check he saus of he wavelenghs o his poin in ime. As all offses are he same, i is impossible ha burss have been already reserved and sar ransmission laer han his burs. This fac leads o quie simple soluions for he resource reservaion [4]. As QoS-classes canno be differeniaed by heir offses, oher mechanisms for QoS are needed. One approach wih few addiional complexiy is an admission conrol scheme o he reservaion process [8]. WL0 WL1 Fig. 1 Offse Offse Reservaion wih nework wide consan offses Few discree offse values in he nework The inroducion of discree offse values could be useful from several poins of view. I would be possible o add a specific offse O for each node o be passed in he nework. In his scenario, he edge node mus know he number of hops for each burs and calculae a basic offse ime for he burs. Anoher scenario for discree offse lengh disribuions could be he inroducion of offse-based QoS suppor. The longer he offse of a burs, he higher he probabiliy o ge a successful resource reservaion. In his case, he core node should no change he offse of passing burss bu compensae he processing delay. As he order of conrol packes does no represen he order of daa burss, he resource reservaion becomes more complex. No only a single sar ime of he burs ransmission bu he enire ransmission inerval mus be checked. Therefore, he reservaion module could deermine he saus of he wavelenghs a all possible offse imes inersecing wih he burs ransmission inerval (marked wih bold lines in Fig. 2). All poins on he ime line a which burss could sar are deermined by he acual sysem ime and he possible offses. The saus informaion a hese poins on he ime line could be precalculaed by he reservaion module for geing a shorer searching period. WL0 WL1 possible Offses Fig. 2 Reservaion wih discree offse disribuion Which scenarios would lead o discree burs lengh disribuions? The main premises are, ha edge nodes only generae offses as muliples of O (E1, E2, E3) and core nodes only change offses by muliples of O (C2). The generaion of differen offses for compensaing processing ime in core nodes and for supporing offse based QoS could only be used in parallel, when he QoS offses QoS are much larger han he offses for processing ime compensaion Comp. This would lead o a spreading of offse imes around muliples of QoS. This resuls o accumulaions of offse lenghs which could probably be used by a reservaion module by caching some precalculaed informaion abou he wavelengh saus. Bu he more core nodes are in he nework, he more he offses will be spread which resuls in a coninuous offse ime disribuion which is described nex. WL0 WL1 possible Offses Fig. 3 Comp O QoS Spreading of offse lengh disribuion

4 Coninuous offse values in he nework The coninuous offse ime disribuion is he mos general case. All scenarios in which core nodes reduce offse lenghs by variable amouns (C3) or scenarios wih generaion of variable offse lenghs in edge nodes need his flexibiliy. Of course, realizaion complexiy for a reservaion module for coninuos offse disribuion is he highes, oo. A reservaion module which can cope wih coninuously disribued offse imes can be used in all scenarios, even if heses scenarios do no need his generaliy. The original inroducion of OBS by Qiao [2] includes an offse based QoS suppor wih variable offses. The reservaion mechanism called jus-enough-ime (JET) reserves wavelenghs jus for he acual ransmission inerval and allows he use of gaps beween burss. 2.5 The JET reservaion scheme The reservaion algorihm JET is a very general algorihm. JET is a so-called RFD-algorihm (reserve-a-fixed-duraion). This means, ha he arrival ime of he burs a he node and he duraion of he ransmission is well known a he node. The wavelengh is reserved exacly for he ime, he burs really needs he ransmission link, bu no longer and no earlier (Aside from some guard imes a beginning and end of he burs for compensaing jier and swiching ime of he opical cross connec.). When a conrol packe arrives, he reservaion module searches for a free wavelengh for he burs in he desired ime inerval. As wavelengh converers are expensive componens, i would be desirable o assign he burs o he same wavelengh on he oupu link like on he inpu link. If he wavelengh is already busy, oher wavelenghs are checked for heir saus in he ime inerval of ransmission. If one or more wavelenghs are no occupied o he allocaed ime one of hem is seleced and reserved for he belonging burs. Special orders for selecing one among he free wavelenghs are no proposed by Qiao and no considered in his paper. Fig. 4 shows a reservaion procedure. A conrol packe for a burs arrives. WLx provides a gap in is ime line and he new burs fis in his gap. The wavelengh is reserved for he corresponding burs. WLx Fig. 4 Offse Reservaion wih JET 3 Realizaion of JET reservaion modules As JET is he mos flexible of he inroduced reservaion mechanisms, we decided o sudy he realizaion of suiable reservaion modules. Oher reservaion algorihms wih less complexiy could also use his reservaion module. As processing of he conrol packes has o be performed very fas, hardware based soluions are needed. Secion 4 will discuss he requiremens of processing ime and he performance of he proposed archiecures. In his secion, we inroduce wo possible archiecures for reservaion modules. One is based on he idea o represen he saus of all wavelenghs like a imeable. Each column in he able represens a specific small ime inerval and each row conains he saus of one wavelengh. This able is implemened as embedded memory in a Field Programmable Gae Array (FPGA). The second archiecure works as a combinaorial circui. For each burs, a specific eniy is used, which sores sar and end ime in regisers and uses comparaors o deermine, wheher a new burs would inersec wih he burs reservaion represened by his module. 3.1 Memory-based soluion of a JET reservaion module As explained in Secion 2.5 he JET reservaion scheme needs informaion abou sar and end ime of a burs. Even when we assume an unsloed media wih coninuous disribuion of burs arrivals, a real sysem will always work wih discree imes. For he realizaion of he reservaion module we inroduce an archiecure, which divides he coninuous ime scale ino discree slos. As we will show in Secion 4 a minimum resoluion in ime of wo percen of he mean burs lengh does no limi he performance of he sysem significanly. In he conex of he conrol packes he offse and lengh of he burs can be measured wih a high resoluion in ime, he allocaion of wavelenghs in he reservaion module is realized wih a sloed srucure wih lower resoluion in ime. The saus of wavelenghs is sored in a able. Each column represens a ime slo, each row represens one wavelengh of he sysem. Fig. 5 shows an example of a able for four wavelenghs. An empy square represens free slos, an in he square shows ha a slo is already reserved for a burs. Wih he arrival of a conrol packe he slos wih sar and end ime of he burs and all inermediae slos are idenified and free wavelenghs for he specified slos are searched (a). A free wavelengh will hen be seleced for he burs and marked as reserved (b). Afer he idenificaion of he new

5 wavelengh a new conrol packe for he nex node is generaed and a reques wih all swiching informaions for he opical cross connec is sen o is conroller. (a) WL 0 WL 1 WL 2 WL 3 New Burs (b) WL 0 WL 1 WL 2 WL 3 New Burs Fig. 5 As all conrol packes in he sysem mus be swiched o heir corresponding oupu por and be processed here one by one, hey mus be buffered on heir way hrough he sysem. As he conrol packes will experience a delay which is no consan bu varies in a relaively wide range, he offse informaion in he packes wouldn be exac anymore. A cenral sysem clock will provide a unique ime informaion for all sysem componens. A he arrival of a conrol packe a imesamp is assigned o i. The imesamp, he offse and he lengh informaion in he packe enable he sysem o compue absolue values for sar and end ime of he burs. These absolue values are used in he reservaion module for a unique addressing of he reservaion able. When a new conrol packe for he nex OBS node is generaed, he offse is recalculaed as difference beween absolue burs arrival ime and acual sysem ime. As reservaion requess will cover a finie ime inerval from he acual sysem ime o a maximum offse plus a maximum burs lengh he reservaion able needs o buffer only his ime inerval. As informaions abou he pas are no relevan for he reservaion anymore all enries which are older han he acual ime could be reused for enries in he fuure. So he reservaion map can be realized as a classical ring buffer where he acual sysem ime is he earlies poin in ime. We call he ime inerval from acual ime o he las possible enry in he able he reservaion horizon. Reservaion able and deerminaion of free wavelenghs Archiecure of he memory-based soluion Fig. 6 shows he basic archiecure of he memory-based version of he reservaion module. Fig. 6 Search Check incoming CTRL-Packes Inpu Buffer Memory Map Daa Buffer Mark Buffer o OC CTRL- Packes Mark nex node Memory-based reservaion module All incoming conrol packes are buffered in a FIFO-queue Inpu buffer and processed one afer anoher. The search module reads he memory conen of he respecive ime inerval and checks for free wavelenghs. In parallel, he memory daa is sored in a FIFO queue daa buffer. The read module selecs one of he free wavelenghs for he burss and forwards he daa o he mark buffer and o he generaor for new conrol packes. The daa buffer is needed, because normal memory devices do no allow wriing single bis. A read-modifywrie-cycle would be necessary for seing he bis in he memory. As he memory daa is already read by he search module, i buffers hese daa in he daa buffer. As he search module does only read he memory and he mark module does only wrie o i, hese modules can work in parallel if dual pored memory is used. The mark module reads he informaion of assigned burss from he mark buffer, combines hem wih he informaions from he daa buffer and wries hese daa ses back o he memory map.

6 The check module is necessary for daa consisency. Burss which are assigned o wavelenghs bu sill in he mark buffer mus be considered by he search module. The check module moniors he addresses of he search module and overrides he memory bis for he burss in he mark buffer. The reservaion of a burs is spli in wo phases: Searching of a wavelengh and marking of he reserved wavelengh in memory. Boh phases are performed in sequence. As hese phases are nearly independen, he reservaion module is realized as a wo sage pipeline. During he marking of one burs, free wavelenghs for oher burss can already be searched. As he burss occupy several slos in he memory map, boh sages need a processing ime proporional o he burs lengh, one clock cycle per slo. As plaform for he implemenaion of he reservaion module we used Field Programmable Gae Arrays (FPGA) of he APE 20KC family from Alera. They have large numbers of logic elemens for realizing regiser banks and sae machines and provide inernal memory blocks for he reservaion able. The memory blocks can be used in several operaing modes and can be combined ogeher for geing wider or deeper memory. Like all digial sysems he reservaion module works wih a cenral clock. Each calculaion or memory access is performed wihin one clock cycle of 10 ns which resuls in an operaion frequency of 100 MHz. The design needs abou 1800 logic elemens and 17 embedded memory blocks for a reservaion module which suppors 64 wavelenghs. These values correspond o 11% of logic elemens and 16% of memory blocks of an APE20KC400. The performance of his archiecure will be discussed in Secion Combinaorial Soluion The processing ime of a conrol packe for reserving a burs should be as shor as possible. The memory-based soluion from he previous secion needs a number of clock cycles proporional o he burs ime. In his secion a soluion wih a consan and deerminisic processing ime is inroduced. I is based on a large combinaorial nework which performs a lo of comparisons of ime informaions for deermining free wavelenghs and reserving one for he corresponding burs. All informaions abou he saus of wavelenghs are represened in a decenralized fashion in dedicaed logical burs reservaion eniies BRes. Each of hese eniies sores informaion for one reserved burs, i. e. sar and end ime of he burs ransmission. The archiecure of he reservaion module is shown in Fig. 7. WL ime sar end WL0 WL1 Fig. 7 BRes (0,0) BRes (1,0) res block res block BRes (0,1) BRes (1,1) block block res(1) res(0) Res- Mgr newwl sar end Archiecure of he combinaorial reservaion module In Fig. 7 wo wavelenghs are used and wo reservaions per wavelengh can be sored simulaneously. There are several dedicaed reservaion eniies BRes per wavelengh which are arranged in one row in he figure. As simulaion resuls show, he probabiliy of having more han eigh burss reserved for one wavelengh is very small even for high load. So we assume four o eigh eniies of BRes per wavelengh for he synhesis of he reservaion module. The daa of he new burs and he sysem ime are indicaed o all eniies. Each eniy carrying a valid reservaions compares is reservaion daa wih he daa of he newly indicaed burs. When he burss overlap, he eniy ses he block signal. All block signals of one wavelengh are logically combined. The reservaion manager ResMgr analyses he block signals of all wavelenghs and deermines one free wavelengh for soring he reservaion of he new burs. I ses he reserve signal for he deermined wavelengh. All eniies of a wavelengh are daisy chained. The reserve signal is forwarded if he reservaion eniy is already soring a valid reservaion iself. The firs free BRes eniy in he chain buffers he reservaion informaions. The acual sysem ime is provided o all reservaion enies. Wih his ime informaion he eniy checks, if he sored reservaion is sill valid or oudaed. In he laer case he informaion is deleed and a new reservaion can be sored in he eniy. An addiional combinaorial nework signals, if all eniies of a wavelengh are in use. In his case i is no possible o buffer an addiional burs, even when he wavelengh is no used during he ransmission inerval of he new burs. The complee procedure is performed in one clock cycle. Wih he following clock edge all reservaion informaion is sored and he nex comparison can be done. Wih his design i would be possible o reserve a burs wihin each clock cycle. Wih growing number of wavelenghs and BRes eniies per wavelengh he size of he combinaorial

7 Logic Elemens Delay / ns BpWL 8 BpWL 4 BPWL 8 BPWL Fig Wavelenghs Usage of Logic Elemens 0 Fig Wavelenghs Combinaorial Delays neworks increase and he criical pahs ge longer. So he maximum clock frequency of he reservaion module decreases wih increasing number of reservaion eniies Logic resources of he reservaion module The reservaion module was modeled and described in VHDL and synhesized for a Field Programmable Gae Array (FPGA). The APE 20KC family from Alera was seleced as arge plaform. The synhesis was made for differen numbers of wavelenghs and for four and eigh reservable burss per wavelengh. The ime informaion was represened as 20 bi words which cover a ime range of 1 ms wih a resoluion of 1 ns. Fig. 8 shows he number of used logic elemens in he APE device. The devices provide a maximum capaciy of logic elemens, which is represened by he dashed line in he graph. Fig. 9 shows he combinaorial delays which limi he maximum operaion frequency of he reservaion sysem. The reservaion akes several ens of nanoseconds. 4 Scalabiliy analysis 4.1 Model of he reservaion module In a reservaion module conrol packes for burss direced o all wavelenghs of his oupu fiber have o be processed. While several daa burss are ransmied on he fiber in parallel using WDM, all conrol packes are processed by a single module. The single server queue wih Poisson arrivals and generally disribued service ime (M/G/1) can be used as a simple model for analyzing performance and scalabiliy of reservaion modules. In case he waiing and processing ime of conrol packes exceeds he offse ime or he delay of he FDL he corresponding burs is los. In a well dimensioned sysem, burs loss probabiliy due o violaion of conrol packe ime consrains has o be much smaller han he loss probabiliy due o oupu conenion. Here, we assume ha i should be less han 1% of he burs loss probabiliy caused by conenion of he opical resources. While burs loss probabiliy can be calculaed according o he well-known Erlang B formula [6], he probabiliy for exceeding a cerain value of he waiing ime is described by he complemenary waiing ime disribuion. The arrival rae of he M/G/1 queue depends on he number of wavelenghs w, he load per wavelengh ρ λ and he mean burs lengh ET [ burs ]: λ res = ( wρ λ ) ( ET [ burs ]) In all evaluaions, burss arrive according o a Poisson process and burs lengh is negaive-exponenially disribued. 4.2 Memory-based soluion In case of he memory-based reservaion module, he ime for reservaion is proporional o he burs lengh and depends on he ime resoluion of he memory map. The resoluion r is defined as he number of slos which comprise an average burs. Fig. 10 shows he impac of differen resoluions of he memory map on burs loss probabiliy for load ρ λ = 0.5. I can be seen ha he raio of he blocking probabiliies for discree and coninuous ime operaion, i. e. operaion wih and wihou slos, decreases wih increasing resolu-

8 raio of blocking probabiliies of discree and coninuous ime Fig % penaly 8 WL 16 WL 32 WL resoluion r Impac of resoluion of he memory map ion. Depending on he number of wavelenghs, i can be seen ha a resoluion of slos per average burs is necessary in order o limi he penaly o approximaely 1%. In order o assess he scalabiliy of he memory-based soluion, he minimum offse ime per node is defined: For he minimum offse ime he probabiliy of a conrol packe o wai longer han his offse equals 1% of he burs loss probabiliy. As described above, for he memory-based soluion, reservaion ime is proporional o he burs lengh and calculaed as r ET [ r1 ] = f clock where f clock is he clock rae of he chip. In order o simplify he analysis, reservaion ime T r1 is assumed o also be negaive-exponenially disribued, i. e. he complemenary disribuion funcion of he waiing ime is calculaed for he M/M/1 queue. Fig. 11 shows he impac of he number of wavelenghs on he minimum offse ime for resoluion 50 and load ρ λ = 0.5. I can be seen ha he minimum offse increases wih increasing number of wavelenghs and decreasing mean burs lengh. The memory-based soluion only leads o minimum offses which are well below he mean burs duraion for a mean burs duraion in he range from 50 µs o 100 µs. In hese cases, up o 60 wavelenghs can be suppored. For a mean burs lengh of 10 µs, he memorybased soluion can manage a very small number of wavelenghs. minimum offse ime [s] Fig MHz 200 MHz number of wavelenghs 4.3 Combinaorial soluion E[T burs ] = 10µs E[T burs ] = 50µs E[T burs ] = 100µs Scalabiliy of he memory-based soluion From a performance poin of view, he combinaorial soluion differs from he memory-based soluion by having a reservaion ime which is consan and independen of he burs duraion. Therefore, he complemenary disribuion funcion of he waiing ime is calculaed for he M/D/1 queue wih a mean reservaion ime T r2 = 60 ns. [9] In Fig. 12, he probabiliy ha he waiing ime of a conrol packe exceeds 1 µs is ploed versus he number of wavelengh for differen values of load (righ). Curves represening 1% of he burs loss probabiliy are also included (lef). I can be seen ha almos independen of he load respecive curves inersec for a wavelengh coun greaer han 120. For a smaller wavelengh coun he blocking probabiliy due o conenion in he opical domain domi- 1% of burs loss probabiliy load 0.5 load 0.7 load number of wavelenghs Fig. 12 Scalabiliy of he combinaorial soluion probabiliy for conrol packe waiing ime > 1 µs

9 naes. Summarizing, his soluion provides very low probabiliies for violaing conrol packe ime consrains even for a large number of wavelenghs and high load. 5 Conclusion Two archiecures for reservaion modules have been proposed and evaluaed wih respec o realizaion, performance and scalabiliy. The archiecures were modeled and described in VHDL and synhesized for deerminaion of he needs of hardware resources. The memory-based soluion needs only few resources of oday s available FPGA devices bu is limied in he number of wavelenghs. Even for faser nex generaion FPGA families he number of suppored wavelenghs will remain raher limied. The usabiliy of his archiecure depends on he size of offse imes or he lengh of fibre delay lines a he inpus, respecively. The combinaorial soluion needs much higher hardware resources. However, as he processing ime for he reservaion of one burs is much faser, his archiecure scales much beer. Even wih oday s devices and wihou addiional opimizaion up o 100 wavelenghs can be realized. Wih nex generaion FPGAs allowing lower delays wih higher complexiy, he wavelengh coun will increase wihou any addiional opimizaion. In general, boh soluions could be used for higher numbers of wavelenghs if several insances of he modules are used in parallel. Several wavelengh bands could be defined and for each of hem, one reservaion module is used. Only if no reservaion in he same band is possible, he reques is forwarded o oher reservaion modules. This approach suis ideally o wavelengh converers wih limied conversion range. As he acual realizaion of he combinaorial soluion is a very large combinaorial ne, he segmenaion in several pipeline sages could increase he maximum frequency and performance of he sysem. Wih a higher hroughpu of conrol packes he maximum number of wavelenghs per oupu por is increased. From our poin of view, he use of JET in core nodes is possible and won limi he nework performance. So, in principle he JET reservaion proocol is no prohibiively complex. 6 Bibliography [1] M. YOO, C. QIAO: A new opical burs swiching proocol for supporing QoS. Proceedings of SPIE Phoonics Eas 98 Conference on All-Opical Neworking, Boson, Nov. 1998, pp [2] C. QIAO, M. YOO: Opical burs swiching (OBS) a new paradigm for an opical Inerne. Journal of High Speed Neworks, Vol. 8, No. 1, Jan. 1999, pp [3] M. YOO, C. QIAO, S. DIIT: QoS Performance of Opical Burs Swiching in IP-over-WDM neworks. IEEE Journal on Seleced Areas in Communicaions, Vol. 18, No. 10, Oc. 2000, pp [4] J. S. TURNER: Terabi burs swiching. Journal of High Speed Neworks, Vol. 8, No. 1, Jan. 1999, pp I. [5] BALDINE, G. ROUSKAS, H. PERROS, D. STEVENSON: JumpSar: A jus-in-ime signalling archiecure for WDM burs swiched neworks. IEEE Communicaions Magazine, Vol. 40, No. 2, Feb. 2002, pp [6] K. DOLZER, C. M. GAUGER, J. SPÄTH, S. BODAMER: Evaluaion of reservaion mechanisms for opical burs swiching. AEÜ Inernaional Journal of Elecronics and Communicaions, Vol. 55, No. 1, Jan [7] K. DOLZER, C. M. GAUGER: On burs assembly in opical burs swiching neworks a performance evaluaion of Jus-Enough-Time. Proceedings of he 17h Inernaional Teleraffic Congress (ITC 17), Salvador da Bahia, Brazil, Sep. 2001, pp [8] K. DOLZER: Assured Horizon A new combined framework for burs assembly and reservaion in opical burs swiched neworks. Proceedings of he European Conference on Neworks and Opical Communicaions (NOC 2002), Darmsad, June [9] H. AKIMARU, K. KAWASHIMA: Teleraffic Theory and Applicaions, Springer, London, 1999.