for Multimedia Applications over All-Optical WDM Multi-Rings M. Ajmone Marsan, A. Bianco, E. Leonardi, A. Morabito, F. Neri

SR 3 : a Bandwidth-Reservation MAC Protocol for Multimedia Applications over All-Optical WDM Multi-Rings M. Ajmone Marsan, A. Bianco, E. Leonardi, A. Morabito, F. Neri Dipartimento di Elettronica { Politecnico di Torino { Italy E-mail: fajmone,bianco,leonardi,nerig@polito.it Abstract The paper describes SR 3 (Synchronous Round Robin with Reservations), a collision-free medium access control protocol for all-optical slotted packet networks based on WDM multi-channel ring topologies where nodes are equipped with one xed-wavelength receiver and one wavelength-tunable transmitter. SR 3 is derived from the SRR and MMR protocols previously proposed by the same authors for the same class of all-optical networks. SRR and MMR already achieve an ecient exploitation of the available bandwidth, while guaranteeing a throughput-fair access to each node. SR 3, in addition, allows nodes to reserve slots, thereby achieving a stronger control on access delays; it is thus well suited to meet tight delay requirements, as it is the case for multimedia applications. Simulation results show that SR 3 provides very good performance to guaranteed quality trac, but also brings signicant performance improvements for besteort trac. 1 Introduction In all-optical networks, data generated in the electronic domain by the source user are converted into the optical domain at the ingress UNI (User/Network Interface), and transported to the destination with no conversion into an electronic format before the egress UNI. Wavelength division multiplexing (WDM) [1] is normally used to partition the very large bandwidth available on the ber into a number of channels whose rates match the speeds of electronic interfaces. In this paper we consider all-optical networks where one logical channel is associated with each destination node. Logical channels are assumed to be obtained with a combination of WDM and space diversity, since in this case the feasibility of the proposed approach was veried [, 3]. Since one channel for each destination exists, nodes are equipped with one tunable This work was supported in part by the Italian National Research Council, and the Italian Ministry for University and Scientic Research. transmitter and one xed receiver. We focus on the particular case of ring topologies (more precisely on multi-channel rings), which have become an attractive solution for all-optical LANs and MANs, mainly thanks to the successful progress achieved in optical ampliers to compensate for insertion losses at intermediate nodes. With respect to other topologies (e.g., stars) rings ease the synchronization at extremely high data rates, permitting a slotted access, hence an ecient and exible use of the available optical bandwidth for packet communications. Providing one channel for each destination means that all that need to transmit to the same destination must share the corresponding channel. Medium Access Control (MAC) protocols are thus needed to arbitrate the access to the shared channels. This paper proposes a collision-free access protocol, which, in spite of its simplicity, allows an almost optimal exploitation of the available re while guaranteeing a fair access to all nodes. In addition, it permits a very ecient integration of guaranteed quality and best-eort services. For these reasons, it is well suited to support multimedia applications. Network Topology We consider ring networks providing one logical channel for transmissions to each destination. If we denote by M the number of nodes in the network, a total of M logical channels is available in the ring. All nodes that need to transmit to the same destination share the same logical channel. Fig. 1 depicts the logical network architecture for the case M =. The M logical channels run in parallel; they are slotted and synchronized, so that M slots (one for each destination node) simultaneously arrive at a node every slot time. Full connectivity is achieved by tuning the transmitter at the source node on a packet-by-packet basis. Fixed-length data packets are transmitted, and the slot size is such that one packet exactly ts into one slot. Since each channel

SR 3 : a Bandwidth-Reservation MAC Protocol for Multimedia Applications to the possible transmissions of M? upstream nodes. We assume nodes not to transmit to themselves. 3 In Fig. 1, node has the best access opportunity on channel 3 (leading to node 3), and the worst access opportunity on channel 1, where it must defer to transmissions of nodes and 3. 1 We say that the trac directed to node j from node i has lower access priority than trac directed to j 3 1 from nodes jj + 1j M ; jj + j M ; ; ji? j M ; ji? 1j M. In particular, we say that node i has access priority ji? jj M when transmitting to node j, 1 being the highest access priority, and M? 1 the lowest. When a packet is ready for transmission from node i to node j, the access must be delayed if the channel leading to j is already occupied with a transmission by one of the nodes with higher access priority with respect to Figure 1: The logical network topology when M = destination j. 3 Access Protocol The access protocol that we propose, called Synchronous Round Robin with Reservations (SR 3 ), can brings information to one destination, we assume that be subdivided in three hierarchical layers: the access the transmitted information is removed from the channel by the destination node: each logical ring is inter- strategy, the fairness control scheme, and the reservation scheme. rupted at a (dierent) node, and the multi-channel 3.1 Access strategy ring can be viewed as a set of staggered logical busses folded in a physical ring layout. The aim of the access strategy is to minimize the In order to avoid collisions (i.e., simultaneous accesses to the same slot) among packets directed to the number of contentions among packets whose transmission is attempted in the same slot to the same receiver. same destination, nodes exploit a channel inspection We consider access strategies with a priori packet capability on each logical channel. Thus the source selection: each transmitter selects the packet to be node, prior to transmitting a new packet with a given transmitted in a given slot without the knowledge of destination, inspects the activity on the channel associated with such destination. If an empty slot is ob- the channel status. This comes from the assumption that the selection of the packet whose transmission served, the packet can be transmitted, and it will reach is attempted in a given slot is normally performed in its destination without collision; otherwise, contention the electronic domain, hence at low speed, during the arises, and priority is given to in-transit packets: the previous slot interval, when the channel status is unknown. We further assume that only one packet can transmission attempt is delayed. The channel inspection operation can be implemented with the subcarrier be transmitted per slot time, since only one transmitter is available at each node. signalling technique (see []). It can be observed that, due to ring symmetries, The access strategy called SRR (Synchronous each node has a better-than-average access to the Round Robin), presented in [5], avoids collisions and channels leading to some, and a worsethan-average access to other channels, leading to other arbitrates contentions, permitting a good exploitation of the available bandwidth.. If we assume node numbers to be increasing in the transmission direction, the channel on In order to implement the SRR protocol, nodes must keep one separate packet queue for every possible destination. Since we assume that a node does which node i, i = ; 1; ; M? 1, has the best access chance is the one leading to node 1 ji? 1j M, since to not transmit to itself, M?1 FIFO queues exist at each access this channel, node i needs not defer to transmissions of any other node; the channel on which node i node. Each queue stores packets at a dierent priority in the channel access; such priority can be associated has the worst access chance is the one leading to node with the queue. ji+1j M, since, to access this channel, node i must defer With SRR, in an arbitrary time slot identied by the label s, node i schedules for transmission a packet 1 The notation j jm indicates the modulo M operator. with destination ji + k + 1j M, with k = jsj M?1. If the

INFOCOM 1997, Kobe Japan 3 corresponding queue is empty, the transmission of the ported in []. In this paper we only point out that rst packet from the longest queue is attempted in slot fairness in MMR is granted by the circulation on the s. If more then one longest queues exists (a tie for the ring of M control messages, named SAT's (standing longest queue occurs) the lowest priority among the for SATised). Each node has a transmission quota longest queues is selected. for each channel, which is restored when a SAT is received (i.e., only a given amount of packets can be In any case, if transmission in slot s is not possible because it would generate a collision, a new packet is transmitted between two SAT visits). Each SAT regulates the access on one channel: the j-sat restores selected in the following slot (labeled s + 1). We call SRR frame one SRR scheduling period, i.e., the quotas for channel j. the time between two slots in which the same queue Each node forwards SAT messages to the upstream at one node has a preferential access. The duration of node, thus SAT messages logically rotate in the opposite direction with respect to data (although the the SRR frame is M? 1 slots. The selection of the next packet to be transmitted requires the information about the lengths of all physical propagation is obviously co-directional). queues, as well as a common knowledge at all nodes about the slot label s, i.e., a global (network-wide) synchronization on the slot sequence. Moreover, the ring latency must be equalized in order to be a multiple of the SRR frame duration. SRR tends to a Time Division Multiple Access (TDMA) scheme under high load, but allows a dynamic partitioning (i.e., statistical multiplexing) of the channel bandwidth among transmitting nodes under asymmetric trac, and an unregulated access to the channel under light trac. SRR can cope, with minor modications, also with several trac classes at dierent user priorities (Upriorities ). In this case, nodes must be equipped with one queue for each trac class for each destination. In an arbitrary time slot identied by the label s, node i schedules for transmission a packet from the highest-u-priority non-empty queue with destination d = ji + k + 1j M, where k is dened as above. If all queues associated with destination d are empty, the transmission of the rst packet from the longest among non-empty highest-u-priority queues is attempted in slot s. If more than one non-empty highest-u-priority queue exists, the queue at lowest priority with respect to the access to the ring (i.e., packets directed to the closest destination) is selected. 3. Fairness scheme SRR can lead to starvation of low access priority nodes in overloaded channels under unbalanced traf- c patterns. We thus use on top of SRR a fairness control algorithm called MMR (Multi-MetaRing) [5, ], derived from the Metaring fairness control algorithm proposal for high speed metropolitan area networks [7]. A complete description of the MMR scheme is re- We use the term \U-priority" to make clear the dierence with the access priorities due to the considered network topology. 3.3 Reservation scheme In the considered network packet access delays (i.e., the time elapsing from the instant when a packet reaches the head of its queue until transmission starts) experienced under non-uniform high-load conditions can be very large (see Fig. in Sect. 5) when the SRR+MMR protocols are used. In order to obtain a stronger control on access delays, we rely on slot reservation. We propose a reservation scheme to be used in conjunction with SRR+MMR, which only requires a marginal algorithmic complexity increase, with no additional signaling messages. A slot reserved by node i on the channel leading to destination j cannot be used by any higher access priority node (but can be eectively used by nodes at a lower priority, i.e., downstream on the logical bus leading to the intended destination). Nodes have therefore a guaranteed access to slots that were reserved. Only slots to which node i has preferential access according to the SRR frame can be reserved on the channel leading to destination j, i.e., slots in position jj? ij M inside the SRR frame. Time is subdivided into successive periods called reservation frames. Each reservation frame comprises P SRR frames. Each node can reserve up to P slots with a given destination per reservation frame, i.e., at most one slot per destination per SRR frame. Reservations are eective if all network nodes are aware of other nodes' reservations. We use SAT messages to broadcast the reservation information: each SAT distributes information regarding current reservations on the channel it regulates. We assume that each SAT contains a Reservation Field (SAT-RF), which is subdivided in M? 1 sub- elds; each sub-eld is assigned to a particular node for reservations. If node i needs to reserve h slots per reservation frame on channel j, it waits until it receives the j-sat; it then forwards the reservation request by properly

SR 3 : a Bandwidth-Reservation MAC Protocol for Multimedia Applications setting the i-th SAT-RF subeld to the value h. The j-sat visits all nodes during the next tour of the multi-ring. By the time node i receives again the j-sat, all nodes in the network are aware of the request of node i. Node i can thus update its reservation request on channel j every time it releases the j-sat. According to our reservation scheme, each SAT must carry channel reservation requests for every node in its SAT-RF. However, this does not limit the network scalability. Indeed, only the number of reserved slots, i.e., an integer h P, per reservation frame must be notied by a node to other nodes: the position of the h reserved slots in the reservation frame is automatically established by running the SRR access strategy in every node (remember that a global synchronization on the slot sequence is assumed). Reservation Strategies The reservation scheme can be very eective in controlling queuing delays (i.e., the time from packet generation until packet transmission). We will see that this performance improvement is benecial to besteort datagram trac transmitted on low priority channels in non-uniform trac scenarios, and that it is even more eective when guaranteed-quality services are oered, or multiple trac classes are to be handled. In both cases, nodes must be able to evaluate whether they need to reserve slots: algorithms to determine the number of slots to be reserved must be provided. A very aggressive reservation policy, where nodes tend to reserve all reservable slots, would lead to performance limitations similar to those observed for the TDMA scheme. A very conservative reservation policy, instead, would leave us with delay and contention problems exhibited by SRR..1 Reservation strategy for guaranteedquality trac Applications with Quality of Service (QoS) requirements usually exploit connection-oriented protocols. At call set-up, the desired performance and the connection characteristics are specied. A Call Admission Control (CAC) algorithm can be introduced, using concepts similar to those developed in the ATM community (see, for example, [, Sect..3]); this would control the amount of statistical multiplexing in the network. Since SR 3 allows nodes to reserve only preferential slots, no interference is observed among slots reserved by dierent nodes on the same channel; each node can thus implement a call admission control policy based on local information, without any need for coordination with other nodes. It is outside the aims of this paper to dene an optimized bandwidth allocation strategy; we will show in Sect. 5 that by reserving slots it is possible to have a really strong control of access and queuing delays. The nice characteristic of the proposed network is that the bandwidth allocation and connection establishment procedures can be naturally introduced in a system whose \native" mode of operation is connectionless. No hybrid solutions (i.e., dierent algorithms and access protocols) nor partitionings of the available re between connection-oriented and datagram trac are necessary. Packets belonging to guaranteed QoS streams do not necessarily have to be transmitted in reserved slots. Moreover, reserved slots not used by the node that made the reservation can be \reused" for transmission by downstream nodes. As a consequence, in multi-class trac scenarios, network performances are not degraded because of too aggressive allocation policies. Furthermore, connections with Variable Bit Rate (VBR) trac, for which the average and peak bandwidth requirements are often not known, are more easily accommodated in our system. The amount of re allocated to connections can be easily varied during the lifetime of each connection, thereby providing good exibility.. Reservation strategy for datagram applications In the case of datagram (best-eort) trac, when a node observes a degraded access (i.e., large access delays) to a channel, it can improve the quality of its access on that channel by reserving slots. Nodes must however be able to evaluate the needed amount of reserved slots: if too many slots are reserved, they may go unused, leading to a degradation of the overall network performance. We use packet queuing delays as a quality index of the channel access. When a node receives a SAT, the queuing delay experienced by the packet at the head of the corresponding transmission queue is used to compute the number of slots to be reserved. We assume that nodes can increase, or decrease, at most by one 3 the number of reserved slots per reservation frame every time they update the reservation requests, i.e., every time they receive the SAT for that channel. Nodes use two thresholds in order to decide whether the number of reserved slots must be increased, de- 3 We could use only two bits to specify a unit reservation change in the SAT-RF subelds. We chose to load every time the current number h of reserved slots for reliability considerations: dierential systems are much more error prone.

HT = C M?1 (1? s) when < h < P, and HT = 1 when h = P. M?1 (1? s) is the expression for the expected TDMA delay with node load s, and C1, C, and LT P are constants, with C1 < C. When h = P, hence s = 1, LT cannot be computed as a function of the expected TDMA delay, since the latter grows to in- nity. Thus, the LT P constant was introduced, which can be computed similarly to LT for < h < P, with s very close to 1 (e.g., s = :9). By varying the values taken by constants C1, C and LT P, it is possible to make the allocation strategy more or less aggressive. It must be noted that an under-estimation of the locally oered load (i.e., of s ) leads to an under- INFOCOM 1997, Kobe Japan 5 creased, or left unchanged. estimation of thresholds values; similarly, an overestimation of the local load leads to an over-estimation Let us dene a High Threshold (HT) and a Low Threshold (LT), with (HT LT). Let D be the queuing delay for the packet at the head of the queue; if tem tends to compensate dierences between s and of the thresholds values. As a consequence, the sys- D < LT, the number of reserved slots is decreased by the real oered load. one; if LT D < HT, the number of reserved slots is left unchanged; nally, if D HT, the number of 5 Simulation Results reserved slots is increased by one. This section presents simulation results for a 1- Threshold values must be chosen so that nodes reserve only the number of slots necessary to improve slots (i.e., km at 1 Gb/s for slots of bits). Each node multi-ring. The length of the multi-ring is 3 their performance when experiencing problems in the node transmission queue can store up to packets, access to the channel, without needlessly subtracting and the MMR transmission quota K (see []) is equal slots to other nodes. to. The reservation frame is set to SRR frames. As a rst approximation, we can assume that a We have already shown in [5] that SRR+MMR node has a degraded access to a given channel if, on achieves an almost optimal exploitation of the available bandwidth, while guaranteeing a throughput-fair the average, it has less than one transmission opportunity (i.e., it gets less than one free slot) per SRR access to every node. In this paper we focus on frame on that channel. In this case, the queuing delay the benets achievable by superposing the reservation is expected to be greater than the one that the node scheme to SRR+MMR. would experience under a TDMA scheme. Each node In the simulation model, packets are assumed to could thus compare the experienced delay with the expected delay under TDMA, in order to decide whether ing to a Poisson process with xed rate P i at node i. always t into one slot, and to be generated accord- a degraded access to the channel is being experienced, The total trac in the network is = M?1 i= i. We and the reservation of extra slots is necessary. assume that nodes do not transmit to themselves. To compute the delay expected under TDMA, the We consider a simple client-server scenario: one amount of load oered at a node must be known. `server' (node ) is present in the network; the remaining M?1 `client' nodes direct half of their trac to the Nodes should therefore locally measure the trac offered to each channel. Local load measures entail a server, while the other half of the trac is uniformly non-negligible algorithmic complexity, and cannot be distributed among clients. The server generates an \cheaply" performed by nodes. amount of trac equal to half the total trac generated by clients, and evenly distributes it to clients. We therefore use the bandwidth allocation parameter s = h=p, equal to the ratio between the number 5.1 Queuing delays for guaranteed quality h of slots reserved per reservation frame and the maximum number P of reservable slots, as an indicator This subsection presents simulation results to assess trac of the local trac oered to the network, and we de- the performance of guaranteed quality trac (which ne threshold values as a function of s. In particular can be considered to belong to a trac class at higher we let LT = when h =, LT = C1 M?1 (1? s) when < h < P, and LT = LT P when h = P ; we let U-priority). We consider the following simple traf- c pattern. All nodes generate datagram (best-eort) trac according to the client-server scenario described above. In addition to this trac, the client node S l at the lowest priority on the channel leading to the server also transmits to the server a trac stream with a higher U-priority, which is meant to belong to a stream with QoS guarantee requirements. To study the eect of the reservation strategy on guaranteed trac, we assume that reservations are issued only by node S l. The packets at higher U-priority are assumed to always t into one slot, and to be generated according to a Poisson process with xed rate hp = :=(M? 1). The Poisson assumption for the stream with QoS requirements was introduced for simplicity, and can be considered unrealistic, since the generation of packets with QoS requirements is usually strongly correlated.

SR 3 : a Bandwidth-Reservation MAC Protocol for Multimedia Applications s = : s = : s = :3 s = :5 d q 11.9. 9.79 39.91 d a.7 5.93 1.3.13 q 7.1.59 5.59 31.9 a 3.1 55.9.91 5.99 s = : s = : s = :3 s = :5 d q 15.7 7.93 5. 3.7 d a 1. 35. 3. 3.73 q 1.3 7.5 9.9 9. a. 1.5 35. 5.1 Table 1: Queuing and access delays for dierent values of the bandwidth allocation parameter s in overload conditions ( = 3:3) We present simulation results for the average d and standard deviation of queueing (subscript q) and access delays (subscript a) experienced by node S l. The average and standard deviation of access delays provide meaningful information about the network behavior. In overload conditions the average access delay is equal to the reciprocal of the throughput, hence it tells whether the channel bandwidth is eciently and fairly shared by nodes. The standard deviation of the access delay provides an even more important information. Large standard deviations mean that certain packets are facing much larger access delays than other packets; this also implies that queuing delays are large. Table 1 refers to overload conditions. We can observe that, if no slot is reserved, node S l experiences large queuing delays on the channel leading to the server. S l does not experience losses of packets at the higher U-priority, because the MMR fairness scheme guarantees the bandwidth necessary for transmissions. However, the experienced standard deviation of the access delay is very large, since node S l gets free slots only when most of the nodes at higher priority on the server channel exhausted their quota. By reserving slots, node S l can improve its access on the server channel. Table 1 shows that the average, and even more the standard deviation, of the access delay can be very eectively controlled by reserving slots. As mentioned above, the standard deviation of the access delay is the key parameter, and it can be reduced by more than an order of magnitude with slot reservations. Table shows results when the load on the server channel is slightly smaller than 1 (i.e., = :93), i.e., just below channel saturation. Also in this case, node S l experiences large queuing delays on the channel leading to the server if it does not reserve any slot. The standard deviation of the access delay is still large, although appreciably smaller than in Table 1. By reserving slots, node S l can again signicantly reduce the average and the standard deviation of its access delay, obtaining a large improvement of queuing delay Table : Queuing and access delays for dierent values of the bandwidth allocation parameter s ( = :95) access delay 1 1 1 1 Figure : Average access delays for a client-server scenario with = :95 when s = : performance. The results presented in the tables show that, by reserving slots, nodes can quite eectively control their channel access, managing to almost completely decouple the performance of trac streams with QoS guarantee requirements from the amount of best-eort trac. 5. Queuing delays for best-eort trac In this subsection we show that slot reservation can improve the queuing delay performance of datagram (best-eort) applications as well. All results presented in this subsection refer to a client-server trac scenario in which the trac from and to the server is very close to one ( = :93). Figs., 3, and show, respectively, the average access delays, the standard deviations of the access delays, and the average queuing delays, when nodes do not reserve slots for any source/destination pair. Focus rst on the server performance. The server experiences an average access delay approximately equal to M? 1 = 15 slots for the client-bound channel at the highest priority, i.e., for the channel leading to node 15. Since the load is rather high (almost 1),

INFOCOM 1997, Kobe Japan 7 For transmissions to the server, the client at the highest priority experiences a very small average and standard deviation of the access delay, and a very small average queuing delay, since it always observes a free channel, and can transmit packets to the server as soon as they are generated. Instead, clients at lower priorities towards the server experience larger average access delays (e.g., the average access delay of the lowest priority client is very close to 15), and access delay standard deviations, because they can transmit only if higher priority clients have not transmitted on the server channel. Also queuing delays are rather large. 1 MMR does not inuence the system behavior (i.e., the MMR blocking eect is negligible), as already shown in [5], in non-overload conditions. A dierent behavior among client nodes is instead Figure 3: Standard deviations of the access delay for observed on client-bound channels. Clients with the a client-server scenario with = :95 when s = : higher priorities on the channel leading to the server exhibit a negligible access delay on client channels, while clients at lower priority when transmitting to the server face access delays slightly larger than (M? 1)= = 7:5 slots, i.e., slightly larger than half the SRR frame. The transmission queues storing packets directed to the server at higher priority clients are empty most of the time, since these clients experience a small queuing delay on the server channel. As a consequence, higher priority clients very often attempt transmission of packets directed to other clients immediately after the packet is generated. 1 On the contrary, the transmission queues towards the server at lower priority clients are almost always non-empty and quite long, because lower priority clients experience large queuing delays on the server Figure : Average queuing delays for a client-server channel. Thus, client-to-client packets are transmitted scenario with = :95 when s = : only when the corresponding transmission queues are selected according to the SRR scheduling algorithm. The queuing delays of client-to-client packets are almost identical (this cannot be seen from Fig. due the server executes an almost perfect round robin service of transmission queues. It may however happen to the very large vertical scale) to access delays, since that the server empties the queues at higher priorities, thus gaining some \extra slot" in which it can the load is very small. By reserving slots, it is possible to improve the performance of the server and of the clients at lower pri- attempt transmissions of lower priority packets. However, these extra slots are not enough to compensate ority on the server channel. Figs. 5,, 7, and, show, for slots \stolen" in a similar way by clients, so that respectively, the average queuing delays, the average the server still suers a slight bandwidth loss on the normalized numbers of reserved slots (i.e., the average values of lower priority channels (the experienced access delay is larger than M? 1 = 15). s ), the average access delays and the standard deviations of the access delays, when nodes The impact of this phenomenon on the queuing delays faced by the server (see Fig. ) is very strong: reserve slots according to the datagram reservation strategy described in Sect... The values :5, :, very large queuing delays, and non-null loss probabilities, are experienced on client-bound channels where and were respectively chosen for the constants C1, C, and LT the server is at low priority. P of the reservation algorithm. standard deviation of access delay queueing delay 1 1

SR 3 : a Bandwidth-Reservation MAC Protocol for Multimedia Applications 35 1 queueing delay 3 5 15 5 1 1 Figure 5: Average queuing delays for a datagram reservation strategy with C1 = :5 and C = : Similar queuing delays are experienced by the server on every client-bound channel; the values taken by these queuing delays are close to those experienced by the server on the highest priority channel when it does not reserve slots. No packet losses are observed on any client-bound channel. The server, indeed, reserves almost all reservable slots (see Fig. ) on every client-bound channel, so that no client can \steal" slots from it. Average access delays are, in this case, below the M? 1 = 15 value, and standard deviations are very small. Also clients at lower priorities on the server channel experience smaller queuing delays with respect to the case in which no reservations are allowed. Lower priority clients, indeed, by reserving slots can eectively control the standard deviation of the access delay on the server channel (see Fig. ), thereby improving the queuing delay performance. It must be noted that, while lower priority clients are not successful in reducing their average access delay on the server channel by slot reservation, they achieve a benecial reduction of the standard deviation of the access delay on the same channel. Note that the average queuing delays experienced by lower priority clients on the server channel are smaller than those achievable in a pure TDMA system (which are approximately 5 slots). The average queuing delays, and the average and standard deviations of the access delays, instead, increase with respect to TDMA because of the unavailability of slots reserved by lower priority nodes. Other simulation results not reported here, showed that the reservation strategy is not really sensitive to reserved slots.9..7..5..3..1 1 1 Figure : Normalized numbers of reserved slots for a datagram reservation strategy with C1 = :5 and C = : the values taken by its parameters C1, C, and LT P. As already mentioned, the behavior is self-adjusting, and the number of reserved slots rapidly adapts to the status of local queues. Conclusions The paper described SR 3, an access protocol designed for all-optical packet networks based on WDM multi-channel ring topologies devoting one logical channel for transmissions to each destination. SR 3 is derived from the protocols already presented in [5, ], which exhibit an ecient exploitation of the available bandwidth, and guarantee a throughput-fair access to each node. With respect to these protocols, SR 3 introduces a novel slot reservation scheme, which allows nodes to dynamically allocate a portion of the available bandwidth, and can thus provide quality of service guarantees for the support of multimedia applications. The interesting feature of the proposed protocol is that, although its \native" mode of operation is the provision of connectionless best-eort services, connection-oriented guaranteed services are naturally supported without aecting the basic access mechanisms. The bandwidth left unused by guaranteed services can be very eectively shared by best-eort traf- c, with a full exploitation of the advantages oered by statistical multiplexing. Even for the basic best-eort service, that requires no service guarantee, the reservation scheme can be very benecial: the averages and variances of access delays are greatly reduced when slots are reserved, leading to improved performance and fairness.

INFOCOM 1997, Kobe Japan 9 1 5 access delay 1 1 1 Figure 7: Average access delays for a datagram reservation strategy with C1 = :5 and C = : References [1] C.A. Brackett, \Dense Wavelength Division Multiplexing Networks: Principles and Applications", IEEE Journal on Selected Areas in Communications, vol., n., pp. 9-9, Aug. 199. [] I. Chlamtac, A. Fumagalli, L.G. Kazovsky, P.T. Poggiolini, \A Multi-Gbit/s WDM Optical Packet Network with Physical Ring Topology and Multi-Subcarrier Header Encoding", ECOC '93, Montreux, Switzerland, Sept. 1993. [3] M. Ajmone Marsan, A. Fumagalli, E. Leonardi, F. Neri, \R-Daisy: an All-Optical Packet Network," EUROPTO European Symposium on Advanced Networks and Services, Amsterdam, The Netherlands, March 1995. [] A. Budman et al., \Multigigabit optical packet switch for self-routing networks with subcarrier addressing", OFC'9, paper Tu, San Jose (CA), Febr. 199. [5] M. Ajmone Marsan, A. Bianco, E. Leonardi, M. Meo, F. Neri, \MAC Protocols and Fairness Control in WDM Multirings with Tunable Transmitters and Fixed Receivers", IEEE/OSA Journal of Lightwave Technology, vol., n., June 199 standard deviation of access delay 15 5 1 1 Figure : Standard deviations of the access delay for a datagram reservation strategy with C1 = :5 and C = : Transactions on Communications, vol. 1, n. 1, pp. 1-, Jan. 1993. [] R.O. Onvural, Asynchronous Transfer Mode Networks - Performance Issues, Artech House, 199 [] M. Ajmone Marsan, A. Bianco, E. Leonardi, F. Neri, S. Toniolo, \Metaring Fairness Control Schemes in All-Optical WDM Rings", IN- FOCOM '97, Kyoto, Japan, April 1997. [7] I. Cidon, Y. Ofek, \MetaRing { a Full-Duplex Ring with Fairness and Spatial Reuse", IEEE