Using Explicit Reservations to Arbitrate Access to a Metropolitan ...

1576

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 23, NO. 4, APRIL 2005

Using Explicit Reservations to Arbitrate Access to a Metropolitan System of Slotted Interconnected Rings Combining TDMA and WDMA Charalambos Linardakis, Student Member, IEEE, Helen C. Leligou, Alexandros Stavdas, Member, IEEE, and John D. Angelopoulos, Member, IEEE

Abstract—A system of slotted interconnected wavelength-division-multiplexing (WDM) rings controlled by a medium access control (MAC) protocol is shown to offer very high utilization within a queuing delay less than a few round-trip times, by means of a very-fast-reacting explicit reservation mechanism. The system can be used to provide interconnectivity in a metropolitan area transferring optical payloads on-the-fly without buffering or converting from the optical domain. All necessary control information is transferred on a dedicated wavelength and is processed in the electrical domain to provide both collision-free packet access at the ring nodes and contention resolution at the hub interconnecting the rings. The end result is a flexible and efficient metropolitan network suitable for bursty data services. Index Terms—Access arbitration, medium access control (MAC) protocol, metropolitan rings, reservation-based MAC, wavelengthdivision multiplexing (WDM).

I. INTRODUCTION

P

HOTONICS took the transmission plant by storm becoming the dominant technology within a decade in the 1980s. An exception was the access part, but the final assault on this last bastion of copper is well under way with the passive optical network (PON) initiatives in the form of gigabit-per-second PON (GPON) of the full-service access network/International Telecommunications Union (FSAN/ITU) [1] and Ethernet PON (EPON) of IEEE [2]. In contrast, the penetration into the switching plant has proven much more elusive, leaving as the most prevalent approach the conversion of payload to the electrical domain just before switching, only to be reconverted back for optical transmission at the output interface. Obviously, the advantages of a possible direct switching of optical payloads are such that it seems unlikely that research efforts will wane before this objective has been achieved. The motivation becomes stronger as the advances in electrical switching do not seem to keep pace with traffic Manuscript received March 23, 2004; revised October 22, 2004. This work was supported in part by the European Union project IST-1999-11742 “DAVID.” C. Linardakis, H. C. Leligou, and J. D. Angelopoulos are with the Department of Electrical and Computer Engineering, National Technical University of Athens, 10682 Athens, Greece (e-mail: [email protected]; [email protected]). A. Stavdas was with the Department of Electrical and Computer Engineering, National Technical University of Athens, Athens, Greece. He is now with the University of Peloponnese, 22100 Tripolis, Greece. Digital Object Identifier 10.1109/JLT.2005.844198

volume growth, while new breakthroughs emerge every year in optical technology [3], [4]. Already slow optical switching is a practical reality, but this can only provide satisfactory performance for traffic with smooth temporal profiles and a degree of predictability, which is not the case in today’s data applications. Lacking the required switching speed for present and emerging services, as well as flexible ways to store optical packets in order to resolve contention for output ports, research turned to novel approaches such as optical burst switching (OBS) [5]. This approach allows preconfiguring the optical switching elements using electronic processing of control signals that carry all the necessary information for the switching of the optical payload bursts. The result is all optical payload switching at the expense of significant inefficiency and traffic loss compared with the electronic switches. However, for medium-sized networks where round-trip times are small enough to allow medium access control (MAC) by closed-loop reservations, optical packet switching can be affected with predictable performance and similar efficiency and loss probabilities as in electronic switches. Such a system was designed and developed for the metropolitan network part of the European Information Societies Technologies (IST) DAVID project [6]–[8]. The medium control of this network is presented and evaluated in this paper. Although the project dealt also with a wide-area network (WAN) part figuring novel optical switching experiments [6], in this paper, the scope is restricted to the metropolitan part of the network, which consisted of a set of slotted wavelength-division-multiplexed (WDM) rings interconnected through a hub which features an optical space switch without internal buffering. The DAVID network targets longer term solutions using packet-over-WDM platforms [6], [9], [10] with optical packet switching capabilities and control allowing data access/forwarding decisions to be made slot by slot, based on new dynamic medium control schemes. The result is a much more efficient metro network with far better performance for packet traffic than current synchronous digital hierarchy (SDH) systems. However, due to the still high cost of burst-mode components, which are currently available only as experimental and expensive devices, this performance improvement carries in the short term a heavy system cost penalty. Yet, the pace of developments in optical technology can soon cause the break-even point to be reached since in the metropolitan area, where access reservations are possible, the efficiency of the

0733-8724/$20.00 © 2005 IEEE

LINARDAKIS et al.: SLOTTED INTERCONNECTED RINGS COMBINING TDMA AND WDMA

Fig. 1.

1577

System conceptual overview.

proposed method provides a significant margin even when starting with higher capital costs. The remainder of the paper is organized as follows. The architecture of the metropolitan network is presented in Section II, while the design of the hub controller (which schedules traffic exchanges among rings) is described in Section III and evaluated by computer simulation in Section IV. Conclusions are drawn in Section V.

II. THE NETWORK ARCHITECTURE The architecture featuring optical packet switching that was designed and implemented in its basic components in the metro part of the DAVID project is shown in Fig. 1. Several alternatives were studied in the project [4], [6], [8], [11], but for the needs of the explicit reservation control protocol, which is the focus of this paper, the system model described below will be used. The network consists of a system of interconnected slotted WDM rings enough to cover a metropolitan area and collect traffic toward the gateway to the core network (WAN) but also able to deliver traffic from any node of any ring to any other node. The nodes, i.e., optical packet add–drop multiplexers (OPADMs) interface to edge routers/switches toward access networks via a variety of interfaces (e.g., Gigabit Ethernet in business areas, PONs in mixed or residential areas, cable head-ends, or other legacy systems). Each ring can own up to eight active nodes and can carry from eight to 32 wavelengths with a rate of 10 Gb/s or more in each wavelength (10 Gb/s was the rate used in the demonstrator [12]). Although in many areas, one ring may be enough, given that limitations imposed by the physical layer do not allow more than about eight nodes per ring, for large metropolitan areas, several

rings may be needed and up to eight can be readily accommodated. The medium is slotted with all slots in all wavelengths synchronized to form a data multislot, but each is accessed independently. To transfer the necessary control information to allow distributed slot access but also fairness, reservations, and is exall the other control functions, one of the wavelengths clusively devoted to control and is the only one that is translated to the electrical domain and processed at each intermediate node. It is assumed that each OPADM can receive and transmit in all wavelengths, i.e., the number of transceivers equals the number of wavelength channels in the ring. On this issue, several solutions were investigated in DAVID that trade performance for lower cost, devoting bands of wavelengths to groups of nodes creating logical rings, thus reducing the number of transceivers. The issue of logical rings will be revisited in more detail at the end of this section. The operation of the protocol is independent of the actual slot size, provided its duration is enough for the control information processing and the accomplishment of the relevant actions. When taking into account the packet size distributions, the impact of the slot size can be significant, but no established traffic model for the future services at the time of potential deployment of such a metro network can be taken for granted. In the DAVID demonstrator, a slot size of 1 s (i.e., 10 000 b at 10 Gb/s) was actually used, but since the slot is the unit of time in all evaluations, the actual size becomes irrelevant, and the results hold for any other slot size deemed more appropriate in the future. As depicted in Fig. 1, control slots have equal duration and are time-locked to the data multislots preceding them by a fixed 1 s (one time slot) to allow ample time for offset time processing the contents of the control slot. To relieve speed of the processing logic in the demo implementation, the control

1578

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 23, NO. 4, APRIL 2005

Fig. 2. Control slot format.

channel operated at 2.5 Gb/s since the resulting 2500 b of information were quite enough even for 32 wavelengths. The detailed format of the control slot is shown in Fig. 2. contains information for all wavelengths of Each slot at the multislots transported in the payload wavelengths , etc. Two regions of information can be discerned in the figure. The wavelength information region carries information about the data multislots while the rest of the fields enable the ring-toring fairness, indicate the destination ring of the multislot (i.e., to which ring it will be switched at the hub), and carry the queue report for the reservation function. The role of each field will become apparent below as each relevant function is discussed later. No local MAC address is used inside the payload. Each node that has data to transmit must search for an empty slot destined for the desired ring in any wavelength by inspecting the corbit in Fig. 2) responding control slot, where a status bit ( indicates its status (“0 empty” or for every slot in every “1 occupied”). The node can grab the slot by setting the relevant bit in the control slot to “1” and inserting the destination address DA and the source address SA in the relevant fields. Nodes can only write in empty slots, i.e., traffic in-transit has precedence over traffic queued in the node. The nodes also monitor the control slots in search of instances of their address in the destination address (DA) fields of the control slot so as to receive the data and remove the payload (slot reuse is employed). The source address field (SA) is used for the reassembly process at the end node. Thus, access in each ring is governed by this simple “empty slot access” protocol. A well-known inherent problem in ring networks with slot reuse is unfairness since the nodes after heavy sinks are favored finding more empty slots. This problem, which has received significant attention, is solved by special mechanisms, such as the SAT (satisfied) [13], which makes sure that an equal number of slots have been inserted by each node on every round of the SAT signal. In addition, in our multi-ring system, fairness must be guaranteed toward each of the rings separately [14]; therefore, the mechanism has been extended to include one such SAT rings, SATs are used in signal for each ring pair, so for the system using the relevant fields shown in Fig. 2. (For more details on the fairness issue using the SAT mechanism, see [9], [13].) The rings have one common point where they can exchange traffic: the hub, which contains an optical space switch operating

on a slot-by-slot basis, and can pass any slot coming from any ring to any other ring. To avoid electronic buffering, the output ring to which each multislot will be switched, when it arrives back at the hub (one round-trip time later), is marked by the hub in the “destination ring” field of the corresponding control slot. The hub determines the most suitable destination for each multislot by running a scheduling algorithm based on collective reservations, which will be described in the next section. In addition, the hub connects via a dedicated ring (or two if dimensioning requires more capacity toward the core) to the gateway that sends/receives traffic to/from the WAN [4], [6], [7]. The connection to the gateway is, for the hub, similar to any ring although all the capacity goes to just one node, which acts as source/sink for all traffic. The DAVID project is considering a combination of optical and electrical buffers for the gateway, but these are outside the scope of the present paper, which focuses on the bufferless metropolitan system mechanisms. The switching at the hub is controlled by a centralized controller, which creates compatible combinations of input/output ring pairs. The objective is to create those combinations that reflect real bandwidth needs among rings, and to achieve this, reservations will be employed. Since each ring can appear only once as an input or output in each slot, the output of the controller is a permutation of the set of existing rings, matched against another permutation of the same set. Fixing the input permutation to an ascending order (i.e., ring 1, 2, 3, etc.), the mandate of the hub controller is to calculate a suitable permutation for the outputs that reflects the dynamic fluctuations of traffic among the rings in an efficient, fair and quick manner. To this end, explicit reservations made by the nodes of each ring are collected a round trip ahead of the moment they are needed. These reservations are based on the length (measured in slots) of the queue destined for each ring (queuing at the nodes takes place per destination ring) and are reported in the control slot (see Fig. 2). The detailed operation is described in the next section. The overall system operates as a distributed wavelength-division multiple-access/ time-division multiple-access (WDMA/TDMA) multiplexer achieving statistical multiplexing in the optical layer, exploiting both the time and wavelength dimensions. The two control mechanisms, i.e., the hub ring switching mechanism and the distributed empty slot MAC described previously (and in [15], [16], and [17]), operate independently but in concert to create a flexible, fair, and

LINARDAKIS et al.: SLOTTED INTERCONNECTED RINGS COMBINING TDMA AND WDMA

efficient overall metropolitan system. Once a payload waiting in a node queue is converted to optical form and inserted into one ring, it will travel with zero collision probability to the final destination node via the hub (unless destined for a node on the same ring before the hub). Variable-size packets are accommodated by use of a train of slots, not necessarily concatenated. Consequently, traffic has to be segmented and/or assembled to fit into the fixed-size slots. Thus, while resort to a fixed slot greatly simplifies the control offering higher utilization, than if attempting to handle variable-length packets in all rings simultaneously, some of the resulting efficiency is sacrificed to the process of creating the fixed slots, since inevitably some slots are not filled in full. Before leaving the architectural issues, it is worth elaborating on the issue of logical rings mentioned in the beginning of this section. To reduce the number of required transceivers and provide a flexible migration path, one of the studied alternatives in DAVID envisaged the separation of wavelengths into bands with nodes receiving in one band only and not in all wavelengths. This can be seen as creating logical rings with some nodes of a physical ring participating in one logical ring and some in other logical rings. For example, a physical ring with eight nodes and two bands of four wavelengths can be considered as two logical rings with, say, three nodes participating in one band and the other five to the other. To send from one logical ring to another, traffic still has to go through the hub to be switched by wavelength conversion this time and not space switching (unless they lie on different physical rings needing both) to the correct logical ring. This can extend from having just one wavelength per logical ring (only one fixed transceiver) to having all wavelengths, thus coinciding with the physical ring. (For more details on the hub switching, see [7]; however, what matters in the scope of this paper is that it makes no other difference on the design of the control algorithms of the system except that the controller should drive the proper converters for switching logical rings and the space switch for switching physical rings.) The control algorithms are, however, the same, so in the remainder of the paper, we will only be concerned with the logical abstraction of rings, regardless to what they correspond in each implementation instance of the system.

III. PERMUTATION SCHEDULING ON THE BASIS OF RESERVATIONS Before elaborating on the scheduling mechanisms, the basic assumptions for the operation of the hub controller should be presented. 1) All rings have equal length/delay and contain the same number of fixed size slots. (This is accomplished by adding fiber delay lines to shorter rings to equalize delay and storage capacity). 2) The timing of the multislots in each ring is controlled centrally by the hub so that slots coincide completely when they arrive at the hub allowing switching among them without buffering. 3) All slots of a data multislot in each logical ring are switched together to a new destination ring.

1579

To create an aggregate report of all traffic waiting in all the nodes of each ring (per destination ring), each node adds to the relevant queue report fields of Fig. 2, the length of its own relevant queue, expressed in slots. This implicitly places a reservation for an equal number of slots (per each destination ring). The report fields are cleared every time they exit the hub, and when they arrive back after a full round, they contain the aggregate of all slots waiting at the queues of the traversed ring and destined for each relevant ring. It is reminded that the “destination ring” field in the control slot announces the destination of the current multislot so that nodes know where to insert payload guaranteed to reach the correct destination ring and have been decided one round earlier by the hub. Once switched to that ring at the next pass through the hub, the payload will be received by the appropriate destination node in the way described previously for the operation of each single ring i.e., on the basis of the DA fields. To be able to know the future destinations means that the relevant permutation, which will be executed when the multislot will arrive at the hub, must have been already prepared at the launch of the multislot from the hub and kept in a queue of precalculated permutations, which are executed one round trip after they are prepared. Attention will be turned next on how the permutations are calculated. The permutation scheduling is conceptually illustrated at the bottom of Fig. 3, which shows successive permutations for rings 1, 2, 3, and 4, which, e.g., for slot are 2, 3, 1, and 4, i.e., ring 1 goes to ring 2, 2 to 3, and so on. Focusing on ring 1, are destined for ring 2, allowing the payloads of slots and for inter-ring communication between ring 1 and 2, while the and are destined for ring 1 (serving payloads of slots intra-ring traffic), i.e., in these two slot times the nodes of ring 1 are acting as single ring nodes. The next instance is again for inter-ring traffic toward ring 3. As mentioned, the scheduling is based on the reports/reservations so that they can follow traffic fluctuations by repeating more often combinations of ring pairs in high demand against those in less demand. The arriving reservations are organized in the hub in a matrix called the reservation matrix, where each element represents the number of optical slots that want to move from ring to ring . The matrix is updated continuously with the arriving reports of each incoming control slot from all rings. Since the field carries aggregate queue length, the old value is overwritten with the updated one, providing robustness to the system, since any corruption in the reservation value would only cause a transient insignificant disturbance until it is overwritten by a new correct value. Having available the information of the reservation matrix, reflecting the queue status in the distributed extensive metro system, a situation quite similar to any input buffered centralized switch arises. In other words, the system of the interconnected rings can be viewed as a very large distributed, input buffered switch with virtual output queuing. All queue information is available at the controller of the switch, albeit with a delay due to the propagation time resulting from the geographical separation of the queues. Because of this inevitable delay, the status is not really accurate since some new arrivals and departures have already occurred, but the available information in

1580

Fig. 3.

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 23, NO. 4, APRIL 2005

Scheduling of permutations.

the reservation matrix is as best as can be had. Our scheduling problem then can borrow from a broad literature of schedulers for virtual output queuing in input buffered switches [18]. Of interest is the relevant discussion in [16] for the same architecture although the control of that system does not resort to explicit reservations. However, it is worth emphasizing that in this case, because of the distributed nature of the system, the performance is much more dependent on the mechanisms for fast acquisition of queue status information (a trivial problem in a normal switch where all queues are local and all information is instantly available to the scheduler). In other words, the merit and novelty of the proposed control system lies more in the information collection method of explicit reservations, than in the permutation scheduling algorithm where many alternatives could be adapted trading performance for complexity. Optimality was not pursued because, even if it could be achieved at reasonable cost, it should be considered as just an “apparent optimality” resting on the assumption that the reservations that are stored in the matrix reflect the current queuing situation at the nodes, which is not really the case due to the distributed nature of the system. For reasons of implementation simplicity and to demonstrate thevalueofthereservations,thesimplegreedyheuristicalgorithm described hereafter was used. As it will be shown in the performance evaluation, this simple scheduling algorithm can lead to an efficient system, exhibiting fair and guaranteed performance in terms of queuing delay and throughput due to the elaborate collection of arrival information via the explicit reservations. The adopted scheduling algorithm works as follows. The current instance of the matrix is inspected row by row, and for each row (i.e., input ring) it is assigned as output the ring for which is assigned to most reservations exist, i.e., the output ring , exinput ring if cluding in each step the already assigned rings. The row inspection sequence is different for every slot and chosen randomly possible infrom a uniform probability distribution of the spection combinations.

Therefore, using the example shown in Fig. 3 depicting a system of four rings and the following matrix

assuming that the row inspection sequence is 1, 2, 3, 4 for a specific time slot , the resulting ring permutation would be 2, 3, 1, 4 , i.e., ring 1 to ring 2, 2 to 3, 3 to 1, 4 to 4. Note that input ring 4 is assigned to output ring 4 (self-assignment), although this combination has minimum number of slots waiting, since no other output is available. This will be rectified later when the inspection sequence will change. Note that self-assignments are necessary to send to upstream nodes sitting after the hub in unidirectional rings, and in any case access to slots for sending data to a node on the same ring, is only possible on slots so marked by the hub. Therefore, the intra-ring traffic is just one combination that must be explicitly granted. The effect of the presented overall control mechanisms of the system can be summarized in the following way. While the offered load is low, most elements of the matrix are zero, and all ring destination assignments appear with equal probability, i.e., in any ring, about one slot in four is for ring 1, one in four for ring 2, and so on (although not in round-robin fashion because of the random inspection sequence). Once more traffic destined for a ring is offered, the build-up of a relevant queue will force the mechanism to start assigning more times the desired destination ring compared with the other rings, responding to demand. The value of the solution is that this happens within just few round-trip times. It must be noted that although the SAT mechanism regulates the bandwidth usage among the nodes belonging to a source/destination ring pair, it is up to the hub scheduler to produce fair assignments for ring pairs. Thus, even simpler mechanisms, such a round-robin row inspection sequence,

LINARDAKIS et al.: SLOTTED INTERCONNECTED RINGS COMBINING TDMA AND WDMA

Fig. 4.

1581

Throughput in ring 1 for each other ring under uniform traffic.

produce unfairness among rings under nonuniform spatial distribution of loads. However, under uniform selection of source/destination pairs, the round robin and other simple greedy algorithms produce similar low-delay results, which strengthen the conclusion that the fast collection of queuing information in combination with the distributed empty slot access is more important than the scheduling algorithm, per se, in this distributed queuing system. IV. PERFORMANCE EVALUATION To evaluate the performance of the system under the explicit reservation scheme, the metropolitan area network (MAN) architecture was modeled using computer simulations where more nodes could be used than in the real demonstrator [12], which had a limited configuration for cost reasons. The network model consists of four rings, each connecting eight nodes and four data wavelengths operating at 10 Gb/s plus a fifth one for control. The ring round-trip time of all rings was equal to 135 s (i.e., 135 time slots, since the slot duration is 1 s). Each node is as600 signed the same number of SAT quota and equal to for all nodes in all rings (this is the number of slots it can send at each SAT round before releasing the SAT signal; see [13] for details). In all simulation scenarios, packets are generated from bursty ON–OFF sources with fixed size equal to 10 kb (fitting this way exactly in the available time slots), i.e., packet aggregation/segmentation was not simulated so as to evaluate the scheme independent of the effect of client-layer packet size. Thus, the evaluation of the explicit reservation mechanism could be decoupled from the choice of traffic model, since no agreement on a definite model for metropolitan traffic is available and particularly for the future services at the time of possible deployment of such a system. The ON and OFF periods of the sources follow an exponential distribution with a ratio between the mean OFF and mean ON period equal to 10, leading to significant bursti-

ness. The mean duration of the ON period is chosen to be three times below the ring round-trip time to exercise the scheme in a challenging time scale. The interarrival times of the packets generated during the ON period also follow exponential distribution whose mean values are derived according to the total offered load in each simulation scenario. Finally, the packet destinations were uniformly distributed over all nodes in all rings. To describe the traffic patterns used in each scenario, resort , with is made to the traffic loading matrix of size representing the traffic generated on ring every element toward ring expressed as a fraction of the ring capacity, i.e., of 40 Gb/s (four wavelengths at 10 Gb/s). The sum of the rows and columns in the matrix is 1, so it corresponds to 100% loading. To achieve a lower loading (keeping the same relationship among ring loads), the matrix is multiplied by the loading factor (e.g., to study the performance at a load equal to 20% of capacity, all coefficients are multiplied by 0.20, for 80% by 0.80, and so on). In the first scenario, a uniform loading matrix is used, with (i.e., all ), so all ring nodes are all generating ON-OFF offered traffic with uniform mean values and the packet destinations uniformly distributed to all metro rings and nodes. In Fig. 4, the throughput for each destination ring on ring 1 is plotted as a function of the offered load. The throughput curves for all destination rings overlap since the allocated bandwidth is equally shared among the rings due to load symmetry. The total ring throughput is also shown. This scenario least of all stresses the system, but it is useful as a yardstick against more asymmetric situations. The overall network throughput reaches saturation just 5% below the total system capacity due to the inevitable slight inefficiency of the SAT mechanism employed to guarantee fairness among the ring nodes. In this uniform scenario, using a simple cyclic repetition of all permutations without reservations would provide the same average throughput but would iron down the bursts. This, however, would also mean extreme delays and buffer overflows consequently violating service quality in most cases due to the very

1582

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 23, NO. 4, APRIL 2005

Fig. 5. Throughput in ring 1 for each ring under traffic scenario 2.

bursty traffic. In contrast, the explicit reservation mechanism reacts very fast to the bursts of the incoming traffic triggering the modification of the ring permutation schedule so as to service the congested ring nodes in preference to others. This is more obvious in the queuing delay performance, which stayed below the one ring round-trip time for low loads rising to just above this at high loads. This is expected since many arrivals find almost immediately empty slots passing by, without needing to wait for the round-trip time of their own reservations. Of course, these may not suffice, and some other payload slots have to wait for the needed assignment patterns to be created by the scheduler. Delay curves are not shown, however, in this scenario for lack of space, preferring to show them in the next asymmetric and therefore more demanding scenarios. Obviously, in this symmetric loading, even better delays are exhibited than the ones shown below under asymmetric loads. In the second scenario, the matrix describing the traffic load pattern is

It is observed from the matrix that the ratio between intra-ring and inter-ring traffic is 3. In Fig. 5, the throughput for each destination ring on ring 1 is depicted as in the previous scenario. Now the total is almost the same as before, but the curves and overlap, while the curve representing the throughput of the intra-ring traffic is at much higher levels due to the higher offered load. For every offered traffic load, the 50% of the ring’s total throughput corresponds to the intra-ring traffic while the remaining available bandwidth is equally shared among all other destination rings. For the most typical load value in scenario 2, i.e., the one where the total offered load is equal to 80% of the system capacity, the queuing delays versus time are also provided in Fig. 6

for a node in ring 1. The queuing delay curves of all four queues corresponding to the four destination rings are characterized by peaks and troughs reflecting the arrival of bursts that cause queue build-up. It is worth mentioning that when traffic sources with same average but lower burstiness are used, the delay remains at lower levels, also presenting a much lower variation. With bursty sources, a build-up of waiting slots is created by successive arrivals, which however is quickly relieved as reservations start creating the required ring assignments. Therefore, the impact of burstiness is a temporary increase of delay which of course adversely affects the overall mean value of delay. In any case, despite the variation of the delay at each queue, the mean queuing delay is the same for all destination rings. The mean value is almost equal to 135 s (i.e., one ring round-trip time), which is not surprising given the communal character of the ring assignments. In other words, the queued traffic can grab slots to the desired destination created by previous reservations while causing future assignments for the next in line. The reservations rectify the unbalances, but their action is at the same time retroactive and proactive. For example, starting in an unloaded system, an arrival will find an empty slot within on average just two slots, without needing reservations to create the required destination. Ring assignments circulate all the time, and the reservations only cause to increase inside each ring the frequency of destinations that are in high demand against the rest of destinations. In the final scenario, a highly unbalanced traffic pattern is used, described by the following load matrix:

The loading factor used was 80% of ring capacity. There is no symmetry between intra-ring and inter-ring traffic; in some rings, the intra-ring dominates in the total ring load but in others

LINARDAKIS et al.: SLOTTED INTERCONNECTED RINGS COMBINING TDMA AND WDMA

Fig. 6.

1583

Queuing delays in simulation scenario 2 at 80% of offered load

not. Only the queuing delays of a ring node in ring 1 are presented, since similar results were observed for all other nodes. As shown in Fig. 7, the slot queuing delay for all destinations remains bounded. Under this asymmetric load, the mean value of queuing delay in each ring depends not only on the total offered load destined to it but also on how this load is distributed among the source rings. For example, on ring 1, the experienced queuing delay of slots destined to ring 1 (intra-ring traffic) has a limited variation, since packets with the same destination are generated only from ring 3 and ring 4 with low peak 1/4 . On the other hand, the delay for destinarate values tion ring 3 varies more than in ring 4, although the offered load destined to these rings is the same, since the peak rates of the incoming traffic of ring 3 are much larger than in the case of ring 4. The important conclusion is that despite the significant asymmetry, still queuing delays in all ring nodes stay within expected limits, with the mean values ranging from 110 s (less than one round-trip time) for destination ring 1 where slot reuse allows a higher available effective bandwidth, up to 240 s for destination ring 3.

V. CONCLUSION An essential requirement of the emerging bursty real-time services is the ability to guarantee delay bounds and buffer limits. This puts a challenge on optical burst switching (OBS) systems where the switching speed and optical buffer limitations present insurmountable problems leading to intolerable inefficiencies. However, within a metropolitan domain where a round-trip time is small enough to allow accurate prearrangement of slot allocations on the basis of queue status, it is possible to achieve on-the-fly optical payload delivery without converting to the electrical domain. For the general case of interconnected wavelength-division-multiplexing (WDM) rings that can cover any size of metropolitan network, the proposed method can collect the necessary information for intra-ring switching of slots relying on a simple medium access control (MAC) for the inter-ring traffic. The efficiency of the system is similar to any shared medium, while the total queuing delay remains within the order of a few round trip times by means of an explicit reservation-based control mechanism. This is of

1584

Fig. 7.

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 23, NO. 4, APRIL 2005

Queuing delays from ring 1 under unbalanced traffic at 80% load.

prime importance since any delay in adapting to traffic demand would be translated in long delay bounds and enormous buffer sizes at the high rates of optical WDM systems.

REFERENCES [1] J. D. Angelopoulos, H.-C. Leligou, T. Argyriou, S. Zontos, E. Ringoot, and T. Van Caenegem, “Efficient transport of packets with QoS in an FSAN-aligned GPON,” IEEE Commun. Mag., vol. 42, no. 2, pp. 92–98, Feb. 2004. [2] Media Access Control Parameters, Physical Layers and Management Parameters for Subscriber Access Networks, Aug. 2002. IEEE Draft P802.3ah/D1.0TM. [3] P. Green, “Progress in optical networking,” IEEE Commun. Mag., vol. 39, no. 1, pp. 54–61, Jan. 2001. [4] D. Chiaroni, “Packet switching matrix: A key element for the backbone and the metro,” IEEE J. Sel. Areas Commun., vol. 21, no. 7, pp. 1018–1025, Sep. 2003. [5] C. Qiao and M. Yoo, “Optical burst switching-a new paradigm for an optical internet,” J. High Speed Netw., vol. 8, pp. 69–84, 1999. [6] L. Dittmann et al., “The European IST project DAVID: A viable approach toward optical packet switching,” IEEE J. Sel. Areas Commun., vol. 21, no. 7, pp. 1026–1040, Sep. 2003.

[7] A. Stavdas, S. Sygletos, M. O’Mahoney, H. Lee, and C. Matrakidis, “IST-DAVID: Concept presentation and physical layer modeling of the metropolitan area network,” J. Lightw. Technol., vol. 21, no. 2, pp. 372–383, Feb. 2003. [8] N. LeSauze, A. Dupas, E. Dotaro, L. Ciavaglia, M. Nizm, A. Ge, and L. Dembeck, “A novel low cost optical packet metropolitan ring architecture,” presented at the 27th Eur. Conf. Optical Communications (ECOC), Amsterdam, The Netherlands, Oct. 1–3, 2001. [9] M. A. Marsan, A. Bianco, E. Leonardi, M. Meo, and F. Neri, “MAC protocols and fairness control in WDM multirings with tunable transmitters and fixed receivers,” J. Lightw. Commun., vol. 14, no. 6, pp. 58–66, Jun. 1996. [10] A. Jourdan, D. Chiaroni, E. Dotaro, G. Eilenberger, F. Masetti, and M. Renaud, “Perspective of optical packet switching in IP dominant backbone and MAN,” IEEE Commun. Mag., vol. 39, no. 3, pp. 136–141, Mar. 2001. [11] A. Bianco, G. Galante, E. Leonardi, F. Neri, and M. Rundo, “Access control protocols for interconnected WDM rings in the DAVID metro network,” presented at the Int. Workshop Digital Communications (IWDC 2001), Taormina, Italy, Sep. 17–20, 2001. [12] B. B. Mortensen and M. S. Berger, “Optical packet switching demonstrator,” presented at the 28th Eur. Conf. Optical Communication (ECOC 2002), Copenhagen, Denmark, Sep. 2002. [13] I. Cidon and Y. Ofek, “Metaring-A full duplex ring with fairness and spatial reuse,” IEEE Trans. Commun., vol. 41, no. 1, pp. 110–119, Jan. 1993.

LINARDAKIS et al.: SLOTTED INTERCONNECTED RINGS COMBINING TDMA AND WDMA

[14] A. Bianco, J. M. Finochietto, G. Galante, F. Neri, and V. Sarra, “A fairness enforcement protocol for interconnected WDM rings,” presented at the Optical Network Design and Modeling 2004 (ONDM 2004), 8th IFIP Working Conf., Gent, Belgium, Feb. 2–4, 2004. [15] Ch. Linardakis, H.-C. Leligou, A. Salis, and J. D. Angelopoulos, “Control of slotted traffic among interconnected WDM rings by explicit reservation,” presented at the Photonics in Switching 2003 (PS 2003), Versailles, Paris, France, Sep. 28–Oct. 2 2003. [16] A. Bianco, J. Finochietto, E. Leonardi, P. Mitton, F. Marigigliano, F. Neri, and L. Quarello, “Multiclass resource allocation in interconnected WDM rings,” presented at the Optical Network Design and Modeling 2003 (ONDM 2003), IFIP TC-6 Working Conf., Budapest, Hungary, Feb. 3–5, 2003. [17] H.-C. Leligou, J. D. Angelopoulos, C. Linardakis, and A. Stavdas, “A MAC protocol for efficient multiplexing QoS-sensitive and best-effort traffic in dynamically configurable WDM rings,” Comput. Netw., vol. 44, no. 3, pp. 305–317, Feb. 20, 2004. [18] N. McKeown, A. Mekkittikul, V. Anantharam, and J. Walrand, “Achieving 100% throughput in an input-queued switch (extended version),” IEEE Trans. Commun., vol. 47, no. 8, pp. 1260–1267, Aug. 1999.

Charalambos Linardakis (S’02) received the Dipl.Ing. and Ph.D. degrees, both in electrical and computer engineering, from the National Technical University of Athens (NTUA), Athens, Greece, in 2000 and 2004, respectively. He is a Hardware Designer at the Telecommunications Laboratory in NTUA. His interest focus on wavelength-division-multiplexed access networks, passive optical network access networks, and convergence of TCP/IP and broad-band telecommunications.

1585

Helen C. Leligou received the Dipl.Ing. and Ph.D. degrees, both in electrical and computer engineering, from the National Technical University of Athens (NTUA), Athens, Greece, in 1995 and 2002, respectively. Her Ph.D. dissertation was in the area of access control mechanisms in broad-band networks. She is currently investigating access mechanisms for high-speed networks for tree-topology hybrid fiber/coax (HFC) and GPON systems and wavelengthdivision-multiplexed metro-ring networks.

Alexandros Stavdas (M’00) received the B.Sc. degree in physics from the University of Athens, Athens, Greece; the M.Sc. degree in optoelectronics and laser devices from Heriot–Watt University/St. Andrews University, Scotland, U.K.; and the Ph.D. degree from the University College of London, London, U.K., in the field of wavelength-routed wavelength-division-multiplexed networks. He was recently appointed Associate Professor of Optical Networking at the Department of Telecommunications Science and Technology of the University of Peloponnese, Tripolis, Greece. He is the author or coauthor of more than 50 journal publications and conference articles. Current interests include physical-layer modeling of optical networks, ultra-high-capacity end-to-end optical networks, optical cross-connect architectures, and optical packet switching.

John D. Angelopoulos (S’92–M’95) received the electrical engineering degree from the National Technical University of Athens (NTUA), Athens, Greece in 1973; the M.Sc. degree in automatic control from Nottingham University, Nottingham, U.K., in 1977; and the Ph.D. degree in telecommunications from NTUA in 1992. He worked for 11 years in the research and development section of the aerospace and telecommunication industries before joining the Technological Institute of Piraeus in 1989, where he is Professor in the Automation Engineering Department teaching data-network-related courses. He is also participating in NTUA research activities in the areas of high-speed networks. He has been involved in several RACE, ESPRIT, ACTS, and IST projects with emphasis on broad-band access systems (passive optical networks, hybrid fiber/coax (HFC), and wireless). His research interests cover most aspects of high-speed communications with special focus on shared medium access protocols and implementations. Dr. Angelopoulos is a Member of the Technical Chamber of Greece.