WONDER: A resilient WDM packet network for ... - Semantic Scholar

Optical Switching and Networking 5 (2008) 19–28 www.elsevier.com/locate/osn

WONDER: A resilient WDM packet network for metro applications A. Antonino a , A. Bianco a , A. Bianciotto b , V. De Feo a , J.M. Finochietto a , R. Gaudino a , F. Neri a,∗ a Dipartimento di Elettronica, Politecnico di Torino C.so Duca degli Abruzzi 24, 10129, Torino, Italy b Siemens IT Solutions and Services PSE, Hofmannstrae 51, D-81379, Munchen, Germany

Received 25 July 2007; accepted 15 September 2007 Available online 2 October 2007

Abstract This paper presents the architecture of a Wavelength Division Multiplexing (WDM) optical packet network, called WONDER, that was designed and prototyped in the PhotonLab at Politecnico di Torino, Italy. The design and implementation of the WONDER network aim to assess the effectiveness of optical technologies with respect to electronic ones, trying to identify an optimal mix of the two technologies. The architecture shows interesting resilience properties that enable the design of fast fault-recovery schemes. In this paper, we present the physical topology and node structure of the prototype, and discuss the implementation and performance of a fault-recovery algorithm. c 2007 Elsevier B.V. All rights reserved.

Keywords: WDM; Optical testbed; Optical packet switching; Fault recovery

1. Introduction The OptCom and Telecommunication Networks groups at Politecnico di Torino, Italy, have designed and prototyped Wavelength Division Multiplexing (WDM) network architectures taking an approach based on optical packets, but limiting the optical complexity to a minimum and trying to use only commercially available components. These research activities were funded by national research projects (dubbed RingO [1], WONDER [2–4], and OSATE [5,6]), and also supported by the European FP6 Network of Excellence ∗ Corresponding author.

E-mail addresses: [email protected] (A. Antonino), [email protected] (A. Bianco), [email protected] (A. Bianciotto), [email protected] (V. De Feo), [email protected] (J.M. Finochietto), [email protected] (R. Gaudino), [email protected] (F. Neri). c 2007 Elsevier B.V. All rights reserved. 1573-4277/$ - see front matter doi:10.1016/j.osn.2007.09.003

e-Photon/ONe [7]. The architecture presented in this paper was defined in the WONDER project, and was aimed at applications in metropolitan area networks; however, it can find interesting applications also as interconnection network inside large packet switching devices, or in data centers, as investigated in the OSATE project, though not addressed in this paper. The network architecture was designed in order to achieve a cost-effective compromise between optical and electronic technologies. To this extent, the bulk of raw data transmission is kept in the optical domain using WDM, while network control functions are mostly implemented in the electronic domain. Each network interface must process electronically only one WDM channel, while the aggregate network capacity can be scaled up by increasing the number of WDM channels without directly affecting the complexity of interfaces. In the following sections we introduce the rationale and design of the WONDER network, providing some

20

A. Antonino et al. / Optical Switching and Networking 5 (2008) 19–28

details of its physical topology, node architecture, Medium Access Control (MAC) protocol, and faultrecovery mechanisms. In particular, Section 2 presents the network architecture, details the main components of network nodes, and discuss the main features of the architecture. Among these features, the fault-recovery properties of the network are described in Section 3, where a distributed fault-recovery algorithm is proposed and its performance is evaluated. Finally, Section 4 concludes the paper and briefly describes ongoing activities. 2. System overview 2.1. Network architecture Fig. 1. The physical architecture.

The physical network topology, as shown in Fig. 1, is based on two unidirectional and counter-rotating WDM fiber rings connecting N nodes. One of the rings, called the “transmission” (TX) ring, is devoted to the transmission of data, while the other one, called the “reception” (RX) ring, is devoted to the reception of data. The two rings are physically interconnected by a fiber shortcut that can be implemented by any node in the network using optical switches. The shortcut is realized by optically closing a loopback fiber between the transmission and reception rings at the output of a node, which is conventionally called the “folding” (F) node. The node located at the other side of the shortcut turns out to be the first one on the transmission ring and the last one on the reception ring. This node is conventionally called the “master” (M) node since, by preceding all the other nodes, it is devoted to the transmission of suitable synchronization signals to the whole network (using a dedicated wavelength λs ). At any time, the network must have only one pair of active master and folding nodes in order to operate properly, all the other nodes being in a generic “through” (T) configuration. The resulting topology can also be viewed as a (multichannel) folded bus on which each node has two connections to the two arms of the bus, as shown in Fig. 2. The transmission of packets is time-slotted and synchronized on all wavelengths, and each packet has a fixed duration corresponding to one time slot. In our current implementation, each slot is 1 µs long. Data slots are operated with 8B/10B Gigabit Ethernet transmission at 1 Gbit/s data rate and 1.25 Gbit/s line rate (such relatively low bit rate has been chosen because this is the maximum bit rate that commercial Field Programmable Gate Arrays (FPGAs) and burstmode receivers can process at the moment). The main

Fig. 2. The logical architecture.

target of the current testbed is the demonstration of a “proof-of-concept”; evolutions to higher bit rates are planned as a next step. The network is synchronized by the master node, which sends to all other nodes, in a broadcast fashion, a “synchronization” (sync) signal containing both bit frequency and slot timing (frequency and phase) on a dedicated wavelength λs . This channel is currently operated as a standard Fast Ethernet digital channel at 100 Mbit/s data rate (125 Mbit/s line rate). Addressing in WONDER is based on a broadcastand-select approach, in which nodes broadcast packets on the wavelength where the destination node receives data, which may be shared by other node receivers. When one node has to send a packet, it tunes its transmitter to the corresponding wavelength. Thus, tuning latencies are required to be small compared to slot durations since no useful data can be transmitted during transmitters’ tuning. In the current experimental implementation, tuning is obtained in less than 30 ns. The optical filter selecting the received wavelength can be fixed or (slowly) tuneable. In this latter case the


21

Fig. 3. Functional description of WONDER nodes.

Fig. 4. Prototype boards (a) λ-Monitor, (b) Burst-Mode Transmitter, (c) Burst-Mode Receiver.

received wavelength can be dynamically reconfigured thus increasing the flexibility in the allocation of traffic between nodes [8] to the available channels. However, receiver tunability does not need to be achieved on a packet-by-packet basis. 2.2. Node architecture The structure of a WONDER node is shown in Fig. 3. Optical signals are represented as grey arrows while electrical signals, both digital and analogue, are displayed in black. The transmission fiber runs in the upper part of the figure from left to right, while the reception fiber runs at the bottom from right to left. The optical signals entering the node on the TX fiber are first power-balanced by a Variable Optical Attenuator (VOA) and then amplified by a Gain-Clamped Optical Amplifier (GCOA). The VOA is placed before the amplifier to make sure that the GCOA works under predictable saturation conditions. The amplified light signal is then partially split toward an Array Waveguide Grating (AWG) filter demultiplexing the WDM channels, including the synchronization channel λs .

The data channels are sent to a “λ-Monitor” block (see Fig. 4(a) for a picture of the board) detecting the busy/free state of each channel on a slot-by-slot basis for the implementation of the empty-slot MAC protocol, while the sync channel is received by a “sync RX” block for slot and bit timing extraction (see later) as well as upstream failure detection. The λ-Monitor has been realized using low-cost components. Each channel is made of a monitor photodiode followed by analogue conditioning circuitry (implementing low-pass filtering and level shifting) and a threshold comparator stage featuring a logic output towards the node controller (realized on an FPGA). Each node is equipped with a sync transmitter (sync TX) coupled to the TX ring so that each node can act as a master (i.e., generate the sync signal) if needed. Sync transmission has been obtained by using a fixedwavelength Distributed Feedback (DFB) laser featuring an integrated Mach–Zehnder modulator driven by the 8B/10B coded bit-stream running at 125 Mbit/s line rate and generated by the node controller. A delay fiber placed on the data path of the optical signal makes sure that the locally generated traffic is inserted on the TX fiber in sync with the in-line traffic already running on

22


it. In practice, the length of the delay fiber is trimmed in such a way that it causes a propagation delay exactly matching the processing delay of the MAC logic inside the FPGA. On the right-hand side of Fig. 3, a pair of optical switches (OSW) allows the node to implement the folding functionality in any node, if needed, by closing the TX fiber on the RX fiber. The sync receiver (sync RX), coupled on λs , is a conventional receiver operating on a continuous-wave, 8B/10B coded, bit-stream running at 125 Mbit/s. The received sync bit-stream is fed to the node controller, which integrates a Clock and Data Recovery (CDR) circuitry recovering a clock running at 125 MHz. This clock reference is then multiplied by a factor of 10 by an integrated Phase Locked Loop (PLL) in order to obtain the bit clock frequency of 1.25 GHz for the data channels (1.25 Gbit/s line rate). In addition, the slot timing (both in frequency and phase) is also recovered by detecting specific 8B/10B “unique words” that are periodically placed in the sync bit-stream once every 1 µs. Based on the output of the MAC protocol control logic, the node enables the transmission of optical packets by means of the wavelength-agile Burst-Mode Transmitter (BMT), which has been implemented as an array of on/off switched, fixed-wavelength DFB lasers whose output is then coupled and externally modulated by the 8B/10B coded data bits running at 1.25 Gbit/s (see Fig. 4(b)). The output of the BMT block is first coupled with the output of the sync TX, and then coupled to the TX ring. The use of an array of fixed lasers, where a single external modulator is used, allows efficient multicast transmission in the network, a required feature for the delivery of multimediaoriented services. Other solutions, based on fast-tunable lasers [9] were also investigated. On the reception side, the signal simply goes through a VOA + GCOA for controlled amplification and then is partially split toward the AWG filter which demultiplexes the WDM channels. The sync channel is received by a second sync RX block for downstream failure monitoring while a Burst-Mode Receiver (BMR) is devoted to the reception of optical packets on a given wavelength λ R X (alternative reconfigurable receiver architectures were also studied). If required, the other data channels can be fed to a second λ-Monitor for traffic monitoring functionalities. The node controller (NC) is at the heart of the node’s operation and implements all the intelligence of the network. It is realized on an FPGA board exchanging data with all the other subsystems, taking care of the implementation of the MAC protocol,

packet queuing, fault-recovery mechanism, burst-mode reception etc. We decided to base all our (electronic) architecture on an FPGA board due to the great increase in functionalities and reduction in costs that FPGA devices have shown during the last years. The Altera Stratix FPGA currently used in our testbed is able to handle several input/output data streams running at up to 3.2 Gbit/s, with advanced CDR capabilities. Currently, two different approaches to burst-mode reception are under investigation. The first approach is based on a commercial optical front-end receiver designed for GPON applications (Zenko SLT, see [10]) whose serial data output is directly fed to the NC which implements burst-mode CDR functionalities on the incoming discontinuous bit-stream. This is a very low-cost solution because it uses a 1.25 Gbit/s CDR and a deserializer integrated into the FPGA. Since this integrated CDR was designed for SONET/SDH or Gigabit Ethernet applications, and not for a burstmode operation, we used an additional logic that allows this CDR to work in burst-mode with acceptable performance. The second approach (see Fig. 4(c)) is based on the same optical front-end as above, followed by a commercial CDR chip designed for GPON burst-mode operation (Zenko ZB1G2CDR). In this last case, the NC receives a parallel (deserialized) bit-stream synchronized with the local clock. Both approaches represent low-cost BMR solutions, being based on commercially available chips designed for GPON applications [11]. We highlight that the required characteristics for a WONDER receiver perfectly match the GPON standard with minor modifications, and can thus take advantage of GPON components, that are becoming increasingly available from different vendors. The FPGA board interfaces the WONDER network with a Personal Computer (PC) through the Peripheral Component Interconnect (PCI) bus (actually, PCI-X or PCI-Express busses). A custom driver for the Linux operating system was developed, so that standard networking protocols (such as FTP, HTTP or SMTP) can be used as a source and a sink of the information transported by the WONDER network. Thus, standard networking applications running on PCs can use the WONDER network as a networking infrastructure. WONDER FPGA boards are viewed by the Linux operating system as network interface units (similar to Ethernet cards), with null impact on network applications. Segmentation of packets in fixed-size data units is achieved in the Linux TCP/IP stack and handled by the interface driver. These data units are received by the FPGA board via the PCI bus, and stored in a set


of queues, which implement a Virtual Output Queuing (VOQ) buffering structure. The entire WONDER network behaves as an input queuing system, and VOQ is necessary to avoid the drawbacks of headof-line (HOL) blocking. Since multicasting is possible in WONDER, more than a queue per destination is maintained in the FPGA board. The MAC protocol chooses on each time slot a packet to transmit from the head of non-empty queues, using the in-transit traffic information (empty/busy slots) on the TX ring provided by the λ-Monitor. 2.3. Features of the architecture The proposed architecture exhibits interesting features when compared with current SONET/SDH circuitswitched solutions for metro applications. First of all, each node is required to transmit, receive and process only the capacity of one wavelength, instead of the aggregate bit rate of the network (all wavelengths), thus significantly saving the amount of processing power required on each node. In addition, the WONDER architecture eases some of the issues that were not addressed by previous architectures [1,12], as detailed below. • Global synchronization is made easier by the linearity of the topology, and by the presence of the master node, which distributes bit frequency and slot phase and frequency on the dedicated broadcast synchronization wavelength λs . Burstmode receivers have a simplified structure, since they must recover only the bit phase. • Due to the equivalent folded bus topology, transmission and reception are separated in the network and no add/drop filters are required to inject/extract optical packets from the rings. As a consequence, the optical path is completely transparent (free of filtering elements), with significant advantages in terms of power budget and scalability of the network. Issues deriving from the recirculation of noise and packets within the network (often negatively affecting all optical ring topologies) are thus solved by design in WONDER. • Due to the absence of filtering elements along the optical data path, two or more nodes can share the same wavelength in reception, thus allowing a finer granularity in the allocation of the available bandwidth. This is achieved by adding a specific field to the header of the optical packet to indicate its destination among the various nodes sharing the same wavelength.

23

• The Medium Access Control (MAC) protocol can be a bare empty-slot access, according to which each node can use any empty slot in the transmission ring. Fairness among nodes can be enforced by simple mechanisms: in the WONDER project, extensions to the “classical” Fasnet [13] fairness control scheme were studied in detail [14]. We remark that these schemes enable an unslotted operation of the WONDER network, using variable-size packets, although this option was not explored yet in our experimental efforts. • The network is able to recover from different failures by simply rearranging the location of the loopback fiber with a reconfiguration of the optical switches available on nodes. However, there is the need that each node learns the new configuration in a distributed way as no signaling may be possible during a failure state. In the next section, we formulate the fault-recovery problem and propose a distributed algorithm that can be run on nodes to find out the new node configuration that brings the network back to an operational state. 3. Fault recovery 3.1. The problem Although all nodes can behave either as master, through or folding, each node must be assigned a specific configuration in order to set up a fully working network. In fact, a proper configuration must comprise only one master and one folding, the remaining nodes being through. Once all nodes have been properly configured at startup, the network is fully operational. Thereafter, if any fault occurs, it is necessary to reconfigure nodes in order to restore a fully working configuration, and this is addressed by the faultrecovery algorithm proposed in what follows. The key idea of the algorithm is to detect the failure and to isolate it by properly reconfiguring the nodes. Faults can happen either at fiber spans that interconnect two nodes (e.g., a fiber cut) or at nodes (e.g., power off), and must be detected by all nodes in order to trigger the recovery procedure. In particular, we investigate the following possible failure events: • Single Fiber Cut (SFC). It considers a single cut on both TX and RX rings. This is considered by far the most likely fault event [15] and, consequently, most of our results focus on this situation. • Multiple Fiber Cuts (MFC). It assumes that there is already a fiber cut between the current folding and master and that a new cut takes place. Although

24


given. When more than one fiber cut (on different spans) is present, it is not possible to restore full network connectivity. However, it is feasible to build two isolated subnetworks so that at least the connectivity among subnetwork nodes is recovered. • Node Failure (NF). It considers a complete node failure, i.e., a situation in which all optoelectronic functions of a node are completely shut down. We assume in this case that the node becomes optically opaque on both transmission and reception fibers. This event can also be seen as a simultaneous fiber cut on the four fibers the node is connected to.

Fig. 5. A fault event in the network.

Fig. 6. New configuration after a fault event.

It is easy to notice that all the possible faults have basically the same effect on the network: they split the topology into two sets of nodes that will be conventionally indicated as the upstream and the downstream nodes (relative to the cut, using the TX ring propagation direction as a reference), as shown in Fig. 5. After a fault event, according to the reference direction as above, the upstream nodes actually “precede” the fault and sense at first a RX signal loss that triggers them to be part of the upstream “group”. In contrast, the nodes located “after” the cut, sense a TX signal loss at first (then followed by a RX loss), and this event triggers them in the downstream “group”. The upstream group includes the master and all through nodes located between the master and the fault, while the downstream group comprises the folding and all through nodes between the fault and the folding itself. It is worth to notice that by properly selecting a new “folding” among upstream nodes and a new “master” among downstream nodes, two fully working and independent subnetworks can be created. If the two subnetworks can then be merged (if a SFC or NF event happened), then these newly elected master and folding nodes can become the master and folding nodes of the whole (merged) network; see Fig. 6. If it is not possible to merge the two subnetworks (i.e., in a MFC case), then they remain isolated but fully operational. Therefore, the fault-recovery problem can be solved basically by selecting a new folding among the upstream nodes and a new master among the downstream nodes. 3.2. The algorithm

Fig. 7. Election of new master and folding nodes after a fault event.

the probability of this event is much lower than the SFC one, it could be very harmful if no protection is

The proposed scheme is totally distributed, and requires null coordination among nodes. In fact, our algorithm is based on a state-machine independently running at each node. The final goal of this algorithm is to find out which configuration each node must adopt in order to set the network back to a working condition. A detailed analysis of this algorithm can be found in [16].


Each node makes its own decisions based on local information. This information consist of its current configuration (i.e., master, through or folding) plus the state of the synchronization signal on both the TX and RX rings. Based on these inputs, the node can set the optical switches (the two OSW in Fig. 3) as either open or close, and the sync transmitter as on or off. Moreover, other network parameters are assumed to be known by each node, such as: • N : the total number of nodes in the network; • T p,max : the maximum propagation time between two adjacent nodes; • Ts,max : the worst switching speed among OSW. The fault-recovery algorithm works as follows: as the failure of the synchronization channel propagates inside the network topology, the nodes enter either the upstream or the downstream groups according to the previously defined criteria on TX and RX failure monitors. As shown in Fig. 7, all upstream nodes switch to a temporary folding (switch closed) configuration (the master ceases transmitting the sync), while the downstream nodes preserve their current configuration. After this initial phase, the downstream nodes start to contend to become the new master of the network. During this contention, each downstream node tries to transmit the sync signal, but immediately ceases to do so if it senses such a signal on the TX ring, because if this happens it means that there is a node that is nearer to the fault, thus better positioned to become the master. The contention phase ends within a time Tmc (Master Contention Time) which is given, in the worst case, by: Tmc = (N − 1) T p,max . It is worth mentioning that we neglect here electronic processing times needed to actually sense the signal loss, and assume that propagation times dominate over processing ones. The result of the contention phase is the selection of a “new” master that will be the downstream node which is closest to the failure. After at most Tmc , the “current” folding node (i.e. the node that was folding before the failure event) detects the fault and waits for a time Tsw (Switch Wait Time) for all the upstream nodes to switch to a temporary folding configuration. Tsw is given, in the worst case, by: Tsw = Ts,max + (N − 1) T p,max . After Tsw , the current folding node switches to a through (or master if it is the only node in the downstream group) configuration. By switching to this configuration, the current folding tries to join the upstream and downstream subnetworks. Since the

25

current master node is now in a folding configuration, the sync signal that is propagating towards the current master node, is expected to come back on the RX ring within 2 T p,max . If no signal comes back, the network is in a MFC condition (there is already a fiber cut between the original master and folding noes). In this case, the subnetworks cannot be merged and the current folding returns to be so inside the downstream subnetwork topology. On the other side of the failure, the current master node, after sensing the fault, waits a time T jw (Join Wait Time) for a signal coming from the downstream group. This time, in the worst case, is given by: T jw = Tmc + Tsw + Ts,max + 2 T p,max . If no signal comes within T jw , the current master senses an MFC condition and, from the present folding configuration, switches back to a full master condition, transmitting the sync signal again in order to rebuild the “upstream subnetwork”. Otherwise, if the signal arrives within T jw , the network is in an SFC/NF configuration, and the full connectivity between nodes can be restored. In both the cases (either MFC or SFC/NF) the faultrecovery algorithm now enters a new phase, where the temporary folded nodes sequentially enter the network to complete the topology. When a temporary folded node senses the sync signal to be stable on the TX ring for a propagation time T p,max , it switches to a through configuration and waits for the signal to come back on the RX ring within 2 T p,max . If this happens, the node can retain its through configuration because there is another temporary folded node closer to the failure. If instead no signal comes back within 2 T p,max , it means that the actual node directly precedes the failure, so it must switch back to a folding configuration. This last phase takes a time Tse (Sequential Entrance Time) which is given, in the worst case, by: Tse = N Ts,max + 2 (N − 1) T p,max measured from when T jw expires (MFC case) or from when the sync signal from the current folding reaches the current master (SFC/NF case). When this last phase ends, the proper end topology has been built, either as a single network or as two independent subnetworks. The time that elapses between the failure event and the restoration of the data transmission by all the nodes in the network is actually called the “fault-recovery time” and, in the worst case, it is given by: T f r = T jw + Tse = (4 N − 2) T p,max + (N + 2) Ts,max .

26


Fig. 8. Performance of the fault-recovery schemes vs. span length.

network (between the current folding and master) and that a new cut occurs on a different span. The NF event considers a complete shutdown of any of the N nodes. We assume time slots of 1 µs duration, thus, time slots that can carry 1 kbits at 1 Gbit/s and 10 kbits at 10 Gbit/s. We choose, as a reference for our simulations, a network with 8 nodes spaced 1 km (this is called “span length”) and switch latencies of 1 µs (N = 8, T p = 5 µs and Ts = 1 µs). Simulations were then run by varying network parameters (N , T p and Ts ) aiming at testing the different aspects of our proposed strategy. Next, we show results concerning the impact on the recovery time when considering different span lengths, optical switch latencies and number of network nodes. Span length In Fig. 8 we show results based on our reference network but considering different span lengths. When the switching latency is small with respect to the propagation delay (Ts T p ), the impact of the span length on the recovery time is directly proportional. The larger the spans, the larger the propagation delays; thus both the fault and the synchronization signal propagation are directly affected. In all cases, recovery times are in the order of few milliseconds, which is very low for most applications.

Fig. 9. Performance of the fault-recovery schemes vs. switch latency.

This analytical result represents an upper bound of fault-recovery times and has been confirmed by several simulation runs under different scenarios. Therefore, it can be used as a design formula to derive the recovery time as a function of the network’s key parameters. Due to space constraints, we only report in the next section some results obtained by simulation. 3.3. Performance evaluation In order to test our fault-recovery strategy we modeled the WONDER network with MATLAB and implemented the fault-recovery procedure described in Section 3.2. The network, the node architecture, and the considered faults (SFC, MFC and NF) were modeled with MATLAB Simulink, while the node controller logic (i.e., the state-machine) with MATLAB Stateflow. The SFC fault can occur on any of the N fiber spans of the network, even in the one between the folding and the master (i.e., no fault is detected). The MFC assumes that there is already a fiber cut in the

Switch latency If Ts T p then switching latencies have a significant impact on the recovery time. In Fig. 9 we plot in logarithmic scales the recovery time when considering different switch latencies. In particular, we simulated 1 µs, 10 µs, 1 ms, 2 ms, 5 ms and 10 ms. Besides, two different fiber spans have been considered, 1 and 10 km. On the leftmost part of the curve it is possible to appreciate the impact of span lengths (as we did in Fig. 8), but from 1 ms onwards there is almost no difference between the two fiber span scenarios. In fact, switch latency becomes dominant and has a linear impact on the recovery time. Network nodes The last parameter we considered is the impact of the number of nodes on the fault-recovery strategy. In other words, this can be seen as the scalability of our algorithm. As previously described, the switch technology has an important impact on the recovery time. In Fig. 10 we consider three possible switch latencies representing different technologies, and plot the recovery time when increasing the number of nodes (N ). In addition, we plot a line representing the typical


Fig. 10. Performance of the fault-recovery schemes vs. network nodes.

50 ms fault-recovery performance of SONET/SDH networks. Slow switches (10 ms) limit our strategy to less than 5 nodes on the network, thus, do not seem to be a suitable technology. However, fast ones (1 µs) provide excellent scalability, allowing more than 50 nodes on the network. A cost-effective technology for a metropolitan area network seems to be the use of switches of 1 ms latency, which permits to recover failures in less than 50 ms up to a network of 40 nodes.

27

lowing the exchange of data between three PCs running the Linux operating system and standard Internet applications. The final experimental demonstrator will provide automatic recovery from failures by implementing the fault-recovery strategy described above. Finally, we remark that the WONDER architecture can also be used as an interconnection network. For example, in the framework of the Italian OSATE project, our group is currently investigating its use as a switching matrix (or “optical backplane”) inside a packet switch. For either application context, we are currently assessing the network scalability in terms of number of nodes and aggregate capacity. Variations to the architecture presented in this paper were derived to improve network scalability (preliminary results have been discussed in [6]). Acknowledgments This work was supported by the European FP6 ePhoton/ONe Network of Excellence (NoE) and by the Italian national project (PRIN) OSATE. Preliminary partial versions of the work reported in this paper were presented in [16,4]. References

4. Conclusions Our work was motivated by the trust that optical packet transmission, in a configuration similar to what is often today called “Ethernet over optics”, though not yet standardized and commercially available, may become in the medium term a promising alternative to the current approach of building static WDM networks with (at most) some limited degree of optical reconfigurability, but where packet switching is still completely handled at the electronic level. At the same time, we do not believe that all packet switching functions can be completely moved from the electrical to the photonic domain in a reliable way without fundamental advances in optical components technology. A good compromise between the two domains (optical and electrical) is the major goal of the research work presented in this paper. Our prototyping efforts, based on commercially available components, give us strong indications on which are the most complex subsystems either in the optical or in the electronic domain. At present, we are completing the realization of a full-featured experimental demonstrator in our PhotonLab [17] in Turin, Italy. The current prototype comprises three nodes, mounted in separate racks, and al-

[1] A. Carena, V. De Feo, J.M. Finochietto, R. Gaudino, F. Neri, C. Piglione, P. Poggiolini, RingO: An experimental WDM optical packet network for metro applications, IEEE Journal on Selected Areas in Communications 22 (8) (2004) 1561–1571. [2] WONDER Project, Online: http://www.tlc-networks.polito.it/ wonder. [3] A. Bianciotto, R. Gaudino, WONDER: Overview of a packetswitched MAN architecture, in: Proceedings OpNeTec 2004, Pisa, Italy, October 2004. [4] A. Bianciotto, V. De Feo, R. Gaudino, A. La Porta, F. Neri, M. Petracca, Experimentation of single-hop WDM rings for metro and distributed switching applications, in: Photonics in Switching 2006, Herakleion, Crete, Greece, October2006. [5] OSATE Project, Online: http://www.tlc-networks.polito.it/ projects/osate. [6] A. Bianco, E. Carta, D. Cuda, J.M. Finochietto, F. Neri, An optical interconnection architecture for large packet switches, in: International Conference on Transparent Optical Networks, ICTON 2007, Rome, Italy, June 2007. [7] e-Photon/ONe, Online: http://www.e-photon-one.org. [8] A. Bianco, J.M. Finochietto, G. Giarratana, F. Neri, C. Piglione, Measurement based reconfiguration in optical ring metro networks, IEEE/OSA Journal of Lightwave Technology 23 (10) (2005) 3156–3166. [9] A. Bianciotto, B. Puttnam, P. Bayvel, et al. Burst-mode operation of a DS-DBR fast tuneable laser for wavelength agile system applications, in: Proc. OFC 2006, Anaheim, CA, USA, March 2006. [10] Zenko, Online: http://www.zenkotech.com.

28


[11] AA. VV., Gigabit-capable Passive Optical Networks (GPON): General characteristics, ITU-T standard G984.1-4. [12] K.V. Shrikhande, I.M. White, D. Wonglumsom, S.M. Gemelos, M.S. Rogge, Y. Fukashiro, M. Avenarius, L.G. Kazovsky, HORNET: A packet-over-WDM multiple access metropolitan area ring network, IEEE Journal on Selected Areas in Communications 18 (10) (2000) 2004–2016. [13] J.O. Limb, C. Flores, Description of fasnet—A unidirectional local-area communication network, The Bell Systems Technical Journal 61 (7) (1982).

[14] A. Bianco, D. Cuda, J.M. Finochietto, F. Neri, C. Piglione, Multi-fasnet protocol: Short-term fairness control in WDM slotted MANs, in: Proc. GLOBECOM 2006, San Francisco, CA, USA, November 2006. [15] R. Ramaswami, K.N. Sivarajan, Optical Networks—A Practical Perspective, Morgan Kaufman, 1988. [16] A. Bianciotto, J.M. Finochietto, R. Gaudino, F. Neri, C. Piglione, A. Samperi, Fast and efficient fault-recovery strategies in the WONDeR metro architecture, in: Proc. NOC 2005, London, UK, July 2005. [17] PhotonLab, Online: http://www.photonlab.org.